Voice restoration model adaptation for noise suppression and voice conversion systems

Abstract

Systems and techniques are provided for audio processing. For instance, a process can include obtaining an enhanced audio frame, wherein the enhanced audio frame is based on an audio frame; encoding a first set of content embeddings and a second set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame, and wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly projecting the encoded first set of content embeddings to generate a first predicted spectrogram; linearly projecting the encoded second set of content embeddings to generate a second predicted spectrogram; and determining a first loss value based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame; and determining a second loss value based on a difference between the second predicted spectrogram and the spectrogram.

Description

PATENT Qualcomm Ref. No.2407249WO 1 VOICE RESTORATION MODEL ADAPTATION FOR NOISE SUPPRESSION AND VOICE CONVERSION SYSTEMS FIELD

[0001] This application is related to speech restoration systems. For example, aspects of the application relate to adaptation techniques for a voice restoration model for noise suppression and voice conversion systems and related techniques. BACKGROUND

[0002] Speech restoration systems may be used to convert degraded speech signals, such as a speech signal accompanied by background noise, audio obtained using a microphone too far away from a speech source, speech signal with interference, etc., into high-quality speech signals. As examples, speech restoration systems may be used to reduce severe background noise for a phone call. Speech restoration systems may differ from noise suppression as speech restoration systems may replace damaged vocal elements of speech with synthesized vocal elements. As speech signals transmitted between devices becomes ever more important in daily lives, speech signals may be received in ever more challenging scenarios, such as locations with severe background noise, using degraded hardware, under low bandwidth conditions, and the like. Techniques for convert degraded speech signals into high quality speech signals may be useful. SUMMARY

[0003] Systems and techniques are described herein for audio processing. For example, an apparatus for audio processing training is provided. The apparatus includes one or more memories configured to store one or more audio frames and one or more processors coupled to the one or more memories. The one or more processors are configured to: reconstruct target speech using a voice conversion generative model, wherein the voice conversion generative model is pretrained, and wherein pretraining the voice conversion generative model causes the voice conversion generative model to: obtain an audio frame of the one or more audio frames, the audio frame including speech from a speaker; obtain an enhanced audio frame, wherein the enhanced audio frame is Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 2 based on the audio frame; encode a first set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encode a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly project the encoded first set of content embeddings to generate a first predicted spectrogram; linearly project the encoded second set of content embeddings to generate a second predicted spectrogram; and adjust the voice conversion generative model based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame and a difference between the second predicted spectrogram and the spectrogram.

[0004] As another example, a method for audio processing training is provided. The method includes: reconstructing target speech using a voice conversion generative model, wherein the voice conversion generative model is pretrained by: obtaining an audio frame of a set of one or more audio frames, the audio frame including speech from a speaker; obtaining an enhanced audio frame, wherein the enhanced audio frame is based on the audio frame; encoding a first set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encoding a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly projecting the encoded first set of content embeddings to generate a first predicted spectrogram; linearly projecting the encoded second set of content embeddings to generate a second predicted spectrogram; and adjusting the voice conversion generative model based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame and a difference between the second predicted spectrogram and the spectrogram.

[0005] In another example, a non-transitory computer-readable medium is provided. The non-transitory computer-readable medium includes stored thereon instructions that, when executed by at least one processor, cause the at least one processor to: reconstruct target speech using a voice conversion generative model, wherein the voice conversion generative model is pretrained, and wherein pretraining the voice conversion generative model causes the voice conversion generative model to: obtain an audio frame of the one or more audio frames, the audio frame including speech from a speaker; obtain an Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 3 enhanced audio frame, wherein the enhanced audio frame is based on the audio frame; encode a first set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encode a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly project the encoded first set of content embeddings to generate a first predicted spectrogram; linearly project the encoded second set of content embeddings to generate a second predicted spectrogram; and adjust the voice conversion generative model based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame and a difference between the second predicted spectrogram and the spectrogram.

[0006] As another example, an apparatus for audio processing is provided. The apparatus includes: means for reconstructing target speech using a voice conversion generative model, wherein the voice conversion generative model is pretrained by: obtaining an audio frame of a set of one or more audio frames, the audio frame including speech from a speaker; obtaining an enhanced audio frame, wherein the enhanced audio frame is based on the audio frame; encoding a first set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encoding a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly projecting the encoded first set of content embeddings to generate a first predicted spectrogram; linearly projecting the encoded second set of content embeddings to generate a second predicted spectrogram; and adjusting the voice conversion generative model based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame and a difference between the second predicted spectrogram and the spectrogram.

[0007] 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 Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 4 apparatuses can include at least one camera for capturing one or more images or video frames. For example, the one or more apparatuses can 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.

[0008] 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.

[0009] 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

[0010] Illustrative examples of the present application are described in detail below with reference to the following figures:

[0011] FIG.1 is a block diagram illustrating a speech restoration system using voice conversion techniques, in accordance with aspects of the present disclosure.

[0012] FIG. 2 is a block diagram illustrating a speech restoration system including a diffusion-based decoder and self-supervised learning (SSL) based encoder, in accordance with aspects of the present disclosure. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 5

[0013] FIG.3 is a is a block diagram illustrating a speech restoration system including a diffusion-based decoder, SSL based encoder, and SSL-based noise suppression, in accordance with aspects of the present disclosure.

[0014] FIG. 4 is a block diagram illustrating a technique for training a voice conversion engine 402, in accordance with aspects of the present disclosure.

[0015] FIG. 5 is a block diagram illustrating a technique for training 500 with a content encoder adaption model, in accordance with aspects of the present disclosure.

[0016] FIG. 6 is a block diagram illustrating a technique for training 600 a noise suppression model with a content encoder adaption model, in accordance with aspects of the present disclosure.

[0017] FIG. 7 is a flow diagram illustrating a process for audio processing training, in accordance with aspects of the present disclosure.

[0018] FIG. 8 provides two sets of images that show the forward diffusion process (which is fixed) and the reverse diffusion process (which is learned) of a diffusion model.

[0019] FIG. 9 is a diagram illustrating how diffusion data is distributed from initial data to noise using a diffusion model in the forward diffusion direction, in accordance with some aspects.

[0020] FIG.10 is a diagram illustrating a U-Net architecture for a diffusion model, in accordance with some aspects.

[0021] FIG. 11 is a diagram illustrating an illustrative example of a neural network (e.g., a deep-learning neural network, in accordance with some examples.

[0022] FIG. 12 is a diagram illustrating an illustrative example of a convolutional neural network (CNN), in accordance with some examples.

[0023] FIG. 13 is a diagram illustrating an example of a system for implementing certain aspects of the present technology. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 6 DETAILED DESCRIPTION

[0024] 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.

[0025] 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.

[0026] Devices may receive audio from a variety of sources, such as from microphones, from a networked device, from a stored file, etc., and often this received audio may include voice signal (e.g., speech signals). In some cases, a device may perform (e.g., execute software) to perform certain audio tasks to process the speech signals (e.g., speech waveform). As an example, speech signals in received audio may sometimes include background noise and a noise suppression audio task may be used to reduce an amount of background noise. In some cases, noise suppression may attempt to remove background noise to improve a perceived quality of the speech signal. For example, for noise suppression, a noisy speech signal may be input to a noise suppression model, which may use machine learning (ML) models, masks or other subtractive techniques, to produce clean speech (e.g., with less or no background noise). However, noise suppression techniques may remove portions of the voice signal in addition to the background noise, for example, if too much noise suppression is applied or if the signal- to-noise ratio (SNR) is too low (e.g., the noise signal is louder than the voice signal). Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 7

[0027] Speech restoration may be another audio task which may be performed by a device to help improve the perceived quality of the speech signal. Speech restoration may convert a degraded speech signal to a higher quality speech signal. Degraded speech signals may be degraded / damaged due to, for example, high levels of background noise, interference, hardware and / or networking issues, etc. In some cases, speech restoration may include aspects of noise suppression, such as removing background noise. Speech restoration may differ from noise suppression in that speech restoration may generate and / or replace damaged vocal elements of speech signals.

