Machine-learned generative audio models for video-synchronized audio generation

Abstract

Video data is processed with a machine-learned generative audiovisual synchrony model to obtain audio data, wherein the video data depicts occurrence of a sound-producing event at a particular time, wherein the audio data comprises an event sound predicted to be produced by the occurrence of the sound-producing event, and wherein occurrence of the event sound within the audio data is synchronized to the particular time at which the sound-producing event is depicted to occur by the video data. An optimization function is evaluated that evaluates at least the event sound. One or more parameters of the machine-learned generative audiovisual synchrony model are adjusted based on the optimization function.

Description

MACHINE-LEARNED GENERATIVE AUDIO MODELS FOR VIDEO- SYNCHRONIZEDAUDIO GENERATIONFIELD

[0001] The present disclosure relates generally to generative audio machine-learned models. More particularly, the present disclosure relates to training and utilizing a generative audio model for generating audio data synchronized to visual model inputs.BACKGROUND

[0002] A computer can receive input(s). The computer can execute instructions to process the input(s) to generate output(s) using a parameterized model. The computer can obtain feedback on its performance in generating the outputs with the model. The computer can generate feedback by evaluating its performance. The computer can receive feedback from an external source. The computer can update parameters of the model based on the feedback to improve its performance. In this manner, the computer can iteratively “learn” to generate the desired outputs. The resulting model is often referred to as a machine-learned model.SUMMARY

[0003] Aspects and advantages of embodiments of the present disclosure will be set forth in part in the following description, or can be learned from the description, or can be learned through practice of the embodiments.

[0004] One example aspect of the present disclosure is directed to a computer-implemented method for training generative audio models to generate audio that is temporally synchronized to sound-producing events depicted by a corresponding video input. The method includes processing, by a computing system comprising one or more computing devices, video data with a machine-learned generative audiovisual synchrony model to obtain audio data, wherein the video data depicts occurrence of a sound-producing event at a particular time, wherein the audio data comprises an event sound predicted to be produced by the occurrence of the sound-producing event, and wherein occurrence of the event sound within the audio data is synchronized to the particular time at which the sound-producing event is depicted to occur by the video data. The method includes evaluating, by the computing system, an optimization function that evaluates atleast the event sound. The method includes adjusting, by the computing system, one or more parameters of the machine-learned generative audiovisual synchrony model based on the optimization function.

[0005] Another example aspect of the present disclosure is directed to a computing system that includes one or more processor devices and one or more non-transitory computer-readable media that store instructions that, when executed by the one or more processors, cause the computing system to perform operations. The operations include obtaining video data and one or more conditioning inputs, wherein the video data depicts occurrence of a sound-producing event at a particular time, wherein a first conditioning input of the one or more conditioning inputs controls inclusion of an event sound type of one or more sound types, wherein the event sound type comprises sound produced by the occurrence of the sound-producing event. The operations include processing the video data and the one or more conditioning inputs with a machine- learned generative audiovisual synchrony model to obtain audio data, wherein the machine- learned generative audiovisual synchrony model is trained to generate audio that is temporally synchronized to events depicted by video inputs, wherein the audio data comprises one or more sounds comprising a first sound of the event sound type of the one or more sound types, and wherein occurrence of the first sound within the audio data is synchronized to the particular time at which the sound-producing event is depicted to occur by the video data.

[0006] Another example aspect of the present disclosure is directed to one or more non- transitory computer-readable media that store instructions that, when executed by one or more processors, cause the one or more processors to perform operations. The operations include processing video data with a machine-learned generative audiovisual synchrony model to obtain audio data, wherein the video data depicts occurrence of a sound-producing event at a particular time, wherein the audio data comprises an event sound predicted to be produced by the occurrence of the sound-producing event, and wherein occurrence of the event sound within the audio data is synchronized to the particular time at which the sound-producing event is depicted to occur by the video data. The operations include evaluating an optimization function that evaluates at least the event sound. The operations include adjusting one or more parameters of the machine-learned generative audiovisual synchrony model based on the optimization function.

[0007] Other aspects of the present disclosure are directed to various systems, apparatuses, non-transitory computer-readable media, user interfaces, and electronic devices.

[0008] These and other features, aspects, and advantages of various embodiments of the present disclosure will become better understood with reference to the following description and appended claims. The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate example embodiments of the present disclosure and, together with the description, serve to explain the related principles.BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Detailed discussion of embodiments directed to one of ordinary skill in the art is set forth in the specification, which makes reference to the appended figures, in which:

[0010] Figure 1 is an overview block diagram of a machine-learned generative audiovisual synchrony model for generating audio temporally synchronized to video inputs according to some implementations of the present disclosure.

[0011] Figure 2 is a detailed block diagram of a machine-learned generative audiovisual synchrony model for generating audio temporally synchronized to video inputs according to some implementations of the present disclosure.

[0012] Figure 3 is a data flow diagram of a method for training a machine-learned generative audiovisual synchrony model for generating audio temporally synchronized to video inputs according to some implementations of the present disclosure.

[0013] Figure 4 depicts a flow chart diagram of an example method to perform training of generative audio models to generate audio that is temporally synchronized to sound-producing events depicted by a corresponding video input according to some implementations of the present disclosure.

[0014] Figure 5 depicts a flow chart diagram of an example method to perform inference with generative audio models trained to generate audio that is temporally synchronized to soundproducing events depicted by a corresponding video input according to some implementations of the present disclosure.

[0015] Figure 6A depicts a block diagram of an example computing system that performs generative audiovisual tasks according to example embodiments of the present disclosure.

[0016] Figure 6B depicts a block diagram of an example computing device that performs according to example embodiments of the present disclosure.

[0017] Figure 6C depicts a block diagram of an example computing device that performs according to example embodiments of the present disclosure.

[0018] Reference numerals that are repeated across plural figures are intended to identify the same features in various implementations.DETAILED DESCRIPTIONOverview

[0019] Generally, the present disclosure is directed to training and utilizing generative audio models for generating audio data synchronized to visual model inputs. More specifically, machine-learned models have recently been utilized to perform generative tasks, such as video data generation, audio data generation, etc. For example, a conventional model can process a text input describing a scene, and in response, generate video data that depicts the described scene. For generative audiovisual tasks, many conventional approaches have attempted to train a single model to jointly generate both audio and video based on an input. To follow the previous example, the model may also generate audio data that includes music to accompany the scene.

[0020] However, such approaches have highlighted the very asymmetric data densities between the two modalities, which ultimately hinder realizing the capabilities of state-of-the-art audio generation (e.g. short duration, low framerate). To follow the previous example, a 10- second video output can include a quantity of information that is orders of magnitude larger than the corresponding 10-second audio output. The substantial difference in data density between generative video and audio tasks exacerbates the difficulty of model training, and often leads to sub-optimal model performance for generative audio tasks.

[0021] Moreover, at inference, generative video models often require substantial quantities of computational resources, and the joint inclusion of a generative audio model can increase the quantity of necessary computing resources past the threshold of computing resources typically available. In other words, the computing resource requirements of joint audio-visual generative models can be greater than the computing resources typically available to users. As such, usage of joint audio-visual generative models is often restricted to those with access to datacenters or the like. However, previous approaches to sequentially generating audio after video generation is complete have failed due to a lack of synchronization between the audio data and video data.Thus, an approach capable of generating high-quality audio data synchronized to a video input is desired.

[0022] Accordingly, implementations described herein propose a sequential generative approach for leveraging generative audio models to generate audio data synchronized to visual model inputs. More specifically, a computing system can obtain video data. For example, the computing system may generate the video data by processing a textual prompt with a machine- learned generative video model. The video data can depict occurrence of a sound-producing event at a particular time (or period of time). For example, if the video data includes ten seconds of video, the video data may depict a drinking glass shattering on the ground between seconds two and three. The computing system can process the video data with a machine-learned generative audiovisual synchrony model to obtain audio data. The audio data can include an “event sound,” which, as described herein, can refer to a sound predicted to be produced by the occurrence of the sound-producing event.

[0023] To follow the previous example, if the sound-producing event depicted by the video data is a drinking glass shattering on the floor, the event sound can refer to the sound produced by the glass shattering on the floor. The occurrence of the event sound within the audio data can be synchronized to the particular time at which the sound-producing event is depicted to occur by the video data. For example, the audio data can be synchronized to the video data such that the event sound occurs at the same time that the sound-producing event is depicted occurring by the video data.

