Improved pitch content in neural speech synthesis

Abstract

Techniques and systems are provided for speech synthesis. For instances, a process can include receiving a first predicted audio sample; causing the first predicted audio sample to be stored in one or more memories; and generating a second predicted audio sample based on the first predicted audio sample, wherein the second predicted audio sample is from a pitch period after the first predicted audio sample.

Description

IMPROVED PITCH CONTENT IN NEURAL SPEECH SYNTHESISFIELD

[0001] The present application is generally related to audio coding (e.g., audio encoding and / or decoding). For example, aspects of the application relate to systems and techniques for improved pitch content in neural speech synthesis.BACKGROUND

[0002] Audio coding (also referred to as voice coding and / or speech coding) is a technique used to represent a digitized audio signal using as few bits as possible (thus compressing the speech data), while attempting to maintain a certain level of audio quality. An audio or voice encoder is used to encode (or compress) the digitized audio (e.g., speech, music, etc.) signal to a lower bit-rate stream of data. The lower bit-rate stream of data can be input to an audio or voice decoder, which decodes the stream of data and constructs an approximation or reconstruction of the original signal. The audio or voice encoder-decoder structure can be referred to as an audio coder (or voice coder or speech coder) or an audio / voice / speech coderdecoder (codec).

[0003] Audio coders exploit the fact that speech signals are highly correlated waveforms. Some speech coding techniques are based on a source-filter model of speech production, which assumes that the vocal cords are the source of spectrally flat sound (an excitation signal), and that the vocal tract acts as a filter to spectrally shape the various sounds of speech. The different phonemes (e.g., vowels, fricatives, and voice fricatives) can be distinguished by their excitation (source) and spectral shape (filter).SUMMARY

[0004] Systems and techniques are described herein for improved pitch content in neural speech synthesis. The following presents a simplified summary relating to one or more aspects disclosed herein. Thus, the following summary should not be considered an extensive overview relating to all contemplated aspects, nor should the following summary be considered to identify key or critical elements relating to all contemplated aspects or to delineate the scope associated with any particular aspect. Accordingly, the following summary presents certain concepts relating to one or more aspects relating to the mechanisms disclosed herein in a simplified form to precede the detailed description presented below

[0005] Disclosed are systems, apparatuses, methods and computer-readable media for speech synthesis. In one illustrative example, an apparatus for speech synthesis is provided. The apparatus includes one or more memories configured to store one or more audio samples and one or more processors coupled to the one or more memories. The one or more processors are configured to: receive a first predicted audio sample; cause the first predicted audio sample to be stored in the one or more memories; and generate a second predicted audio sample based on the first predicted audio sample, wherein the second predicted audio sample is from a pitch period after the first predicted audio sample.

[0006] As another example, a method for speech synthesis is provided. The method includes: receiving a first predicted audio sample; causing the first predicted audio sample to be stored in one or more memories; and generating a second predicted audio sample based on the first predicted audio sample, wherein the second predicted audio sample is from a pitch period after the first predicted audio sample.

[0007] In another example, a non-transitory computer-readable medium having stored thereon instructions is provided. The instructions, when executed by one or more processors, cause the one or more processors to: receive a first predicted audio sample; cause the first predicted audio sample to be stored in one or more memories; and generate a second predicted audio sample based on the first predicted audio sample, wherein the second predicted audio sample is from a pitch period after the first predicted audio sample.

[0008] As another example, an apparatus for speech synthesis is provided. The apparatus includes means for receiving a first predicted audio sample; means for causing the first predicted audio sample to be stored in one or more memories; and means for generating a second predicted audio sample based on the first predicted audio sample, wherein the second predicted audio sample is from a pitch period after the first predicted audio sample.

[0009] In some aspects, one or more of the apparatuses described herein can include or be part of an extended reality device (e g., a virtual reality (VR) device, an augmented reality (AR) device, or a mixed reality (MR) device), a mobile device (e.g., a mobile telephone or other mobile device), a wearable device (e.g., a network-connected watch or other wearable device), a personal computer, a laptop computer, a server computer, a television, a video game console, or other device. In some aspects, the one or more apparatuses can include at least one camera for capturing one or more images or video frames. For example, the one or more apparatusescan include a camera (e.g., an RGB camera) or multiple cameras for capturing one or more images and / or one or more videos including video frames. In some aspects, the one or more apparatuses can include a display for displaying one or more images, videos, notifications, or other displayable data. In some aspects, the one or more apparatuses can include at least one transmitter configured to transmit data or information over a transmission medium to at least one device. In some aspects, at least one processor of the one or more apparatuses can include a central processing unit (CPU), a digital signal processor (DSP), a graphics processing unit (GPU), a neural processing unit (NPU), a neural signal process (NSP). or other processing device or component.

[0010] This summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used in isolation to determine the scope of the claimed subject matter. The subject matter should be understood by reference to appropriate portions of the entire specification of this patent, any or all drawings, and each claim.

[0011] The foregoing, together with other features and examples, will become more apparent upon referring to the following specification, claims, and accompanying drawings.BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Illustrative examples of the present application are described in detail below with reference to the following figures:

[0013] FIG. 1 is a block diagram illustrating an example of a voice coding system, in accordance with some examples;

[0014] FIG. 2 is a block diagram illustrating an example of a CELP-based voice coding system, in accordance with some examples;

[0015] FIG. 3 is a block diagram of an example voice decoding system, in accordance with aspects of the present disclosure;

[0016] FIG. 4 is a block diagram illustrating a sample-by-sample autoregressive ML-based speech synthesizer, in accordance with aspects of the present disclosure.

[0017] FIG. 5 is a block diagram illustrating a postprocessing process for improving pitch content, in accordance with aspects of the present disclosure.

[0018] FIG. 6 is a block diagram illustrating an integrated process for improving pitch content, in accordance with aspects of the present disclosure.

[0019] FIG. 7A is a block diagram illustrating vector-based process for improving pitch content, in accordance with aspects of the present disclosure.

[0020] FIG. 7B is a block diagram illustrating an example of an attention mechanism, in accordance with aspects of the present disclosure;

[0021] FIG. 7C is a block diagram illustrating another example of an attention mechanism, in accordance with aspects of the present disclosure;

[0022] FIG. 8 is a block diagram illustrating a postprocessing process for improving pitch content, in accordance with aspects of the present disclosure;

[0023] FIG. 9 is a block diagram illustrating an integrated process for improving pitch content, in accordance with aspects of the present disclosure;

[0024] FIG. 10 is a block diagram illustrating an example of an audio codec system that can be used to generate reconstructed audio (e.g., synthesized speech) using a neural speech synthesizer, in accordance with aspects of the present disclosure;

[0025] FIG. 11 illustrates an example of an audio codec system (e.g., generative voice codec) that includes a decoder configured to implement an NHV-based neural speech synthesizer and a postprocessing process for improving pitch content, in accordance with aspects of the present disclosure;

[0026] FIG. 12 illustrates an example of an audio codec system (e.g., generative voice codec) that includes a decoder configured to implement an NHV-based neural speech synthesizer and including an integrated process for improving pitch content, in accordance with aspects of the present disclosure;

[0027] FIG. 13 is a flow diagram illustrating a process for speech synthesis, in accordance with aspects of the present disclosure;

[0028] FIG. 14 is an illustrative example of a neural network that can be used to implement machine-leaming-based speech synthesis, in accordance with aspects of the present disclosure;

[0029] FIG. 15 is an illustrative example of a convolutional neural network (CNN);

[0030] FIG. 16 is a block diagram illustrating an example of a deep convolutional network;

[0031] FIG. 17 is a diagram illustrating an example of a system for implementing certain aspects of the present technology;

[0032] FIG. 18 is a block diagram illustrating an example of a decoder 1018 of a voice coding signal synthesis system, in accordance with aspects of the present disclosure; and

[0033] FIG. 19 illustrates an example of an audio codec system, in accordance with aspects of the present disclosure.DETAILED DESCRIPTION

[0034] Certain aspects and examples of this disclosure are provided below. Some of these aspects and examples may be applied independently and some of them may be applied in combination as would be apparent to those of skill in the art. In the following description, for the purposes of explanation, specific details are set forth in order to provide a thorough understanding of subject matter of the application. However, it will be apparent that various examples may be practiced without these specific details. The figures and description are not intended to be restrictive.

[0035] The ensuing description provides illustrative examples only, and is not intended to limit the scope, applicability, or configuration of the disclosure. Rather, the ensuing description will provide those skilled in the art with an enabling description for implementing the illustrative examples. It should be understood that various changes may be made in the function and arrangement of elements without departing from the spirit and scope of the application as set forth in the appended claims.

[0036] In some cases, a neural network-based voice decoder may be used to decode digitized audio signals (e.g., a speech signal) to synthesize the audio signal (e.g., speech). The speech signal may be decoded audio-sample by audio sample, such as from an audio stream. In some cases, a neural network-based voice decoder can be trained to predict a residual of a linear predicative filter (LPC). Such LPC residual prediction can include complex operations that require the use of a large amount of computing resources by the neural network-based voice decoder. For example, sample prediction for neural network-based voice decoders mayuse multiple gated recurrent units (GRUs). These GRUs may be used to, in part, remember previous samples, such as those from a previous pitch period, along with previously predicted sample(s), pitch lag. and gain information to predict a current sample. This can result in a relatively complex GRU that may be computationally expensive. However, simplifying the GRUs may result in not remembering or remembering fewer previously predicted samples can result in a warbling sound in the synthesized speech. Thus, techniques to improve pitch content in neural speech synthesis, for example, by reducing GRU complexity while avoiding warble may be useful.

[0037] Systems, apparatuses, electronic devices, methods (also referred to as processes), and computer-readable media (collectively referred to herein as “systems and techniques”) are described for improving pitch content in neural network-based speech synthesis. For example, there is a predictable pitch periodicity and temporal structure in human speech (e.g., based on vocal cord movements). A pitch lag may be incorporated to help bring in previously estimated samples around a previous pitch period to process a current predicted sample. In some cases, processing the current predicted sample may be performed as a post-processing process to the neural network-based speech synthesis process. In some cases, a refined current predicted sample may be generated for output. This refined current predicted sample may be fed back to the neural network-based speech synthesis process to help process a next predicted sample. In some cases, predicted samples may be stored in a delay circuit and the delay circuit may determine a set of previously predicted samples to use based on information about a pitch lag (e.g., an amount of time between pitch periods). The set of previously predicted samples may be within a certain time window around the previous pitch period, or all of the samples from the previous pitch period may be included. In some cases, rather than post-processing the set of previously predicted samples, the set of previously predicted samples, or information about the set previously predicted samples, may be passed into the neural network-based speech synthesis process for use in synthesizing the current predicted sample.

[0038] Autoencoders have gained popularity in recent years as they are able to learn efficient representations of input data without the need of labels (e.g., based on performing unsupervised learning). Various types of autoencoders exist and are well explained in “Autoencoder and its various variants” 2018 IEEE International Conference on Systems, Man, and Cybernetics, Zhang et. al. In speech coding, autoencoders conditioned on log Mel Cepstrum inputs and / or spectrogram inputs have been used for speech compression in"Enhancing into the Codec: Noise Robust Speech Coding with Vector-Quantized Autoencoders" arXiv:2102.06610vl [eess.AS] 12 Feb 2021, Casebeer et. al.

[0039] Speech coding is a lossy compression process. Autoencoders perform dimensionality reduction (e.g., an N-dimensional vector input vector is passed into the encoder of the autoencoder, and lossy compression is performed to represent the important aspects of the input vector in an M-dimensional vector, where M is smaller than N, e.g. by an order of magnitude). A drawback of using an autoencoder for speech coding is that it cannot necessarily exploit the temporal relationship between sets of input data. U.S. Patent No. 11,526,734 (‘’the ‘734 patent’’, assigned to Qualcomm, Inc.), "Method and Apparatus for recurrent auto encoding", Yang et. al. improves upon an autoencoder to exploit temporal redundancies and correlations and describes a feedback recurrent autoencoder ("FRAE"). The FRAE in the '734 patent can be used for training and application of compression of sequential data with temporal correlation. The recurrent structure of the FRAE can be used to efficiently extract the redundancy embedded along the time-dimension of sequential data and enables compact discrete representation of the data at the bottleneck in a sequential fashion. For example, Table 1 of the '734 patent illustrates the MSE (Mean Squared Error) used as Mel-scale mean-square- error configured as a reconstruction loss for training of the FRAE, where the MSE of each frequency bin is scaled according to its weight at Mel-frequency both for latent feedback and output feedback.

[0040] The FRAE described in the ‘734 patent has two advantageous features not present in autoencoders: (1) recurrent layers, e.g. LSTM or GRU layers, have memory of the past; and (2) feedback from the decoder of the autoencoder to the encoder of the autoencoder. The feedback connection 150 in Fig. 2, Fig. 3 and Fig. 7 of the ‘734 patent provides additional historical information from a state (ht) in the decoder to the encoder indicative of how reconstruction of a prior input has fared. This feedback loop is present during training and inference, which allows the encoder to be trained to respond to particular feedback conditions in a manner that improves reproduction of the input data by the decoder. Thus, the feedback loop is analogous to a mode switch input in that it is not encoded by the encoder, but it is used as an input that influences how the encoder operates on the next input vector.

[0041] In addition, the ‘734 patent describes a second feedback connection (152) from the state (ht) of the decoder for a first iteration of series of inputs Xt, to the next iteration ofseries of inputs Xt+1. This second feedback connection (152) enables the decoder to learn from its previous reconstruction attempts, providing additional historical context about how reconstruction of a prior input has fared. This second feedback loop is also present during both training and inference, which allows the decoder to be trained to respond to particular feedback conditions in a manner that improves reproduction of the input data by the decoder.

[0042] The ‘734 patent also describes other optional feedback connections. For example, an embedding vector (z) may be fed back via a third feedback connection (356 in Fig. 3 of the ‘734 patent) to the encoder. As another optional example, the '734 patent describes that the FRAE can be a variational autoencoder. In this example, an output of the decoder (754 in Fig. 7 of the ‘734 patent) is sample and parameterized to a generate an autoregressive prior which can be used as a fourth feedback connection to condition a prior model for the next latent space embedded vector (zt+1 ). Similar to the previously described feedback connection, each of these optional connections is present both training and inference, and thus are trained as part of the training process thereby improving functionality of the FRAE.

[0043] Various aspects of the present disclosure will be described with respect to the figures.

[0044] Audio coding (e.g., speech coding, music signal coding, or other type of audio coding) can be performed on a digitized audio signal (e.g., a speech signal) to compress the amount of data for storage, transmission, and / or other use. FIG. 1 is a block diagram illustrating an example of a voice coding system 100 (which can also be referred to as a voice or speech coder or a voice coder-decoder (codec)). A voice encoder 102 of the voice coding system 100 can use a voice coding algonthm to process a speech signal 101. The speech signal 101 can include a digitized speech signal generated from an analog speech signal from a given source. For instance, the digitized speech signal can be generated using a filter to eliminate aliasing, a sampler to convert to discrete-time, and an anal og-to-digi tai converter for converting the analog signal to the digital domain. The resulting digitized speech signal (e.g., speech signal 101) is a discrete-time speech signal with sample values (referred to herein as samples) that are also discretized.

[0045] Using the voice coding algorithm, the voice encoder 102 can generate a compressed signal (including a lower bit-rate stream of data) that represents the speech signal 101 using as few bits as possible, while attempting to maintain a certain quality level for the speech. Thevoice encoder 102 can use any suitable voice coding algorithm, such as a linear prediction coding algorithm (e.g., Code-excited linear prediction (CELP), algebraic-CELP (ACELP), or other linear prediction technique) or other voice coding algorithm.

[0046] The voice encoder 102 can compress the speech signal 101 in an attempt to reduce the bit-rate of the speech signal 101. The bit-rate of a signal is based on the sampling frequency and the number of bits per sample. For instance, the bit-rate of a speech signal can be determined as follows:BR = S * b,

[0047] Where BR is the bit-rate, 5 is the sampling frequency, and b is the number of bits per sample. In one illustrative example, at a sampling frequency (5) of 8 kilohertz (kHz) and at 16 bits per sample (b), the bit-rate of a signal would be a bit-rate of 128 kilobits per second (kbps).

[0048] The compressed speech signal can then be stored and / or sent to and processed by a voice decoder 104. In some examples, the voice decoder 104 can communicate with the voice encoder 102, such as to request speech data, send feedback information, and / or provide other communications to the voice encoder 102. In some examples, the voice encoder 102 or a channel encoder can perform channel coding on the compressed speech signal before the compressed speech signal is sent to the voice decoder 104. For instance, channel coding can provide error protection to the bitstream of the compressed speech signal to protect the bitstream from noise and / or interference that can occur during transmission on a communication channel.

[0049] The voice decoder 104 can decode the data of the compressed speech signal and construct a reconstructed speech signal 105 that approximates the original speech signal 101. The reconstructed speech signal 105 includes a digitized, discrete-time signal that can have the same or similar bit-rate as that of the original speech signal 101. The voice decoder 104 can use an inverse of the voice coding algorithm used by the voice encoder 102, which as noted above can include any suitable voice coding algorithm, such as a linear prediction coding algorithm (e.g., CELP. ACELP, or other suitable linear prediction technique) or other voice coding algorithm. In some cases, the reconstructed speech signal 105 can be converted tocontinuous-time analog signal, such as by performing digital-to-analog conversion and antialiasing fdtering.

[0050] Voice coders can exploit the fact that speech signals are highly correlated waveforms. The samples of an input speech signal can be divided into blocks of N samples each, where a block of N samples is referred to as a frame. In one illustrative example, each frame can be 10-20 milliseconds (ms) in length.