[0028] Another audio task may include voice conversion. In voice conversion, the speech of a source speaker may be modified to sound like a target speaker without changing the contents of the speech from the source speaker. For example, voice conversion may be performed by a voice conversion model (also referred to as a voice conversion generative model), which may be a ML model, that may receive information about a target speaker, such as their speaking style (e.g., speaker embedding). The voice conversion model may also receive speech from the source speaker (e.g., source speech) and extract linguistic information from the source speech. The voice conversion model may apply the information about the target speaker (e.g., the speaking style) to the extracted linguistic information (from the source speech) to generate target speech which sounds like speech from the target speaker. The linguistic information from the source speech may be preserved in voice conversion. In some cases, the voice conversion model may be trained using clean speech and a loss generated based on an output of the voice conversion model and a spectrogram generated based on the clean speech. In some cases, training with clean speech may have a high character error rate, indicating that the voice conversion engine may be hallucinating for noisy / enhanced input.

[0029] Systems, apparatuses, processes (also referred to as methods), and computer- readable media (collectively referred to as “systems and techniques”) are described herein for adaptation techniques for a voice restoration model for noise suppression and voice conversion systems and related techniques. For example, in some cases, training for a noise suppression and voice conversion system may be enhanced. The training for a noise suppression and voice conversion system may include obtaining a clean audio frame Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 8 along with an enhanced audio frame based on the clean audio frame. In some cases, the enhanced audio frame may be a noisy audio frame (e.g., the clean audio frame plus noise). In other cases, the enhanced audio frame may be a noise suppressed audio frame generated by a noise suppression model. For example, the noise suppression model may obtain a noisy audio frame and the noise suppression model may perform noise suppression on the noisy audio frame to generate the enhanced audio frame. In some cases, a loss value between the enhanced audio frame (e.g., noise suppressed noisy audio frame) and the clean audio frame and the noise suppression model may be trained based on this loss value.

[0030] Based on the clean audio frame, a first set of content embeddings may be generated and the first set of content embeddings may be encoded. The encoded first set of content embeddings may be linearly projected to generate a first predicted spectrogram. Similarly, based on the enhanced audio frame, a second set of content embeddings may be generated and the second set of content embeddings may be encoded and the encoded second set of content embeddings may be linearly projected to generate a second predicted spectrogram. A clean spectrogram may be generated based on the clean audio frame. For example, the clean audio frame may be passed through a mel filterbank to produce the clean spectrogram. A first loss value may then be determined based on a difference between the first predicted spectrogram and the clean spectrogram. A second loss value may also be determined based on a difference between the second spectrogram and the clean spectrogram. A voice conversion generative model may then be trained based on the first loss value and the second loss value. In some examples, target speech may be reconstructed using the linearly projected encoded first set of content embeddings. This target speech may be compared to the clean audio frame to determine a third loss value. The voice conversion generative model may also be trained based on this third loss value.

[0031] In some cases, the voice conversion generative model may include an SSL encoder extract content embeddings from the clean audio frame, a content encoder to encode content embeddings, a linear projection engine to linearly project the encoded Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 9 content embeddings, and a diffusion decoder to reconstruct target speech using the linearly projected encoded content embeddings.

[0001] Various aspects of the techniques described herein will be discussed below with respect to the figures. Further aspects and examples related to the present disclosure are included in Appendix A attached hereto, the contents of which is hereby incorporated by reference in its entirety and for all purposes.

[0032] FIG. 1 is a block diagram illustrating a speech restoration system 100 using voice conversion techniques, in accordance with aspects of the present disclosure. In FIG. 1, noisy (e.g., degraded) input speech 102 (e.g., speech waveform) from a reference speaker in an audio frame may be input to a first stage noise suppression model 104 of a noise suppression engine 120. The noise suppression model 104 may be any technique for noise suppression, such as a ML model for performing noise suppression. Noise suppressed audio output 105 from the noise suppression model 104 may be input to a second stage voice conversion generative model 106 of a voice conversion engine 122. The voice conversion generative model 106 may also receive a speaker embedding 108 for the reference person. In some cases, the voice conversion generative model 106 may be any voice conversion model.

[0033] In some cases, the speaker embedding 108 may be generated using clean speech 110 provided by the reference person, for example, during an enrollment process. The clean speech 110 may be input to a speaker encoder 112 trained to extract information about a speaking style of the reference person as the speaker embedding 108. In some cases, clean speech 110 from the reference person may not be available, and noisy input speech 102 or noise suppressed audio output 105 may be input to the speaker encoder 112 to generate the speaker embedding 108. In such cases, the speaker encoder 112 may be trained to generate speaker embeddings 108 based on generic speakers. In some cases, the speaker encoder 112 may be a pre-trained speaker embedding extraction model such as ECAPA-TDNN. In some cases ECAPA-TDNN may include blocks of time delay neural blocks (TDNNs) and squeeze and excite (SE) layers unified with blocks of Res2Block layers. The speaker encoder 112 may receive, as input, a spectrogram, such as a mel-spectrogram, and outputs a vector including speaker style information. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 10

[0034] In some cases, the noise suppressed audio output 105 may be input to a voice quality evaluator 114. The voice quality evaluator 114 may evaluate a quality of the noise suppressed audio output 105 to dynamically adjust an iterative process between the noise suppression model 104 and the voice conversion generative model 106. In some cases, the quality of the noise suppressed audio output 105 may be evaluated based on one or more metrics. These metrics may include a signal-to-noise ratio (SNR), word error count rates, or another quality measure. In some cases, the iterative process may be adjusted by the voice quality evaluator by adjusting a number of iterations between the noise suppression model 104 and the voice conversion generative model 106 may be performed. For example, where the quality of the noise suppressed audio output 105 is evaluated based on SNR, a low SNR may indicate that the quality of the noise suppressed audio output 105 is low and more dynamic iterations 116 of the noise suppression model 104 and the voice conversion generative model 106 may be performed. As an example, where the metric is below a first threshold (e.g., where SNR < 0), a maximum number of dynamic iterations 116 may be performed. If the metric is between the first threshold and a second threshold a reduced number of dynamic iterations 116 may be performed. If the metric is above a second threshold, a minimum number of dynamic iterations 116 may be performed. As another example, dynamic iterations 116 may be performed until the metric is above a certain threshold. For an iteration of the dynamic iterations 116, output from the voice conversion generative model 106 may be input to the noise suppression model 104 to generate another noise suppressed audio output 105 which may be passed into the voice conversion generative model 106 to generate higher quality output samples. After a number of dynamic iterations 116, restored speech 118 may be output from the voice conversion generative model 106.

[0035] FIG. 2 is a block diagram illustrating a speech restoration system 200 including a diffusion-based decoder and self-supervised learning (SSL) based encoder, in accordance with aspects of the present disclosure. In FIG.2, noisy input speech 202 from a reference speaker may be input to a first stage noise suppression model 204 of a noise suppression engine 220. In some cases, the input speech 202 and noise suppression model 204 may be substantially similar to input speech 102 and noise suppression model 104, Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 11 respectively, of FIG.1 as discussed above. In some cases, output of the noise suppression model 204 may be passed to a neural vocoder 224. The neural vocoder 224 may convert a representation of an audio signal (e.g., a vector representation of speech) to a waveform of the audio signal. In some cases, the noise suppression engine 220 may be used in place of noise suppression engine 120 of FIG.1. Noise suppressed audio output 205 of the noise suppression engine 220 may be input to a voice conversion engine 222.

[0036] In some cases, the noise suppressed audio output 205 may be input to an SSL based encoding engine 226 and a feature extractor 228 of the voice conversion engine 222. The SSL based encoding engine 226 may include an SSL encoder 230 and a content encoder 232. In some cases, the SSL encoder 230 may be ML based, such as based on a Hidden-Unit BERT (HuBERT) ML model for speech representation learning. In some cases, the SSL encoder 230 may learn to, using unlabeled training data, extract content embeddings from the noise suppressed audio output 205. In some cases, output of the SSL encoder 230 may be quantized based on a vector quantization (VQ) codebook. Vector quantization (VQ) may model a probability density of the encoded signal, for example, to characterize the signal with a description. In some cases, the VQ based on a VQ codebook may be skipped.

[0037] These content embeddings may be passed to the content encoder 232 for encoding and the encoded content embeddings passed to a diffusion decoder 234 The f0 extractor 228 may extract f0 features of the noise suppressed audio output 205 and pass the extracted audio features to the diffusion decoder 234. For instance, a content embedding can be a vector with shape including a number of frames, HuBERT feature dimension, etc.

[0038] The speech restoration system 200 may include clean speech 210 from the reference person and a speaker encoder 212 to produce speaker embeddings 208 for the reference person in a manner substantially similar to that discussed above with respect to clean speech 110, speaker encoder 112 and speaker embedding 108 of FIG.1.