[0024] The computing system can evaluate an optimization function that evaluates at least the event sound. For example, the video data processed by the model can be obtained alongside matching “ground-truth” audio. The optimization function can evaluate differences between the ground-truth audio and the event audio of the audio data. Based on the optimization function, the computing system can adjust parameter(s) of the machine-learned generative audiovisual synchrony model based on the optimization function. In such fashion, implementations described herein can train a machine-learned model for generating audio data that is temporally synchronized to input video data while mitigating, or avoiding, inefficiencies associated with performance of joint audiovisual generative tasks.

[0025] Aspects of the present disclosure provide a number of technical effects and benefits. As one example technical effect and benefit, implementations described herein can be utilized togenerate high-quality temporally synchronous audio data while avoiding inefficiencies associated with performance of joint audiovisual generative tasks. For example, performing joint audiovisual generative tasks generally requires large quantities of computing resources (e.g., compute cycles, memory, storage, Input / Output (I / O) bandwidth, etc.) typically found only in datacenters or a very small proportion of user devices. Furthermore, previous attempts to separate audio generation from video generation have suffered from a lack of temporal synchronization (e.g., synchronizing sound-producing events with sounds). However, implementations described herein can generate audio data based on a video input that is temporally synchronized to the video input, thus reducing the computing resource requirements necessary for performing generative audiovisual tasks.

[0026] As another example technical effect and benefit, implementations described herein enable synthetic data generation for training machine-learned models used in robotics, autonomous vehicles, and similar industries. For example, many autonomous vehicle systems often listen for audio signals to associate audio signals with captured video to identify, or confirm, the occurrence of events (e.g., occurrence of a collision sound confirming the occurrence of an automobile collision, etc.). However, such models require training examples with both audio data and video data that is temporally synchronous, which is prohibitively difficult to obtain. Due to this lack of training data, conventional models exhibit sub-optimal performance at temporally associating audio signals to the occurrence of events depicted in video data. As such, by enabling the generation of synthetic training data that is temporally synchronous, implementations described herein can facilitate substantial improvements in robotics, autonomous vehicles, and other technical areas that leverage temporal synchronicity as a signal for models.

[0027] With reference now to the Figures, example embodiments of the present disclosure will be discussed in further detail.Example Model Arrangements

[0028] Figure 1 is an overview block diagram of a machine-learned generative audiovisual synchrony model for generating audio temporally synchronized to video inputs according to some implementations of the present disclosure. More specifically, a computing system 100 can include a machine-learned generative audiovisual synchrony model 102. The machine-learnedgenerative audiovisual synchrony model 102 can be, or include, any type or manner of machine- learned model, such as an encoder model, decoder model, attention-based model, neural network(s), etc. The machine-learned generative audiovisual synchrony model 102 can be trained to generate audio data based on a video input that is temporally synchronized to the video input.

[0029] More specifically, the computing system 100 can obtain video data 104. The video data 104 can include a plurality of image frames. The video data 104 can be a particular duration (e.g., ten seconds, etc.). A portion 106 of the video data 104 can depict occurrence of a soundproducing event. As described herein, a “sound-producing event” refers to any type or manner of event that produces sound, such as an object contacting a surface (e.g., hitting a baseball with a bat, a glass shattering, etc.), a sound-producing device being activated (e.g., a doorbell ringing, a fax machine receiving data, etc ), a person moving (e.g., walking, snapping their fingers, playing an instrument, humming, etc.), etc.

[0030] It should be noted that a sound-producing event may or may not be explicitly depicted by the portion 106 of the video data 104. For example, the portion 106 of the video data 104 may explicitly depict a glass shattering on the ground by literally depicting the glass falling, contacting the floor, and then shattering. Alternatively, the portion 106 of the video data 104 may implicitly depict the glass shattering by depicting a human’s reaction to the glass shattering (e.g., a look of surprise followed by a scene depicting the shattered glass). For another example, the portion 106 of the video data 104 may explicitly depict a doorbell ringing by depicting a human activating a doorbell with their fingers. Alternatively, the portion 106 of the video data 104 may implicitly depict the doorbell ringing by depicting a human approaching a doorbell and then depicting a human inside a house with a facial reaction typically associated with hearing a doorbell ring (without explicitly depicting the doorbell being activated).

[0031] The machine-learned generative audiovisual synchrony model 102 can process the video data 104 to obtain audio data 108. The audio data 108 can be any type or manner of audio information, and can be encoded using any encoding scheme or format (e.g., mp3, WAV, etc.). In some implementations, the audio data 108 can be an implicit or intermediate representation of audio data that can be further processed to generate audio data that is compatible with audio output devices.

[0032] Tt should be noted that the video data 104 can be any type or manner of video data, or can be an intermediate representation that is derived from the video data 104. For example, the video data 104 can encode a plurality of image frames. Each image frame can include a plurality of pixel values (e.g., color values, radiance or light values, etc.). In some implementations, the machine-learned generative audiovisual synchrony model 102 can process the video data 104 by processing individual video frames, or sets of video frames, to obtain the audio data 108.

[0033] Further, it should be noted that audio data 108 can be, or otherwise include, any type or manner of audio data, such as a waveform signal, encoded audio data, etc. For example, assume that the audio data 102 can be raw audio data (e.g., a non-compressed file that losslessly records the audio), compressed audio data (e.g., audio data that has been encoded using an encoding schema, such as an mp3 encoding scheme), etc. In some implementations, the decoder portion 218 can produce an intermediate output from which the audio data can be derived.

[0034] A portion 110 of the audio data 108 can include an event sound. The event sound can be a sound produced by the sound-producing event depicted by the portion 106 of the video data 104. In addition, occurrence of the event sound within the portion 110 of the audio data 108 can be synchronized to occurrence of the sound-producing event depicted by the video data 104. To follow the depicted example, the audio data 108 can be of the same length as the video data 104, and the sound-producing event depicted by the portion 106 of the video data 104 can occur at the same time (or during the same time period) as the event sound within the portion 110 of the audio data 108 (e.g., between 0 and 1 seconds). In such fashion, implementations described herein can generate audio data that is temporally synchronized to video data inputs.

[0035] Figure 2 is a detailed block diagram of a machine-learned generative audiovisual synchrony model for generating audio temporally synchronized to video inputs according to some implementations of the present disclosure. More specifically, a computing system 200 can include a machine-learned generative audiovisual synchrony model 202 as described with regards to the computing system 100 and machine-learned generative audiovisual synchrony model 102 of Figure 1. The computing system 200 can obtain video data 204 as described with regards to the video data 104 of Figure 1.

[0036] The computing system 200 can obtain conditioning input(s) 206. Each of the conditioning input(s) 206 can control whether or not a particular type of sound is included in the audio output of the machine-learned generative audiovisual synchrony model 202. Additionally,or alternatively, in some implementations, the conditioning input(s) 206 can control a degree of inclusion of a particular sound type. Examples of sound types controlled by the conditioning input(s) 206 can include a silence sound type (e.g., portions of the audio data in which no audible sound is included), a speech sound type (e.g., sound produced by a spoken utterance), a music sound type (e.g., a sound track to accompany a video clip, etc.), an ambient sound type (e g., ambient “background” noises for a particular environment), etc.

[0037] For example, assume that the conditioning input(s) 206 include a conditioning input for an ambient sound type. If the conditioning input 206 for the ambient sound type has a value of zero, the audio data produced by the machine-learned generative audiovisual synchrony model 202 can exclude any manner of ambient noise. Conversely, if the conditioning input(s) 206 for the ambient sound type has a value of one, the audio data produced by the machine-learned generative audiovisual synchrony model 202 can include any manner of ambient noise. If the conditioning input(s) 206 for the ambient sound type have a scalar value between 0 and 1, the quantity, intensity, etc. of ambient sound included in the audio data produced by the machine- learned generative audiovisual synchrony model 202 can vary based on the magnitude of the scalar value (e.g., a value of 0.3 would cause inclusion of quieter and / or less frequent ambient sounds while a value of 0.9 would cause inclusion of louder and / or more frequent ambient sounds. Additionally, in some implementations, the conditioning input(s) 206 can include an audiovisual temporal synchrony type. Similar to the example described above, the audiovisual temporal synchrony type can control a degree of temporal synchrony exhibited by the audio data output by the (s) 206 for the ambient sound type has a value of one, the audio data produced by the machine-learned generative audiovisual synchrony model 202. The audiovisual temporal synchrony type will be discussed in greater detail subsequently within the present disclosure.

[0038] In some implementations, the conditioning input(s) 206 can be processed alongside the video data 204. Alternatively, in some implementations, the conditioning input(s) 206 can be inserted or appended to an intermediate representation, such as a sequence of video encodings 208 produced by an encoder portion 210 of the machine-learned generative audiovisual synchrony model 202, which will be discussed subsequently.

