Active speaker detection in videos using cross-signal reprogramming

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for performing active speaker detection in a video. In one aspect, the active speaker detection is performed using cross-signal reprogramming on audio and video feature vectors.

Description

[0001] ACTIVE SPEAKER DETECTION IN VIDEOS USING CROSS-SIGNAL REPROGRAMMING

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS

[0003] This application claims the benefit of priority to U.S. Provisional Application No. 63 / 605.441 filed on December 1, 2023, the contents of which are hereby incorporated by reference.

[0004] BACKGROUND

[0005] This specification relates to processing audio and video inputs using neural networks.

[0006] Neural networks are machine learning models that employ one or more layers of nonlinear units to predict an output for a received input. Some neural networks include one or more hidden layers in addition to an output layer. The output of each hidden layer is used as input to the next layer in the network, i.e., the next hidden layer or the output layer. Each layer of the network generates an output from a received input in accordance with current value inputs of a respective set of parameters.

[0007] SUMMARY

[0008] This specification describes a system implemented as computer programs on one or more computers in one or more locations that performs active speaker detection on input videos.

[0009] The subject matter described in this specification can be implemented in particular embodiments so as to realize one or more of the following advantages.

[0010] This specification describes techniques for performing active speaker detection on input videos using an active speaker detection neural network. Active speaker detection is the task of determining which person is speaking in any given frame of a given video.

[0011] Some systems perform active speaker detection on input videos by generating audio feature vectors from the audio track of the video and video features from the video frames of the video and then directly providing the audio and video features as input to an active speaker detection neural network, e.g., a Transformer-based neural network. However, these techniques struggle to achieve reliably high-quality performance, e.g., because the neural network struggles to grasp the subtle nuances of audio-visual interactions solely from the uni- modal features. Moreover, the distributions of these features can be susceptible to noise, e.g., harming training and inference performance of the active speaker detection neural network.

[0012] Unlike these systems, the described techniques use cross-signal reprogramming (CSR) to generate reprogrammed audio and video feature vectors. In other words, the described techniques use the reprogrammed audio and video feature vectors as input to the active speaker detection neural network instead of the audio and video feature vectors.

[0013] Because both the audio and video feature vectors are susceptible to noise, using crosssignal reprogramming (CSR) stabilizes the respective input distributions and allows the active speaker detection neural network to more accurately perform active speaker detection.

[0014] Moreover, making use of CSR allows the active speaker detection neural network to achieve state-of-the-art results despite being significantly more computationally efficient than conventional high-performing active speaker detection neural networks. For example, the neural network can be three times smaller (in terms of number of parameters) than conventional high-performing neural networks but still achieve comparable or better results.

[0015] The details of one or more embodiments of the subject matter of this specification are set forth in the accompanying drawings and the description below.

[0016] Other features, aspects, and advantages of the subj ect matter will become apparent from the description, the drawings, and the claims.

[0017] BRIEF DESCRIPTION OF THE DRAWINGS

[0018] FIG. 1 shows an example active speaker detection system.

[0019] FIG. 2 is a flow diagram of an example process for generating an active speaker detection output.

[0020] FIG. 3 shows an example of the operation of the system.

[0021] FIG. 4 is a flow diagram of an example process for performing cross-signal reprogramming.

[0022] FIG. 5 is a flow diagram of an example process for processing the reprogrammed feature vectors to generate the speaker detection output.

[0023] FIG. 6 shows an example of the performance of the described techniques relative to existing approaches.

[0024] Like reference numbers and designations in the various draw ings indicate like elements. DETAILED DESCRIPTION

[0025] FIG. 1 shows an example active speaker detection system 100. The active speaker detection system 100 is an example of a system implemented as computer programs on one or more computers in one or more locations, in which the systems, components, and techniques described below can be implemented.

[0026] The system 100 performs active speaker detection on input videos.

[0027] In particular, the system 100 receives an input video 110 that includes a respective video frame 112 at each of multiple time steps and a corresponding audio signal 114.

[0028] In at least some of the video frames 112, the faces of one or more candidate speakers are depicted within the video frame 112. The speakers are referred to as “candidate’' speakers because, while a given candidate speaker may be depicted in a given video frame 112 at a given time step, a different speaker may be speaking at the given time step or there may be no one speaking at the given time step. That is, in any given frame 112, multiple people may be depicted, while only one of the people is speaking, so that some of the “candidate” speakers are not actually speaking in the frame 112. Similarly, one or more people may be depicted, while none are speaking, so that none of the “candidate” speakers are actually speaking in the frame 112.