[0039] Output of the content encoder 232 and feature extractor 228 may be input to the diffusion decoder 234 along with the speaker embeddings 208. The diffusion decoder Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 12 234 may be a diffusion based ML model and the diffusion decoder 234 may generate / reconstruct the target speech by adding random noise to input o and then reversing the encoding to construct speech spectrograms. The diffusion decoder 234 may operate iteratively 236 to reconstruct the speech spectrograms. In some cases, the diffusion decoder 234 may dynamically iterate 236 cased on a metric, such as the SNR of a restored spectrogram 238. In some cases, the diffusion decoder 234 may determine how many times to dynamically iterate 236 in a manner similar to that described above with respect to the voice quality evaluator 114 of FIG. 1. The diffusion decoder 234 may dynamically iterate 236 by inputting the output of the diffusion decoder 234 back into the diffusion decoder 234 along with the extracted features from the feature extractor 228 and output of the content encoder 232. The diffusion decoder 234 may output a restored spectrogram 238 (e.g., as a mel-spectrogram) to a neural vocoder 240 of the voice conversion engine 222. The neural vocoder 240 may convert a representation of an audio signal (e.g., spectrogram, mel-spectrogram, etc.) to a waveform of the audio signal to generate the restored speech 242. An example of a neural vocoder 240 may be a high fidelity generative adversarial network (HiFi-GAN), which may use multi-receptive field fusion to upsample mel-spectrograms into waveforms. In some cases, the voice conversion engine 222 may be used in place of the voice conversion engine 122 of FIG. 1.

[0040] FIG. 3 is a block diagram illustrating a speech restoration system 300 including a diffusion-based decoder, SSL based encoder, and SSL-based noise suppression, in accordance with aspects of the present disclosure. The speech restoration system 300 includes a voice conversion engine 322 including an SSL encoder 330, content encoder 332, feature extractor 328, diffusion decoder 334, and neural vocoder 340 which may be substantially similar to the voice conversion engine 222, SSL encoder 230, content encoder 232, feature extractor 228, diffusion decoder 234, and neural vocoder 240, respectively, of FIG. 2. Similarly, the speech restoration system 300 may receive input speaker embeddings generated based on clean speech 310 from a speaker encoder 312 in a manner substantially similar to that described above with respect to the clean speech 210 and speaker encoder 212 of FIG.2. The speech restoration system 300 Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 13 includes a VQ noise suppression engine 350 which may perform a similar role as the noise suppression engine 220 of FIG. 2. The VQ noise suppression engine 350 includes an SSL encoder 352, a content encoder 354, a VQ noise suppressor 356, and a neural vocoder 324. The SSL encoder 352 and a content encoder 354 may be similar to SSL encoder 330 and content encoder 332, respectively, which allows the VQ noise suppression engine 350 to perform some in-painting / correction and improve content embedding with the SSL encoder 352. In some cases, the speaker embeddings from the speaker encoder 312 may be passed to the VQ noise suppressor 356 to allow for a speaker- dependent noise suppression stage. The speaker-dependent noise suppression stage allows the noise suppression to be tailored for speaker to help avoid suppressing vocal elements of the speaker. In some cases, the neural vocoder 324 may be substantially similar to neural vocoder 224 of FIG.2.

[0041] FIG. 4 is a block diagram illustrating a technique for training 400 a voice conversion engine 402, in accordance with aspects of the present disclosure. The voice conversion engine 402 may correspond to voice conversion generative model 106 of FIG. 1, voice conversion engine 222 of FIG.2, voice conversion engine 322 of FIG.3, etc. The voice conversion engine 402 includes an SSL encoder 404, content encoder 406, linear projection engine 408, and a diffusion decoder 410. The SSL encoder 404 may correspond to SSL encoder 230 of FIG.2, SSL encoder 330 of FIG. 3, etc. The content encoder 406 may correspond to content encoder 232 of FIG.2, content encoder 332 of FIG.3, etc. The diffusion decoder 410 may correspond to diffusion decoder 234 of FIG. 2, diffusion decoder 334 of FIG.3, etc. During training, clean source speech 412 may be input to the SSL encoder 404, which may process the source speech 412 as described above with respect to FIG. 2 and / or FIG. 3. The content encoder 406 may process output from the SSL encoder 404 in a manner as described above with respect to FIG.2 and / or FIG.3. In this example, the VQ based on a VQ codebook is skipped. The encoded content embeddings output from the content encoder 406 may be input to the linear projection engine 408. The linear projection engine 408 may perform a linear projection of the encoded content embeddings to, for example, adjust content embedding dimensions to have the same dimension of the mel filterbank 416 output, for input to the diffusion Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 14 decoder 410. In some cases, the linear projection engine 408 may be considered as part of the content encoder 406 (e.g., performed by a layer of the content encoder 406, such as a final layer of the content encoder 406). In some cases, an encoder loss 414 ℒ^୬^may be determined based on the output of the linear projection engine 408 and an output of a mel filterbank 416. The mel filterbank 416 may used during training and the mel filterbank 416 may receive the source speech 412 and generate a clean (e.g., based on non-noisy audio frames) mel spectrogram ^^^from the source speech 412. This clean mel spectrogram ^^^may be used as a ground truth for comparison with the encoder loss 414. The encoder loss may be expressed as ℒ^୬^൫^^^,^^^ ൯ ൌ ^^൫^^^, ^^^ ൯, where ^^^may be the output of the linear projection ^^ metricbetween two vectors.

[0042] The output of the linear projection engine 408 may also be input to the diffusion decoder 410. In some cases, clean source speech 412 may be input to a speaker encoder 420. The speaker encoder 420 may correspond to speaker encoder 112, of FIG. 1, speaker encoder 212 of FIG.2, speaker encoder 312 of FIG.3, etc. The speaker encoder 420 may produce speaker embeddings for the reference person in a manner substantially similar to that discussed above with respect to clean speech 110, speaker encoder 112 and speaker embedding 108 of FIG.1. The speaker embeddings may be input to the diffusion decoder 410. The diffusion decoder 410 may generate / reconstruct target speech, as described above with respect to FIGs. 2-3. As a part of training, the generated / reconstructed target speech may be used to determine a diffusion loss 422 ℒୈ୧^^^^^^^. The diffusion loss 422 may be based on a difference between a clean mel spectrogram from clean source speech 412 (e.g., as a ground truth) and the output of the diffusion decoder 410. In some cases, the diffusion loss 422 may be described as ଶ ℒୈ୧^^^^^^^ ൌ ^^ఢ^ ^ฮ^^^௧^^ఏ൫^^௧, ^^|^^^ , ^^൯ ^where ^^௧is a noisy Mel-spectrogram atfrom the speaker encoder, ^^௧represents a scaling factor at time step t, ^^௧represents the actual added noise, and ^^ఏa score estimator from the diffusion decoder 410, such as a U-net decoder. In some cases, a total loss may be expressed as ℒ ൌ ℒୈ୧^^൫^^^,^^^ ൯ ^ ^^ℒ^୬ୡ൫^^^, ^^^ ൯, where^^ represents a scale factor for the loss terms.Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 15

[0043] In some cases, training with the technique for training 400 may provide extended in-painting of noise suppressed speech that goes beyond looking at neighboring speech regions to reconstruct overly suppressed speed components, even with speaker agonistically trained noise suppression and voice conversion engines. In some cases, training with the technique for training 400 may have a high character error rate, indicating that the voice conversion engine may be hallucinating for noisy / enhanced input. In some cases, a content encoder adaption model may be used for training voice conversions may be useful.

[0044] FIG. 5 is a block diagram illustrating a technique for training 500 with a content encoder adaption model, in accordance with aspects of the present disclosure. The technique for training 500 may include a VC engine 502 with a SSL encoder 504, content encoder 506, linear projection engine 508, and diffusion decoder 510, along with a speaker encoder 520 and mel filterbank 516, for determining a diffusion loss 522 and encoder loss 514 from source speech 512 in a manner that is substantially similar to the technique for training 400, including the VC engine 402 , SSL encoder 404, content encoder 406, linear projection engine 508, diffusion decoder 410, speaker encoder 420, mel filterbank 416, diffusion loss 422, and encoder loss 414, respectively.