[0051] Various voice coding algorithms can be used to encode a speech signal. For instance, code-excited linear prediction (CELP) is one example of a voice coding algorithm. The CELP model is based on a source-filter model of speech production, which assumes that the vocal cords are the source of spectrally flat sound (an excitation signal), and that the vocal tract acts as a fdter to spectrally shape the various sounds of speech. The different phonemes (e.g., vowels, fricatives, and voice fricatives) can be distinguished by their excitation (source) and spectral shape (filter).

[0052] In general, CELP uses a linear prediction (LP) model to model the vocal tract, and uses entries of a fixed codebook (FCB) as input to the LP model. For instance, long-term linear prediction can be used to model pitch of a speech signal, and short-term linear prediction can be used to model the spectral shape (phoneme) of the speech signal. Entries in the FCB are based on coding of a residual signal that remains after the long-term and short-term linear prediction modeling is performed. For example, long-term linear prediction and short-term linear prediction models can be used for speech synthesis, and a fixed codebook (FCB) can be searched during encoding to locate the best residual for input to the long-term and short-term linear prediction models. The FCB provides the residual speech components not captured by the short-term and long-term linear prediction models. A residual, and a corresponding index, can be selected at the encoder based on an analysis-by-synthesis process that is performed to choose the best parameters so as to match the original speech signal as closely as possible. The index can be sent to the decoder, which can extract the corresponding LTP residual from the FCB based on the index.

[0053] FIG. 2 is a block diagram illustrating an example of a CELP-based voice coding system 200, including a voice encoder 202 and a voice decoder 204. The voice encoder 202 can obtain a speech signal 201 and can segment the samples of the speech signal into frames and sub-frames. For instance, a frame of N samples can be divided into sub-frames. In oneillustrative example, a frame of 240 samples can be divided into four sub-frames each having 60 samples. For each frame, sub-frame, or sample, the voice encoder 202 chooses the parameters (e.g.. gain, filter coefficients or linear prediction (LP) coefficients, etc.) for a synthetic speech signal so as to match as much as possible the synthetic speech signal with the original speech signal.

[0054] The voice encoder 202 can include a short-term linear prediction (LP) engine 210, a long-term linear prediction (LTP) engine 212. and a fixed codebook (FCB) 214. The shortterm LP engine 210 models the spectral shape (phoneme) of the speech signal. For example, the short-term LP engine 210 can perform a short-term LP analysis on each frame to yield linear prediction (LP) coefficients. In some examples, the input to the short-term LP engine 210 can be the original speech signal or a pre-processed version of the original speech signal. In some implementations, the short-term LP engine 210 can perform linear prediction for each frame by estimating the value of a current speech sample based on a linear combination of past speech samples. For example, a speech signal s(n) can be represented using an autoregressive (AR) model, such as s(n) = ™=iaksCn~ k) + e(n), where each sample is represented as a linear combination of the previous m samples plus a prediction error term e(n). The weighting coefficients al, a2, through am are referred to as the LP coefficients. The prediction error term e(n) can be found as follows: e(n — s(n) — k=1aks(n — k). By minimizing the mean square prediction error with respect to the filter coefficients, the short-term LP engine 2f0 can obtain the LP coefficients. The LP coefficients can be used to form an analysis filter:= 1 - P(z).

[0055] The short-term LP engine 210 can solve for P(z) (which can be referred to as a transfer function) by computing the LP coefficients (ak) that minimize the error in the above AR model equation (s(n)) or other error metric. In some implementations, the LP coefficients can be determined using a Levinson-Durbin method, a Leroux-Gueguen algorithm, or other suitable technique. In some examples, the voice encoder 202 can send the LP coefficients to the voice decoder 204. In some examples, the voice decoder 204 can determine the LP coefficients, in which case the voice encoder 202 may not send the LP coefficients to the voicedecoder 204. In some examples, Line Spectral Frequencies (LSFs) can be computed instead of or in addition to LP coefficients.

[0056] The LTP engine 212 models the pitch of the speech signal. Pitch is a feature that determines the spacing or periodicity of the impulses in a speech signal. For example, speech signals are generated when the airflow from the lungs is periodically interrupted by movements of the vocal cords. The time between successive vocal cord openings corresponds to the pitch period. The LTP engine 212 can be applied to each frame or each sub-frame of a frame after the short-term LP engine 210 is applied to the frame. The LTP engine 212 can predict a current signal sample from a past sample that is one or more pitch periods apart from a current sample (hence the term “long-term”). For instance, the current signal sample can be predicted as pr(n) = gpr(n — T), where T denotes the pitch period. gpdenotes the pitch gain, and r(n — T) denotes an LP residual for a previous sample one or more pitch periods apart from a current sample. Pitch period can be estimated at every frame. By comparing a frame with past samples, it is possible to identify the period in which the signal repeats itself, resulting in an estimate of the actual pitch period. The LTP engine 212 can be applied separately to each subframe.

[0057] The FCB 214 can include a number (denoted as L) of long-term linear prediction (LTP) residuals. An LTP residual includes the speech signal components that remain after the long-term and short-term linear prediction modeling is performed. The LTP residuals can be, for example, fixed or adaptive and can contain deterministic pulses or random noise (e.g., white noise samples). The voice encoder 202 can pass through the number L of LTP residuals in the FCB 214 a number of times for each segment (e.g.. each frame or other group of samples) of the input speech signal, and can calculate an error value (e g., a mean-squared error value) after each pass. The LTP residuals can be represented using codevectors. The length of each codevector can be equal to the length of each sub-frame, in which case a search of the FCB 214 is performed once every sub-frame. The LTP residual providing the lowest error can be selected by the voice encoder 202. The voice encoder 202 can select an index corresponding to the LTP residual selected from the FCB 214 for a given sub-frame or frame. The voice encoder 202 can send the index to the voice decoder 204 indicating which LTP residual is selected from the FCB 214 for the given sub-frame or frame. Again associated with the lowest error can also be selected and send to the voice decoder 204.

[0058] The voice decoder 204 includes an FCB 224, an LTP engine 222, and a short-term LP engine 220. The FCB 224 has the same LTP residuals (e.g., codevectors) as the FCB 214. The voice decoder 204 can extract an LTP residual from the FCB 224 using the index transmitted to the voice decoder 204 from the voice encoder 202. The extracted LTP residual can be scaled to the appropriate level and filtered by the LTP engine 222 and the short-term LP engine 220 to generate a reconstructed speech signal 205. The LTP engine 222 creates periodicity in the signal associated with the fundamental pitch frequency, and the short-term LP engine 220 generates the spectral envelope of the signal.

[0059] Other linear predictive-based coding systems can also be used to code voice signals, including enhanced voice services (EVS), adaptive multi -rate (AMR) voice coding systems, mixed excitation linear prediction (MELP) voice coding systems, linear predictive coding- 10 (LPC-10), among others.

[0060] A voice codec for some applications and / or devices (e.g., Intemet-of-Things (loT) applications and devices) is required to deliver higher quality coding of speech signals at low bit-rates, with low complexity, and with low memory requirements. Existing linear predictivebased codecs cannot meet such requirements. For example, ACELP-based coding systems provide high quality, but do not provide low bit-rate or low complexity / low memory. Other linear-predictive coding systems provide low bit-rate and low complexity / low memory, but do not provide high quality.

[0061] In some cases, machine learning systems (e.g., using one or more neural network models) can be used to generate reconstructed voice or audio signals. For example, using features extracted from a frame of audio data, a neural network-based voice decoder can generate coefficients for at least one linear filter. In some examples, a neural network-based voice decoder can be trained to estimate the parameters (e.g., linear filter coefficients) of a speech synthesis pipeline. The linear filter(s) that are generated or tuned using a neural network-based voice decoder can then be used to generate a reconstructed signal.

[0062] However, such a neural network-based voice decoder can be highly complex and resource intensive. In some examples, a neural network-based voice decoder can be trained to directly predict speech samples. For instance, the neural network-based voice decoder can be trained to perform the operations of a linear predictive filter (LPC), such as the short-term LPengine 210 of FIG. 2. Such LPC operations can include complex operations that require the use of a large amount of computing resources by the neural network-based voice decoder.

[0063] In some examples, a neural network-based voice decoder can be trained to perform one or more non-linear operations (e.g., a non-linear transform, a non-linear likelihood model, etc.). The non-linear operations can be performed directly on input features obtained for a voice or audio signal and / or can be performed on a signal that was itself generated directly from the input features. Non-linear transforms and other non-linear operations can provide greater flexibility, accuracy, and output quality when used in a neural network-based voice decoder.

[0064] FIG. 3 is a block diagram of an example voice decoding system 300, in accordance with aspects of the present disclosure. As shown in FIG. 3, a set of features 302 may be provided as input to a neural network-based non-linear engine 304 (e.g., neural network-based likelihood speech model). The voice decoding system 300 can be configured to decode data of a compressed speech signal, shown in FIG. 3 as features 302 including a speech feature vector S[m] for a frame m, to generate a reconstructed speech signal x[n] (also referred to as a synthesized speech sample) for a cunent time instant n that approximates an original speech signal that was previously compressed by a voice encoder (not shown) and where m indicates a frame index. In some cases, speech features may be analyzed on a per frame basis for frames of reconstructed speech signal xfnf. In some examples, the speech feature vector S[m] can be a log-Mel spectrogram generated by a feature extraction engine (not shown). The neural network-based non-linear engine 304 can receive as input one or more of the features 301 (e.g., from speech feature vector S[mJ). In some cases, the neural network-based non-linear engine 304 may be trained to perform a non-linear operation over a received input and provide an output such as the time-domain excitation signal y[n], For instance, the features 302 can be input to the neural network-based non-linear engine 304, which is trained to generate a probability distribution (e.g., likelihood model) P(s[n] | s[n — l], s[n — 2], ... , s[n — M], features) The neural network-based non-linear engine 304 can be neural network-based. In some examples, neural network-based non-linear engine 304 can be trained directly and subsequently utilized as a trained component. In some examples, the neural network-based non-linear engine 304 can be based on a Wavenet, WaveRNN, LPCNet, and / or other likelihood models.

[0065] Output from the neural network-based non-linear engine 304 may be input to a sampling process 306 to sample from the probability7distribution P(s[n] | s[n — l], s[n — 2], ... , s[n — M], features') and output 308 a sampled signal s(h) that is in the time domain. The type of output 308 generated therefore can, in some examples, be dependent upon the nature and type of the sampling that is performed in sampling process 306. By adjusting sampling process 306, the output 308 can also be adjusted.

[0066] FIG. 4 is a block diagram illustrating a sample-by-sample autoregressive ML-based speech synthesizer 400, in accordance with aspects of the present disclosure. In FIG. 4, features 402 may be input to a LPC engine 404 and a frame rate network 406. The frame rate network 406 may be implemented by a as a ML model, such as a neural network. The frame rate network 406 may output frame conditioning features from the features 402 to a sample rate network 408. The LPC engine 404 may output LPC information to a prediction engine 410 which may generate a prediction based on the LPC information and a previously predicted sample from the sample rate network 408. In some cases, the prediction engine 410 may be implemented as a ML model, such as a neural network, trained to predict a current sample (pt) based on linear prediction co-efficients (ak) and previous samples (st-k) as shown by the equation Pt= ' k=iakst-k- The predicted current sample along with previously predicted sample and previously predicted excitation may be input to the sample rate network 408. The sample rate network 408 may predict an output probability distribution for the excitation. In some cases, the frame rate network 406 may operate based on frame level features, while the sample rate network 408 handles sample-to-sample predictions. With a goal of predicting the excitation, frame rate network parameters may be learned together with the sample rate parameters. In some cases, the frame rate network 406 and the sample rate network 408 may be trained together to predict the target excitation. If the excitation is designed to be discrete sample values [e.g. 1, 2, ... 256], then the training can be done using cross-entropy loss function (e.g., between target and predicted excitation).

[0067] In some cases, the sample rate network 408 may include multiple gated recunent units (GRUs). In some cases, the GRUs may be used to remember previous samples, such as those from a previous pitch period, along with previously predicted sample(s), pitch lag, and gain information to predict a current sample. This can result in a relatively complex GRU that may be computationally expensive. In some cases, reducing the complexity of the GRU may result in warble in the pitch pulses in reconstructed speech. Thus, techniques to improve pitchcontent in neural speech synthesis, for example, by reducing GRU complexity while avoiding warble may be useful.

[0068] In some cases, as there is a predictable pitch periodicity and temporal structure in human speech (e.g., based on vocal cord movements) a pitch lag may be incorporated to help bring in previously estimated samples around a previous pitch period to post-process a current predicted sample. For example, a pitch of a current sample may be consistent with a pitch of a previous sample generated around the pitch period. The pitch lag (e.g.. pitch delay, pitch period) may be the amount of time between repetitions of a speech signal (e.g., based on movements of the vocal cords).

[0069] FIG. 5 is a block diagram illustrating a postprocessing process for improving pitch content 500, in accordance with aspects of the present disclosure. In FIG. 5, a sample rate network 502 may output predicted samples to a post-processing engine 504. The sample rate network 502 may be a part of a sample-by-sample autoregressive ML -based speech synthesizer, such as one shown in FIG. 4. In some cases, the sample rate network 502 may include simplified GRUs, as compared to sample rate network 408 of FIG. 4. The postprocessing engine 504 may also receive a set of previously predicted samples 506 from a delay circuit 510. The delay circuit 510 may receive information about a pitch lag 512 indicating an amount of time (T) of the pitch lag (e.g., pitch delay, pitch period). In some cases, the set of previously predicted samples 506 may include previously predicted samples 506 from a window around time (T), such as five samples from T-2 - T+2. For example, the delay circuit 510 may save previously predicted samples and select previously predicted samples to include in the window of time based on the pitch lag. In some cases, this window of time may be fixed. In other cases, the previously predicted samples 506 may be all of the previously predicted samples 506 from the previous pitch period. In yet other cases, the previously predicted samples 506 may be from a variable window around time (T). In some cases, previously predicted samples from around the pitch period (e.g., time (T)) may be included as the exact pitch period may vary slightly between one pitch period and a next pitch period.

[0070] In some cases, the post-processing engine 504 may receive the predicted sample 516 (S't) from the sample rate netw ork 502 along with the set of previously predicted samples 506, and a pitch gain 508 to estimate an improved version of the predicted sample 516 (S't) from the sample rate network 502 for output 514 (St). The post-processing engine 504 may betrained to predict relationships between the predicted sample 516 (S't) and previously predicted samples from the set of previously predicted samples 506 from the previous pitch period (e.g., at time (n-T), where n is a current time and S(n) represents a sample at the current time n) as a function of the pitch gain 508 to improve the predicted sample 516 (S't). The pitch gain 508 may indicate how correlated one pitch period is with another pitch period and may indicate how well the previous sample can be used to influence the current sample. In some cases, the post-processing engine 504 may be ML based. For example, the post-processing engine 504 may be structured as a convolutional neural network (CNN), feed-forward network. RNN (if implemented as a time-series), among other ML network structures.

[0071] In some cases, the output 514 (St) of the post-processing engine 504 may be stored in the delay circuit 510 and input to the sample rate network 502 for predicting a next sample. In some cases, storing the set of previously predicted samples 506 from the previous pitch period in the delay circuit may help relieve a burden from the sample rate network 502 to remember (e g., store) previously predicted samples. In some cases, the pitch period may be determined, for example, by any technique for determining a pitch period (T), such as a non- exhaustive open loop delay search as described in Enhanced Variable Rate Codec, Speech Service Options 3, 68, 70, and 73 for Wideband Spread Spectrum Digital Systems 3GPP2 C.S0014-D. Similarly, the pitch gain 508 may also be determined using any technique for determining the pitch gain, such as based on refining the open loop delay estimate as described in Enhanced Variable Rate Codec, Speech Service Options 3. 68. 70. and 73 for Wideband Spread Spectrum Digital Systems 3GPP2 C.S0014-D. While shown in context with a samplerate network 502, it may be understood that the techniques described herein may be applied to any sample-by-sample autoregressive ML-based speech synthesizer that can accept information about samples from the previous pitch period.

[0072] FIG. 6 is a block diagram illustrating an integrated process for improving pitch content 600, in accordance with aspects of the present disclosure. In some cases, a separate post-processing process may be skipped and a set of previously predicted samples 602 may be directly input to a sample rate network 604. The sample rate network 604 may be a part of a sample-by-sample autoregressive ML-based speech synthesizer, such as one shown in FIG. 4.

[0073] Output 608 (Sn) of the sample rate network 604 may be passed into a first delay circuit 606. The first delay circuit 606 may be substantially similar to delay circuit 510 of FIG.5. For example, the output 608 (Sn) of the sample rate network 604 may be stored in the first delay circuit 606. The first delay circuit 606 may also receive information about a pitch lag 610. In some cases, information about the pitch lag 610 may be substantially similar to information about pitch lag 512 of FIG. 5. The first delay circuit 606 may output the set of previously predicted samples 602 to the sample rate network 604. As describe above, the set of previously predicted samples 602 may include samples based on the previous pitch period (e.g., n-T), and include samples near the previous pitch period (e.g., n-T, n+T) within a fixed or variable window, or samples for an entire previous pitch period.

[0074] Output of the sample rate network 604 may also be passed into a second delay circuit 612. The second delay circuit 612 may provide an immediate past predicted sample (e.g., n-1) to the sample rate network 604. In some cases, the sample rate network 604 may be trained to use the set of previously predicted samples 602 and immediate past predicted sample to predict relationships between the previously predicted samples and a current sample being processed by the sample rate network 604 to generate a predicted sample for output 608. While shown in context with a sample-rate network 604, it may be understood that the techniques described herein may be applied to any sample-by-sample autoregressive ML-based speech synthesizer that can accept information about samples from the previous pitch period.

[0075] FIG. 7A is a block diagram illustrating vector-based process for improving pitch content 700, in accordance with aspects of the present disclosure. FIG. 7A includes a sample rate network 702 that may be similar to sample rate network 604 of FIG. 6. The sample rate network 702 may be a part of a sample-by-sample autoregressive ML-based speech synthesizer, such as one shown in FIG. 4.