[0039] The machine-learned generative audiovisual synchrony model 202 can include an encoder portion 210. The encoder portion 210 can process the video data 204 to obtain a sequence of video encodings 208 (e g., token encodings, etc.). In some implementations, theencoder portion 210 of the machine-learned generative audiovisual synchrony model 202 can include a pre-trained vision encoder network 212. The pre-trained vision encoder network 212 can be any type or manner of vision encoder with sufficiently high degree of spatial and temporal resolution (e.g., to facilitate object and semantic awareness). For example, the pre-trained vision encoder network 212 can be an attention-based model such as a vision transformer model.

[0040] The pre-trained vision encoder network 212 can process the video data 204 to obtain a plurality of patch encodings 214. For example, the computing system 200 can sample sets of image frames from the video data 204 (e.g., a 12 frames per second sampling rate, etc.). Assume that the images are resized to 224 x 224 pixel images prior to processing with the pre-trained vision encoder network 212. The pre-trained vision encoder network 212 can process the resized images to generate the patch encodings 214, which can be 768-dimensional encodings for corresponding patches of the 224 x 224 images (e.g., 16 x 16 patches, 8 x 8 patches, etc.). As such, the patch encodings 214 can be, or otherwise include, 768-dimensional encodings derived from patches of image frames extracted from the video data 204.

[0041] The encoder portion 210 can include a linear time encoder 216. The linear time encoder 216 can be any type or manner of encoder that processes the patch encodings 214 to generate the sequence of video encodings 208 with a linear time complexity or similar, although, in some implementations, the linear time encoder 216 can be an encoder with a time complexity that is greater than a linear time complexity. The linear time encoder 216 can include efficient video encoding layers, such as a sliding window attention layer, conv3D layer, etc ). In some implementations, the linear time encoder 216 can be or otherwise include a tokenizer. For example, the linear time encoder 216 can process the patch encodings 214 to obtain the sequence of video encodings 208, and the sequence of video encodings 208 can be a sequence of token embeddings.

[0042] The machine-learned generative audiovisual synchrony model 202 can include a decoder portion 218. The decoder portion 218 can process the sequence of video encodings 208 to obtain audio data 220. For example, the decoder portion 218 can be a transformer-based decoder trained to process token embeddings of video data to generate audio data. In some implementations, the decoder portion 218 can include a tokenizer 222. The tokenizer 222 can process the sequence of video encodings 208 to obtain a corresponding sequence of audio encodings 221.

[0043] In some implementations, each audio encoding of the sequence of audio encodings 221 can be associated with a corresponding video encoding of the sequence of video encodings. As described previously, a video encoding can generally possess a higher degree of data density than an audio encoding. As such, to reduce the encoding rate (i.e., tokenization rate), the decoder portion 218 of the machine-learned generative audiovisual synchrony model 202 can limit the sequence length of the sequence of audio encodings (i.e., the number of audio tokens) and encode as much high-level unimodal information into the sequence of audio encodings 221 as possible. For example, the tokenizer 222 of the decoder portion 218 can, in some implementations, utilize a low-token-rate / high-bit-rate factorized tokenization scheme to produce the sequence of audio encodings 221.

[0044] In some implementations, the tokenizer 222 can include a temporal submodel 224 and a depth submodel 226. The temporal submodel 224 can, in conjunction with the depth submodel 226, process the sequence of video encodings 208 of shape Tvx D (i.e., time by depth) conditioned via cross-attention to generate a corresponding sequence of audio encodings 221 of shape T x D. Specifically, in some implementations, rather than the sequence of audio encodings 221 being a flat sequence of length T o D, the decoder portion 218 can utilize two autoregressive decoding stages: a temporal stage with sequence length T and a depth stage with sequence length D. In some implementations, since predicting tokens across depth can utilize fewer computing resources than predicting tokens across time, the depth submodel 226 can be smaller than the temporal submodel 224. In particular, as the depth model 226 is smaller, and operates over the D depth tokens for a single time step, calculating self-attention across the depth model is much less computationally expensive.

[0045] The depth submodel 226 can sample tokens depthwise and pass encodings (i.e., tokens) between time steps in conjunction with the temporal submodel 224. Any conventional technique can be utilized to pass the depth sampled encodings between time steps, such as simply taking the mean across depth of the token encodings, concatenating the token encodings and projecting them to the smaller model dimension, etc.

[0046] In some implementations, the temporal submodel 224 can output an encoding of each timestep, which is dependent on the sequence of video encodings 208 via the crossattention layer of the decoder portion 218 (to be discussed in greater detail with regards to Figure 3). The depth submodel 226 can generate the sequence of audio tokens 220 by sampling thesequence of audio tokens 220 conditioning on the output encoding from the temporal submodel 224.

[0047] In some implementations, the decoder portion 218 can include a detokenizer 228 and / or a vocoder 230. The detokenizer 228 can process the sequence of audio encodings 221 to obtain a detokenized (and / or disentangled) intermediate representation 229. The vocoder 230 can process the intermediate representation 229 to obtain the audio data 220. In conjunction, the detokenizer 228 and vocoder 230 can be leveraged to apply neural Product Quantization (PQ) to process a sequence of perceptually invertible, self-supervised audio encodings (e.g., the sequence of audio encodings 221) into a sequence of token D-tuples (e.g., the intermediate representation 229) that are each drawn from D vector quantizer codebooks. Specifically, in some implementations, a transformer model of the decoder portion 218 can generate the token D- tuples, and the detokenizer (e.g., the detokenizer 228 of Figure 2) can convert the tokens back to intermediate representations that can be converted to audio with the vocoder (e.g., the vocoder 230 of Figure 2). In general, the detokenizer can be applied to tokens to receive a continuous output. It should be noted that the audio data 220 can be, or otherwise include, any type or manner of audio data, such as a waveform signal, encoded audio data, etc. In some implementations, the decoder portion 218 can produce an intermediate output from which the audio data can be derived.

[0048] By utilizing neural PQ and a product codebook structure, the decoder portion 218 can enable effective vocabulary growth necessary for high quality semantic encoding reconstruction while only introducing a linear time complexity increase in embedding lookup table and prediction head sizes. For PQ with C codebooks of V entries each, the effective vocabulary size can be VC, while the embedding lookup table and prediction head would only require V • C entries each. For example, the tokenizer 222 may utilize values of C = 16 and V = 1024, thus requiring a prediction head output of 16384 but an effective vocabulary size of 1048. As described previously, the decoder complexity impact of C product codebooks can be mitigated via use of the temporal submodel 224 and the depth submodel 226.

[0049] The vocoder 230 can be any type or manner of model that can generate the audio data 220 (e.g., waveforms, etc.), such as a neural network, transformer network, generative adversarial network, etc., based on the intermediate representation(s) 229. For example, the detokenizer 228 can produce the intermediate representation(s) 229 by processing the sequenceof audio encodings 221 (i.e., tokens) to generate reconstructed transformer encodings. The vocoder 230 can process the reconstructed transformer encodings to obtain the audio data 220.

[0050] In some implementations, the tokenizer 222 can be used to map audio to tokens during training. For example, given training data, the tokenizer 222 can map audio to tokens so that the temporal submodel 224 and the depth submodel 226 can be trained to predict the next token from the sequence of video encodings 208 and the previous tokens. The submodels 224 and 226 can output the sequence of audio encodings 220 (i.e., tokens, etc.), the detokenizer 228 can map those audio tokens to a continuous audio recording, and the vocoder 230 can map those encodings to the generated audio data 220.

[0051] In some implementations, the encoder portion 210 and the decoder portion 218 of the machine-learned generative audiovisual synchrony model 202 can include cross-attention layers that calculate attention based on inputs to the paired cross-attention layers at the encoder portion 210 and decoder portion 218. For example, the encoder portion 210 can include a cross-attention encoder layer that calculates attentional keys and values from an input to the cross-attention encoder layer, and the decoder portion 218 can include a cross-attention decoder layer that calculates attentional queries from an input to the cross-attention decoder layer. Specifically, in some implementations, the decoder portion 218 can be limited to a single cross-attention layer to substantially reduce time complexity associated with processing operations of the decoder portion 218. The cross-attention layer of the decoder portion 218 will be discussed in greater detail with regards to Figure 3.