[0045] In the technique for training 500, noisy speech or enhanced noisy speech 530 (e.g., noisy speech substantially similar to noisy speech 102 of FIG. noisy speech 202 of FIG. 2, etc., or enhanced noisy speech substantially similar to noise suppressed audio output 105 of FIG.1, noise suppressed audio output 205 of FIG.2, etc.) may be received by an SSL encoder 532. In some cases, the SSL encoder 532 may be substantially similar to SSL encoder 404 and the SSL encoder 532 may extract content embeddings from the noisy speech / enhanced noisy speech 530. The extracted content embeddings may be passed to a content encoder 534. The content encoder 534 may be substantially similar to content encoder 406, and the content encoder 534 may encode the content embeddings. The encoded content embeddings may be passed to a linear projection engine 536. The linear projection engine may be substantially similar to linear projection engine 508 and the linear projection engine 508 may perform a linear projection of the encoded content embeddings to, for example, adjust content embedding dimensions to have the same Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 16 dimension of the mel filterbank 516 output, for input to the diffusion decoder 510. The output 538 ^^^ ௩^ of the content encoder after linear projection may be used to determinea noise robustness loss 540 ℒேோ. The noise robustness loss 540 may be used to reduce a content embedding distance between clean and noisy / enhanced speech. The noise robustness loss 540 may be expressed as ℒேோ൫^^^,^^^ ௩^൯ ൌ ^^൫^^^,^^^ ௩^൯. Where thenoise robustness loss 540 is used along with the diffusion loss 522 and encoder loss 514, the total loss for training may be expressed as: ℒ ൌ ℒୈ୧^^൫^^^,^^^ ൯ ^ℒ^୬ୡ൫^^^,^^^ ൯+^^ℒ^ୖ൫^^^,^^^ ௩^൯.technique for training 500 may be finetuned to train the noise suppression model as well. FIG. 6 is a block diagram illustrating a technique for training 600 a noise suppression model with a content encoder adaption model, in accordance with aspects of the present disclosure. The technique for training 600 may include a VC engine 602 with a SSL encoder 604, content encoder 606, linear projection engine 608, and diffusion decoder 610, speaker encoder 620 mel filterbank 616, SSL encoder 632, content encoder 634, output of the content encoder 634, and linear projection engine 636 for determining a diffusion loss 622, encoder loss 614, and noise robustness loss 640 from source speech 612 and enhanced noisy speech 630 (e.g., based on outputs ^^^ , ^^^, and ^^^ ௩^) in a manner that is substantially similar to the technique for training500, including the VC engine 502 , SSL encoder 504, content encoder 506, linear projection engine 608, diffusion decoder 510, speaker encoder 520, mel filterbank 516, SSL encoder 532, content encoder 534, output 538 of the content encoder 534, and linear projection engine 536, diffusion loss 522, encoder loss 514, noise robustness loss 540, source speech 512 and enhanced noisy speech 530, respectively.

[0047] In the technique for training 600, noisy speech 650 may be passed into a noise suppression model 652. The noise suppression model 652 may be substantially similar to noise suppression model 104 of FIG. 1 and noise suppression model 204 of FIG. 2, VQ noise suppression engine 350 of FIG. 3, etc.). The noise suppression model 652 may perform noise suppression on the audio frame to generate a noise suppressed audio output ^^^^^as enhanced noisy speech 630. This enhanced noisy speech 630 may be passed into Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 17the SSL encoder 632. In some cases, a noise suppression loss ℒ^ௌ൫^^^,^^^ ^^൯ may bedetermined between the enhanced noisy speech 630 (e.g., noise suppressed audio output^^^^^) and the clean mel spectrogram ^^^. In some cases, the noise suppression loss maybe used to train the noise suppression model 652. The total loss for training may then beexpressed as ℒ ൌ ℒୈ୧^^൫^^^,^^^ ൯ ^ ^^ℒ^୬ୡ൫^^^, ^^^ ൯+^^ℒ^ୖ൫^^^,^^^ ௩^൯ ^ ^^ℒ^ୗ൫^^^,^^^^^൯,where ^^ ^^

[0048] FIG. 7 is a flow diagram illustrating a process 700 for audio processing training, in accordance with aspects of the present disclosure. The process 700 may be performed by a computing device (or apparatus) or a component (e.g., a chipset, codec, etc.) of the computing device (e.g., computing system 1300 of FIG. 13, 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 700 may be implemented as software components that are executed and run on one or more processors (e.g., processor 1310 of FIG.13, etc.).

[0049] At block 702, the computing device (or component thereof) may obtain an audio frame (e.g., source speech 512 of FIG.5, source speech 612 of FIG.6, etc.) of the one or more audio frames, the audio frame including speech from a speaker. In some cases, the computing device (or component thereof) may include one or more microphones configured to capture the one or more audio frames.

[0050] At block 704, the computing device (or component thereof) may obtain an enhanced audio frame (e.g., noisy / enhanced speech 530 of FIG.5, enhanced noisy speech 630 of FIG.6, noisy speech 650 of FIG.6, etc.). In some cases, the enhanced audio frame is based on the audio frame. For example, the noisy / enhanced speech may be generated from the source speech. In some examples, the enhanced audio frame is a noisy audio frame. In some cases, the enhanced audio frame is generated by a noise suppression model (e.g., noise suppression model 652 of FIG. 6). In some examples, the computing device (or component thereof) may perform noise suppression on an obtained noisy audio frame Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 18 (e.g., noisy speech 650 of FIG. 6) using the noise suppression model. In some cases, the computing device (or component thereof) may determine a fourth loss value (e.g., noise suppression loss) based on a difference between the audio frame and the enhanced audio frame; and train the noise suppression model based on the fourth loss value.

[0051] At block 706, the computing device (or component thereof) may encode (e.g., by content encoder 506 of FIG.5, content encoder 606 of FIG.6, etc.) a first set of content embeddings. In some cases, the first set of content embeddings are generated based on the audio frame. For example, an SSL encoder may extract content embeddings from input audio data and a content encoder may encode the extracted content embeddings.

[0052] At block 708, the computing device (or component thereof) may encode (e.g., by content encoder 534 of FIG. 5, content encoder 634 of FIG. 6, etc.) a second set of content embeddings. In some cases, the second set of content embeddings are generated based on the enhanced audio frame.

[0053] At block 710, the computing device (or component thereof) may linearly project (e.g., by linear projection engine 508 of FIG. 5, linear projection engine 608 of FIG. 6, etc.) the encoded first set of content embeddings to generate a first predicted spectrogram.

[0054] At block 712, the computing device (or component thereof) may linearly project (e.g., by linear projection engine 536 of FIG. 5, linear projection engine 636 of FIG.6, etc.) the encoded second set of content embeddings to generate a second predicted spectrogram. In some cases, the computing device (or component thereof) may reconstruct target speech (e.g., by diffusion decoder 510 of FIG.5, diffusion decoder 610 of FIG. 6, etc.) using the linearly projected encoded first set of content embeddings; determine a third loss value (e.g., diffusion loss 522 of FIG.5, diffusion loss 622 of FIG. 6, etc.) based on a difference between the reconstructed target speech and the audio frame; and train the voice conversion generative model based on the third loss value.

[0055] At block 714, the computing device (or component thereof) may determine a first loss value (e.g., encoder loss 514 of FIG. 5, encoder loss 614 of FIG.6) based on a Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 19 difference between the first predicted spectrogram and a spectrogram generated based on the audio frame. In some cases, the computing device (or component thereof) may generate the spectrogram based on the audio frame using a mel filterbank (e.g., mel filterbank 516 of FIG.5, mel filterbank 616 of FIG.6, etc.)

[0056] At block 716, the computing device (or component thereof) may determine a second loss value (e.g., noise robustness loss 540 of FIG.5, noise robustness loss 640 of FIG. 6) based on a difference between the second predicted spectrogram and the spectrogram.

[0057] At block 718, the computing device (or component thereof) may train a voice conversion generative model (e.g., VC engine 502 of FIG. 5, VC engine 602 of FIG. 6) based on the first loss value and the second loss value.

[0058] As noted herein, the techniques or processes described herein (e.g., the process 700) 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 type 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. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 20

[0059] 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.

[0060] In some cases, the devices or apparatuses configured to perform the operations of the process 700 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 700 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.

[0061] The components of the device or apparatus configured to carry out one or more operations of the process 700 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 Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 21 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.

[0062] The process 700 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 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.

[0063] Additionally, the processes described herein (e.g., the process 700 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.

[0064] 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 Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 22 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.