[0076] Output 704 (Sn) of the sample rate network 702 may be passed into a first delay circuit 706. The first delay circuit 706 may be substantially similar to delay circuit 510 of FIG.5 and / or delay circuit 606 of FIG. 6. For example, the output 704 (Sn) of the sample rate network 702 may be stored in the first delay circuit 706. The first delay circuit 706 may output a set of previously predicted samples 708 to a vector processing engine 710. The vector processing engine 710 may also receive a pitch gain 712. In some cases, the pitch gain 712 may be similar to pitch gain 508 of FIG. 5.

[0077] The vector processing engine 710 may generate a context vector 714 describing the input set of previously predicted samples 708 and pitch gain 712 from the previous pitch period(e.g., n-T as pitch lag). In some cases, the vector processing engine 710 may also receive an immediately previous pitch sample (e.g., S(n-l)) and include a description of the immediately previous pitch sample in the context vector 714. The context vector 714 may be passed to the sample rate network 702. In some cases, the vector processing engine 710 may be implemented as a post-processing operation.

[0078] The context vector 714 may be a consolidated vector based on the previous pitch samples and the context vector 714 may be generated based on an attention mechanism 718. For example, the attention mechanism 718 may be ML based, such as with an RNN. The attention mechanism 718 may look at how the current sample attends to samples around the previous pitch period based on the pitch gain 712. A weighted combination / sum may be determined based on how closely matching the previously generated samples may be to the current sample. For example, the attention mechanism 718 may apply a greater weight to those previously predicted samples that more closely match the current sample, for example, those samples closest to S(n-T) and less weights to previously predicted samples further away in time from the pitch period. These weighted combination / sums may be included in the context vector. As describe above, the set of previously predicted samples 708 may include samples based on the previous pitch period (e.g., n-T), and include samples near the previous pitch period (e.g., n-T, n+T) within a fixed or variable window; or samples for an entire previous pitch period.

[0079] Output of the sample rate network 702 may also be passed into a second delay circuit 716. The second delay circuit 716 may be substantially similar to second delay circuit 612 of FIG. 6 and may provide an immediate past predicted sample (e.g., n-1) to the sample rate network 702. In some cases, the sample rate network 702 may be trained to use the context vector 714 and immediate past predicted sample to predict relationships between the previously predicted sample and a current sample being processed by the sample rate network 702 to generate a predicted sample for output 704.

[0080] FIG. 7B is a block diagram illustrating an example of an attention mechanism 720, in accordance with aspects of the present disclosure. In some cases, the attention mechanism 720 may be used in place of the attention mechanism 718 of FIG. 7A. The attention mechanism 720 may be a cross-attention mechanism. The attention mechanism 720 may include a concatenation layer 722 which may concatenate the output 704 (Sn) of the sample rate network 702 with the set of previously predicted samples 708 to generate concatenated data. In somecases, the pitch gain 712 may also be concatenated. The concatenated data may then be passed through a first linear layer 724, non-linear activation layer 726, second linear layer 728, and softmax layer 730 to obtain attention weights 732. The attention weights may then be applied to the samples of the set of previously predicted samples 708 to weight the samples.

[0081] FIG. 7C is a block diagram illustrating another example of an attention mechanism 750, in accordance with aspects of the present disclosure. In some cases, the attention mechanism 750 may be used in place of the attention mechanism 718 of FIG. 7A. In FIG. 7C. the output 704 may be input as a vector and multiplied by the set of previously predicted samples 708, and then scaled, optionally masked, and a softmax function applied to obtain a set of weights that may then be multiplied by the set of previously predicted samples 708 to weight the samples.

[0082] FIG. 8 is a block diagram illustrating a postprocessing process for improving pitch content 800, in accordance with aspects of the present disclosure. In some cases, the postprocessing process for improving pitch content may be applied independently of a sample- by-sample autoregressive ML-based speech synthesizer 802, such as a neural homomorphic vocoder (NHV). In some cases, the NHV may use a neural network, such as a neural network (NN) filter estimator 804 to control one or more filters, such as LTV filters of the speech synthesizer 802. In some cases, predicted samples 806 (S't) output by the speech synthesizer 802 may be passed to a post-processing engine 808. In some cases, the post-processing engine 808 may be similar to post-processing engine 504 of FIG. 5. In some cases, the post-processing engine 808 may be ML based. The post-processing engine 808 may receive a set of previously predicted samples 810 from a delay circuit 812 and a pitch gain 814 and the post-processing engine 808 may estimate an improved version of the predicted sample 806 (S't) from the speech synthesizer 802 for output 816 (St). In some cases, the pitch gain 814 may be substantially similar to pitch gain 508 of FIG. 8. The delay circuit 812 may be substantially similar to delay circuit 510 of FIG. 5. The delay circuit 812 may store the output 816 (St) of the post-processing engine 808 for input back to the post-processing engine at a later pitch period. The delay circuit may also receive pitch lag information 818, which may be used to determine the set of previously predicted samples 810. As describe above, the set of previously predicted samples 810 may include samples based on the previous pitch period (e.g., n-T), and include samples near the previous pitch period (e.g., n-T, n+T) within a fixed or variable window, or samples for an entire previous pitch period. Of note, while speech synthesizer 802 is used in thisexample, it may be understood that the postprocessing process for improving pitch content 800 may be used with other sample-by-sample autoregressive ML-based speech synthesizers, such as Lyra.

[0083] FIG. 9 is a block diagram illustrating an integrated process for improving pitch content 900, in accordance with aspects of the present disclosure. FIG. 9 includes a speech synthesizer 902, such as a NHV based speech synthesizer. As indicated above, the speech synthesizer 902 may include an NN filter estimator 904 to control one or more filters of the speech synthesizer 902. The speech synthesizer 902 may output predicted samples 906 S(n). The output predicted samples 906 may be input to a delay circuit 908. In some cases, delay circuit 908 may receive pitch lag information 910. The delay circuit 908 may be substantially similar to first delay circuit 606 of FIG. 6. The delay circuit 908 may output a set of previously predicted samples 912 to the NN filter estimator 904 based on the pitch lag information 910. As describe above, the set of previously predicted samples 912 may include samples based on the previous pitch period (e.g.. n-T), and include samples near the previous pitch period (e.g., n-T, n+T) within a fixed or variable window, or samples for an entire previous pitch period. In some cases, the NN filter estimator 904 may be trained to use the set of previously predicted samples 912 to help predict relationships between the previously predicted samples and a current sample being processed by the NN filter estimator 904 to generate a predicted sample for output (e.g., as output predicted samples 906).

[0084] FIG. 10 is a block diagram illustrating an example of an audio codec system 1000 that can be used to generate reconstructed audio (e.g., synthesized speech) using a neural speech synthesizer. In some aspects, the audio codec system 1000 can also be referred to as a voice coding system (e.g., vocoder), a signal synthesis system, a voice coding signal synthesis system, etc. In one illustrative example, the audio codec system 1000 can be a generative audio codec (e.g., a generative voice codec) including one or more feedback recurrent autoencoders (FRAEs) that may be used to generate encoded audio data, and one or more neural synthesizers that may be used to generate reconstructed audio data based on the encoded audio data.

[0085] For example, the audio codec system 1000 (e.g., a generative voice codec) can include an encoder 1005 (e.g., a transmitter) and a decoder 1010 (e.g., a receiver). The encoder 1005 can be used to generate encoded features that are transmitted to the decoder 1010 over a channel 1040. For example, the encoder 1005 can receive audio 1015 and generate encodedfeatures corresponding to the audio 1015. The encoded features can be generated using a feedback recurrent autoencoder (FRAE) 1025 included in the encoder 1005. In some examples, the encoder 1005 can include one or more FRAEs, where each FRAE of the one or more FRAEs is configured to generate a corresponding one or more encoded features associated with the audio 1015. In some aspects, the audio 1015 may include and / or comprise a speech signal, a voice signal, etc.

[0086] The decoder 1010 can receive encoded features (e.g., from the encoder 1005. over the channel 1040) and generate reconstructed audio 1055 based at least in part on the encoded features received from the encoder 1005. The reconstructed audio 1055 may include and / or comprise a reconstructed speech signal, a reconstructed voice signal, etc. The reconstructed audio 1055 can be generated using a neural network-based speech synthesizer 1050 (e.g., also referred to as a neural speech synthesizer and / or a neural synthesizer) included in the decoder 1010. The reconstructed audio 1055 generated using the neural speech synthesizer 1050 may also be referred to as synthesized speech. In some examples, the decoder 1010 can include an FRAE decoder 1045, which can be used to decode the encoded features received from the encoder 1005. In some aspects, the FRAE decoder 1045 can be the same as or similar to a decoder implemented by the FRAE 1025 included in the encoder 1005. In some examples, the decoder 1010 can include one or more FRAE decoders 1045, which can correspond to one or more FRAEs 1025 included in the encoder 1005. For example, a number of FRAE decoders 1045 included in the decoder 1010 can be equal to the number of FRAE 1025 included in the encoder 1005.

[0087] In some examples, the audio 1015 includes speech. The audio 1015 may be an example of the speech signal 101 of FIG. 1, and / or the speech signal 201 of FIG. 2, etc. The encoder 1005 of the audio codec system 1000 can be configured to generate one or more encoded representations of the input audio 1015. For example, the one or more encoded representations can include spectral envelope information associated with the audio 1015, and / or can include pitch information associated with the audio 1015, etc. In some aspects, the encoder 1005 can extract spectral envelope features from the audio 1015 using spectral envelope feature extraction engine 1020, and can encode the extracted spectral envelope features using the FRAE 1025. The encoded spectral features z can be transmitted from the FRAE 1025 to the decoder 1010, using the channel 1040. In some examples, the encoder 1005 can extract pitch information from the audio 1015 using pitch extraction engine 1030, and canperform quantization 1035 to generate quantized pitch information associated with the audio 1015. The quantized pitch information can be transmitted from the pitch quantizer 1035 to the decoder 1010, using the channel 1040.

[0088] In some aspects, the encoder 1005 can be configured to extract spectral features (e.g., spectral envelope features) from the audio 1015 using the spectral envelope feature extraction engine 1020. In one illustrative example, the spectral envelope feature extraction engine 1020 can perform a cepstrum computation to determine cepstrum information and / or one or more cepstral coefficients corresponding to the audio 1015. In some cases, the spectral features extracted from the audio 1015 using the spectral envelope feature extraction engine 1020 can include mel-frequency cepstral coefficients (MFCC), such as MFCC-24 features (e.g., 24-dimensional MFCC features). In some examples, the spectral envelope feature extraction engine 1020 applies one or more mel-scaled filter(s) (e.g., a mel-scaled filterbank), a logarithmic compression, and / or a discrete cosine transform (DCT) to the audio 1015 (e.g., and / or to a magnitude spectrum associated with the audio 1015). The encoder 1005 processes the extracted spectral features using a feedback recurrent autoencoder (FRAE) 1025 to generate encoded features z. The FRAE 1025 includes a decoder and an encoder. The FRAE 1025 can implement feedback of state information h between the decoder and the encoder included in the FRAE 1025. For example, the state information h can be determined or obtained at the decoder of the FRAE 1025, and feedback of the state information h can be performed for the decoder of the FRAE 1025 (e.g.. the state information h is fed back to the decoder of the FRAE 1025) and for the encoder of the FRAE 1025 (e.g., the state information h is fed back from the decoder of the FRAE 1025 to the encoder of the FRAE 1025).

[0089] In some examples, the spectral envelope feature extraction engine 1020 can be configured to generate and / or determine (e.g., extract) one or more types of spectral envelope features. For example, the extracted spectral envelope features may include cepstrum information and / or cepstral coefficients. In some cases, cepstral liftering can be performed based on DCT truncation to exclude or remove pitch information and only capture spectral envelope information in the extracted cepstrum or cepstral coefficients. In some examples, the extracted spectral envelope features can include companded (e.g., log) filterbank energies associated with the input audio 1015. For example, companded filterbank energies can be determined based on applying an inverse DCT (e.g.. IDCT) to a liftered cepstrum. In some cases, the extracted spectral envelope features can include filterbank energies determined basedon uncompanding (e.g., exp) the companded filterbank energies. In some examples, the extracted spectral envelope features can include a full resolution spectrum (e.g., DFT domain with companded (e.g., log. linear amplitude, etc.) information, etc ). The full resolution spectrum can be smoothed to the envelope of the spectrum, for example based on interpolating the companded or uncompanded filterbank energies. In some aspects, the extracted spectral envelope features can include one or more linear prediction (LP) coefficients, for example determined based on processing one or more speech frames of the input audio 1015 using the Levinson-Durbin algorithm (e.g., autocorrelation technique), and / or using a covariance technique. In some examples, the extracted spectral envelope features can include one or more of line spectral frequency (LSF) information and / or line spectral pair (LSP) information.

[0090] In some aspects, the encoded features or encoded feature information transmitted from the generative voice codec system encoder 1005 can comprise a latent representation z between the encoder of the FRAE 1025 and the decoder of the FRAE 1025. The encoder 1005 can be configured to pass (e.g., transmit) the encoded features z through a channel 1040 to the decoder 1010.

[0091] The encoder 1005 also processes the audio 1015 using a pitch extraction engine 1030 to extract pitch information from the audio 1015. In some aspects, the audio 1015 can be processed in parallel by the spectral envelope feature extraction engine 1020 and the pitch extraction engine 1030. In some cases, the encoder 1005 can include a denoiser 1018 that processes the input audio 1015 and provides a de-noised audio to the spectral envelope feature extraction engine 1020 and / or to the pitch extraction engine 1030.

[0092] Based on the audio 1015. the pitch extraction engine 1030 can generate one or more types of pitch information. For example, the pitch extraction engine 1030 can output pitch information in the frequency -domain (e.g., fo pitch information of the audio 1015, in units of Hertz (Hz)) and / or can output pitch information in the time-domain (e.g., pitch lag in samples, with or without a fractional component, and / or pitch lag in milliseconds). In some cases, the pitch extraction engine 1030 can be used to generate pitch estimation indicative of a pitch lag (or pitch delay) from the audio 1015, and / or to identity’ a pitch correlation from the audio 1015. In some aspects, the pitch information generated by the pitch extraction engine 1030 can include a pitch lag and a pitch correlation.

[0093] The encoder 1005 can be configured to processes the pitch information of the audio 1015 (e.g., pitch, pitch lag, and / or pitch correlation, determined using the pitch extraction engine 1030) using a quantizer 1035 < () to generate a quantized pitch signal. The encoder 1005 passes the quantized pitch signal through the channel 1040 to the decoder 1010. In some examples, the quantizer 1035 Q() may also be referred to as a pitch quantizer. In some aspects, the quantizer 1035 O() can perform vector quantization (VQ) to generate the quantized pitch signal for transmission to the decoder 1010 over the channel 1040. In some examples, the quantizer 1035 < () can perform single-stage VQ and / or can perform multi-stage VQ (e.g., MSVQ). In some cases, the quantizer 1035 Q() can be implemented using one or more FRAEs. In some aspects, the quantizer 1035 Q() can be configured to implement forward error correction (FEC) for the quantized pitch signal that is transmitted to the decoder 1010 over the channel 1040. For example, the quantizer 1035 <9() can implement multiple description coding (MDC) for the transmission of the quantized pitch signal over the channel 1040, can implement full-redundancy FEC for the transmission of the quantized pitch signal over the channel 1040, etc.

[0094] The generative voice codec system decoder 1010 can be configured to receive the encoded features z from the FRAE 1025 included in the generative voice codec system encoder 1005. For example, the decoder 1010 can receive the encoded features z via the channel 1040. The decoder 1010 decodes the encoded features z using a FRAE 1045 to generate decoded features. In some examples, the FRAE 1045 of the decoder 1010 includes only a decoder, without an encoder. In some examples, the FRAE 1045 of the decoder 1010 can include an encoder. In some examples, the decoded features generated by the FRAE 1045 include mel- frequency cepstral coefficients (MFCC), such as MFCC-24 features (e.g., 24-dimensional MFCC features).

[0095] The decoded features determined using the FRAE decoder 1045 can be provided to a neural speech synthesizer 1050, which can be configured to generate a reconstructed audio 1055 (e.g., synthesized speech) based at least in part on the decoded spectral envelope features from the FRAE decoder 1045. In one illustrative example, the neural speech synthesizer 1050 receives as input the decoded spectral envelope features (e.g., from the FRAE decoder 1045) and the received pitch encoding information (e.g., the quantized pitch signal received by the decoder 1010 over the channel 1040 and from the pitch quantizer 1035 of the encoder 1005). In some aspects, the neural speech synthesizer 1050 can generate the reconstructed audio 1055(e.g., synthesized speech) based on processing the decoded spectral envelope features along with a pitch signal (e.g., the quantized pitch signal or a reconstructed variant thereof).

[0096] For example, the decoder 1010 can receive the quantized pitch signal (e.g., indicative of the pitch, the pitch lag, and / or the pitch correlation of the audio 1015) from the encoder 1005 via the channel 1040. In some examples, the decoder 1010 passes the quantized pitch signal to the neural speech synthesizer 1050, and the neural speech synthesizer 1050 processes the decoded spectral envelope features along with the quantized pitch signal to generate reconstructed audio 1055 (e.g., synthesized speech).