[0052] It should be noted that, although only a single sound-producing event and corresponding event sound are illustrated in the examples herein, any number of soundproducing events and corresponding event sounds can be included in the video data and audio data described herein, respectively. In some implementations, if video data depicts three soundproducing events, the audio data can include three corresponding event sounds. Additionally, or alternatively, in some implementations, if the video data depicts a single sound-producing event, the audio data may include multiple event sounds for different sounds produced during the sound-producing event. For example, multiple event sounds can be generated for a soundproducing event in which a hammer falls through a sheet of glass and lands on the floor (e.g., a sound of glass breaking, a sound of the hammer landing on the floor, etc.).

[0053] Figure 3 is a data flow diagram of a method for training a machine-learned generative audiovisual synchrony model for generating audio temporally synchronized to video inputs according to some implementations of the present disclosure. More specifically, a computing system 300 (e.g., the computing system 100 of Figure 1, the computing system 200 of Figure 2, etc.) can include a machine-learned generative audiovisual synchrony model 302 (e g., the machine-learned generative audiovisual synchrony model 202 of Figure 2, the machine- learned generative audiovisual synchrony model 102 of Figure 1, etc.).

[0054] The computing system 300 can include a model trainer 304 for training the machine- learned generative audiovisual synchrony model 302. To do so, the model trainer 304 can obtain video data 306. The video data 306 can be one item of a supervised training example that also includes a ground-truth event sound 308. The video data 306 can depict the occurrence of a sound-producing event 310. The ground-truth event sound 308 can be or otherwise include audio data that includes an event sound produced by the sound-producing event 310 depicted by the video data 306. In some implementations, the video data 306 can be a length of video extracted from an audiovisual media repository (e.g., an audiovisual hosting site, a social media site, a program or file that includes audiovisual information, etc.), and the ground-truth event sound 308 can be audio that was originally accompanied the video data 306 (e.g., recorded at the same time as the video data, added during production of the video data 306, etc.).

[0055] For example, assume that the video data 306 depicts a drinking glass falling and shattering on the floor. The ground-truth event sound 308 may include audio that was originally produced by the glass shattering when the video data 306 was captured. For another example, the ground-truth event sound 308 may include audio that was manually added (e.g., during production or post -production stages, etc.) to match the video data 306. As such, the ground-truth event sound 308 can serve as an “ideal” or “optimal” example of temporal synchronization between a sound-producing event depicted by video data and an event sound that is produced by the occurrence of the sound-producing event.

[0056] The machine-learned generative audiovisual synchrony model 302 can include an encoder portion 312 and a decoder portion 314 as described with regards to the encoder portion 210 and the decoder portion 218 of Figure 2. The encoder portion 312 can generate an intermediate representation 316 of the video data 306 (e.g., a sequence of video encodings, etc.), and the decoder portion 314 can process the intermediate representation 316 of the video data306 to obtain audio data 318. The audio data 318 can include an event sound 320 predicted to be caused by the sound-producing event 310 depicted by the video data 306.

[0057] In some implementations, the decoder portion 218 of the machine-learned generative audiovisual synchrony model 202 can include attention layers 322A - 322N (generally, attention layers 322). Similarly, the encoder portion 312 can include self-attention layer(s) 324 and crossattention layer(s) 326. It should be noted that, although only one self-attention layer 322A is illustrated as preceding the cross-attention layer 322B, any number of self-attention layer(s) can precede and succeed the cross-attention layer 322B. As described herein, an attention layer refers to a layer of the machine-learned generative audiovisual synchrony model 302 including an attention mechanism that assigns a particular weight or “importance” (i.e., attention) to specific portions of an input. Attention mechanisms can include self-attention layers and cross-attention layers. A self-attention layer refers to a layer where attention is calculated across certain portions of the input to the self-attention layer. A cross-attention layer refers to a layer that is paired to other cross-attention layer(s), and attention is calculated using the inputs to the paired crossattention layer(s).

[0058] Many conventional attention mechanisms calculate keys, values, and queries based on an input to the layer. A self-attention layer can determine keys, values, and queries based on inputs to the layer. Conversely, given a pair of cross-attention layers, the keys and values can be calculated based on the input to one cross-attention layer while the queries are calculated based on the input to another layer. For example, the attention layers 322 of the decoder portion 314 can include a cross-attention layer 322B. The computing system 300 can perform cross-attention between the cross-attention layer 322B of the decoder portion and the cross-attention layer(s) 326 of the encoder portion 312. For example, the cross-attention layer(s) 326 of the encoder portion 312 can calculate keys and values while the cross-attention layer 322B of the decoder portion 314 can calculate queries.

[0059] It should be noted that, conventionally, encoder and decoder portions of generative models have typically utilized multiple cross-attention layers. However, as described with regards to Figure 2, the intermediate representation 316 (e.g., the sequence length of video encodings) has a long sequence length that substantially exacerbates the computational complexity of performing cross-attention. As such, in some implementations, the cross-attention layer 322B can be the only cross-attention layer included in the attention layers 322 of thedecoder portion 314. Since the sequence length Tv of the video encodings (e.g. 14 height x 14 width x 12 fps x 10 seconds = 23520 for 10s of video frames at 12 FPS) is substantially greater than the sequence length T of the audio encodings, cross-attention can be limited to the crossattention layer 322B that sits at the “middle” of the decoder portion 314. In other words, in some implementations, the cross-attention layer 322B can be the only cross-attention layer included in the decoder portion 314 and as such can be immediately preceded and succeeded by selfattention layers (e.g., layers 322A and 322C). By limiting performance of cross-attention to a single cross-attention layer 322B of the decoder portion 314, the computational complexity can be limited to O(T-Tv), as the other self-attention layers 314 in the decoder portion 314 utilize self-attention mechanisms which only use O(T2) self-attention over the audio sequence length T.

[0060] Additionally, it should be noted that the computational complexity of the encoder portion 312 of the machine-learned generative audiovisual synchrony model 302 can be further reduced by minimizing the number of the self-attention layer(s) 324 included in the encoder portion 312. For example, to avoid O(Ty) self-attention cost of a transformer-based encoder, the pretrained semantic video encodings (e.g., the pre-trained vision encoder network 212 and the patch encodings 214) described with regards to Figure 2 can be utilized either directly by the cross-attention layer after processing with the linear time encoder 216 of Figure 2.

[0061] The model trainer 304 can include an optimization function 327. The optimization function 327 can evaluate at least the event sound 320. For example, the optimization function 327 can evaluate difference(s) between the event sound 320 and the ground-truth event sound 308. The model trainer 304 can train the machine-learned generative audiovisual synchrony model 302 based on the optimization function. For example, the model trainer 304 can adjust value(s) of parameter(s) of the machine-learned generative audiovisual synchrony model 302 based on the optimization function 327.

[0062] In some implementations, the model trainer 304 can obtain conditioning input(s) 328. As described previously, the video data 306 of the training example can be sampled from an audiovisual data repository, such as an online audiovisual hosting site. However, such samples often include substantial quantities of speech sound, music sound, and silent sound (e.g., a lack of any sound). The conditioning input(s) 328 can be utilized to prevent the machine-learned generative audiovisual synchrony model 302 from outputting speech and / or silence, and to enable control of whether music is generated by the model 302 in addition to the event sound320. In other words, the conditioning input(s) 328 can be utilized by the model trainer 304 to train the machine-learned generative audiovisual synchrony model 302 to generate additional sound(s) 330 in addition to the audio data 318. Additionally, the model trainer 304 can train the machine-learned generative audiovisual synchrony model 302 so that generation of the additional sound(s) 330 is controllable based on the inclusion (or lack thereof) and / or value of the conditioning inputs 328 during inference.

[0063] To do so, the model trainer 304 can first obtain the conditioning input(s) 328. The conditioning input(s) 328 can be any type or manner of input, such as an embedding, encoding, token, parameter, value, etc. Although not illustrated, in some implementations, the conditioning input(s) 328 can be hyperparameters of the model 302, or some other manner of configurable parameter, rather than an input. In some implementations, the model trainer 304 can obtain the conditioning input(s) 328 by extracting the conditioning input(s) 328 with an audio classifier model or the like. For example, the model trainer 304 may process training data with an audio classifier model to obtain classification information for a large quantity of sound types, and then pool the sound types to obtain a speech sound type, a silence sound type, and a music sound type. However, it should be noted that additional, or alternative, classes can be identified by the conditioning input(s) 328, such as an ambient sound type (e.g., ambient background noise that is agnostic to a depicted environment, such as humans conversing in the background), an environmental sound type (e.g., environmental background noise that is specific to a depicted environment, such as the noise of a coffee machine in a coffee shop environment, insect noises in an outdoor environment, etc.)