[0065] FIG. 8 provides two sets of images 800 that show the forward diffusion process (which is fixed) and the reverse diffusion process (which is learned) of a diffusion model. As shown in the forward diffusion process of FIG.8, noise 803 is gradually added to a first set of images 802 at different time steps for a total of T time steps (e.g., making up a Markov chain), producing a sequence of noisy samples X1 through XT.

[0066] Diffusion models from a training perspective will take an image and will slowly add noise to the image to destroy the information in the image. In some aspects, the noise 803 is Gaussian noise. Each time step can correspond to each consecutive image of the first set of images 802 shown in FIG.8. The initial image X0 of FIG.8 is of a cat. Addition of the noise 803 to each image (corresponding to noisy samples X1 to XT) results in gradual diffusion of the pixels in each image until the final image (corresponding to sample XT) essentially matches the noise distribution. For example, by adding the noise, each data sample X1 through XT gradually loses its distinguishable features as the time step becomes larger, eventually resulting in the final sample XT being equivalent to the target noise distribution, for instance a unit variance zero- Gaussian ^^^0,^^^.

[0067] The second set of images 804 shows the reverse diffusion process in which XT is the starting point with a noisy image (e.g., one that has Gaussian noise). The diffusion model can be trained to reverse the diffusion process (e.g., by training a model pθ(xt-1 | xt)) to generate new data. In some aspects, a diffusion model can be trained by finding the reverse Markov transitions that maximize the likelihood of the training data. By traversing backwards along the chain of time steps, the diffusion model can generate the new data. For example, as shown in FIG.8, the reverse diffusion process proceeds to generate X0 as the image of a vase. In other cases, the input data and output data can vary based on the task for which the diffusion model is trained. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 23

[0068] As noted above, the diffusion model is trained to be able to denoise or recover the original image X0 in an incremental process as shown in the second set of images 804. In some aspects, the neural network of the diffusion model can be trained to recover Xt given Xt-1, such as provided in the below example equation: ^^^^^௧|^^௧ି^^ ൌ ^^൫^^௧; ^1 െ ^^௧^^௧ି^,^^௧^^൯.

[0069] Define ∝^ ൌ ∏௧௧ ^ୀ^ ^1 െ ^^^^ → ^^^^^௧|^^^^ ൌ ^^^^^௧; ^∝^௧ ^^^ , ^1 െ ∝^௧ ^^^^

[0070] ^^௧ ൌ ^∝^௧ ^^^ ^ ^1 െ ∝^௧ ^^ where ε ∼ ^^^^^, ^^^.

[0071] In as a noise schedule) isdesigned such that ∝^ఁ ^ 0 and ^^^^^ఁ|^^^^ ^ ^^^^^ఁ ; ^^, ^^^.

[0072] Theto incrementally generate the input image X0. In one example, the model may have twenty steps. However, in other examples, the number of steps can vary.

[0073] FIG. 9 is a diagram 900 illustrating how diffusion data is distributed from initial data to noise using a diffusion model in the forward diffusion direction, in accordance with some aspects. Note that the initial data q(X0) is detailed in the initial stage of the diffusion process. An illustrative example of the data q(X0) is the initial image of the vase shown in FIG.8. As the diffusion model iterates and iteratively adds sampled noise to the data from t = 0 to t = T, as shown in FIG.9, the data becomes nosier and may ultimately result in pure noise (e.g., at q(XT)). The example of FIG. 9 illustrates the progression of the data and how it becomes diffused with noise in the forward diffusion process.

[0074] In some aspects, the diffused data distribution (e.g., as shown in FIG. 9) can be as follows: Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 24 ^^^^^௧^ ൌ ^ ^^^^^^, ^^௧^ ^^^^^ ൌ ^^^^^^^^ ^^^^^௧|^^^^ ^^^^^.

[0075] In data distribution,^^^^^^, ^^ ^௧data distribution,and ^^^^^௧|^^^^ is the diffusion kernel. In this regard, the model can sample ^^௧ ∼ ^^^^^௧^by first sampling ^^^ ∽ ^^^^^^^ and then sampling ^^௧ ∼ ^^^^^௧|^^^^ (which may bereferred to as ancestral sampling). The diffusion kernel takes the and returns a vector or other data structure as output.

[0076] The following is a summary of a training algorithm and a sampling algorithm for a diffusion model. A training algorithm can include the following steps: 1: repeat 2: ^^^ ∼ ^^^^^^^3: ^^ ∼ Uniform ^^1, ... ,^^ ^^4:∈ ~ ^^^^^, ^^^5: Take gradient descent step on ∇∅ ∥ ∈ െ ∈∅ ^^∝^ ^^ ^ 1 െ ∝^ ଶ௧ ^ ^ ௧∈, ^^^ ∥

[0077] A sampling algorithm can include the following steps: 1: ^^ఁ ∼ ^^^^^, ^^^2: for ^^ ൌ ^^, ... , 1 do3: ^^ ∼ ^^^^^, ^^^4:^ ^^^^ି∝^^ ∈ ^^^^ ^6: return ^^^

[0078] FIG. 10 is a diagram illustrating a U-Net architecture 1000 for a diffusion model, in accordance with some aspects. The initial image 1002 (e.g., of a cat) is provided Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 25 to the U-Net architecture 1000 which includes a series of residual networks (ResNet) blocks and self-attention layers to represent the network ^^ (xt, t). The U-Net architecture 1000 also includes fully connected layers 1008. In some cases, time representation 1010 can be sinusoidal positional embeddings or random Fourier features. Noisy output 1006 from the forward diffusion process is also shown.

[0079] The U-Net architecture 1000 includes a contracting path 1004 and an expansive path 1005 as shown in FIG.10, which gives it the U-shaped architecture. The contracting path 1004 can be a convolutional network that includes repeated convolutional layers (that apply convolutional operations), each followed by a rectified linear unit (ReLU) and a max pooling operation. When images are being processed (e.g., the image 1002) during the contracting path 1004, the spatial information of the image 1002 is reduced as features are generated. The expansive path 1005 combines the features and spatial information through a sequence of up-convolutions and concatenations with high-resolution features from the contracting path 1004. Some of the layers can be self- attention layers, which leverage global interactions between semantic features at the end of the encoder to explicitly model full contextual information.

[0080] FIG. 11 is an illustrative example of a neural network 1100 (e.g., a deep- learning neural network) that can be used to implement machine-learning-based 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.

[0081] An input layer 1102 includes input data. Neural network 1100 includes multiple hidden layers hidden layers 1106a, 1106b, through 1106n. The hidden layers 1106a, 1106b, through hidden layer 1106n 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 many layers as needed for the given application. Neural network 1100 further includes an output layer 1104 that provides an output resulting from the processing performed by the hidden layers 1106a, 1106b, through 1106n. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 26

[0082] Neural network 1100 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 1100 can include a feed-forward network, in which case there are no feedback connections where outputs of the network are fed back into itself. In some cases, neural network 1100 can include a recurrent neural network, which can have loops that allow information to be carried across nodes while reading in input.

[0083] Information can be exchanged between nodes through node-to-node interconnections between the various layers. Nodes of input layer 1102 can activate a set of nodes in the first hidden layer 1106a. For example, as shown, each of the input nodes of input layer 1102 is connected to each of the nodes of the first hidden layer 1106a. The nodes of first hidden layer 1106a 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 1106b, 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 1106b can then activate nodes of the next hidden layer, and so on. The output of the last hidden layer 1106n can activate one or more nodes of the output layer 1104, at which an output is provided. In some cases, while nodes (e.g., node 1108) in neural network 1100 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.

[0084] 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 1100. Once neural network 1100 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 tunable numeric weight that can be tuned (e.g., based on a Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 27 training dataset), allowing neural network 1100 to be adaptive to inputs and able to learn as more and more data is processed.

[0085] Neural network 1100 may be pre-trained to process the features from the data in the input layer 1102 using the different hidden layers 1106a, 1106b, through 1106n in order to provide the output through the output layer 1104. In an example in which neural network 1100 is used to identify features in images, neural network 1100 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 [0010000000].

[0086] In some cases, neural network 1100 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 1100 is trained well enough so that the weights of the layers are accurately tuned.

[0087] For the example of identifying objects in images, the forward pass can include passing a training image through neural network 1100. The weights are initially randomized before neural network 1100 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 intensity 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).

[0088] As noted above, for a first training iteration for neural network 1100, the output will likely include values that do not give preference to any particular class due to Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 28 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 can be 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 1100 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 as . The loss can be set to be equal to the value of Etotal.