[0029] The system 100 processes the input video 110 and the audio signal 114 to generate an active speaker detection output 120 that indicates, for each candidate speaker face at a given time step within the video 110, whether the corresponding candidate speaker is speaking at the given time step.

[0030] For example, the speaker detection output 120 can include, for each candidate speaker at a given time step, a score, e.g., a probability, that represents a predicted likelihood that the candidate speaker is speaking at the time step.

[0031] Once generated, the system 100 can use the active speaker detection output 120 for any of a variety of purposes.

[0032] For example, the system 100 can post-process the video 110 to visually identify, in some or all of the video frames 112, the face of the speaker that is speaking in the video frame 112. For example, the system can annotate the video to place bounding boxes around or otherwise highlight the speakers that are speaking in various video frames 112 within the video.

[0033] As another example, the system 100 can post-process the video 110 to crop out or otherwise edit portions of the video frames 112 that are not relevant to the active speaker, e.g., by removing the faces of non-speakers from a given video frame 112. As yet another example, the system 100 can use the active speaker detection output 120 to perform speaker diarisation of the audio signal.

[0034] As yet another example, the system 100 can use the active speaker detection output to perform speaker tracking on the input video.

[0035] To perform active speaker detection, the system 100 generates, from the video 110, a set of video feature vectors 130. Each video feature vector 130 corresponds to a respective face of a respective candidate speaker at a respective time step in the video. Generally, each video feature vector 130 includes a first number of features.

[0036] The system 100 also generates, from the audio signal 114, a respective audio feature vector 140 for each of the time steps in the video 110. Each of the audio feature vectors 140 includes a second number of features, which can be the same as or different from the first number of features.

[0037] The system 100 then preforms cross-signal reprogramming 150 on both the video feature vectors 130 and the audio feature vectors 140 to generate, for each audio feature vector 140. a reprogrammed audio feature vector 142. and, for each video feature vector, a reprogrammed video feature vector 132.

[0038] In particular, as part of performing cross-signal reprogramming 150, for each audio feature vector 140, the system 140 generates a reprogrammed audio feature vector 142 that includes a source number of features by applying, to the audio feature vector 140, a learned reprogramming transformation. Generally, the source number of features is larger than the first and second numbers of features.

[0039] For each video feature vector 130, the system 100 generates a reprogrammed video feature vector 132 that includes the source number of features by applying, to the video feature vector 130. the learned reprogramming transformation.

[0040] Generally, applying the learned reprogramming transformation enriches the audio and video feature vectors with additional information so that they can be more effectively processed to perform active speaker detection.

[0041] Performing cross-signal reprogramming will be described in more detail below with reference to FIGS. 2-4.

[0042] The system 100 then processes the reprogrammed audio feature vectors 132 and the reprogrammed video feature vectors 142 using an active speaker detection neural network 160 to generate an active speaker detection output 120 for the video 110. As described above, the active speaker detection output 120 indicates, for each video feature vector 130. whether the corresponding candidate speaker is speaking at the corresponding time step in the video 1 10.

[0043] In other words, the system 130 uses the reprogrammed audio and video feature vectors 142 and 132 as input to the active speaker detection neural network 160 instead of the audio and video feature vectors 140 and 130.

[0044] Because both the audio and video feature vectors 140 and 130 are susceptible to noise, using cross-signal reprogramming (CSR) 150 stabilizes the respective input distributions and allows the active speaker detection neural network 160 to more accurately perform active speaker detection.

[0045] Moreover, making use of CSR allows the active speaker detection neural network 160 to achieve state-of-the-art results despite being significantly more computationally efficient than conventional high-performing active speaker detection neural networks. For example, the neural network 160 can be three times smaller (in terms of number of parameters) than conventional high-performing neural networks but still achieve comparable or better results.

[0046] FIG. 2 is a flow diagram of an example process 200 for performing active speaker detection on a video. For convenience, the process 200 will be described as being performed by a system of one or more computers located in one or more locations. For example, an active speaker detection system, e.g., the active speaker detection system 100 depicted in FIG. 1. appropriately programmed in accordance with this specification, can perform the process 200.

[0047] The system receives a new input (step 202).

[0048] The new input contains (i) a video that has a respective video frame at each of a plurality of time steps and (ii) a corresponding audio signal. For example, the system can receive the new input as an upload from a user or through an application programming interface (API) call.

[0049] The system can obtain the audio signal represented in any of a variety of ways. For example, the system can receive the audio signal as a waveform or as a spectrogram.

[0050] The system generates, from the video, a set of video feature vectors that each have a first number of features (step 204). That is, each feature vector is a vector in Hkdv‘deo, where c / video is the first number of features.