[0097] In some cases, the decoder 1010 can receive the quantized pitch signal (e.g., indicative of the pitch, the pitch lag, and / or the pitch correlation of the audio 1015) from the encoder 1005 via the channel 1040, and can perform dequantization of the quantized pitch signal to obtain a reconstructed pitch signal. The reconstructed pitch signal can be passed to the neural speech synthesizer, and used in combination with the decoded spectral envelope features from the FRAE decoder 1045 to generate the reconstructed audio 1055 (e.g., synthesized speech). In some examples, the decoder 1010 can include a reconstruction engine that is separate from the neural speech synthesizer 1050. The reconstruction engine processes the quantized pitch signal to reconstruct the pitch signal (e.g., including the pitch, the pitch lag, and / or the pitch correlation) before the neural speech synthesizer 1050 using the reconstructed pitch signal (e.g.. the pitch, the pitch lag, and / or the pitch correlation) to generate the reconstructed audio 1055.

[0098] In some examples, the output of the neural speech synthesizer 1050 can be provided to one or more linear predictive coding (LPC) layers 1052 included in the decoder 1010. The LPC layers 1052 can be used to perform linear prediction analysis and / or linear prediction synthesis, based on the output of the neural speech synthesizer 1050. For example, the LPC layers 1052 can perform linear prediction based on the output of the neural speech synthesizer 1050, to generate the reconstructed audio 1055 (e.g., synthesized speech). In some aspects, the decoder 1010 does not include the LPC layers 1052, and the neural speech synthesizer 1050 can be trained and / or configured to generate the reconstructed audio 1055 directly (e.g., the output of the neural speech synthesizer 1050 can be the reconstructed audio 1055).

[0099] In some aspects, the LPC layers 1052 can be used to implement a linear prediction (LP) synthesis filter, based on:

[0100] Here, x[n] represents the input signal to the LP filter (e.g., the input signal to the LPC layers 1052), y[n] represents the output signal (e.g., the output of the LPC layers 1052, which can be the reconstructed audio 1055), akrepresents the linear prediction coefficients associated with implementing the LP synthesis filter, and p is a value corresponding to the LP filter order. For example, the LP filter order p can have a value of 16 or 12, etc., among various other LP filter order values.

[0101] In some aspects, the LP synthesis filter associated with the LPC layers 1052 can be implemented in the time domain or in the frequency domain. For example, in the time domain, the LP synthesis filter can be implemented based on a difference equation. In some cases, the LP synthesis filter can be implemented in the time domain by convolving the input signal x[n] with the LP filter impulse response or an approximation of the LP filter impulse response. For example, the LP filter can be associated with an infinite impulse response (HR), which can be approximated using a finite segment of the IIR (e.g., such as the first N samples of the IIR for a configured integer value of N. etc ). The LP synthesis filter associated with the LPC layers 1052 can be implemented in the frequency domain based on multiplying the FFT of the input signal x[n ] with the frequency response of the LP filter, and then determining the IFFT of the result to convert the output of the frequency domain LP synthesis filter into a time domain signal corresponding to the reconstructed audio 1055.

[0102] In some examples, the LP coefficients akcan be estimated from the decoded spectral envelope features obtained using the FRAE decoder 1045. In some cases, the spectral envelope features can be LP coefficients (e.g., determined by the spectral envelope feature extraction engine 1020 based on using the Levinson-Durbin algorithm or other autocorrelation technique, or using a covariance technique, to process the input speech frames of the audio 1015), and the decoded LP coefficients on the decoder side can be used for the LP filter implemented by the LPC layers 1052.

[0103] In some examples, the extracted spectral envelope features may be LSFs or LSPs, and the decoded LSFs or LSPs obtained using the FRAE decoder 1045 can be converted to the corresponding LP coefficients ak. In some cases, the LP coefficients akof the LPC layers 1052 can be estimated from the spectral envelope features using a neural network-based and / or aDSP-based technique. For example, the extracted spectral envelope features may comprise a cepstrum, and a DSP-based technique can be used to convert the decoded cepstrum into filterbank energies, and subsequently interpolate the filterbank energies to estimate an FFT square magnitude at all FFT bins (e.g., including bins beyond the centers of filterbank filters). After estimating the FFT square magnitude based on the interpolated filterbank energies, the DSP-based technique can include applying an IFFT to obtain an estimate of the autocorrelation sequence, and utilizing the Levinson-Durbin algorithm on the estimated autocorrelation sequence to obtain the LP coefficients ak.

[0104] In some examples, the spectral envelope feature extraction engine 1020 can be performed based at least in part on spectral feature learning. For example, spectral feature learning can be performed by one or more machine learning models configured and / or trained to generate as output the one or more extracted spectral envelope features. In some cases, a cepstrum or cepstrum information may be an optional input to a machine learning spectral feature learning model used to implement the spectral envelope feature extraction engine 1020. In some cases, the spectral envelope feature extraction engine 1020 can be implemented using one or more machine learning spectral feature learning models or engines, which can be jointly trained with the neural speech synthesizer 1050, an NHV used to implement the neural speech synthesizer (e.g., such as the NHV-based neural speech synthesizer 1150 of FIG. 11, NHV- based neural speech synthesizer 1250 of FIG. 12, etc.), an LPC network used to implement the neural speech synthesizer (e.g., such as the LPC network-based neural speech synthesizer 1150 of FIG. 11, NHV-based neural speech synthesizer 1250 of FIG. 12, etc.), etc.

[0105] In some cases, joint training of the spectral envelope feature extraction engine 1020 and the neural speech synthesizer 1050 can be performed based on a short-time Fourier transform (STFT) loss and / or STFT loss function. In some aspects, joint training of the spectral envelope feature extraction engine 1020 and the neural speech synthesizer 1050 can be performed based on an adversarial and / or generative adversarial network (GAN) loss. For example, joint training can be performed using a multi-resolution STFT loss, based on calculating STFT amplitude spectrograms from ground truth speech information and the synthesized speech information. The multi-resolution STFT loss can then be determined as a sum of mean absolute error and mean absolute error in the log domain. The STFT amplitude spectrograms can be calculated at different window lengths to capture representations of the error and / or loss at different time and / or frequency resolutions. In some aspects, joint trainingbased on adversarial or GAN loss can be used to learn temporal fine structures in speech signals. An adversarial or GAN loss can be used to match the distribution of real speech and the synthesized speech (e.g., in the reconstructed audio 1055). For example, the neural speech synthesizer 1050 (and / or an NHV-based neural speech synthesizer) can be used as the generator network. A separate discriminator network can be used to train based on the adversarial loss, with both the NHV and the discriminator jointly trained using adversarial training techniques. In some examples, the discriminator network can be implemented as a classifier configured to classify whether an input audio represents real speech or synthesized speech. The generator network can attempt to fool the discriminator network to classify synthesized speech as real speech. Over the course of training, the generator network improves and begins producing (e.g., generating or synthesizing) speech that is very similar to the real speech. In some cases, the discriminator network can be implemented using WaveNet.

[0106] The neural speech synthesizer 1050 can be implemented using one or more trained neural networks. For example, the one or more trained neural networks can be trained to perform speech synthesis and / or signal synthesis. In some examples, the neural speech synthesizer 1050 can be implemented using or based on an LPCNet machine learning architecture, a WaveNet machine learning architecture, a WaveRNN machine learning architecture, etc. In one illustrative example, the neural speech synthesizer 1050 can be provided as a neural homomorphic vocoder (NHV).

[0107] For example, FIG. 10 illustrates an example of an audio codec system 1000 (e.g., generative voice codec) that includes a decoder 1010 configured to implement an NHV-based neural speech synthesizer. In some examples, an NHV is a type of neural vocoder that can synthesize speech with source-filter models controlled by one or more neural networks. An NHV-based neural vocoder may include one or more neural networks in a source-filter model that can synthesize speech based on filtering impulse trains and noise with linear time-varying (LTV) filters, with the one or more neural networks used to control the LTV filters by estimating complex cepstrums of time-varying impulse responses given acoustic features. Traditional or non-neural vocoders may operate based on decomposing speech into various parameters such as pitch, timbre, rhythm, etc., which can subsequently be manipulated and resynthesized to generate a desired output audio or voice signal. Neural vocoders (e.g., such as NHVs, etc.) can apply transformations and manipulations to speech signals directly within a learned feature space of the neural network. The learned feature space used by neural vocoders may capturemore complex relationships and characteristics of speech than non-neural network-based signal processing and / or vocoder techniques. For example, neural vocoders can be trained to leam a mapping between raw speech waveforms and the spectral or cepstral representations of the speech. An NHV system can apply one or more homomorphic processing techniques within the learned space corresponding to the mapping.

[0108] FIG. 11 illustrates an example of an audio codec system 1100 (e.g.. generative voice codec) that includes a decoder 1110 configured to implement an NHV-based neural speech synthesizer and a postprocessing process for improving pitch content, in accordance with aspects of the present disclosure. In some examples, an NHV is a type of neural vocoder that can synthesize speech with source-filter models controlled by one or more neural networks. An NHV-based neural vocoder may include one or more neural networks in a source-filter model that can synthesize speech based on filtering impulse trains and noise with linear time-varying (LTV) filters, with the one or more neural networks used to control the LTV filters by estimating complex cepstrums of time-varying impulse responses given acoustic features. Traditional or non-neural vocoders may operate based on decomposing speech into various parameters such as pitch, timbre, rhythm, etc., which can subsequently be manipulated and resynthesized to generate a desired output audio or voice signal. Neural vocoders (e.g., such as NHVs, etc.) can apply transformations and manipulations to speech signals directly within a learned feature space of the neural network. The learned feature space used by neural vocoders may capture more complex relationships and characteristics of speech than non-neural network-based signal processing and / or vocoder techniques. For example, neural vocoders can be trained to leam a mapping between raw speech waveforms and the spectral or cepstral representations of the speech. An NHV system can apply one or more homomorphic processing techniques within the learned space corresponding to the mapping.

[0109] In some aspects, the decoder 1110 may also be referred to as an NHV synthesizer decoder. In some examples, the audio codec system 1100 of FIG. 11 may be similar to the audio codec system 1000 of FIG. 10. In some cases, the channel 1140 of FIG. 11 can be the same as or similar to the channel 1040 of FIG. 10, and may be associated with an encoder that is the same as or similar to the encoder 1005 of FIG. 10, etc.

[0110] In some aspects, the decoder 1110 can include a neural speech synthesizer 1150 that is the same as or similar to the neural speech synthesizer 1050 included in the decoder 1010 ofFIG. 10. For example, the neural speech synthesizer 1150 can receive a first input comprising decoded spectral envelope features (e.g., determined by a FRAE decoder 1145 that is the same as or similar to the FRAE decoder 1045 of FIG. 10). The neural speech synthesizer 1150 can receive a second input comprising a received pitch encoding (e.g.. received pitch information), that is the same as or similar to the received pitch encoding provided to the neural speech synthesizer 1050 of FIG. 10.

[0111] In one illustrative example, the NHV synthesizer decoder 1110 can include the FRAE decoder 1145, which may be the same as or similar to the FRAE decoder 1045 of FIG. 10, and may include a pitch dequantization engine 1138 (e.g., de Q()). The pitch dequantization engine 1138 can be used to process a received pitch encoding obtained by the decoder 1110 over the channel 1140. For example, the received pitch encoding information can be quantized pitch information or a quantized pitch signal transmitted to the decoder 1110 over the channel 1140 by a corresponding encoder (e.g., an encoder associated with the decoder 1110, which may be the same as or similar to the encoder 1005 of FIG. 10, etc.). In some aspects, the pitch dequantization engine 1138 can generate reconstructed pitch information based on performing a codebook lookup to dequantize the received pitch encoding obtained from the channel 1140 (e.g., to dequantize the quantized pitch encoding received by the decoder 1110 over the channel 1140).

[0112] The dequantized (e.g.. reconstructed) pitch information determined by the pitch dequantization engine 1138 can include pitch information in the frequency-domain (e.g., fo pitch frequency information in Hz), pitch information in the time-domain (e.g., pitch lag or pitch delay information in samples, milliseconds, etc.), and / or pitch correlation information. In some aspects, the dequantized (e.g.. reconstructed) pitch information determined by the pitch dequantization engine 1138 can include information indicative of a voiced or unvoiced (V / UV) classification.

[0113] In one illustrative example, the neural speech synthesizer 1150 can be implemented as a neural homomorphic vocoder (NHV) speech synthesizer and / or an NHV-based neural speech synthesizer. For example, the neural speech synthesizer 1150 can include a neural filter estimator 1152 configured to generate respective filter specification or filter configuration information to parameterize one or more linear time-varying (LTV) filters of the NHV speech synthesizer 1150. In some aspects, the neural filter estimator 1152 can be trained to generaterespective filter coefficients for a noise linear time-varying filter (LTVF) 1175 and to generate respective filter coefficients for a harmonic LTVF 1165.

[0114] The neural network model of the neural filter estimator 1152 can include any neural network architecture that can be trained to model the filter coefficients for the noise LTVF 1175 and the harmonic LTVF 1165 (e.g., and / or that can be trained to model the filter coefficients for one or more additional filters implemented by the NHV speech synthesizer 1150, in either the time-domain, the frequency -domain, or combinations thereof). Examples of neural network architectures that may be included in the neural network filter estimator 1152 can include a generative neural network (e.g., a generative-adversarial network (GAN)), convolutional neural networks (CNN), an autoencoder, and / or other type(s) of neural networks, etc.

[0115] In some examples, the neural filter estimator 1152 can be configured to receive a stream of decoded spectral envelope features from the FRAE decoder 1145. Based on the decoded spectral envelope features, the neural filter estimator 1152 can generate corresponding filter characterization parameters for each filter of one or more filters included in the NHV speech synthesizer 1150. The respective filter characterization parameters can be output from the neural filter estimator 1152 and used to parameterize and / or configure corresponding learned filters (e.g., learned linear filters, etc.) for each respective set of filter characterization parameters. For example, the neural filter estimator 1152 can generate a first set of filters corresponding to the noise LTVF 1175, can generate a second set of filter characterization parameters corresponding to the harmonic LTVF 1165, etc. The one or more filters included in the NHV speech synthesizer 1150 (e.g., the noise LTVF 1175, the harmonic LTVF 1165, etc.) can be specified in various different forms. For example, the noise LTVF 1175, the harmonic LTVF 1165, and / or various other filters that may be included in the NHV speech synthesizer 1150 can be specified or characterized (e.g., using the filter characterization parameters determined by the neural filter estimator 1152) as cepstrums, as frequency responses, as timedomain impulse responses, as difference equation coefficients, etc.

[0116] In some aspects, the neural filter estimator 1152 can be configured to receive as input the decoded spectral envelope features from the FRAE decoder 1145, and may additionally receive as input at least a portion of the dequantized pitch information generated by the pitch dequantization engine 1138. For example, the neural filter estimator 1152 may receive as input the decoded spectral envelope features and pitch dequantization information(e.g., such as fo pitch information in the frequency domain, pitch lag or pitch delay in the time domain (e.g., in units of samples or milliseconds, etc.), etc.). In some examples, the pitch dequantization information provided as input to the neural filter estimator 1152 may include voiced / unvoiced (V / UV) classification information indicating whether the underlying audio represented in the encoded information received by the decoder 1110 over the channel 1140 corresponds to voiced or unvoiced sounds, speech, etc.

[0117] In examples where the neural filter estimator 1152 receives spectral envelope features and dequantized pitch information as inputs, the neural filter estimator 1152 can generate the corresponding filter characterization parameters for each NHV filter (e.g., noise LTVF 1175, harmonic LTVF 1165, etc.) based on the decoded spectral envelope features and the dequantized pitch information.

[0118] The noise LTVF 1175 and the harmonic LTVF 1165 can be implemented as time domain filters or frequency domain filters. In some examples, the noise LTVF 1175 and the harmonic LTVF 1165 can be implemented in the time domain or the frequency domain, independent of whether the neural filter estimator 1152 is configured to generate the corresponding filter characterization parameters in the time domain or the frequency domain. For example, in some aspects, the neural filter estimator 1152 can output impulse responsebased filter characterization parameters, and the operation of noise LTVF 1175 and / or harmonic LTVF 1165 can be implemented in the time domain as convolutions with the impulse response. Using the same impulse response-based filter characterization parameters, the operation of noise LTVF 1175 and / or harmonic LTVF 1165 may be implemented in the frequency domain based on converting the impulse response to frequency response (e.g., using an FFT transform) and multiplying with the FFT of the input signal, and subsequently converting back to the time domain using an IFFT transform.

[0119] In one illustrative example, the NHV speech synthesizer 1150 can include a pulse train generator 1160. The dequantized pitch information (e.g., generated using the pitch dequantization engine 1138) can be provided to the pulse train generator 1160. In some aspects, the pulse train generator 1160 can generate a pulse train based at least in part on the pitch frequency (e.g., fo) and / or pitch lag or pitch delay information obtained from the pitch dequantization engine 1138 of the decoder 1110. In some examples, the pulse train generator 1160 can be implemented as a cosine sum pulse generator, which can be configured to processthe dequantized pitch information (e.g., obtained from the pitch dequantization engine 1138) to generate a pulse train p\n\. The pulse train may also be referred to as an impulse train. For example, the pulse train generator 1160 can be a differentiable cosine sum pulse generator, and / or can be a non-differentiable cosine sum pulse generator (e.g.. among various other pulse generators).

[0120] The pulse train generated by the pulse train generator 1160 can be processed by the harmonic LTVF 1165. using the corresponding harmonic filter characterization parameters determined by the neural filter estimator 1152 for the harmonic LTVF 1165. In some aspects, the harmonic LTVF 1165 can generate a harmonic output based on processing the pulse train from the pulse train generator 1160.

[0121] In some examples, the pulse train generator 1160 may receive an additional input from the pitch dequantization engine 1138, indicative of a voice or unvoiced (e.g., V / UV) classification. For example, based on receiving an unvoiced (UV) indication or classification from the pitch dequantization engine 1138, the pulse train generator 1160 can be configured to generate a 0 output or a noise output that is provided to the harmonic LTVF 1165 instead of the pulse train p[n ] p[n ] provided from the pulse train generator 1160 to the harmonic LTVF 1165 in response to a voiced (V) indication or classification from the pitch dequantization engine 1138).