[0089] 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 1100 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 as , where w denotes a weight, widenotes the initial weight, and η denotes alearning 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.

[0090] Neural network 1100 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 1100 can include any other deep network Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 29 other than a CNN, such as an autoencoder, a deep belief nets (DBNs), a Recurrent Neural Networks (RNNs), among others.

[0091] FIG. 12 is an illustrative example of a convolutional neural network (CNN) 1200. The input layer 1202 of the CNN 1200 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 1204, an optional non-linear activation layer, a pooling hidden layer 1206, and fully connected layer 1208 (which fully connected layer 1208 can be hidden) to get an output at the output layer 1210. While only one of each hidden layer is shown in FIG.12, one of ordinary skill will appreciate that multiple convolutional hidden layers, non-linear layers, pooling hidden layers, and / or fully connected layers can be included in the CNN 1200. 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.

[0092] The first layer of the CNN 1200 can be the convolutional hidden layer 1204. The convolutional hidden layer 1204 can analyze image data of the input layer 1202. Each node of the convolutional hidden layer 1204 is connected to a region of nodes (pixels) of the input image called a receptive field. The convolutional hidden layer 1204 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 1204. 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 28×28 array, and each filter (and corresponding receptive field) is a 5×5 array, then there will be 24×24 nodes in the convolutional hidden layer 1204. Each connection between a node and a receptive field for that node learns a weight and, in some cases, an overall bias such that each node learns to analyze its particular local receptive field in the input image. Each node of the Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 30 convolutional hidden layer 1204 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.

[0093] The convolutional nature of the convolutional hidden layer 1204 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 1204 can begin in the top-left corner 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 1204. 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 corner 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 1204. For example, a filter can be moved by a step amount (referred to as a stride) 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 1204.

[0094] The mapping from the input layer to the convolutional hidden layer 1204 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 1204 can include several activation maps in Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 31 order to identify multiple features in an image. The example shown in FIG. 12 includes three activation maps. Using three activation maps, the convolutional hidden layer 1204 can detect three different kinds of features, with each feature being detectable across the entire image.

[0095] In some examples, a non-linear hidden layer can be applied after the convolutional hidden layer 1204. 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. A ReLU 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 1200 without affecting the receptive fields of the convolutional hidden layer 1204.

[0096] The pooling hidden layer 1206 can be applied after the convolutional hidden layer 1204 (and after the non-linear hidden layer when used). The pooling hidden layer 1206 is used to simplify the information in the output from the convolutional hidden layer 1204. For example, the pooling hidden layer 1206 can take each activation map output from the convolutional hidden layer 1204 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 1206, 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 1204. In the example shown in FIG. 12, three pooling filters are used for the three activation maps in the convolutional hidden layer 1204.

[0097] 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 1204. The output from a max-pooling 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 2×2 nodes in the previous layer (with each node being a Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 32 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 1204 having a dimension of 24x24 nodes, the output from the pooling hidden layer 1206 will be an array of 12x12 nodes.

[0098] 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 2×2 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.

[0099] 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 1200.

[0100] The final layer of connections in the network is a fully-connected layer that connects every node from the pooling hidden layer 1206 to every one of the output nodes in the output layer 1210. Using the example above, the input layer includes 28 x 28 nodes encoding the pixel intensities of the input image, the convolutional hidden layer 1204 includes 3×24×24 hidden feature nodes based on application of a 5×5 local receptive field (for the filters) to three activation maps, and the pooling hidden layer 1206 includes a layer of 3×12×12 hidden feature nodes based on application of max-pooling filter to 2×2 regions across each of the three feature maps. Extending this example, the output layer 1210 can include ten output nodes. In such an example, every node of the 3x12x12 pooling hidden layer 1206 is connected to every node of the output layer 1210. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 33

[0101] The fully connected layer 1208 can obtain the output of the previous pooling hidden layer 1206 (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 1208 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 1208 and the pooling hidden layer 1206 to obtain probabilities for the different classes. For example, if the CNN 1200 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).

[0102] In some examples, the output from the output layer 1210 can include an M- dimensional vector (in the prior example, M=10). M indicates the number of classes that the CNN 1200 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 [000.050.80 0.150000], 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.

[0103] In some aspects, training of one or more of the machine learning systems or neural networks described herein (e.g., such as the technique for training 400 of FIG. 4, technique for training 500 of FIG.5, technique for training 600 of FIG.6, among various other machine learning systems or neural 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 (e.g., such as the noisy input speech 102 of FIG. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 34 1, etc.) is processed, for instance for performance of the speech restoration 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 pre-trained 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.

[0104] FIG. 13 is a diagram illustrating an example of a system for implementing certain aspects of the present technology. In particular, FIG. 13 illustrates an example of computing system 1300, which can be for example any computing device making up internal computing system, 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 1305. Connection 1305 can be a physical connection using a bus, or a direct connection into processor 1310, such as in a chipset architecture. Connection 1305 can also be a virtual connection, networked connection, or logical connection.

[0105] In some examples, computing system 1300 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.

[0106] Example system 1300 includes at least one processing unit (CPU or processor) 1310 and connection 1305 that couples various system components including system Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 35 memory 1315, such as read-only memory (ROM) 1320 and random access memory (RAM) 1325 to processor 1310. Computing system 1300 can include a cache 1312 of high-speed memory connected directly with, in close proximity to, or integrated as part of processor 1310.

[0107] Processor 1310 can include any general purpose processor and a hardware service or software service, such as services 1332, 1334, and 1336 stored in storage device 1330, configured to control processor 1310 as well as a special-purpose processor where software instructions are incorporated into the actual processor design. Processor 1310 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.

[0108] To enable user interaction, computing system 1300 includes an input device 1345, 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 1300 can also include output device 1335, which can be one or more of a number of output mechanisms. In some instances, multimodal systems can enable a user to provide multiple types of input / output to communicate with computing system 1300. Computing system 1300 can include communications interface 1340, which can generally govern 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 serial 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), Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 36 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 1340 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 1300 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.

[0109] Storage device 1330 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 Memory Stick® 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), Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 37 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.

[0110] The storage device 1330 can include software services, servers, services, etc., that when the code that defines such software is executed by the processor 1310, 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 1310, connection 1305, output device 1335, etc., to carry out the function.

[0111] 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 carrying 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.

[0112] 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. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 38 However, when mentioned, non-transitory computer-readable storage media expressly exclude media such as energy, carrier signals, electromagnetic waves, and signals per se.

[0113] 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.

[0114] 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.

[0115] 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 and data 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 Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 39 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.

[0116] 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 necessary tasks. 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.

[0117] 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.

[0118] 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 above-described application may be used individually or jointly. Further, examples can be utilized in any number of environments and applications Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 40 beyond those described herein without departing from 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.

[0119] 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.

[0120] 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.

[0121] 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 Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 41 the set of A and B. The phrases “at least one” and “one or more” are used interchangeably herein.

[0122] 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 multiple processors (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 perform X, 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.

[0123] 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.

[0124] 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, Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 42 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).

[0125] 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 functionality is 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.

[0126] 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 Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 43 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 memory (SDRAM), read-only memory (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.

[0127] 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 logic arrays (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).

[0128] Illustrative aspects of the present disclosure include:

[0129] Aspect 1. An apparatus for audio processing training, comprising: one or more memories configured to store one or more audio frames; and one or more processors coupled to the one or more memories and configured to: obtain an audio frame of the one Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 44 or more audio frames, the audio frame including speech from a speaker; obtain an enhanced audio frame, wherein the enhanced audio frame is based on the audio frame; encode a first set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encode a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly project the encoded first set of content embeddings to generate a first predicted spectrogram; linearly project the encoded second set of content embeddings to generate a second predicted spectrogram; determine a first loss value based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame; determine a second loss value based on a difference between the second predicted spectrogram and the spectrogram; and train a voice conversion generative model based on the first loss value and the second loss value.

[0130] Aspect 2. The apparatus of Aspect 1, wherein the one or more processors are configured to: reconstruct target speech using the linearly projected encoded first set of content embeddings; determine a third loss value based on a difference between the reconstructed target speech and the audio frame; and train the voice conversion generative model based on the third loss value.

[0131] Aspect 3. The apparatus of any of Aspects 1-2, wherein the enhanced audio frame is a noisy audio frame.