[0064] In some implementations, the conditioning input(s) 328 can be or otherwise include a binary event roll that represents the three classes. For example, each of the sound types can be represented by a matrix of binary values that is “stretched” across the duration of the video data 306. Since these binary values are constant for the duration of the video data, they can be repeated to the same shape as the sequence of video encodings (e.g., Tv x 3), and can be concatenated onto the shape of the video encoding sequence of shape (e.g., Tv x 768, etc.). During training, the model trainer 304 can train the machine-learned generative audiovisual synchrony model 302 to use the conditioning input(s) 328 to control the sound types of the additional sound(s) 330 present in the audio data 318 during inference. For example, at inference, a binary event roll of all -zeros (speech audio type = 0, music audio type = 0, silenceaudio type = 0, etc.) generally prevents the machine-learned generative audiovisual synchrony model 302 from predicting any speech or music in the audio data 318.

[0065] For example, assume that the model trainer 304 completes the training process for the machine-learned generative audiovisual synchrony model 302. The computing system 300 can obtain additional video data and conditioning inputs for the machine-learned generative audiovisual synchrony model 302 to process at inference. The machine-learned generative audiovisual synchrony model 302 can process the additional video data to obtain additional audio data. The audio data can include event sound(s), and other additional sound(s) based on the additional conditioning input(s).

[0066] As described previously, the video data 306 can be randomly sampled from audiovisual data elements from an audiovisual data repository. In some instances, such audiovisual data elements (e g., movies, television shows, etc.) will clearly depict the occurrence of sound-producing events and corresponding audio data that is synchronized with the depiction of the sound-producing event. However, in many instances, sound-producing events occur offscreen and are not depicted, even if an event sound is still included. For example, a doorbell ringing sound may play even if a doorbell is not visibly depicted being activated (e.g., to imply activation of the doorbell). To account for such discrepancies, the model trainer 304 can include a machine-learned audiovisual coincidence model 332. The machine-learned audiovisual coincidence model 332 can predict a degree of likelihood whether a pair of audio data and video data (or intermediate representations thereof) are coincidental or causative. The output of the machine-learned audiovisual coincidence model 332 can be used as a training signal by the model trainer 304 to train the machine-learned generative audiovisual synchrony model 302. In other words, the optimization function 327 can evaluate the output of the machine-learned audiovisual coincidence model 332.

[0067] Additionally, in some implementations, the machine-learned audiovisual coincidence model 332 can be trained by the model trainer 304 using the optimization function 327 alongside the machine-learned generative audiovisual synchrony model 302. For example, the machine- learned audiovisual coincidence model 332 can generate a coincidence prediction score between the video data 306 and the audio data 318. During inference, a threshold coincidence prediction score can be included in the conditioning input(s) 328. The threshold coincidence prediction score can serve as a conditioning input that controls a degree of temporal synchronizationbetween the video data 306 and the audio data 318. In this manner, users of the model 302 can decide whether synchrony is desired for the desired audio data output. For example, if the user sets the coincidence token (i.e., the conditioning input 328 for the audiovisual temporal synchrony type) to 0, and sets the conditioning input 328 for the music sound type to 1, the audio data 318 generated using the machine-learned generative audiovisual synchrony model 302 can include a music soundtrack that is temporally independent (i.e., de-synchronized) from what is depicted by the video data 306. If the user wishes to include music along with event sounds that correspond to sound-producing events depicted by the video data 306, the user can set the coincidence token to 1 and the music token to 1.

[0068] In some implementations, Classifier-Free Guidance (CFG) can be leveraged as an effective mechanism by the model trainer 304 to improve performance of the machine-learned generative audiovisual synchrony model 302, both in terms of generating audio that is both more synchronous with video and consistent with sound type conditioning. The model trainer 304 can apply CFG by “running” or otherwise processing information with the machine-learned generative audiovisual synchrony model 302 twice. The first stage can process the information with the machine-learned generative audiovisual synchrony model 302 in a typical conditioned state, and the second stage can process the information with the model 302 again in an “unconditioned” state where the encoder inputs (i.e. video encodings + binary classes) are zeroed out, though for each time step the past history of the conditioned branch is used (i.e., “forced aligned”). The logits for sampling can be computed as Lu+ c (L - Lu), where Lurepresents unconditioned logits, L represents conditioned logits, and c > 1 is the CFG value.Example Methods

[0069] Figure 4 depicts a flow chart diagram of an example method 400 to perform training of generative audio models to generate audio that is temporally synchronized to sound-producing events depicted by a corresponding video input according to some implementations of the present disclosure. Although Figure 4 depicts steps performed in a particular order for purposes of illustration and discussion, the methods of the present disclosure are not limited to the particularly illustrated order or arrangement. The various steps of the method 400 can be omitted, rearranged, combined, and / or adapted in various ways without deviating from the scope of the present disclosure.

[0070] At 402, a computing system can process video data with a machine-learned generative audiovisual synchrony model to obtain audio data. The video data can depict occurrence of a sound-producing event at a particular time. The audio data can include an event sound predicted to be produced by the occurrence of the sound-producing event. Occurrence of the event sound within the audio data can be synchronized to the particular time at which the sound-producing event is depicted to occur by the video data.

[0071] In some implementations, the video data that depicts the occurrence of the soundproducing event at the particular time can depict a sound-producing object that produces sound at the particular time. In some implementations, the occurrence of the sound-producing event depicted by the video data can include the sound-producing object contacting a particular type of surface at the particular time. The event sound audio can include a sound predicted to be produced by the sound-producing object contacting the particular type of surface.

[0072] In some implementations, the video data further depicts an environment in which the sound-producing event occurs. Processing the video data with the machine-learned generative audiovisual synchrony model to obtain the audio data can include processing the video data with the machine-learned generative audiovisual synchrony model to obtain the audio data. The audio data can include the event sound, and the audio data can further include ambient sound predicted to correspond to the environment in which the sound-producing event occurs. For example, the video data can depict a person speaking within the environment in which the sound-producing event occurs. The computing system can process the video data with the machine-learned generative audiovisual synchrony model to obtain the audio data, and the audio data can include the event sound, the ambient sound, and a speech sound predicted to correspond to the person speaking within the environment in which the sound-producing event occurs.

[0073] In some implementations, to process the video data, the computing system can process the video data with an encoder portion of the machine-learned generative audiovisual synchrony model to obtain a sequence of video encodings. The computing system can process the sequence of video encodings with a decoder portion of the machine-learned generative audiovisual synchrony model to obtain the audio data. In some implementations, to process the sequence of video encodings, the computing system can process the sequence of video encodings with the decoder portion of the machine-learned generative audiovisual synchrony model to obtain a sequence of audio tokens.

[0074] In some implementations, processing the sequence of video encodings with the decoder portion can include processing the sequence of video encodings with a temporal submodel of the decoder portion of the machine-learned generative audiovisual synchrony model. The computing system can process the sequence of video encodings with a depth submodel of the decoder portion of the machine-learned generative audiovisual synchrony model. In some implementations, the decoder portion of the machine-learned generative audiovisual synchrony model can include a cross-attention layer, a first self-attention layer that immediately precedes the cross-attention layer, and a second self-attention layer that immediately succeeds the cross-attention layer. In some implementations, the cross-attention layer of the decoder portion of the machine-learned generative audiovisual synchrony model can include cross-attention values based on a first input to the cross-attention layer of the decoder portion and a second input to one or more corresponding cross-attention layers of the encoder portion of the machine-learned generative audiovisual synchrony model.

[0075] In some implementations, processing the sequence of video encodings with the decoder portion of the machine-learned generative audiovisual synchrony model to obtain the sequence of audio tokens can include processing the sequence of audio tokens with a generative vocoder portion of the machine-learned generative audiovisual synchrony model to obtain the audio data.

[0076] In some implementations, processing the video data with the encoder portion of the machine-learned generative audiovisual synchrony model to obtain the sequence of video encodings can include processing a plurality of image frames of the video data with a pre-trained vision encoder network of the encoder portion of the machine-learned generative audiovisual synchrony model to obtain a plurality of semantic patch encodings. The plurality of semantic patch encodings can be processed with a linear-time encoder submodel of the encoder portion of the machine-learned generative audiovisual synchrony model to obtain the sequence of video encodings.

[0077] In some implementations, processing the sequence of video encodings with the decoder portion of the machine-learned generative audiovisual synchrony model to obtain the audio data can include appending a plurality of conditioning inputs to the sequence of video encodings. The plurality of conditioning inputs can be indicative of a plurality of threshold scores for plurality of sound types respectively associated with the plurality of conditioninginput(s). The plurality of sound types can include at least one of a silence sound type, a speech sound type, a music sound type, or an audiovisual temporal synchrony type.