[0122] The NHV speech synthesizer 1150 can include a noise generator 1170 that is configured to generate a noise signal (e.g., white noise, etc.) for processing by the noise LTVF 1175. For example, the noise LTVF 1175 can be parameterized based on the respective noise filter characterization parameters generated by the neural filter estimator 1152 for the noise LTVF 1175, and can subsequently be used to process the noise signal generated by the noise generator 1170. Based on processing the noise signal from the noise generator 1170, the noise LTVF 1175 can generate a noise-filtered output.

[0123] The NHV speech synthesizer 1150 can include a combination function 1180 to combine the harmonic output (e.g., from the harmonic LTVF 1165) and the noise-filtered output (e.g., from the noise LTVF 1175) to generate a predicted sample 1182 (S't) that may be input to a post-processing engine 1184. In some examples, the combination function 1180 can include an adder, a multiplier, a divider, weighted sum, a weighted product, a weighted ratio, an average, a weighted average, a weighted mean, a weighted median, a weighted mode, or acombination thereof. The reconstructed audio 1155 may be an example of the reconstructed speech signal 105 of FIG. 1, the reconstructed speech signal 205 of FIG. 2, etc.

[0124] In some cases, the post-processing engine 1184 may be similar to post-processing engine 808 of FIG. 8 and / or post-processing engine 504 of FIG. 5. In some cases, the postprocessing engine 1184 may be ML based. The post-processing engine 1184 may receive a set of previously predicted samples 1186 from a delay circuit 1188 and a pitch gain 1190 and the post-processing engine 1184 may estimate an improved version of the set of previously predicted sample 1186 (S't) from the LPC layers 1154 for output as reconstructed audio 1155. In some cases, the pitch gain 1190 may be substantially similar to pitch gain 508 of FIG. 5 and / or pitch gain 814 of FIG. 8. The delay circuit 1188 may be substantially similar to delay circuit 510 of FIG. 5 and / or delay circuit 812 of FIG. 8. The delay circuit 1188 may store the reconstructed audio 1155 (St) of the post-processing engine 1184 for input back to the postprocessing engine at a later pitch period. The delay circuit may also receive pitch lag information 1192 (e.g., from the pitch dequantization engine 1138), which may be used to determine the set of previously predicted samples 1186. As describe above, the set of previously predicted samples 1186 may include samples based on the previous pitch period (e.g.. n-T). and include samples near the previous pitch period (e.g., n-T, n+T) within a fixed or variable window, or samples for an entire previous pitch period. In some examples, the reconstructed audio 1155 includes synthesized speech. In some aspects, the reconstructed audio 1155 can be the same as or similar to the reconstructed audio 1055 of FIG. 10.

[0125] FIG. 12 illustrates an example of an audio codec system 1200 (e.g., generative voice codec) that includes a decoder 1210 configured to implement an NHV-based neural speech synthesizer and including an integrated process for improving pitch content, in accordance with aspects of the present disclosure. In some cases, the audio codec system 1200 may implement an NHV-based neural speech synthesizer in a manner substantially similar to audio codec system 1100 of FIG. 11. In some examples, the audio codec system 1200 of FIG. 12 may be similar to the audio codec system 1000 of FIG. 10 and audio codec system 1100 of FIG. 11. In some cases, the channel 1240 of FIG. 12 may be substantially similar to the channel 1040 of FIG. 10 and / or channel 1140 of FIG. 11, and may be associated with an encoder that is the same as or substantially similar to the encoder 1005 of FIG. 10, etc.

[0126] The decoder 1210 may also be referred to as an NHV synthesizer decoder and the decoder 1210 can include a neural speech synthesizer 1250 that is substantially similar to the neural speech synthesizer 1050 included in the decoder 1010 of FIG. 10 and / or neural speech synthesizer 1150 of FIG. 11, and may correspond to speech synthesizer 902 of FIG. 9. The decoder 1210 may include a FRAE decoder 1245 and a pitch dequantization engine 1238 that operate in a substantially similar way as FRAE decoder 1145 and pitch dequantization engine 1138 of FIG. 11, respectively. In some cases, the decoder 1210 can include a neural speech synthesizer 1250 that is substantially similar to the neural speech synthesizer 1050 included in the decoder 1010 of FIG. 10 and / or synthesizer 1150 included in the decoder 11100 of FIG. 11. For example, the neural speech synthesizer 1250 can receive a first input comprising decoded spectral envelope features. Aneural filter estimator 1252 of the neural speech synthesizer 1250 can receive a second input comprising a received pitch encoding (e.g., received pitch information), that is the same as or similar to the received pitch encoding provided to the neural speech synthesizer 1050 of FIG. 10 and / or the neural speech synthesizer 1150 of FIG. 11. The neural filter estimator 1252 may be similar to the NN filter estimator 904 of FIG. 9 and neural filter estimator 1152. In some examples, the neural filter estimator 1252 may also receive a set of previously predicted samples 1286 from a delay circuit 1288. The neural filter estimator 1252 can include any neural network architecture that can be trained to model the filter coefficients for the noise LTVF 1275 and the harmonic LTVF 1265 (e.g., and / or that can be trained to model the filter coefficients for one or more additional filters implemented by the NHV speech synthesizer 1250, in either the time-domain, the frequency -domain, or combinations thereof). The neural filter estimator 1252 may receive as input the decoded spectral envelope features from the FRAE decoder 1245, at least a portion of the dequantized pitch information generated by the pitch dequantization engine 1238, and set of previously predicted samples 1286 and generate the corresponding filter characterization parameters for each NHV filter (e.g.. noise LTVF 1275. harmonic LTVF 1265, etc.) based on the decoded spectral envelope features, the dequantized pitch information, and previous samples in a manner substantially similar to that discussed above with respect to neural filter estimator 1152 of FIG. 11 and NN filter estimator 904 of FIG. 9. In some cases, a pulse train generator 1260, noise generator 1270, and combination function 1280 of the neural speech synthesizer 1250 may operate in a manner substantially similar to the pulse train generator 1160, noise generator 1170, and combination function 1280 of the neural speech synthesizer 1150 of FIG. 11.

[0127] In some cases, the combination function 1280 may output a predicted sample (S't) that may be input to the delay circuit 1288. The delay circuit 1288 may receive pitch lag information 1292 (e.g.. from the pitch dequantization engine 1238). The delay circuit 1288 may be substantially similar to first delay circuit 606 of FIG. 6. delay circuit 908 of FIG. 9, etc. The delay circuit 1288 may output a set of previously predicted samples 1286 to the neural filter estimator 1252 based on the pitch lag information 1292. As describe above, the set of previously predicted samples 1286 may include samples based on the previous pitch period (e.g., n-T), and include samples near the previous pitch period (e.g., n-T, n+T) within a fixed or variable window, or samples for an entire previous pitch period. In some cases, the neural filter estimator 1252 may be trained to use the set of previously predicted samples 1286 to help predict relationships between the previously predicted samples and a current sample being processed by the neural filter estimator 1252 to generate a predicted sample for output (e.g., as output by the combination function 1280). Output from the combination function 1280 can be provided to one or more LPC layers 1254. The one or more LPC layers may be substantially similar to LPC layers 1052 of FIG. 10 and LPC layers 1154 of FIG. 11. The LPC layers 1052 can perform linear prediction based on the output of the neural speech synthesizer 1250, to generate the reconstructed audio 1255 (e.g.. synthesized speech).

[0128] FIG. 13 is a flow diagram illustrating a process 1300 for speech synthesis, in accordance with aspects of the present disclosure. The process 1300 may be performed by a computing device (or apparatus) or a component (e.g., a chipset, codec, etc.) of the computing device (e.g., voice coding system 130 of FIG. 1 , voice decoder 204 of FIG. 2, voice decoding system 300 of FIG. 3, computing system 1700 of FIG. 17, etc.). The computing device may be a mobile device (e.g., a mobile phone), a network-connected wearable such as a watch, an extended reality (XR) device such as a virtual reality (VR) device or augmented reality (AR) device, a vehicle or component or system of a vehicle, or other type of computing device. The operations of the process 1300 may be implemented as software components that are executed and run on one or more processors (e.g., processor 1710 of FIG. 17, etc.).

[0129] At block 1302, the computing device (or component thereof) may receive a first predicted audio sample (e.g., from a predication engine 410 of FIG. 4, frame rate network 406, etc.). In some cases, the first predicted audio sample is a part of a set of previously predicted audio samples (e.g., set of previously predicted samples 602 of FIG. 6, set of previously predicted samples 708 of FIG. 7A and 7C, set of previously predicted samples 810 of FIG. 8,set of previously predicted samples 912 of FIG. 9, etc.), and the second predicted audio sample is generated based on the set of previously predicted audio samples. For example, the set of previously predicted samples may include samples based on the previous pitch period (e.g., n- T). and include samples near the previous pitch period (e.g., n-T. n+T) within a fixed or variable window, or samples for an entire previous pitch period. In some examples, the first predicted audio sample is generated by at least one of a sample rate network, neural homomorphic vocoder, or lyra speech synthesizer. In some cases, the computing device (or component thereof) may include a speaker, and the computing device (or component thereof) may output the second predicted audio sample via the speaker.

[0130] At block 1304, the computing device (or component thereof) may cause the first predicted audio sample to be stored in the one or more memories (e.g.. delay circuit 510 of FIG. 5, delay circuit 606 of FIG. 6, delay circuit 706 of FIG. 7A, delay circuit 812 of FIG. 8, delay circuit 908 of FIG. 9, memory 1715 of FIG. 17, etc.).

[0131] At block 1306, the computing device (or component thereof) may generate a second predicted audio sample (e.g., output 608 of FIG. 6, output 704 of FIG. 7A, output 816 of FIG. 8, output predicted samples 906 of FIG. 9, etc.) based on the first predicted audio sample. In some cases, the second predicted audio sample is from a pitch period after the first predicted audio sample. In some cases, the computing device (or component thereof) may generate the second predicted audio sample using a machine learning model (e.g., such as for the postprocessing engine 504 of FIG. 5, sample rate netw ork 604 of FIG. 6, attention mechanism 718 of FIG. 7A, attention mechanism 720 of FIG. 7B, attention mechanism 750 of FIG. 7C, speech synthesizer 802 of FIG. 8, post-processing engine 808 of FIG. 8, neural netw ork filter estimator 904 of FIG. 9, speech synthesizer 902 of FIG. 9, neural network 1400 of FIG. 14, CNN 1500 ofFIG. 15, deep convolutional network 1650 ofFIG. 16, among various other machine learning networks described herein). In some examples, the machine learning model is configured to receive, as input, the set of previously predicted audio samples. In some cases, the computing device (or component thereol) may generate the second predicted audio sample using a machine learning model and generate a context vector describing the set of previously predicted audio samples, where the machine learning model is configured to receive, as input, the context vector. In some examples, the computing device (or component thereol) may generate the context vector based on an attention mechanism (e.g., attention mechanism 718 of FIG. 7A, attention mechanism 720 ofFIG. 7B, attention mechanism 750 ofFIG. 7C). In some cases, thecomputing device (or component thereof) may receive pitch lag information (e.g., pitch lag 610 of FIG. 6, pitch lag information 818 of FIG. 8, pitch lag information 910 of FIG. 9, etc.) and select the first predicted audio sample for the set of previously predicted audio samples based on the pitch lag information. In some examples, the second predicted audio sample is generated by receiving a third predicted audio sample; and post-processing the third predicted audio sample based on the set of previously predicted audio samples to generate the second predicted audio sample. In some cases, the computing device (or component thereof) may provide the second predicted audio sample for generating a next predicted audio sample. In some examples, the computing device (or component thereof) may receive pitch gain information (e.g., pitch gain 712 of FIG. 7A, pitch gain 814 of FIG. 8, etc.) indicating how correlated one pitch period is with another pitch period; and post-process the third predicted audio sample based on the pitch gain information.

[0132] As noted herein, the techniques or processes described herein (e.g., the process 1300) may be performed by a computing device, an apparatus, and / or any other computing device. In some cases, the computing device or apparatus may include a processor, microprocessor, microcomputer, or other component of a device that is configured to carry out the steps of processes described herein. In some examples, the computing device or apparatus may include a camera configured to capture video data (e.g., a video sequence) including video frames. For example, the computing device may include a camera device, which may or may not include a video codec. As another example, the computing device may include a mobile device with a camera (e.g., a camera device such as a digital camera, an IP camera or the like, a mobile phone or tablet including a camera, or other t pe of device with a camera). In some cases, the computing device may include a display for displaying images. In some examples, a camera or other capture device that captures the video data is separate from the computing device, in which case the computing device receives the captured video data. The computing device may further include a network interface, transceiver, and / or transmitter configured to communicate the video data. The network interface, transceiver, and / or transmitter may be configured to communicate Internet Protocol (IP) based data or other network data.

[0133] The processes described herein can be implemented in hardware, computer instructions, or a combination thereof. In the context of computer instructions, the operations represent computer-executable instructions stored on one or more computer-readable storage media that, when executed by one or more processors, perform the recited operations.Generally, computer-executable instructions include routines, programs, objects, components, data structures, and the like that perform particular functions or implement particular data types. The order in which the operations are described is not intended to be construed as a limitation, and any number of the described operations can be combined in any order and / or in parallel to implement the processes.

[0134] In some cases, the devices or apparatuses configured to perform the operations of the process 1300 and / or other processes described herein may include a processor, microprocessor, micro-computer, or other component of a device that is configured to carry out the steps of the process 1300 and / or other process. In some examples, such devices or apparatuses may include one or more sensors configured to capture image data and / or other sensor measurements. In some examples, such computing device or apparatus may include one or more sensors and / or a camera configured to capture one or more images or videos. In some cases, such device or apparatus may include a display for displaying images. In some examples, the one or more sensors and / or camera are separate from the device or apparatus, in which case the device or apparatus receives the sensed data. Such device or apparatus may further include a network interface configured to communicate data.

[0135] The components of the device or apparatus configured to carry out one or more operations of the process 1300 and / or other processes described herein can be implemented in circuitry. For example, the components can include and / or can be implemented using electronic circuits or other electronic hardware, which can include one or more programmable electronic circuits (e.g., microprocessors, graphics processing units (GPUs), digital signal processors (DSPs), central processing units (CPUs), and / or other suitable electronic circuits), and / or can include and / or be implemented using computer software, firmware, or any combination thereof, to perform the various operations described herein. The computing device may further include a display (as an example of the output device or in addition to the output device), a network interface configured to communicate and / or receive the data, any combination thereof, and / or other component(s). The network interface may be configured to communicate and / or receive Internet Protocol (IP) based data or other type of data.

[0136] The process 1300 is illustrated as a logical flow diagram, the operations of which represent sequences of operations that can be implemented in hardware, computer instructions, or a combination thereof. In the context of computer instructions, the operations representcomputer-executable instructions stored on one or more computer-readable storage media that, when executed by one or more processors, perform the recited operations. Generally, computerexecutable instructions include routines, programs, objects, components, data structures, and the like that perform particular functions or implement particular data types. The order in which the operations are described is not intended to be construed as a limitation, and any number of the described operations can be combined in any order and / or in parallel to implement the processes.

[0137] Additionally, the processes described herein (e.g., the process 1300 and / or other processes) may be performed under the control of one or more computer systems configured with executable instructions and may be implemented as code (e.g.. executable instructions, one or more computer programs, or one or more applications) executing collectively on one or more processors, by hardware, or combinations thereof. As noted above, the code may be stored on a computer-readable or machine-readable storage medium, for example, in the form of a computer program including a plurality of instructions executable by one or more processors. The computer-readable or machine-readable storage medium may be non-transitory.

[0138] Additionally, the processes described herein may be performed under the control of one or more computer systems configured with executable instructions and may be implemented as code (e.g.. executable instructions, one or more computer programs, or one or more applications) executing collectively on one or more processors, by hardware, or combinations thereof. As noted above, the code may be stored on a computer-readable or machine-readable storage medium, for example, in the form of a computer program comprising a plurality of instructions executable by one or more processors. The computer-readable or machine-readable storage medium may be non-transitory.

[0139] FIG. 14 is an illustrative example of a neural network 1400 (e.g., a deep-learning neural network) that can be used to implement machine-leaming-based speech synthesis, image generation, feature segmentation, implicit-neural-representation generation, rendering, classification, object detection, image recognition (e.g., face recognition, object recognition, scene recognition, etc.), feature extraction, authentication, gaze detection, gaze prediction, and / or automation.

[0140] An input layer 1402 includes input data. Neural network 1400 includes multiple hidden layers hidden layers 1406a, 1406b, through 1406n. The hidden layers 1406a, 1406b,through hidden layer 1406n include “n” number of hidden layers, where “n” is an integer greater than or equal to one. The number of hidden layers can be made to include as manylayers as needed for the given application. Neural network 1400 further includes an output layer 1404 that provides an output resulting from the processing performed by the hidden layers 1406a, 1406b, through 1406n.

[0141] Neural network 1400 may be, or may include, a multi-layer neural network of interconnected nodes. Each node can represent a piece of information. Information associated with the nodes is shared among the different layers and each layer retains information as information is processed. In some cases, neural network 1400 can include a feed-forward network, in w hich case there are no feedback connections where outputs of the netw ork are fed back into itself. In some cases, neural network 1400 can include a recurrent neural network, which can have loops that allow information to be carried across nodes while reading in input.