[0051] Each video feature vector corresponds to a respective face of a respective candidate speaker at a respective time step in the video. The system can generate the video feature vectors in any of a variety of ways.

[0052] As one example, the system can apply a face detector to the video frame to identify one or more regions within the video frame that each depict a respective face of a respective candidate speaker. The face detector can be any appropriate, pre-trained face detector.

[0053] The system can then generate, using a video feature extraction neural network, a respective video feature vector for each of the one or more regions. For example, the video feature extractor neural network can be a convolutional neural network, e.g., one with a ResNet-based architecture. As will be described in more detail below, the system can train the video feature extraction neural network jointly with the active speaker detection neural network.

[0054] As one example, the system can process the video frames using the feature extractor neural network to generate a feature representation of the video that includes a respective feature vector for each of multiple regions within each of the video frames. The system can then extract, for each region identified by the face detector, the corresponding feature vector from the feature representation.

[0055] The system generates, from the audio signal, a respective audio feature vector for each of the time steps in the video (step 206). Each of the audio feature vectors has a second number of features. That is, each audio vector is a vector in Ukdaudio, where daudiois the second number of features. In some cases, the first and second number of features are the same while, in other cases, the first and second numbers of features are different from one another. That is daudiocan be equal to dvideoor can be different from dvideo.

[0056] For example, the system can generate the audio feature vectors by processing a representation of the audio signal using an audio feature extraction neural network to generate the respective audio feature vectors.

[0057] For example, the representation of the audio signal can be a spectrogram of the audio signal and the audio feature extraction neural network can be a convolutional neural network, e.g., a 2-D convolutional neural network, e.g., a ResNet. As will be described in more detail below, like the video feature extraction neural network, the system can train the audio feature extraction neural network j ointly with the active speaker detection neural network.

[0058] The system then performs cross-signal reprogramming on the audio and video feature vectors to generate reprogrammed audio and video feature vectors (step 208).

[0059] Generally, the reprogrammed audio and video features both include a source number of features. That is, each reprogrammed audio and video feature vector is a vector in IRdsource,where dsourceis the source number of features. Generally, the source number of features is greater than the first number of features and the second number of features. That is, the reprogrammed audio and video feature vectors have more features than both the audio and video feature vectors.

[0060] Performing cross-signal reprogramming is described in more detail below with reference to FIG. 4.

[0061] The system processes the reprogrammed audio feature vectors and the reprogrammed video feature vectors using an active speaker detection neural network to generate an active speaker detection output (step 210).

[0062] As described above, the active speaker detection output indicates, for each video feature vector, whether the corresponding candidate speaker is speaking at the respective time step in the video. For example, the output can include a respective score, e.g., a probability or other score, for each video feature vector that indicates whether the corresponding candidate speaker is speaking at the respective time step in the video that represents the likelihood.

[0063] The active speaker detection neural network can generally have any appropriate architecture that maps each reprogrammed video feature to a score conditioned on the reprogrammed audio feature vectors. Examples of such architectures include self-attention neural networks, convolutional neural networks, and so on.

[0064] One example of the processing performed by the neural network is described below with reference to FIG. 5.

[0065] FIG. 3 shows an example 300 of the operation of the system.

[0066] As shown in the example 300, the system receives a video 310 and corresponding audio 320. The system performs video feature extraction 330 on the video 310 to generate video feature vectors and performs audio feature extraction 340 on the audio signal 320 to generate audio feature vectors.

[0067] The system then performs cross-signal reprogramming 350 on the audio and video feature vectors to generate reprogrammed audio and video feature vectors.

[0068] The system then processes the reprogrammed audio feature vectors and the reprogrammed video feature vectors using an active speaker detection neural network to generate an active speaker detection output 370. In the example 300, the active speaker detection neural network includes a sequence of Transformer blocks 360. some of which perform cross-attention while others of which perform self-attention. This example will be described in more detail below with reference to FIG. 5. FIG. 4 is a flow diagram of an example process 400 for performing cross-signal reprogramming. For convenience, the process 400 will be described as being performed by a system of one or more computers located in one or more locations. For example, an active speaker detection system, e.g., the active speaker detection system 100 depicted in FIG. 1, appropriately programmed in accordance with this specification, can perform the process 400.

[0069] The system obtains the audio feature vectors and video feature vectors (step 402).

[0070] For each audio feature vector, the system generates a reprogrammed audio feature vector that includes the source number of features by applying, to the audio feature vector, a learned reprogramming transformation (step 404).