[0078] In some implementations, the plurality of audio class training tokens are indicative of the audiovisual temporal synchrony classification. The computing system can process the audio data and the video data with a pre-trained machine-learned audiovisual coincidence model to obtain a model output indicative of a probability that events captured by the video data and the audio data occurred during a same period of time. The computing system can train the machine- learned generative audiovisual synchrony model based on the model output.

[0079] At 404, the computing system can evaluate an optimization function that evaluates at least the event sound.

[0080] At 406, the computing system can adjust one or more parameters of the machine- learned generative audiovisual synchrony model based on the optimization function.

[0081] Figure 5 depicts a flow chart diagram of an example method 500 to perform inference with generative audio models trained to generate audio that is temporally synchronized to sound-producing events depicted by a corresponding video input according to some implementations of the present disclosure. Although Figure 5 depicts steps performed in a particular order for purposes of illustration and discussion, the methods of the present disclosure are not limited to the particularly illustrated order or arrangement. The various steps of the method 500 can be omitted, rearranged, combined, and / or adapted in various ways without deviating from the scope of the present disclosure.

[0082] At 502, a computing system can obtain video data and one or more conditioning inputs. In some implementations, the computing system can obtain the video data and the conditioning input(s) from a user. For example, the user can provide the video data to a generative audio service (e.g., via a user device). The user can then select the conditioning input(s) via an interface provided by the generative audio service. The video data can depict occurrence of a sound-producing event at a particular time. A first conditioning input of the one or more conditioning inputs can control inclusion of an event sound type of one or more sound types. The event sound type can include sound produced by the occurrence of the soundproducing event.

[0083] At 504, the computing system can process the video data and the one or more conditioning inputs with a machine-learned generative audiovisual synchrony model to obtainaudio data. The machine-learned generative audiovisual synchrony model can be trained to generate audio that is temporally synchronized to events depicted by video inputs. The audio data can include one or more sounds that include a first sound of the event sound type of the one or more sound types. Occurrence of the first sound within the audio data is synchronized to the particular time at which the sound-producing event is depicted to occur by the video data.

[0084] In some implementations, the one or more conditioning inputs can include the first conditioning input and a second conditioning input that controls inclusion of a second sound type of the one or more sound types. The second sound type can include an ambient sound type, a speech sound type, a music sound type, or a silence sound type. In some implementations, the one or more conditioning inputs further include an audiovisual temporal synchrony input that controls a degree of audiovisual synchrony between the occurrence of the first sound of the event sound type within the audio data and the occurrence of the sound-producing event at the particular time within the video data.Example Devices and Systems

[0085] Figure 6A depicts a block diagram of an example computing system 600 that performs generative audiovisual tasks according to example embodiments of the present disclosure. The system 600 includes a user computing device 602, a server computing system 630, and a training computing system 650 that are communicatively coupled over a network 680.

[0086] The user computing device 602 can be any type of computing device, such as, for example, a personal computing device (e.g., laptop or desktop), a mobile computing device (e.g., smartphone or tablet), a gaming console or controller, a wearable computing device, an embedded computing device, or any other type of computing device.

[0087] The user computing device 602 includes one or more processors 612 and a memory 614. The one or more processors 612 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. The memory 614 can include one or more non-transitory computer-readable storage media, such as RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. The memory 614 can store data 616 and instructions 618 which are executed by the processor 612 to cause the user computing device 602 to perform operations.

[0088] Tn some implementations, the user computing device 602 can store or include one or more machine-learned generative audiovisual synchrony models 620. For example, the machine- learned generative audiovisual synchrony models 620 can be or can otherwise include various machine-learned models such as neural networks (e.g., deep neural networks) or other types of machine-learned models, including non-linear models and / or linear models. Neural networks can include feed-forward neural networks, recurrent neural networks (e.g., long short-term memory recurrent neural networks), convolutional neural networks or other forms of neural networks. Some example machine-learned models can leverage an attention mechanism such as selfattention. For example, some example machine-learned models can include multi-headed selfattention models (e.g., transformer models). Example machine-learned generative audiovisual synchrony models 620 are discussed with reference to Figures 1-3.

[0089] In some implementations, the one or more machine-learned generative audiovisual synchrony models 620 can be received from the server computing system 630 over network 680, stored in the user computing device memory 614, and then used or otherwise implemented by the one or more processors 612. In some implementations, the user computing device 602 can implement multiple parallel instances of a single machine-learned generative audiovisual synchrony model 620 (e.g., to perform parallel generative audiovisual tasks across multiple instances of the machine-learned generative audiovisual synchrony model 620).

[0090] Additionally, or alternatively, one or more machine-learned generative audiovisual synchrony models 640 can be included in or otherwise stored and implemented by the server computing system 630 that communicates with the user computing device 602 according to a client-server relationship. For example, the machine-learned generative audiovisual synchrony models 640 can be implemented by the server computing system 630 as a portion of a web service (e.g., a generative audiovisual service). Thus, one or more models 620 can be stored and implemented at the user computing device 602 and / or one or more models 640 can be stored and implemented at the server computing system 630.

[0091] The user computing device 602 can also include one or more user input components 622 that receives user input. For example, the user input component 622 can be a touch-sensitive component (e.g., a touch-sensitive display screen or a touch pad) that is sensitive to the touch of a user input object (e.g., a finger or a stylus). The touch-sensitive component can serve toimplement a virtual keyboard. Other example user input components include a microphone, a traditional keyboard, or other means by which a user can provide user input.

[0092] The server computing system 630 includes one or more processors 632 and a memory 634. The one or more processors 632 can be any suitable processing device (e.g., a processor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. The memory 634 can include one or more non-transitory computer-readable storage media, such as RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. The memory 634 can store data 636 and instructions 638 which are executed by the processor 632 to cause the server computing system 630 to perform operations.

[0093] In some implementations, the server computing system 630 includes or is otherwise implemented by one or more server computing devices. In instances in which the server computing system 630 includes plural server computing devices, such server computing devices can operate according to sequential computing architectures, parallel computing architectures, or some combination thereof.

[0094] As described above, the server computing system 630 can store or otherwise include one or more machine-learned generative audiovisual synchrony models 640. For example, the models 640 can be or can otherwise include various machine-learned models. Example machine- learned models include neural networks or other multi-layer non-linear models. Example neural networks include feed forward neural networks, deep neural networks, recurrent neural networks, and convolutional neural networks. Some example machine-learned models can leverage an attention mechanism such as self-attention. For example, some example machine-learned models can include multi-headed self-attention models (e.g., transformer models). Example models 640 are discussed with reference to Figures 1-3.

[0095] The user computing device 602 and / or the server computing system 630 can train the models 620 and / or 640 via interaction with the training computing system 650 that is communicatively coupled over the network 680. The training computing system 650 can be separate from the server computing system 630 or can be a portion of the server computing system 630.

[0096] The training computing system 650 includes one or more processors 652 and a memory 654. The one or more processors 652 can be any suitable processing device (e.g., aprocessor core, a microprocessor, an ASIC, an FPGA, a controller, a microcontroller, etc.) and can be one processor or a plurality of processors that are operatively connected. The memory 654 can include one or more non-transitory computer- readable storage media, such as RAM, ROM, EEPROM, EPROM, flash memory devices, magnetic disks, etc., and combinations thereof. The memory 654 can store data 656 and instructions 658 which are executed by the processor 652 to cause the training computing system 650 to perform operations. In some implementations, the training computing system 650 includes or is otherwise implemented by one or more server computing devices.

[0097] The training computing system 650 can include a model trainer 660 that trains the machine-learned models 620 and / or 640 stored at the user computing device 602 and / or the server computing system 630 using various training or learning techniques, such as, for example, backwards propagation of errors. For example, a loss function can be backpropagated through the model(s) to update one or more parameters of the model(s) (e.g., based on a gradient of the loss function). Various loss functions can be used such as mean squared error, likelihood loss, cross entropy loss, hinge loss, and / or various other loss functions. Gradient descent techniques can be used to iteratively update the parameters over a number of training iterations.

[0098] In some implementations, performing backwards propagation of errors can include performing truncated backpropagation through time. The model trainer 660 can perform a number of generalization techniques (e.g., weight decays, dropouts, etc.) to improve the generalization capability of the models being trained.

[0099] In particular, the model trainer 660 can train the machine-learned generative audiovisual synchrony models 620 and / or 640 based on a set of training data 662. The training data 662 can include, for example, the ground-truth event sound 308 of Figure 3, or the like. Additionally, or alternatively, in some implementations, the model trainer 660 can perform unsupervised or weakly supervised training processes to train the model(s) 620 and / or 640.