[0142] Information can be exchanged between nodes through node-to-node interconnections between the various layers. Nodes of input layer 1402 can activate a set of nodes in the first hidden layer 1406a. For example, as shown, each of the input nodes of input layer 1402 is connected to each of the nodes of the first hidden layer 1406a. The nodes of first hidden layer 1406a can transform the information of each input node by applying activation functions to the input node information. The information derived from the transformation can then be passed to and can activate the nodes of the next hidden layer 1406b. which can perform their own designated functions. Example functions include convolutional, up-sampling, data transformation, and / or any other suitable functions. The output of the hidden layer 1406b can then activate nodes of the next hidden layer, and so on. The output of the last hidden layer 1406n can activate one or more nodes of the output layer 1404, at which an output is provided. In some cases, while nodes (e.g., node 1408) in neural network 1400 are shown as having multiple output lines, a node has a single output and all lines shown as being output from a node represent the same output value.

[0143] In some cases, each node or interconnection between nodes can have a weight that is a set of parameters derived from the training of neural network 1400. Once neural netw ork 1400 is trained, it can be referred to as a trained neural network, which can be used to perform one or more operations. For example, an interconnection between nodes can represent a piece of information learned about the interconnected nodes. The interconnection can have a tunablenumeric weight that can be tuned (e.g., based on a training dataset), allowing neural network 1400 to be adaptive to inputs and able to learn as more and more data is processed.

[0144] Neural network 1400 may be pre-trained to process the features from the data in the input layer 1402 using the different hidden layers 1406a, 1406b, through 1406n in order to provide the output through the output layer 1404. In an example in which neural network 1400 is used to identify features in images, neural network 1400 can be trained using training data that includes both images and labels, as described above. For instance, training images can be input into the network, with each training image having a label indicating the features in the images (for the feature-segmentation machine-learning system) or a label indicating classes of an activity in each image. In one example using object classification for illustrative purposes, a training image can include an image of a number 2, in which case the label for the image can be [0 0 1 0 0 0 0 0 0 0],

[0145] In some cases, neural network 1400 can adjust the weights of the nodes using a training process called backpropagation. As noted above, a backpropagation process can include a forward pass, a loss function, a backward pass, and a weight update. The forward pass, loss function, backward pass, and parameter update is performed for one training iteration. The process can be repeated for a certain number of iterations for each set of training images until neural network 1400 is trained well enough so that the weights of the layers are accurately tuned.

[0146] For the example of identifying objects in images, the forward pass can include passing a training image through neural network 1400. The weights are initially randomized before neural network 1400 is trained. As an illustrative example, an image can include an array of numbers representing the pixels of the image. Each number in the array can include a value from 0 to 255 describing the pixel intensify at that position in the array. In one example, the array can include a 28 x 28 x 3 array of numbers with 28 rows and 28 columns of pixels and 3 color components (such as red, green, and blue, or luma and two chroma components, or the like).

[0147] As noted above, for a first training iteration for neural network 1400. the output will likely include values that do not give preference to any particular class due to the weights being randomly selected at initialization. For example, if the output is a vector with probabilities that the object includes different classes, the probability value for each of the different classes canbe equal or at least very similar (e.g., for ten possible classes, each class can have a probability value of 0.1). With the initial weights, neural network 1400 is unable to determine low-level features and thus cannot make an accurate determination of what the classification of the object might be. A loss function can be used to analyze error in the output. Any suitable loss function definition can be used, such as a cross-entropy loss. Another example of a loss function includes the mean squared error (MSE), defined asEf"tQl =equal to the value of Etotal.

[0148] The loss (or error) will be high for the first training images since the actual values will be much different than the predicted output. The goal of training is to minimize the amount of loss so that the predicted output is the same as the training label. Neural network 1400 can perform a backward pass by determining which inputs (weights) most contributed to the loss of the network and can adjust the weights so that the loss decreases and is eventually minimized. A derivative of the loss with respect to the weights (denoted as dL / dW, where W are the weights at a particular layer) can be computed to determine the weights that contributed most to the loss of the network. After the derivative is computed, a weight update can be performed by updating all the weights of the filters. For example, the weights can be updated so that they change in the opposite direction of the gradient. The weight update can be denoted didenotes a weight, wi denotes the initial weight, and q denotes a learning rate. The learning rate can be set to any suitable value, with a high learning rate including larger weight updates and a lower value indicating smaller weight updates.

[0149] In some aspects, training of one or more of the machine learning systems or neural networks described herein (e.g., such as for the post-processing engine 504 of FIG. 5, sample rate network 502 of FIG. 5, sample rate network 604 of FIG. 6, sample rate network 702 of FIG. 7A, attention mechanism 718 of FIG. 7 A, attention mechanism 720 of FIG. 7B, attention mechanism 750 of FIG. 7C, speech synthesizer 802 of FIG. 8, post-processing engine 808 of FIG. 8, neural network filter estimator 904 of FIG. 9, speech synthesizer 902 of FIG. 9, neural network 1400 of FIG. 14, CNN 1500 of FIG. 15, deep convolutional network 1650 of FIG. 16, among various other machine learning networks described herein) can be performed using online training (e.g., in some case on-device training), offline training, and / or various combinations of online and offline training. In some cases, online may refer to time periods during which the input data is processed, for instance for performance of the speech synthesis processing implemented by the systems and techniques described herein. In some examples.offline may refer to idle time periods or time periods during which input data is not being processed. Additionally, offline may be based on one or more time conditions (e.g., after a particular amount of time has expired, such as a day, a week, a month, etc.) and / or may be based on various other conditions such as network and / or server availability, etc., among various others. In some aspects, offline training of a machine learning model (e.g., a neural network model) can be performed by a first device (e.g., a server device) to generate a pretrained model, and a second device can receive the trained model from the second device. In some cases, the second device (e.g., a mobile device, an XR device, a vehicle or system / component of the vehicle, or other device) can perform online (or on-device) training of the pre-trained model to further adapt or tune the parameters of the model.

[0150] Neural network 1400 can include any suitable deep network. One example includes a convolutional neural network (CNN), which includes an input layer and an output layer, with multiple hidden layers between the input and out layers. The hidden layers of a CNN include a series of convolutional, nonlinear, pooling (for downsampling), and fully connected layers. Neural network 1400 can include any other deep network other than a CNN, such as an autoencoder, a deep belief nets (DBNs). a Recurrent Neural Networks (RNNs). among others.

[0151] FIG. 15 is an illustrative example of a convolutional neural network (CNN) 1500. The input layer 1502 of the CNN 1500 includes data representing an image or frame. For example, the data can include an array of numbers representing the pixels of the image, with each number in the array including a value from 0 to 255 describing the pixel intensity at that position in the array. Using the previous example from above, the array can include a 28 x 28 x 3 array of numbers with 28 rows and 28 columns of pixels and 3 color components (e.g., red, green, and blue, or luma and two chroma components, or the like). The image can be passed through a convolutional hidden layer 1504, an optional non-linear activation layer, a pooling hidden layer 1506, and fully connected layer 1508 (which fully connected layer 1508 can be hidden) to get an output at the output layer 1510. While only one of each hidden layer is shown in FIG. 15, one of ordinary skill will appreciate that multiple convolutional hidden layers, nonlinear layers, pooling hidden layers, and / or fully connected layers can be included in the CNN 1500. As previously described, the output can indicate a single class of an object or can include a probability' of classes that best describe the object in the image.

[0152] The first layer of the CNN 1500 can be the convolutional hidden layer 1504. The convolutional hidden layer 1504 can analyze image data of the input layer 1502. Each node of the convolutional hidden layer 1504 is connected to a region of nodes (pixels) of the input image called a receptive field. The convolutional hidden layer 1504 can be considered as one or more filters (each filter corresponding to a different activation or feature map), with each convolutional iteration of a filter being a node or neuron of the convolutional hidden layer 1504. For example, the region of the input image that a filter covers at each convolutional iteration would be the receptive field for the filter. In one illustrative example, if the input image includes a 28x28 array, and each filter (and corresponding receptive field) is a 5x5 array, then there will be 24x24 nodes in the convolutional hidden layer 1504. Each connection between a node and a receptive field for that node leams a weight and, in some cases, an overall bias such that each node leams to analyze its particular local receptive field in the input image. Each node of the convolutional hidden layer 1504 will have the same weights and bias (called a shared weight and a shared bias). For example, the filter has an array of weights (numbers) and the same depth as the input. A filter will have a depth of 3 for an image frame example (according to three color components of the input image). An illustrative example size of the filter array is 5 x 5 x 3. corresponding to a size of the receptive field of a node.

[0153] The convolutional nature of the convolutional hidden layer 1504 is due to each node of the convolutional layer being applied to its corresponding receptive field. For example, a filter of the convolutional hidden layer 1504 can begin in the top-left comer of the input image array and can convolve around the input image. As noted above, each convolutional iteration of the filter can be considered a node or neuron of the convolutional hidden layer 1504. At each convolutional iteration, the values of the filter are multiplied with a corresponding number of the original pixel values of the image (e.g., the 5x5 filter array is multiplied by a 5x5 array of input pixel values at the top-left comer of the input image array). The multiplications from each convolutional iteration can be summed together to obtain a total sum for that iteration or node. The process is next continued at a next location in the input image according to the receptive field of a next node in the convolutional hidden layer 1504. For example, a filter can be moved by a step amount (referred to as a stnde) to the next receptive field. The stride can be set to 1 or any other suitable amount. For example, if the stride is set to 1, the filter will be moved to the right by 1 pixel at each convolutional iteration. Processing the filter at each unique location of the input volume produces a number representing the filter results for that location, resulting in a total sum value being determined for each node of the convolutional hidden layer 1504.

[0154] The mapping from the input layer to the convolutional hidden layer 1504 is referred to as an activation map (or feature map). The activation map includes a value for each node representing the filter results at each location of the input volume. The activation map can include an array that includes the various total sum values resulting from each iteration of the filter on the input volume. For example, the activation map will include a 24 x 24 array if a 5 x 5 filter is applied to each pixel (a stride of 1) of a 28 x 28 input image. The convolutional hidden layer 1504 can include several activation maps in order to identify multiple features in an image. The example shown in FIG. 15 includes three activation maps. Using three activation maps, the convolutional hidden layer 1504 can detect three different kinds of features, with each feature being detectable across the entire image.

[0155] In some examples, a non-linear hidden layer can be applied after the convolutional hidden layer 1504. The non-linear layer can be used to introduce non-linearity to a system that has been computing linear operations. One illustrative example of a non-linear layer is a rectified linear unit (ReLU) layer. AReLU layer can apply the function f(x) = max(0, x) to all of the values in the input volume, which changes all the negative activations to 0. The ReLU can thus increase the non-linear properties of the CNN 1500 without affecting the receptive fields of the convolutional hidden layer 1504.

[0156] The pooling hidden layer 1506 can be applied after the convolutional hidden layer 1504 (and after the non-linear hidden layer when used). The pooling hidden layer 1506 is used to simplify' the information in the output from the convolutional hidden layer 1504. For example, the pooling hidden layer 1506 can take each activation map output from the convolutional hidden layer 1504 and generates a condensed activation map (or feature map) using a pooling function. Max-pooling is one example of a function performed by a pooling hidden layer. Other forms of pooling functions be used by the pooling hidden layer 1506, such as average pooling, L2-norm pooling, or other suitable pooling functions. A pooling function (e.g., a max-pooling filter, an L2-norm filter, or other suitable pooling filter) is applied to each activation map included in the convolutional hidden layer 1504. In the example shown in FIG. 15, three pooling filters are used for the three activation maps in the convolutional hidden layer 1504.

[0157] In some examples, max-pooling can be used by applying a max-pooling filter (e.g., having a size of 2x2) with a stride (e.g., equal to a dimension of the filter, such as a stride of 2)to an activation map output from the convolutional hidden layer 1504. The output from a maxpooling filter includes the maximum number in every sub-region that the filter convolves around. Using a 2x2 filter as an example, each unit in the pooling layer can summarize a region of 2x2 nodes in the previous layer (with each node being a value in the activation map). For example, four values (nodes) in an activation map will be analyzed by a 2x2 max-pooling filter at each iteration of the filter, with the maximum value from the four values being output as the “max” value. If such a max-pooling filter is applied to an activation filter from the convolutional hidden layer 1504 having a dimension of 24x24 nodes, the output from the pooling hidden layer 1506 will be an array of 12x12 nodes.

[0158] In some examples, an L2-norm pooling filter could also be used. The L2-norm pooling filter includes computing the square root of the sum of the squares of the values in the 2x2 region (or other suitable region) of an activation map (instead of computing the maximum values as is done in max-pooling) and using the computed values as an output.

[0159] The pooling function (e.g.. max-pooling, L2-norm pooling, or other pooling function) determines whether a given feature is found anywhere in a region of the image and discards the exact positional information. This can be done without affecting results of the feature detection because, once a feature has been found, the exact location of the feature is not as important as its approximate location relative to other features. Max-pooling (as well as other pooling methods) offer the benefit that there are many fewer pooled features, thus reducing the number of parameters needed in later layers of the CNN 1500.

[0160] The final layer of connections in the network is a fully-connected layer that connects every node from the pooling hidden layer 1506 to every one of the output nodes in the output layer 1510. Using the example above, the input layer includes 28 x 28 nodes encoding the pixel intensities of the input image, the convolutional hidden layer 1504 includes 3x24x24 hidden feature nodes based on application of a 5x5 local receptive field (for the filters) to three activation maps, and the pooling hidden layer 1506 includes alayer of 3x 12x 12 hidden feature nodes based on application of max-pooling filter to 2x2 regions across each of the three feature maps. Extending this example, the output layer 1510 can include ten output nodes. In such an example, every node of the 3x12x12 pooling hidden layer 1506 is connected to every node of the output layer 1510.

[0161] The fully connected layer 1508 can obtain the output of the previous pooling hidden layer 1506 (which should represent the activation maps of high-level features) and determines the features that most correlate to a particular class. For example, the fully connected layer 1508 can determine the high-level features that most strongly correlate to a particular class and can include weights (nodes) for the high-level features. A product can be computed between the weights of the fully connected layer 1508 and the pooling hidden layer 1506 to obtain probabilities for the different classes. For example, if the CNN 1500 is being used to predict that an object in an image is a person, high values will be present in the activation maps that represent high-level features of people (e.g., two legs are present, a face is present at the top of the object, two eyes are present at the top left and top right of the face, a nose is present in the middle of the face, a mouth is present at the bottom of the face, and / or other features common for a person).

[0162] In some examples, the output from the output layer 1510 can include an M- dimensional vector (in the prior example, M=10). M indicates the number of classes that the CNN 1500 has to choose from when classifying the object in the image. Other example outputs can also be provided. Each number in the M-dimensional vector can represent the probability the object is of a certain class. In one illustrative example, if a 10-dimensional output vector represents ten different classes of objects is [0 0 0.05 0.8 0 0.15 0 0 0 0], the vector indicates that there is a 5% probability that the image is the third class of object (e.g., a dog), an 80% probability that the image is the fourth class of object (e.g., a human), and a 15% probability that the image is the sixth class of object (e.g., a kangaroo). The probability for a class can be considered a confidence level that the object is part of that class.

[0163] FIG. 16 is a block diagram illustrating an example of a deep convolutional network 1650. The deep convolutional network 1650 may include multiple different types of layers based on connectivity and weight sharing. As shown in FIG. 16, the deep convolutional network 1650 includes the convolution blocks 1654A, 1654B. Each of the convolution blocks 1654A. 1654B may be configured with a convolution layer (CONV) 1656, a normalization layer (LNorm) 1658, and a max pooling layer (MAX POOL) 1660. Of note, the layers illustrated with respect to convolution blocks 1654A and 1654B are examples of layers that may be included in a convolution layer and are not intended to be limiting and other types of layers may be included in any order.

[0164] The convolution layers 1656 may include one or more convolutional filters, which may be applied to the input data 1652 to generate a feature map. Although only two convolution blocks 1654A, 1654B are shown, the present disclosure is not so limiting, and instead, any number of convolution blocks (e.g., convolution blocks 1654A. 1654B) may be included in the deep convolutional network 1650 according to design preference. The normalization layer 1658 may normalize the output of the convolution filters. For example, the normalization layer 1658 may provide whitening or lateral inhibition. The max pooling layer 1660 may provide down sampling aggregation over space for local invariance and dimensionality reduction.

[0165] The parallel filter banks, for example, of a deep convolutional network may be loaded on a processor such as a CPU or GPU, or any other ty pe of processor 1710 discussed with respect to the computing system 1700 of FIG. 17 to achieve high performance and low power consumption. In alternative aspects, the parallel filter banks may be loaded on a DSP or an ISP of the computing system 1700 of FIG. 17. In addition, the deep convolutional netw ork 1650 may access other processing blocks that may be present on the computing system 1700 of FIG. 17, such as sensor processor and navigation module, dedicated, respectively, to sensors and navigation.

[0166] The deep convolutional network 1650 may also include one or more fully connected layers, such as layer 1662A (labeled “FCU) and layer 1662B (labeled “FC2”). The deep convolutional netw ork 1650 may further include a logistic regression (LR) layer 1664. Between each layer 1656, 1658, 1660, 1662A, 1662B, 1664 of the deep convolutional network 1650 are weights (not shown) that are to be updated. The output of each of the layers (e.g., 1656, 1658, 1660, 1662A, 1662B, 1664) may serve as an input of a succeeding one of the layers (e.g., 1656, 1658, 1660, 1662A, 1662B, 1664) in the deep convolutional network 1650 to learn hierarchical feature representations from input data 1652 (e.g., images, audio, video, sensor data and / or other input data) supplied at the first of the convolution blocks 1654A. The output of the deep convolutional network 1650 is a classification score 1666 for the input data 1652. The classification score 1666 may be a set of probabilities, where each probability is the probability of the input data including a feature from a set of features.

[0167] FIG. 17 is a diagram illustrating an example of a system for implementing certain aspects of the present technology. In particular. FIG. 17 illustrates an example of computing system 1700, which can be for example any computing device making up internal computingsystem, a remote computing system, a camera, or any component thereof in which the components of the system are in communication with each other using connection 1705. Connection 1705 can be a physical connection using a bus. or a direct connection into processor 1710, such as in a chipset architecture. Connection 1705 can also be a virtual connection, networked connection, or logical connection.