[0132] Aspect 4. The apparatus of any of Aspects 1-2, wherein the enhanced audio frame is generated by a noise suppression model.

[0133] Aspect 5. The apparatus of Aspect 4, wherein the one or more processors are configured to perform noise suppression on an obtained noisy audio frame using the noise suppression model.

[0134] Aspect 6. The apparatus of any of Aspects 4-5, wherein the one or more processors are configured to: determine a fourth loss value based on a difference between the audio frame and the enhanced audio frame; and train the noise suppression model based on the fourth loss value. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 45

[0135] Aspect 7. The apparatus of any of Aspects 1-6, wherein the one or more processors are configured to generate the spectrogram based on the audio frame using a mel filterbank.

[0136] Aspect 8. The apparatus of any of Aspects 1-7, further comprising one or more microphones configured to capture the one or more audio frames.

[0137] Aspect 9. A method for audio processing training, comprising: obtaining an audio frame of a set of one or more audio frames, the audio frame including speech from a speaker; obtaining an enhanced audio frame, wherein the enhanced audio frame is based on the audio frame; encoding a first set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encoding a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly projecting the encoded first set of content embeddings to generate a first predicted spectrogram; linearly projecting the encoded second set of content embeddings to generate a second predicted spectrogram; determining a first loss value based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame; determining a second loss value based on a difference between the second predicted spectrogram and the spectrogram; and training a voice conversion generative model based on the first loss value and the second loss value.

[0138] Aspect 10. The method of Aspect 90, further comprising: reconstructing target speech using the linearly projected encoded first set of content embeddings; determining a third loss value based on a difference between the reconstructed target speech and the audio frame; and training the voice conversion generative model based on the third loss value.

[0139] Aspect 11. The method of any of Aspects 9-10, wherein the enhanced audio frame is a noisy audio frame.

[0140] Aspect 12. The method of any of Aspects 9-10, wherein the enhanced audio frame is generated by a noise suppression model. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 46

[0141] Aspect 13. The method of Aspect 12, further comprising performing noise suppression on an obtained noisy audio frame using the noise suppression model.

[0142] Aspect 14. The method of any of Aspects 12-13, further comprising: determining a fourth loss value based on a difference between the audio frame and the enhanced audio frame; and training the noise suppression model based on the fourth loss value.

[0143] Aspect 15. The method of any of Aspects 9-14, further comprising generating the spectrogram based on the audio frame using a mel filterbank.

[0144] Aspect 16. 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: obtain an audio frame of a set of one or more audio frames, the audio frame including speech from a speaker; obtain an enhanced audio frame, wherein the enhanced audio frame is based on the audio frame; encode a first set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encode a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly project the encoded first set of content embeddings to generate a first predicted spectrogram; linearly project the encoded second set of content embeddings to generate a second predicted spectrogram; determine a first loss value based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame; determine a second loss value based on a difference between the second predicted spectrogram and the spectrogram; and train a voice conversion generative model based on the first loss value and the second loss value.

[0145] Aspect 17. The non-transitory computer-readable medium of Aspect 16, wherein the instructions cause the one or more processors to: reconstruct target speech using the linearly projected encoded first set of content embeddings; determine a third loss value based on a difference between the reconstructed target speech and the audio frame; and train the voice conversion generative model based on the third loss value. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 47

[0146] Aspect 18. The non-transitory computer-readable medium of any of Aspects 16-17, wherein the enhanced audio frame is a noisy audio frame.

[0147] Aspect 19. The non-transitory computer-readable medium of any of Aspects 16-17, wherein the enhanced audio frame is generated by a noise suppression model.

[0148] Aspect 20. The non-transitory computer-readable medium of Aspect 19, wherein the instructions cause the one or more processors to perform noise suppression on an obtained noisy audio frame using the noise suppression model.

[0149] Aspect 21. The non-transitory computer-readable medium of Aspect 20 wherein the instructions cause the one or more processors to perform noise suppression on an obtained noisy audio frame using the noise suppression model.

[0150] Aspect 22. The apparatus of any of Aspects 20-21, wherein the instructions cause the one or more processors to: determine a fourth loss value based on a difference between the audio frame and the enhanced audio frame; and train the noise suppression model based on the fourth loss value.

[0151] Aspect 23. The apparatus of any of Aspects 16-22, wherein the instructions cause the one or more processors generate the spectrogram based on the audio frame using a mel filterbank.

[0152] Aspect 24: An apparatus for audio processing, comprising means for performing one or more of operations according to any of Aspects 9 to 15.

[0153] Aspect 25. An apparatus for audio processing, comprising: one or more memories configured to store one or more audio frames; and one or more processors coupled to the one or more memories and configured to: reconstruct target speech using a voice conversion generative model, wherein the voice conversion generative model is pretrained, and wherein pretraining the voice conversion generative model causes the voice conversion generative model to: obtain an audio frame of the one or more audio frames, the audio frame including speech from a speaker; obtain an enhanced audio frame, wherein the enhanced audio frame is based on the audio frame; encode a first set of Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 48 content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encode a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly project the encoded first set of content embeddings to generate a first predicted spectrogram; linearly project the encoded second set of content embeddings to generate a second predicted spectrogram; determine a first loss value based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame; determine a second loss value based on a difference between the second predicted spectrogram and the spectrogram; and adjust the voice conversion generative model based on the first loss value and the second loss value.

[0154] Aspect 26. The apparatus of Aspect 25, wherein pretraining the voice conversion generative model causes the voice conversion generative model to: reconstruct target speech using the linearly projected encoded first set of content embeddings; determine a third loss value based on a difference between the reconstructed target speech and the audio frame; and adjust the voice conversion generative model based on the third loss value.

[0155] Aspect 27. The apparatus of any of Aspects 25-26, wherein the enhanced audio frame is a noisy audio frame.

[0156] Aspect 28. The apparatus of any of Aspects 25-27, wherein the enhanced audio frame is generated by a noise suppression model.

[0157] Aspect 29. The apparatus of Aspect 28, wherein the pretraining the voice conversion generative model causes the voice conversion generative model to perform noise suppression on an obtained noisy audio frame using the noise suppression model.

[0158] Aspect 30. The apparatus of any of Aspects 28-29, wherein the pretraining the voice conversion generative model causes the voice conversion generative model to: determine a fourth loss value based on a difference between the audio frame and the enhanced audio frame; and adjust the noise suppression model based on the fourth loss value. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 49

[0159] Aspect 31. The apparatus of any of Aspects 25-30, wherein the pretraining the voice conversion generative model causes the voice conversion generative model to generate the spectrogram based on the audio frame using a mel filterbank.

[0160] Aspect 32. The apparatus of any of Aspects 25-31, further comprising one or more microphones configured to capture the one or more audio frames.

[0161] Aspect 33. The apparatus of any of Aspects 25-32, wherein a first loss value is determined based on the difference between the first predicted spectrogram and a spectrogram generated based on the audio frame, wherein a second loss value is determined based on the difference between the second predicted spectrogram and the spectrogram, and wherein the voice conversion generative model is trained based on the first loss value and the second loss value.

[0162] Aspect 34. A method for audio processing, comprising: reconstructing target speech using a voice conversion generative model, wherein the voice conversion generative model is pretrained by: obtaining an audio frame of a set of one or more audio frames, the audio frame including speech from a speaker; obtaining an enhanced audio frame, wherein the enhanced audio frame is based on the audio frame; encoding a first set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encoding a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly projecting the encoded first set of content embeddings to generate a first predicted spectrogram; linearly projecting the encoded second set of content embeddings to generate a second predicted spectrogram; determining a first loss value based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame; determining a second loss value based on a difference between the second predicted spectrogram and the spectrogram; and adjusting the voice conversion generative model based on the first loss value and the second loss value.

[0163] Aspect 35. The method of Aspect 34, wherein pretraining the voice conversion generative model further comprises: reconstructing target speech using the linearly projected encoded first set of content embeddings; determining a third loss value based Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 50 on a difference between the reconstructed target speech and the audio frame; and adjusting the voice conversion generative model based on the third loss value.

[0164] Aspect 36. The method of any of Aspects 34-35, wherein the enhanced audio frame is a noisy audio frame.

[0165] Aspect 37. The method of any of Aspects 34-36, wherein the enhanced audio frame is generated by a noise suppression model.

[0166] Aspect 38. The method of Aspect 37, wherein pretraining the voice conversion generative model further comprises performing noise suppression on an obtained noisy audio frame using the noise suppression model.