[0100] In some implementations, if the user has provided consent, the training examples can be provided by the user computing device 602. Thus, in such implementations, the model 620 provided to the user computing device 602 can be trained by the training computing system 650 on user-specific data received from the user computing device 602. In some instances, this process can be referred to as personalizing the model.

[0101] The model trainer 660 includes computer logic utilized to provide desired functionality. The model trainer 660 can be implemented in hardware, firmware, and / or software controlling a general purpose processor. For example, in some implementations, the model trainer 660 includes program files stored on a storage device, loaded into a memory and executed by one or more processors. In other implementations, the model trainer 660 includes one or more sets of computer-executable instructions that are stored in a tangible computer-readable storage medium such as RAM, hard disk, or optical or magnetic media.

[0102] The network 680 can be any type of communications network, such as a local area network (e.g., intranet), wide area network (e.g., Internet), or some combination thereof and can include any number of wired or wireless links. In general, communication over the network 680 can be carried via any type of wired and / or wireless connection, using a wide variety of communication protocols (e g., TCP / IP, HTTP, SMTP, FTP), encodings or formats (e.g., HTML, XML), and / or protection schemes (e.g., VPN, secure HTTP, SSL).

[0103] The machine-learned models described in this specification may be used in a variety of tasks, applications, and / or use cases.

[0104] In some implementations, the input to the machine-learned model(s) of the present disclosure can be image data. The machine-learned model(s) can process the image data to generate an output. As an example, the machine-learned model(s) can process the image data to generate an image recognition output (e.g., a recognition of the image data, a latent embedding of the image data, an encoded representation of the image data, a hash of the image data, etc ). As another example, the machine-learned model(s) can process the image data to generate an image segmentation output. As another example, the machine-learned model(s) can process the image data to generate an image classification output. As another example, the machine-learned model(s) can process the image data to generate an image data modification output (e.g., an alteration of the image data, etc.). As another example, the machine-learned model(s) can process the image data to generate an encoded image data output (e.g., an encoded and / or compressed representation of the image data, etc.). As another example, the machine-learned model(s) can process the image data to generate an upscaled image data output. As another example, the machine-learned model(s) can process the image data to generate a prediction output.

[0105] Tn some implementations, the input to the machine-learned model(s) of the present disclosure can be text or natural language data. The machine-learned model(s) can process the text or natural language data to generate an output. As an example, the machine-learned model(s) can process the natural language data to generate a language encoding output. As another example, the machine-learned model(s) can process the text or natural language data to generate a latent text embedding output. As another example, the machine-learned model(s) can process the text or natural language data to generate a translation output. As another example, the machine-learned model(s) can process the text or natural language data to generate a classification output. As another example, the machine-learned model(s) can process the text or natural language data to generate a textual segmentation output. As another example, the machine-learned model(s) can process the text or natural language data to generate a semantic intent output. As another example, the machine-learned model(s) can process the text or natural language data to generate an upscaled text or natural language output (e.g., text or natural language data that is higher quality than the input text or natural language, etc.). As another example, the machine-learned model(s) can process the text or natural language data to generate a prediction output.

[0106] In some implementations, the input to the machine-learned model(s) of the present disclosure can be latent encoding data (e.g., a latent space representation of an input, etc.). The machine-learned model(s) can process the latent encoding data to generate an output. As an example, the machine-learned model(s) can process the latent encoding data to generate a recognition output. As another example, the machine-learned model(s) can process the latent encoding data to generate a reconstruction output. As another example, the machine-learned model(s) can process the latent encoding data to generate a search output. As another example, the machine-learned model(s) can process the latent encoding data to generate a reclustering output. As another example, the machine-learned model(s) can process the latent encoding data to generate a prediction output.

[0107] In some cases, the machine-learned model(s) can be configured to perform a task that includes encoding input data for reliable and / or efficient transmission or storage (and / or corresponding decoding). For example, the task may be an audio compression task. The input may include audio data and the output may comprise compressed audio data. In another example, the input includes visual data (e.g. one or more images or videos), the output comprisescompressed visual data, and the task is a visual data compression task. In another example, the task may comprise generating an embedding for input data (e.g. input audio or visual data).

[0108] In some cases, the input includes visual data and the task is a computer vision task. In some cases, the input includes pixel data for one or more images and the task is an image processing task. For example, the image processing task can be image classification, where the output is a set of scores, each score corresponding to a different object class and representing the likelihood that the one or more images depict an object belonging to the object class. The image processing task may be object detection, where the image processing output identifies one or more regions in the one or more images and, for each region, a likelihood that region depicts an object of interest. As another example, the image processing task can be image segmentation, where the image processing output defines, for each pixel in the one or more images, a respective likelihood for each category in a predetermined set of categories. For example, the set of categories can be foreground and background. As another example, the set of categories can be object classes. As another example, the image processing task can be depth estimation, where the image processing output defines, for each pixel in the one or more images, a respective depth value. As another example, the image processing task can be motion estimation, where the network input includes multiple images, and the image processing output defines, for each pixel of one of the input images, a motion of the scene depicted at the pixel between the images in the network input.

[0109] In some cases, the input includes audio data representing a spoken utterance and the task is a speech recognition task. The output may comprise a text output which is mapped to the spoken utterance. In some cases, the task comprises encrypting or decrypting input data. In some cases, the task comprises a microprocessor performance task, such as branch prediction or memory address translation.

[0110] Figure 6A illustrates one example computing system that can be used to implement the present disclosure. Other computing systems can be used as well. For example, in some implementations, the user computing device 602 can include the model trainer 660 and the training dataset 662. In such implementations, the models 620 can be both trained and used locally at the user computing device 602. In some of such implementations, the user computing device 602 can implement the model trainer 660 to personalize the models 620 based on userspecific data.

[0111] Figure 6B depicts a block diagram of an example computing device 670 that performs according to example embodiments of the present disclosure. The computing device 670 can be a user computing device or a server computing device.

[0112] The computing device 670 includes a number of applications (e.g., applications 1 through N). Each application contains its own machine learning library and machine-learned model(s). For example, each application can include a machine-learned model. Example applications include a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, etc.

[0113] As illustrated in Figure 6B, each application can communicate with a number of other components of the computing device, such as, for example, one or more sensors, a context manager, a device state component, and / or additional components. In some implementations, each application can communicate with each device component using an API (e.g., a public API). In some implementations, the API used by each application is specific to that application.

[0114] Figure 6C depicts a block diagram of an example computing device 675 that performs according to example embodiments of the present disclosure. The computing device 675 can be a user computing device or a server computing device.

[0115] The computing device 675 includes a number of applications (e.g., applications 1 through N). Each application is in communication with a central intelligence layer. Example applications include a text messaging application, an email application, a dictation application, a virtual keyboard application, a browser application, etc. In some implementations, each application can communicate with the central intelligence layer (and model(s) stored therein) using an API (e.g., a common API across all applications).

[0116] The central intelligence layer includes a number of machine-learned models. For example, as illustrated in Figure 6C, a respective machine-learned model can be provided for each application and managed by the central intelligence layer. In other implementations, two or more applications can share a single machine-learned model. For example, in some implementations, the central intelligence layer can provide a single model for all of the applications. In some implementations, the central intelligence layer is included within or otherwise implemented by an operating system of the computing device 675.

[0117] The central intelligence layer can communicate with a central device data layer. The central device data layer can be a centralized repository of data for the computing device 675. Asillustrated in Figure 6C, the central device data layer can communicate with a number of other components of the computing device, such as, for example, one or more sensors, a context manager, a device state component, and / or additional components. In some implementations, the central device data layer can communicate with each device component using an API (e.g., a private API).Additional Disclosure

[0118] The technology discussed herein makes reference to servers, databases, software applications, and other computer-based systems, as well as actions taken and information sent to and from such systems. The inherent flexibility of computer-based systems allows for a great variety of possible configurations, combinations, and divisions of tasks and functionality between and among components. For instance, processes discussed herein can be implemented using a single device or component or multiple devices or components working in combination. Databases and applications can be implemented on a single system or distributed across multiple systems. Distributed components can operate sequentially or in parallel.

[0119] While the present subject matter has been described in detail with respect to various specific example embodiments thereof, each example is provided by way of explanation, not limitation of the disclosure. Those skilled in the art, upon attaining an understanding of the foregoing, can readily produce alterations to, variations of, and equivalents to such embodiments. Accordingly, the subject disclosure does not preclude inclusion of such modifications, variations and / or additions to the present subject matter as would be readily apparent to one of ordinary skill in the art. For instance, features illustrated or described as part of one embodiment can be used with another embodiment to yield a still further embodiment. Thus, it is intended that the present disclosure cover such alterations, variations, and equivalents.