[0168] In some examples, computing system 1700 is a distributed system in which the functions described in this disclosure can be distributed within a datacenter, multiple data centers, a peer network, etc. In some examples, one or more of the described system components represents many such components each performing some or all of the functions for which the component is described. In some cases, the components can be physical or virtual devices.

[0169] Example computing system 1700 includes at least one processing unit (CPU or processor) 1710 and connection 1705 that couples various system components including system memory 1715, such as read-only memory (ROM) 1720 and random access memory (RAM) 1725 to processor 1710. Computing system 1700 can include a cache 1712 of highspeed memory connected directly with, in close proximity to, or integrated as part of processor 1710.

[0170] Processor 1710 can include any general purpose processor and a hardware service or software service, such as services 1732, 1734, and 1736 stored in storage device 1730, configured to control processor 1710 as well as a special-purpose processor where software instructions are incorporated into the actual processor design. Processor 1710 may be a completely self-contained computing system, containing multiple cores or processors, a bus, memory controller, cache, etc. A multi-core processor may be symmetric or asymmetric.

[0171] To enable user interaction, computing system 1700 includes an input device 1745, which can represent any number of input mechanisms, such as a microphone for speech, a touch-sensitive screen for gesture or graphical input, keyboard, mouse, motion input, speech, camera, accelerometers, gyroscopes, etc. Computing system 1700 can also include output device 1735, which can be one or more of a number of output mechanisms, such as a display, speaker, haptic feedback device, etc. In some instances, multimodal systems can enable a user to provide multiple types of input / output to communicate with computing system 1700. Computing system 1700 can include communications interface 1740, which can generallygovern and manage the user input and system output. The communication interface may perform or facilitate receipt and / or transmission of wired or wireless communications using wired and / or wireless transceivers, including those making use of an audio jack / plug, a microphone jack / plug, a universal senal bus (USB) port / plug, an Apple® Lightning® port / plug, an Ethernet port / plug, a fiber optic port / plug, a proprietary wired port / plug, a BLUETOOTH® wireless signal transfer, a BLUETOOTH® low energy' (BLE) wireless signal transfer, an IBEACON® wireless signal transfer, a radio-frequency identification (RFID) wireless signal transfer, near-field communications (NFC) wireless signal transfer, dedicated short range communication (DSRC) wireless signal transfer, 802.11 Wi-Fi wireless signal transfer, wireless local area network (WLAN) signal transfer, Visible Light Communication (VLC), Worldwide Interoperability for Microwave Access (WiMAX), Infrared (IR) communication wireless signal transfer, Public Switched Telephone Network (PSTN) signal transfer, Integrated Services Digital Network (ISDN) signal transfer, 3G / 4G / 5G / LTE cellular data network wireless signal transfer, ad-hoc network signal transfer, radio wave signal transfer, microwave signal transfer, infrared signal transfer, visible light signal transfer, ultraviolet light signal transfer, wireless signal transfer along the electromagnetic spectrum, or some combination thereof. The communications interface 1740 may also include one or more Global Navigation Satellite System (GNSS) receivers or transceivers that are used to determine a location of the computing system 1700 based on receipt of one or more signals from one or more satellites associated with one or more GNSS systems. GNSS systems include, but are not limited to. the US-based Global Positioning System (GPS), the Russia-based Global Navigation Satellite System (GLONASS), the China-based BeiDou Navigation Satellite System (BDS), and the Europe-based Galileo GNSS. There is no restriction on operating on any particular hardware arrangement, and therefore the basic features here may easily be substituted for improved hardware or firmware arrangements as they are developed.

[0172] Storage device 1730 can be a non-volatile and / or non-transitory and / or computer- readable memory' device and can be a hard disk or other types of computer readable media which can store data that are accessible by a computer, such as magnetic cassettes, flash memory cards, solid state memory’ devices, digital versatile disks, cartridges, a floppy disk, a flexible disk, a hard disk, magnetic tape, a magnetic strip / stripe, any other magnetic storage medium, flash memory, memristor memory', any other solid-state memory; a compact disc read only memory' (CD-ROM) optical disc, a rewritable compact disc (CD) optical disc, digital video disk (DVD) optical disc, a blu-ray disc (BDD) optical disc, a holographic optical disk,another optical medium, a secure digital (SD) card, a micro secure digital (microSD) card, a Memory7Stick® card, a smartcard chip, a EMV chip, a subscriber identity' module (SIM) card, a mini / micro / nano / pico SIM card, another integrated circuit (IC) chip / card, random access memory (RAM), static RAM (SRAM), dynamic RAM (DRAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory' (EEPROM), flash EPROM (FLASHEPROM), cache memory' (L1 / L2 / L3 / L4 / L5 / L#), resistive random-access memory (RRAM / ReRAM). phase change memory (PCM), spin transfer torque RAM (STT-RAM), another memory chip or cartridge, and / or a combination thereof.

[0173] The storage device 1730 can include software services, servers, services, etc., that when the code that defines such software is executed by the processor 1710, it causes the system to perform a function. In some examples, a hardware service that performs a particular function can include the software component stored in a computer-readable medium in connection with the necessary' hardware components, such as processor 1710, connection 1705, output device 1735, etc., to carry out the function.

[0174] FIG. 18 is a block diagram illustrating an example of a decoder 1018 of a voice coding signal synthesis system that can be used to generate reconstructed audio (e.g., synthesized speech) using a neural speech synthesizer comprising a linear prediction coding (LPC) network. The decoder 1810 can be included in an audio codec system (e.g.. generative voice codec system) 1800 that can be used to generate reconstructed audio (e.g., synthesized speech) 1855.

[0175] In some aspects, the generative voice codec system 1800 of FIG. 18 can be the same as or similar to the generative voice codec system 1000 of FIG. 10, etc. In some cases, the channel 1840 of FIG. 18 can be the same as or similar to the channel 1040 of FIG. 10, the 12channel 1140 of FIG. 11, the channel 1240 of FIG. 12, , etc.

[0176] The decoder 1810 of FIG. 18 can be the same as or similar to the decoder 1010 of FIG. 10, etc. For example, the decoder 1010 can include an FRAE decoder 1845 the same as or similar to the FRAE decoder 1045 of FIG. 10, etc. The decoder 1810 can include a neural speech synthesizer 1850 that can be the same as or similar to the neural speech synthesizer 1050 of FIG. 10, etc. The decoder 1810 can generate reconstructed audio 1855 (e.g..synthesized speech) that can be the same as or similar to the reconstructed audio 1055 of FIG. 10, etc.

[0177] In some examples, the neural speech synthesizer 1850 can receive a first input comprising decoded spectral envelope features (e.g., determined by the FRAE decoder 1845). The neural speech synthesizer 1850 can receive a second input comprising a received pitch encoding (e.g., received pitch information), that is the same as or similar to the received pitch encoding provided to the neural speech synthesizer 1050 of FIG. 10. etc. For example, the decoder 1810 may include a pitch de-quantization engine 1838 (e.g., de O()). The pitch dequantization engine 1838 can be used to process a received pitch encoding obtained by the decoder 1810 over the channel 1840. For example, the received pitch encoding information can be quantized pitch information or a quantized pitch signal transmitted to the decoder 1810 over the channel 1840 by a corresponding encoder (e.g., an encoder associated with the decoder 1810, which may be the same as or similar to the encoder 1005 of FIG. 10, etc.). In some aspects, the pitch dequantization engine 1838 can generate reconstructed pitch information based on performing a codebook lookup to dequantize the received pitch encoding obtained from the channel 1840 (e.g., to dequantize the quantized pitch encoding received by the decoder 1810 over the channel 1840). The dequantized (e.g., reconstructed) pitch information determined by the pitch dequantization engine 1838 can include pitch information in the frequency-domain (e.g., fo pitch frequency information in Hz), pitch information in the timedomain (e.g.. pitch lag or pitch delay information in samples, milliseconds, etc.), and / or pitch correlation information. In some aspects, the dequantized (e g., reconstructed) pitch information determined by the pitch dequantization engine 1838 can include information indicative of a voiced or unvoiced (V / UV) classification.

[0178] In one illustrative example, the neural speech synthesizer 1850 can be implemented as a linear predictive coding (LPC) network. For example, the LPC network-based neural speech synthesizer 1850 can be used to implement a linear prediction (LP) synthesis filter, based on y[n] =aky[n~ k] + [n]. Here, %[n] represents the input signal to the LP filter, y[n] represents the output signal, akrepresents the linear prediction coefficients associated with implementing the LP synthesis filter, and p is a value corresponding to the LP filter order. For example, the LP filter order p can have a value of 16 or 12, etc., among various other LP filter order values.

[0179] For example, the LPC-based neural speech synthesizer 1850 can include a frame rate network 1852 configured to process inputs comprising the decoded features obtained using the FRAE decoder 1845 and the decoded pitch information obtained using the pitch dequantization engine 1838. An LPC estimation engine 1862 can process the decoded features from the FRAE decoder 1845 to determine one or more estimated LP coefficients. In some aspects, the LPC estimation engine 1862 can generate estimated LP coefficients ak, based on an input comprising the decoded features determined using the FRAE decoder 1845.

[0180] In some examples, the LP coefficients estimated using the LPC estimation engine 1862 can be provided to an LP prediction engine 1864, configured to generate as output a prediction p[n], where p[n] =— k], In some aspects, the LP prediction engine 1864 can generate the LP prediction p[n] based on a first input comprising the LP coefficients akestimated by the LPC estimation engine 1862, and a feedback input s[n-l],

[0181] The estimated or predicted LP coefficients p[n] can be provided as input to a sample rate network 1872 and a downstream combination or summation operation 1880. The sample rate network 1872 can receive additional inputs comprising the output of the frame rate network 1852, the feedback s[n-l] from the previous step n-1, and an intermediate feedback value e[n- 1] from the same previous step n-1.

[0182] The output of the sample rate network 1872 can be the probability distribution P(e[n]), which is a probability distribution for e[n], A sampling engine 1874 can perform sampling from the probability distribution P(e[n]) to obtain a realization or representation of

[0183] The representation of e[n] determined by the sampling engine 1874 can be combined with the LP prediction p[n] (e g., determined by the LP prediction engine 1864), using the combination or summation operation 1880 to thereby generate as output the signal s[n]. In some aspects, the output signal s[n] can be the same as the reconstructed audio signal 1855 of the LPC network-based neural speech synthesizer 1850 and / or decoder 1810.

[0184] The representation of e[n] determined by the sampling engine 1874 can additionally be provided to a first feedback calculation 1878, which generates as output e[n-l] provided as an additional input to the sample rate network 1872.

[0185] The output signal s[n] of the combination or summation operation 1880 can be output as a predicted sample 1182 and may additionally be provided to a second feedback calculation 1879, which generates the representation s[n-l] based on the input s[n]. The representation s[n-I ] can be provided as a feedback input to the LP prediction engine 1864 and to the sample rate network 1872.

[0186] In some cases, the predicted sample 1182 may be input to a post-processing engine 1884. The post-processing engine 1884 may be similar to post-processing engine 808 of FIG.8 and / or post-processing engine 504 of FIG. 5. In some cases, the post-processing engine 1884 may be ML based. The post-processing engine 1884 may receive a set of previously predicted samples 1886 from a delay circuit 1888 and a pitch gain 1890 and the post-processing engine 1884 may estimate an improved version of the set of previously predicted sample 1886 (S't) from the LPC layers 1854 for output as reconstructed audio 1855. In some cases, the pitch gain 1890 may be substantially similar to pitch gain 508 of FIG. 5 and / or pitch gain 814 of FIG. 8. The delay circuit 1888 may be substantially similar to delay circuit 510 of FIG. 5 and / or delaycircuit 812 of FIG. 8. The delay circuit 1888 may store the reconstructed audio 1855 (St) of the post-processing engine 1884 for input back to the post-processing engine at a later pitch period. The delay circuit may' also receive pitch lag information 1892 (e.g., from the pitch dequantization engine 1838), which may be used to determine the set of previously predicted samples 1886. As describe above, the set of previously predicted samples 1886 may include samples based on the previous pitch period (e.g., n-T), and include samples near the previous pitch period (e.g., n-T, n+T) within afixed or variable window, or samples for an entire previous pitch period. In some examples, the reconstructed audio 1855 includes synthesized speech. In some aspects, the reconstructed audio 1855 can be the same as or similar to the reconstructed audio 1055 of FIG. 10.

[0187] FIG. 19 illustrates an example of an audio codec system 1900 (e.g., generative voice codec) that includes a decoder 1910 configured to implement an LPC-based neural speech synthesizer and including an integrated process for improving pitch content, in accordance with aspects of the present disclosure. In some cases, the audio codec system 1900 may implement a LPC network-based neural speech synthesizer 1950 in a manner similar to LPC networkbased neural speech synthesizer 1850 of FIG. 18. In some examples, the audio codec system 1800 of FIG. 18 may be similar to the audio codec system 1000 of FIG. 10. In some cases, the channel 1840 of FIG. 18 may be substantially similar to the channel 1040 of FIG. 10. and maybe associated with an encoder that is the same as or substantially similar to the encoder 1005 of FIG. 10, etc.

[0188] The LPC-based neural speech synthesizer may generate a predicated sample signal s[n] of the combination or summation operation 1980 in a manner substantially similar to that described above with respect to FIG. 18. The combination function 1280 may generate a predicated sample signal s[n] to a delay circuit 1988. The delay circuit 1988 may receive pitch lag information 1992 (e.g.. from the pitch dequantization engine 1938). The delay circuit 1988 may be substantially similar to first delay circuit 606 of FIG. 6, delay circuit 908 of FIG. 9, etc. The delay circuit 1988 may output a set of previously predicted samples 1986 to the LPC estimation engine 1862 (and / or LP Predication engine 1964) based on the pitch lag information 1992. As describe above, the set of previously predicted samples 1986 may include samples based on the previous pitch period (e.g., n-T), and include samples near the previous pitch period (e.g., n-T, n+T) within a fixed or variable window, or samples for an entire previous pitch period. In some cases, the LPC estimation engine 1862 (and / or LP Predication engine 1964) may be trained to use the set of previously predicted samples 1986 to help predict relationships between the previously predicted samples and a current sample being processed by the LPC estimation engine 1862 (and / or LP Predication engine 1964) to generate a predicted sample for output (e.g., as output by the combination function 1980) as the reconstructed audio 1955 (e.g., synthesized speech).

[0189] As used herein, the term “computer-readable medium” includes, but is not limited to, portable or non-portable storage devices, optical storage devices, and various other mediums capable of storing, containing, or carry ing instruction(s) and / or data. A computer-readable medium may include a non-transitory medium in which data can be stored and that does not include carrier waves and / or transitory electronic signals propagating wirelessly or over wired connections. Examples of a non-transitory medium may include, but are not limited to, a magnetic disk or tape, optical storage media such as compact disk (CD) or digital versatile disk (DVD), flash memory, memory or memory devices. A computer-readable medium may have stored thereon code and / or machine-executable instructions that may represent a procedure, a function, a subprogram, a program, a routine, a subroutine, a module, a software package, a class, or any combination of instructions, data structures, or program statements. A code segment may be coupled to another code segment or a hardware circuit by passing and / or receiving information, data, arguments, parameters, or memory contents. Information,arguments, parameters, data, etc. may be passed, forwarded, or transmitted using any suitable means including memory sharing, message passing, token passing, network transmission, or the like.

[0190] In some examples, the computer-readable storage devices, mediums, and memories can include a cable or wireless signal containing a bit stream and the like. However, when mentioned, non-transitory computer-readable storage media expressly exclude media such as energy, carrier signals, electromagnetic waves, and signals per se.

[0191] Specific details are provided in the description above to provide a thorough understanding of the examples provided herein. However, it will be understood by one of ordinary skill in the art that the examples may be practiced without these specific details. For clarity of explanation, in some instances the present technology may be presented as including individual functional blocks including functional blocks comprising devices, device components, steps or routines in a method embodied in software, or combinations of hardware and software. Additional components may be used other than those shown in the figures and / or described herein. For example, circuits, systems, networks, processes, and other components may be shown as components in block diagram form in order not to obscure the examples in unnecessary detail. In other instances, well-known circuits, processes, algorithms, structures, and techniques may be shown without unnecessary detail in order to avoid obscuring the examples.

[0192] Individual examples may be described above as a process or method which is depicted as a flowchart, a flow diagram, a data flow diagram, a structure diagram, or a block diagram. Although a flowchart may describe the operations as a sequential process, many of the operations can be performed in parallel or concurrently. In addition, the order of the operations may be re-arranged. A process is terminated when its operations are completed, but could have additional steps not included in a figure. A process may correspond to a method, a function, a procedure, a subroutine, a subprogram, etc. When a process corresponds to a function, its termination can correspond to a return of the function to the calling function or the main function.

[0193] Processes and methods according to the above-described examples can be implemented using computer-executable instructions that are stored or otherwise available from computer-readable media. Such instructions can include, for example, instructions anddata which cause or otherwise configure a general purpose computer, special purpose computer, or a processing device to perform a certain function or group of functions. Portions of computer resources used can be accessible over a network. The computer executable instructions may be. for example, binaries, intermediate format instructions such as assembly language, firmware, source code, etc. Examples of computer-readable media that may be used to store instructions, information used, and / or information created during methods according to described examples include magnetic or optical disks, flash memory, USB devices provided with non-volatile memory, networked storage devices, and so on.