[0167] Aspect 39. The method of Aspect 37, wherein pretraining the voice conversion generative model further comprises: determining a fourth loss value based on a difference between the audio frame and the enhanced audio frame; and adjusting the noise suppression model based on the fourth loss value.

[0168] Aspect 40. The method of any of Aspects 34-39, further comprising generating the spectrogram based on the audio frame using a mel filterbank.

[0169] Aspect 41. The method of any of Aspects 34-40, wherein a first loss value is determined based on the difference between the first predicted spectrogram and a spectrogram generated based on the audio frame, wherein a second loss value is determined based on the difference between the second predicted spectrogram and the spectrogram, and wherein the voice conversion generative model is trained based on the first loss value and the second loss value.

[0170] Aspect 42. 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: reconstruct target speech using a voice conversion generative model, wherein the voice conversion generative model is pretrained, and wherein pretraining the voice conversion generative model causes the voice conversion generative model to: obtain an audio frame of a set of one or more audio frames, the audio frame including Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 51 speech from a speaker; obtain an enhanced audio frame, wherein the enhanced audio frame is based on the audio frame; encode a first set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encode a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly project the encoded first set of content embeddings to generate a first predicted spectrogram; linearly project the encoded second set of content embeddings to generate a second predicted spectrogram; determine a first loss value based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame; determine a second loss value based on a difference between the second predicted spectrogram and the spectrogram; and adjust the voice conversion generative model based on the first loss value and the second loss value.

[0171] Aspect 43. The non-transitory computer-readable medium of Aspect 42, wherein the instructions cause the one or more processors to: reconstruct target speech using the linearly projected encoded first set of content embeddings; determine a third loss value based on a difference between the reconstructed target speech and the audio frame; and adjust the voice conversion generative model based on the third loss value.

[0172] Aspect 44. The non-transitory computer-readable medium of any of Aspects 42-43, wherein the enhanced audio frame is a noisy audio frame.

[0173] Aspect 45. The non-transitory computer-readable medium of any of Aspects 42-44, wherein the enhanced audio frame is generated by a noise suppression model.

[0174] Aspect 46. The non-transitory computer-readable medium of Aspect 45, wherein the instructions cause the one or more processors to perform noise suppression on an obtained noisy audio frame using the noise suppression model.

[0175] Aspect 47. The non-transitory computer-readable medium of any of Aspects 42- 46, wherein a first loss value is determined based on the difference between the first predicted spectrogram and a spectrogram generated based on the audio frame, wherein a second loss value is determined based on the difference between the second predicted Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 52 spectrogram and the spectrogram, and wherein the voice conversion generative model is trained based on the first loss value and the second loss value.

[0176] Aspect 48. 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 perform one or more operations according to any of Aspects 42-47.

[0177] Aspect 49. An apparatus for audio processing, comprising means for performing one or more of operations according to any of Aspects 42-47. Polsinelli Ref. No.094922-855580

Claims

PATENT Qualcomm Ref. No.2407249WO 53 CLAIMS What is Claimed Is:

1. An apparatus for audio processing, comprising: one or more memories configured to store one or more audio frames; and one or more processors coupled to the one or more memories and configured to: reconstruct target speech using a voice conversion generative model, wherein the voice conversion generative model is pretrained, and wherein pretraining the voice conversion generative model causes the voice conversion generative model to: obtain an audio frame of the one or more audio frames, the audio frame including speech from a speaker; obtain an enhanced audio frame, wherein the enhanced audio frame is based on the audio frame; encode a first set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encode a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly project the encoded first set of content embeddings to generate a first predicted spectrogram; linearly project the encoded second set of content embeddings to generate a second predicted spectrogram; and adjust the voice conversion generative model based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame and a difference between the second predicted spectrogram and the spectrogram.

2. The apparatus of claim 1, wherein pretraining the voice conversion generative model causes the voice conversion generative model to: reconstruct target speech using the linearly projected encoded first set of content embeddings; Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 54 determine a third loss value based on a difference between the reconstructed target speech and the audio frame; and adjust the voice conversion generative model based on the third loss value.

3. The apparatus of claim 1, wherein the enhanced audio frame is a noisy audio frame.

4. The apparatus of claim 1, wherein the enhanced audio frame is generated by a noise suppression model.

5. The apparatus of claim 4, wherein the pretraining the voice conversion generative model causes the voice conversion generative model to perform noise suppression on an obtained noisy audio frame using the noise suppression model.

6. The apparatus of claim 4, wherein the pretraining the voice conversion generative model causes the voice conversion generative model to: determine a fourth loss value based on a difference between the audio frame and the enhanced audio frame; and adjust the noise suppression model based on the fourth loss value.

7. The apparatus of claim 1, wherein the pretraining the voice conversion generative model causes the voice conversion generative model to generate the spectrogram based on the audio frame using a mel filterbank.

8. The apparatus of claim 1, further comprising one or more microphones configured to capture the one or more audio frames.

9. The apparatus of claim 1, wherein a first loss value is determined based on the difference between the first predicted spectrogram and a spectrogram generated based on the audio frame, wherein a second loss value is determined based on the difference between the second predicted spectrogram and the spectrogram, and wherein the voice Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 55 conversion generative model is trained based on the first loss value and the second loss value.

10. A method for audio processing, comprising: reconstructing target speech using a voice conversion generative model, wherein the voice conversion generative model is pretrained by: obtaining an audio frame of a set of one or more audio frames, the audio frame including speech from a speaker; obtaining an enhanced audio frame, wherein the enhanced audio frame is based on the audio frame; encoding a first set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encoding a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly projecting the encoded first set of content embeddings to generate a first predicted spectrogram; linearly projecting the encoded second set of content embeddings to generate a second predicted spectrogram; and adjusting the voice conversion generative model based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame and a difference between the second predicted spectrogram and the spectrogram.

11. The method of claim 10, wherein pretraining the voice conversion generative model further comprises: reconstructing target speech using the linearly projected encoded first set of content embeddings; determining a third loss value based on a difference between the reconstructed target speech and the audio frame; and adjusting the voice conversion generative model based on the third loss value. Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 56 12. The method of claim 10, wherein the enhanced audio frame is a noisy audio frame.

13. The method of claim 10, wherein the enhanced audio frame is generated by a noise suppression model.

14. The method of claim 13, wherein pretraining the voice conversion generative model further comprises performing noise suppression on an obtained noisy audio frame using the noise suppression model.

15. The method of claim 13, wherein pretraining the voice conversion generative model further comprises: determining a fourth loss value based on a difference between the audio frame and the enhanced audio frame; and adjusting the noise suppression model based on the fourth loss value.

16. The method of claim 10, further comprising generating the spectrogram based on the audio frame using a mel filterbank.

17. The method of claim 10, wherein a first loss value is determined based on the difference between the first predicted spectrogram and a spectrogram generated based on the audio frame, wherein a second loss value is determined based on the difference between the second predicted spectrogram and the spectrogram, and wherein the voice conversion generative model is trained based on the first loss value and the second loss value.

18. 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: reconstruct target speech using a voice conversion generative model, wherein the voice conversion generative model is pretrained, and wherein pretraining the voice conversion generative model causes the voice conversion generative model to: Polsinelli Ref. No.094922-855580PATENT Qualcomm Ref. No.2407249WO 57 obtain an audio frame of a set of one or more audio frames, the audio frame including speech from a speaker; obtain an enhanced audio frame, wherein the enhanced audio frame is based on the audio frame; encode a first set of content embeddings, wherein the first set of content embeddings are generated based on the audio frame; encode a second set of content embeddings, wherein the second set of content embeddings are generated based on the enhanced audio frame; linearly project the encoded first set of content embeddings to generate a first predicted spectrogram; linearly project the encoded second set of content embeddings to generate a second predicted spectrogram; and adjust the voice conversion generative model based on a difference between the first predicted spectrogram and a spectrogram generated based on the audio frame and a difference between the second predicted spectrogram and the spectrogram.

19. The non-transitory computer-readable medium of claim 18, wherein the instructions cause the one or more processors to: reconstruct target speech using the linearly projected encoded first set of content embeddings; determine a third loss value based on a difference between the reconstructed target speech and the audio frame; and adjust the voice conversion generative model based on the third loss value.

20. The non-transitory computer-readable medium of claim 18, wherein the enhanced audio frame is a noisy audio frame. Polsinelli Ref. No.094922-855580

Images