Claims

WHAT IS CLAIMED IS:

1. A computer-implemented method for training a generative audio model to generate audio that is temporally synchronized to sound-producing events depicted by a corresponding video input, comprising: processing, by a computing system comprising one or more computing devices, video data with a machine-learned generative audiovisual synchrony model to obtain audio data, wherein the video data depicts occurrence of a sound-producing event at a particular time, wherein the audio data comprises an event sound predicted to be produced by the occurrence of the soundproducing event, and wherein occurrence of the event sound within the audio data is synchronized to the particular time at which the sound-producing event is depicted to occur by the video data; evaluating, by the computing system, an optimization function that evaluates at least the event sound; and adjusting, by the computing system, one or more parameters of the machine-learned generative audiovisual synchrony model based on the optimization function.

2. The computer-implemented method of claim 1, wherein the video data that depicts the occurrence of the sound-producing event at the particular time depicts a sound-producing object that produces sound at the particular time.

3. The computer-implemented method of claim 2, wherein the occurrence of the soundproducing event depicted by the video data comprises the sound-producing object contacting a particular type of surface at the particular time; and wherein the event sound audio comprises a sound predicted to be produced by the soundproducing object contacting the particular type of surface.

4. The computer-implemented method of claim 1, wherein the video data further depicts an environment in which the sound-producing event occurs; and wherein processing the video data with the machine-learned generative audiovisual synchrony model to obtain the audio data comprises:processing, by the computing system, the video data with the machine-learned generative audiovisual synchrony model to obtain the audio data, wherein the audio data comprises the event sound, and wherein the audio data further comprises ambient sound predicted to correspond to the environment in which the sound-producing event occurs.

5. The computer-implemented method of claim 3, wherein the video data further depicts a person speaking within the environment in which the sound-producing event occurs; and wherein processing the video data with the machine-learned generative audiovisual synchrony model to obtain the audio data comprises: processing, by the computing system, the video data with the machine-learned generative audiovisual synchrony model to obtain the audio data, wherein the audio data comprises the event sound and the ambient sound, and wherein the audio data further comprises speech sound predicted to correspond to the person speaking within the environment in which the sound-producing event occurs.

6. The computer-implemented method of claim 1, wherein processing the video data with the machine-learned generative audiovisual synchrony model to obtain the audio data comprises: processing, by the computing system, the video data with an encoder portion of the machine-learned generative audiovisual synchrony model to obtain a sequence of video encodings; and processing, by the computing system, the sequence of video encodings with a decoder portion of the machine-learned generative audiovisual synchrony model to obtain the audio data.

7. The computer-implemented method of claim 6, wherein processing the sequence of video encodings with the decoder portion of the machine-learned generative audiovisual synchrony model to obtain the audio data comprises: processing, by the computing system, the sequence of video encodings with the decoder portion of the machine-learned generative audiovisual synchrony model to obtain a sequence of audio tokens.

8. The computer-implemented method of claim 7, wherein processing the sequence of video encodings with the decoder portion of the machine-learned generative audiovisual synchrony model to obtain the sequence of audio tokens comprises: processing, by the computing system, the sequence of video encodings with a temporal submodel of the decoder portion of the machine-learned generative audiovisual synchrony model; and processing, by the computing system, the sequence of video encodings with a depth submodel of the decoder portion of the machine-learned generative audiovisual synchrony model.

9. The computer-implemented method of claim 7, wherein the decoder portion of the machine-learned generative audiovisual synchrony model comprises a cross-attention layer, a first self-attention layer that immediately precedes the cross-attention layer, and a second selfattention layer that immediately succeeds the cross-attention layer.

10. The computer-implemented method of claim 9, wherein the cross-attention layer of the decoder portion of the machine-learned generative audiovisual synchrony model comprises cross-attention values based on a first input to the cross-attention layer of the decoder portion and a second input to one or more corresponding cross-attention layers of the encoder portion of the machine-learned generative audiovisual synchrony model.

11. The computer-implemented method of claim 7, wherein processing the sequence of video encodings with the decoder portion of the machine-learned generative audiovisual synchrony model to obtain the sequence of audio tokens further comprises: processing, by the computing system, the sequence of audio tokens with a generative vocoder portion of the machine-learned generative audiovisual synchrony model to obtain the audio data.

12. The computer-implemented method of claim 6, wherein processing the video data with the encoder portion of the machine-learned generative audiovisual synchrony model to obtain the sequence of video encodings comprises:processing, by the computing system, a plurality of image frames of the video data with a pre-trained vision encoder network of the encoder portion of the machine-learned generative audiovisual synchrony model to obtain a plurality of semantic patch encodings; and processing, by the computing system, the plurality of semantic patch encodings with a linear-time encoder submodel of the encoder portion of the machine-learned generative audiovisual synchrony model to obtain the sequence of video encodings.

13. The computer-implemented method of claim 6, wherein processing the sequence of video encodings with the decoder portion of the machine-learned generative audiovisual synchrony model to obtain the audio data comprises: appending, by the computing system, a plurality of conditioning inputs to the sequence of video encodings, wherein the plurality of conditioning inputs are indicative of a plurality of threshold scores for plurality of sound types respectively associated with the plurality of conditioning input(s), comprising at least one of: a silence sound type; a speech sound type; a music sound type; or an audiovisual temporal synchrony type.

14. The computer-implemented method of claim 13, wherein the plurality of conditioning inputs are indicative of the audiovisual temporal synchrony type, and wherein the method comprises: processing, by the computing system, the audio data and the video data with a pre-trained machine-learned audiovisual coincidence model to obtain a model output indicative of a probability that events captured by the video data and the audio data occurred during a same period of time; and training, by the computing system, the machine-learned generative audiovisual synchrony model based on the model output.

15. A computing system, comprising: one or more processor devices; andone or more non-transitory computer-readable media that store instructions that, when executed by the one or more processors, cause the computing system to perform operations, the operations comprising: obtaining video data and one or more conditioning inputs, wherein the video data depicts occurrence of a sound-producing event at a particular time, wherein a first conditioning input of the one or more conditioning inputs controls inclusion of an event sound type of one or more sound types, wherein the event sound type comprises sound produced by the occurrence of the sound-producing event; and processing the video data and the one or more conditioning inputs with a machine- learned generative audiovisual synchrony model to obtain audio data, wherein the machine- learned generative audiovisual synchrony model is trained to generate audio that is temporally synchronized to events depicted by video inputs, wherein the audio data comprises one or more sounds comprising a first sound of the event sound type of the one or more sound types, and wherein occurrence of the first sound within the audio data is synchronized to the particular time at which the sound-producing event is depicted to occur by the video data.

16. The computing system of claim 15, wherein the one or more conditioning inputs comprises the first conditioning input and a second conditioning input that controls inclusion of a second sound type of the one or more sound types, wherein the second sound type comprises: an ambient sound type; a speech sound type; a music sound type; or a silence sound type.

17. The computing system of claim 15, wherein the one or more conditioning inputs further comprise an audiovisual temporal synchrony input that controls a degree of audiovisual synchrony between the occurrence of the first sound of the event sound type within the audio data and the occurrence of the sound-producing event at the particular time within the video data.

18. The computing system of claim 15, wherein processing the video data and the one or more conditioning inputs with the machine-learned generative audiovisual synchrony model to obtain the audio data comprises: processing the video data with an encoder portion of the machine-learned generative audiovisual synchrony model to obtain a sequence of video encodings; and processing the sequence of video encodings with a decoder portion of the machine-learned generative audiovisual synchrony model to obtain the audio data.

19. The computing system of claim 18, wherein processing the sequence of video encodings with the decoder portion of the machine-learned generative audiovisual synchrony model to obtain the audio data comprises: processing the sequence of video encodings with the decoder portion of the machine- learned generative audiovisual synchrony model to obtain a sequence of audio tokens.

20. One or more non-transitory computer-readable media that store instructions that, when executed by one or more processors, cause the one or more processors to perform operations, the operations comprising: processing video data with a machine-learned generative audiovisual synchrony model to obtain audio data, wherein the video data depicts occurrence of a sound-producing event at a particular time, wherein the audio data comprises an event sound predicted to be produced by the occurrence of the sound-producing event, and wherein occurrence of the event sound within the audio data is synchronized to the particular time at which the sound-producing event is depicted to occur by the video data; evaluating an optimization function that evaluates at least the event sound; and adjusting one or more parameters of the machine-learned generative audiovisual synchrony model based on the optimization function.

Images