[0194] Devices implementing processes and methods according to these disclosures can include hardware, software, firmware, middleware, microcode, hardware description languages, or any combination thereof, and can take any of a variety of form factors. When implemented in software, firmware, middleware, or microcode, the program code or code segments to perform the necessary tasks (e.g., a computer-program product) may be stored in a computer-readable or machine-readable medium. A processor(s) may perform the necessarytasks. Typical examples of form factors include laptops, smart phones, mobile phones, tablet devices or other small form factor personal computers, personal digital assistants, rackmount devices, standalone devices, and so on. Functionality described herein also can be embodied in peripherals or add-in cards. Such functionality- can also be implemented on a circuit board among different chips or different processes executing in a single device, by way of further example.

[0195] The instructions, media for conveying such instructions, computing resources for executing them, and other structures for supporting such computing resources are example means for providing the functions described in the disclosure.

[0196] In the foregoing description, aspects of the application are described with reference to specific examples thereof, but those skilled in the art will recognize that the application is not limited thereto. Thus, while illustrative examples of the application have been described in detail herein, it is to be understood that the inventive concepts may be otherwise variously embodied and employed, and that the appended claims are intended to be construed to include such variations, except as limited by the prior art. Various features and aspects of the abovedescribed application may be used individually or jointly. Further, examples can be utilized in any number of environments and applications beyond those described herein without departingfrom the broader spirit and scope of the specification. The specification and drawings are, accordingly, to be regarded as illustrative rather than restrictive. For the purposes of illustration, methods were described in a particular order. It should be appreciated that in alternate examples, the methods may be performed in a different order than that described.

[0197] One of ordinary skill will appreciate that the less than (“<”) and greater than (“>’’) symbols or terminology used herein can be replaced with less than or equal to (“<”) and greater than or equal to (“>”) symbols, respectively, without departing from the scope of this description.

[0198] Where components are described as being “configured to” perform certain operations, such configuration can be accomplished, for example, by designing electronic circuits or other hardware to perform the operation, by programming programmable electronic circuits (e.g., microprocessors, or other suitable electronic circuits) to perform the operation, or any combination thereof.

[0199] The phrase “coupled to” refers to any component that is physically connected to another component either directly or indirectly, and / or any component that is in communication with another component (e.g., connected to the other component over a wired or wireless connection Claim language or other language reciting “at least one of’ a set and / or “one or more” of a set indicates that one member of the set or multiple members of the set (in any combination) satisfy the claim. For example, claim language reciting “at least one of A and B” or “at least one of A or B” means A, B, or A and B. In another example, claim language reciting “at least one of A, B, and C” or “at least one of A, B, or C” means A, B, C, or A and B, or A and C, or B and C, A and B and C, or any duplicate information or data (e.g., A and A, B and B, C and C, A and A and B, and so on), or any other ordering, duplication, or combination of A, B, and C. The language “at least one of’ a set and / or “one or more” of a set does not limit the set to the items listed in the set. For example, claim language reciting “at least one of A and B” or “at least one of A or B” may mean A, B. or A and B, and may additionally include items not listed in the set of A and B. The phrases “at least one” and “one or more” are used interchangeably herein.

[0200] Claim language or other language reciting “at least one processor configured to,” “at least one processor being configured to,” “one or more processors configured to,” “one or more processors being configured to,” or the like indicates that one processor or multipleprocessors (in any combination) can perform the associated operation(s). For example, claim language reciting “at least one processor configured to: X, Y, and Z” means a single processor can be used to perform operations X, Y, and Z: or that multiple processors are each tasked with a certain subset of operations X, Y, and Z such that together the multiple processors performX, Y, and Z; or that a group of multiple processors work together to perform operations X, Y, and Z. In another example, claim language reciting “at least one processor configured to: X, Y, and Z” can mean that any single processor may only perform at least a subset of operations X,Y, and Z.

[0201] Where reference is made to one or more elements performing functions (e.g., steps of a method), one element may perform all functions, or more than one element may collectively perform the functions. When more than one element collectively performs the functions, each function need not be performed by each of those elements (e.g., different functions may be performed by different elements) and / or each function need not be performed in whole by only one element (e g., different elements may perform different sub-functions of a function). Similarly, where reference is made to one or more elements configured to cause another element (e.g., an apparatus) to perform functions, one element may be configured to cause the other element to perform all functions, or more than one element may collectively be configured to cause the other element to perform the functions.

[0202] Where reference is made to an entity (e.g., any entity or device described herein) performing functions or being configured to perform functions (e.g., steps of a method), the entity may be configured to cause one or more elements (individually or collectively) to perform the functions. The one or more components of the entity may include at least one memory, at least one processor, at least one communication interface, another component configured to perform one or more (or all) of the functions, and / or any combination thereof. Where reference to the entity performing functions, the entity' may be configured to cause one component to perform all functions, or to cause more than one component to collectively perform the functions. When the entity is configured to cause more than one component to collectively perform the functions, each function need not be performed by each of those components (e.g., different functions may be performed by different components) and / or each function need not be performed in whole by only one component (e.g., different components may perform different sub-functions of a function).

[0203] The various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the examples disclosed herein may be implemented as electronic hardware, computer software, firmware, or combinations thereof. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps have been described above generally in terms of their functionality. Whether such functionality7is implemented as hardware or software depends upon the particular application and design constraints imposed on the overall system. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.

[0204] The techniques described herein may also be implemented in electronic hardware, computer software, firmware, or any combination thereof. Such techniques may be implemented in any of a variety of devices such as general purposes computers, wireless communication device handsets, or integrated circuit devices having multiple uses including application in wireless communication device handsets and other devices. Any features described as modules or components may be implemented together in an integrated logic device or separately as discrete but interoperable logic devices. If implemented in software, the techniques may be realized at least in part by a computer-readable data storage medium comprising program code including instructions that, when executed, performs one or more of the methods described above. The computer-readable data storage medium may form part of a computer program product, which may include packaging materials. The computer-readable medium may comprise memory or data storage media, such as random access memory (RAM) such as synchronous dynamic random access memory7(SDRAM), read-only memory7(ROM), non-volatile random access memory (NVRAM), electrically erasable programmable read-only memory (EEPROM), FLASH memory, magnetic or optical data storage media, and the like. The techniques additionally, or alternatively, may be realized at least in part by a computer- readable communication medium that carries or communicates program code in the form of instructions or data structures and that can be accessed, read, and / or executed by a computer, such as propagated signals or waves.

[0205] The program code may be executed by a processor, which may include one or more processors, such as one or more digital signal processors (DSPs), general purpose microprocessors, an application specific integrated circuits (ASICs), field programmable logicarrays (FPGAs), or other equivalent integrated or discrete logic circuitry. Such a processor may be configured to perform any of the techniques described in this disclosure. A general purpose processor may be a microprocessor; but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. A processor may also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration. Accordingly, the term “processor,” as used herein may refer to any of the foregoing structure, any combination of the foregoing structure, or any other structure or apparatus suitable for implementation of the techniques described herein. In addition, in some aspects, the functionality described herein may be provided within dedicated software modules or hardware modules configured for encoding and decoding, or incorporated in a combined video encoder-decoder (CODEC).

[0206] Illustrative aspects of the present disclosure include:

[0207] Aspect 1. An apparatus for speech synthesis, the apparatus comprising: one or more memories configured to store one or more audio samples; and one or more processors coupled to the one or more memories, the one or more processors being configured to: receive a first predicted audio sample; cause the first predicted audio sample to be stored in the one or more memories; and generate a second predicted audio sample based on the first predicted audio sample, wherein the second predicted audio sample is from a pitch penod after the first predicted audio sample.

[0208] Aspect 2. The apparatus of Aspect 1, wherein the first predicted audio sample is a part of a set of previously predicted audio samples, and wherein the second predicted audio sample is generated based on the set of previously predicted audio samples.

[0209] Aspect 3. The apparatus of Aspect 2, wherein the one or more processors are configured to generate the second predicted audio sample using a machine learning model, and wherein the machine learning model is configured to receive, as input, the set of previously predicted audio samples.

[0210] Aspect 4. The apparatus of any of Aspects 2-3, wherein the one or more processors are configured to: generate the second predicted audio sample using a machine learning model;and generate a context vector describing the set of previously predicted audio samples, wherein the machine learning model is configured to receive, as input, the context vector.

[0211] Aspect 5. The apparatus of Aspect 4, wherein the one or more processors are configured to generate the context vector based on an attention mechanism.

[0212] Aspect 6. The apparatus of any of Aspects 2-6, wherein the one or more processors are configured to: receive pitch lag information; and select the first predicted audio sample for the set of previously predicted audio samples based on the pitch lag information.

[0213] Aspect 7. The apparatus of any of Aspects 2-7, wherein, to generate the second predicted audio sample, the one or more processors are configured to: receive a third predicted audio sample; and post-process the third predicted audio sample based on the set of previously predicted audio samples to generate the second predicted audio sample.

[0214] Aspect 8. The apparatus of Aspect 7, wherein the one or more processors are configured to provide the second predicted audio sample for generating a next predicted audio sample.

[0215] Aspect 9. The apparatus of any of Aspects 7-8, wherein the one or more processors are configured to: receive pitch gain information indicating how correlated one pitch period is with another pitch period; and post-process the third predicted audio sample based on the pitch gain information.

[0216] Aspect 10. The apparatus of any of Aspects 1 -9, wherein the first predicted audio sample is generated by at least one of a sample rate network, neural homomorphic vocoder, or lyra speech synthesizer.

[0217] Aspect 11. The apparatus of any of Aspects 1-10, further including a speaker, and wherein the one or more processors are configured to output the second predicted audio sample via the speaker.

[0218] Aspect 12. A method for speech synthesis, comprising: receiving a first predicted audio sample; causing the first predicted audio sample to be stored in one or more memories; and generating a second predicted audio sample based on the first predicted audio sample, wherein the second predicted audio sample is from a pitch period after the first predicted audio sample.

[0219] Aspect 13. The method of Aspect 12, wherein the first predicted audio sample is a part of a set of previously predicted audio samples, and wherein the second predicted audio sample is generated based on the set of previously predicted audio samples.

[0220] Aspect 14. The method of Aspect 13, comprising generating the second predicted audio sample using a machine learning model, and wherein the machine learning model is configured to receive, as input, the set of previously predicted audio samples.

[0221] Aspect 15. The method of any of Aspects 13-14, comprising: generating the second predicted audio sample using a machine learning model; and generating a context vector describing the set of previously predicted audio samples, wherein the machine learning model is configured to receive, as input, the context vector.

[0222] Aspect 16. The method of Aspect 15, comprising generating the context vector based on an attention mechanism.

[0223] Aspect 17. The method of any of Aspects 13-1 , comprising: receiving pitch lag information; and selecting the first predicted audio sample for the set of previously predicted audio samples based on the pitch lag information.

[0224] Aspect 18. The method of any of Aspects 13-17, wherein generating the second predicted audio sample comprises: receiving a third predicted audio sample; and postprocessing the third predicted audio sample based on the set of previously predicted audio samples to generate the second predicted audio sample.

[0225] Aspect 19. The method of Aspect 18, comprising providing the second predicted audio sample for generating a next predicted audio sample.

[0226] Aspect 20. The method of any of Aspects 18-19, comprising: receiving pitch gain information indicating how correlated one pitch period is with another pitch period; and postprocessing the third predicted audio sample based on the pitch gain information.

[0227] Aspect 21. The method of any of Aspects 12-20, wherein the first predicted audio sample is generated by at least one of a sample rate network, neural homomorphic vocoder, or lyra speech synthesizer.

[0228] Aspect 22. A non-transitory computer-readable medium having stored thereon instructions that, when executed by at least one processor, cause the at least one processor to: receive a first predicted audio sample; cause the first predicted audio sample to be stored in one or more memories; and generate a second predicted audio sample based on the first predicted audio sample, wherein the second predicted audio sample is from a pitch period after the first predicted audio sample.

[0229] Aspect 23. The non-transitory computer-readable medium of Aspect 22, wherein the first predicted audio sample is a part of a set of previously predicted audio samples, and wherein the second predicted audio sample is generated based on the set of previously predicted audio samples.

[0230] Aspect 24. The non-transitory computer-readable medium of Aspect 23, wherein the instructions cause the one or more processors to generate the second predicted audio sample using a machine learning model, and wherein the machine learning model is configured to receive, as input, the set of previously predicted audio samples.

[0231] Aspect 25. The non-transitory computer-readable medium of any of Aspects 23-24, wherein the instructions cause the one or more processors to: generate the second predicted audio sample using a machine learning model; and generate a context vector describing the set of previously predicted audio samples, wherein the machine learning model is configured to receive, as input, the context vector.

[0232] Aspect 26. The non-transitory computer-readable medium of Aspect 25, wherein the instructions cause the one or more processors to generate the context vector based on an attention mechanism.

[0233] Aspect 27. The non-transitory computer-readable medium of any of Aspects 23-26, wherein the instructions cause the one or more processors: receive pitch lag information; and select the first predicted audio sample for the set of previously predicted audio samples based on the pitch lag information.

[0234] Aspect 28. The non-transitory computer-readable medium of any of Aspects 23-27. wherein, to generate the second predicted audio sample, the instructions cause the one or more processors to: receive a third predicted audio sample; and post-process the third predicted audiosample based on the set of previously predicted audio samples to generate the second predicted audio sample.

[0235] Aspect 29. The non-transitory computer-readable medium of Aspect 28, wherein the instructions cause the one or more processors to provide the second predicted audio sample for generating a next predicted audio sample.

[0236] Aspect 30. The non-transitory computer-readable medium of any of Aspects 28-29, wherein the instructions cause the one or more processors: receive pitch gain information indicating how correlated one pitch period is with another pitch period; and post-process the third predicted audio sample based on the pitch gain information.

[0237] Aspect 31. The non-transitory computer-readable medium of any of Aspects 22-30, wherein the first predicted audio sample is generated by at least one of a sample rate network, neural homomorphic vocoder, or lyra speech synthesizer.

[0238] Aspect 32. The non-transitory computer-readable medium of any of Aspects 22-31, wherein the instructions cause the one or more processors to output the second predicted audio sample via a speaker.

[0239] Aspect 33: An apparatus for speech synthesis, comprising means for performing one or more of operations according to any of Aspects 12 to 21.

Claims

CLAIMSWhat is Claimed Is:

1. An apparatus for speech synthesis, the apparatus comprising: one or more memories configured to store one or more audio samples; and one or more processors coupled to the one or more memories, the one or more processors being configured to: receive a first predicted audio sample; cause the first predicted audio sample to be stored in the one or more memories; and generate a second predicted audio sample based on the first predicted audio sample, wherein the second predicted audio sample is from a pitch period after the first predicted audio sample.

2. The apparatus of claim 1, wherein the first predicted audio sample is a part of a set of previously predicted audio samples, and wherein the one or more processors are configured to generate the second predicted audio sample based on the set of previously predicted audio samples.

3. The apparatus of claim 2, wherein the one or more processors are configured to generate the second predicted audio sample using a machine learning model, and wherein the machine learning model is configured to receive, as input, the set of previously predicted audio samples.

4. The apparatus of claim 2, wherein the one or more processors are configured to: generate the second predicted audio sample using a machine learning model; and generate a context vector describing the set of previously predicted audio samples, wherein the machine learning model is configured to receive, as input, the context vector.

5. The apparatus of claim 4, wherein the one or more processors are configured to generate the context vector based on an attention mechanism.

6. The apparatus of claim 2, wherein the one or more processors are configured to:receive pitch lag information; and select the first predicted audio sample for the set of previously predicted audio samples based on the pitch lag information.

7. The apparatus of claim 2, wherein, to generate the second predicted audio sample, the one or more processors are configured to: receive a third predicted audio sample; and post-process the third predicted audio sample based on the set of previously predicted audio samples to generate the second predicted audio sample.

8. The apparatus of claim 7, wherein the one or more processors are configured to provide the second predicted audio sample for generating a next predicted audio sample.

9. The apparatus of claim 7, wherein the one or more processors are configured to: receive pitch gain information indicating how correlated one pitch period is with another pitch period; and post-process the third predicted audio sample based on the pitch gain information.

10. The apparatus of claim 1, wherein the first predicted audio sample is generated by at least one of a sample rate network, neural homomorphic vocoder, neural speech synthesizer, or lyra speech synthesizer.

11. The apparatus of claim 1 , further including a speaker, and wherein the one or more processors are configured to output the second predicted audio sample via the speaker.

12. A method for speech synthesis, comprising: receiving a first predicted audio sample; causing the first predicted audio sample to be stored in one or more memories; and generating a second predicted audio sample based on the first predicted audio sample, wherein the second predicted audio sample is from a pitch period after the first predicted audio sample.

13. The method of claim 12, wherein the first predicted audio sample is a part of a set of previously predicted audio samples, and wherein the second predicted audio sample is generated based on the set of previously predicted audio samples.

14. The method of claim 13, further comprising generating the second predicted audio sample using a machine learning model, and wherein the machine learning model is configured to receive, as input, the set of previously predicted audio samples.

15. The method of claim 13, further comprising: generating the second predicted audio sample using a machine learning model; and generating a context vector describing the set of previously predicted audio samples, wherein the machine learning model is configured to receive, as input, the context vector.

16. The method of claim 15, further comprising generating the context vector based on an attention mechanism.

17. The method of claim 13. further comprising: receiving pitch lag information; and selecting the first predicted audio sample for the set of previously predicted audio samples based on the pitch lag information.

18. The method of claim 13, wherein generating the second predicted audio sample comprises: receiving a third predicted audio sample; and post-processing the third predicted audio sample based on the set of previously predicted audio samples to generate the second predicted audio sample.

19. The method of claim 18, further comprising providing the second predicted audio sample for generating a next predicted audio sample.

20. A non-transitory computer-readable medium having stored thereon instructions that, when executed by one or more processors, cause the one or more processors to: receive a first predicted audio sample;cause the first predicted audio sample to be stored in one or more memories; and generate a second predicted audio sample based on the first predicted audio sample, wherein the second predicted audio sample is from a pitch period after the first predicted audio sample.

Images