[0071] That is. the learned reprogramming transformation reprograms a feature vector from a target domain, i.e., the domain of vectors that have the second number of features, to a source domain, i.e., the domain of vectors that have the source number of features.

[0072] One example of the learned reprogramming transformation now follows.

[0073] In this example, the system can apply the learned reprogramming transformation such that each reprogrammed audio feature vector includes, (i) for each feature in the source number of features that has a corresponding feature in the second number of features, the value of the corresponding feature from the audio feature vector and (ii) for each feature in the source number of features that does not have a corresponding feature in the second number of features, a respective learned value for the feature.

[0074] In this example, the respective learned values for the features are learned during training of the active speaker detection neural network.

[0075] Thus, in this example, the learned reprogramming transformation does not modify the features in the second number of features but generates additional, learned values for the features that are appended to the audio feature vector to generate the reprogrammed audio feature vectors.

[0076] In other words, given an input feature xt where dtargetis the first number of features or the second number features, the learned reprogramming transformation generates an output feature x'tin ^source where dtarget< dsource. In the above example, x'tcan satisfy: x't= S(xtfl) = P(%t) + BQ . where P(xt) represents a zero-padding function that generates a zero-padded feature vector with dimension of dsource. B is a binary mask that signifies the location of xtwithin F(t). 0 represents the Hadamard (element- wise) product, and fl in IRdsource js a setof learnable parameters used to align the distributions of the source and target domains. In other words, B is a binary mask in which the z-th entry is a) 0 if xtis present at that position, i.e., indicating it as non-reprogrammable and b) is 1 if the position is unoccupied, i.e., designating it as reprogrammable.

[0077] For each video feature vector, the system generates a reprogrammed video feature vector that includes the source number of features by applying, to the video feature vector, the learned reprogramming transformation (step 406).

[0078] That is, the system applies the same learned reprogramming transformation to the video feature vectors as to the audio feature vectors. Thus, continuing with the above example, for each video feature vector, the respective reprogrammed video feature vector includes (i) for each feature in the source number of features that has a corresponding feature in the first number of features, the value of the corresponding feature from the video feature vector; and (ii) for each feature in the source number of features that does not have a corresponding feature in the first number of features, a respective learned value for the feature.

[0079] In some cases, the respective learned values for video and audio features are the same while, in other cases, the respective learned value scan be different between video and audio features.

[0080] FIG. 5 is a flow diagram of an example process 500 for processing the reprogrammed audio feature vectors and the reprogrammed video feature vectors using the active speaker detection neural network to generate an active speaker detection output. For convenience, the process 500 will be described as being performed by a system of one or more computers located in one or more locations. For example, an active speaker detection system, e.g., the active speaker detection system 100 depicted in FIG. 1, appropriately programmed in accordance with this specification, can perform the process 500.

[0081] In particular, the process 500 is one example of the operations performed by the active speaker detection neural network to generate the active speaker detection output.

[0082] The system processes the reprogrammed video feature vectors conditioned on the reprogrammed audio feature vectors to update the reprogrammed video feature vectors (step 502).

[0083] For example, the system can update the reprogrammed video feature vectors by performing cross-attention into the reprogrammed audio feature vectors. That is. the system can process the reprogrammed video feature vectors using a cross-attention layer block that includes one or more cross-attention heads. Each cross-attention head generates queries from the reprogrammed video feature vectors and keys and values from the reprogrammed audio feature vectors and then applies an attention mechanism to the queries, keys, and values to generate the output of the attention head. When there are multiple attention heads, the crossattention layer block can then combine the outputs of the attention heads to generate a combined output, e.g.. by concatenating each of the output vectors and then optionally applying a linear transformation to the concatenation. Optionally, the layer block can also include other operations, e.g., one or more of position-wise feed-forward layers, normalization layers, and residual connections.

[0084] The system processes the reprogrammed audio feature vectors conditioned on the reprogrammed video feature vectors to update the reprogrammed audio feature vectors (step 504).

[0085] For example, the system can update the reprogrammed audio feature vectors by performing cross-attention into the reprogrammed video feature vectors. That is. the system can process the reprogrammed audio feature vectors using a cross-attention layer block that includes one or more cross-attention heads. Each cross-attention head generates queries from the reprogrammed audio feature vectors and keys and values from the reprogrammed video feature vectors and then applies an attention mechanism to the queries, keys, and values to generate the output of the attention head. When there are multiple attention heads, the crossattention layer block can then combine the outputs of the attention heads to generate a combined output, e.g., by concatenating each of the output vectors and then optionally applying a linear transformation to the concatenation. Optionally, the layer block can also include other operations, e.g., one or more of position- wise feed-forward layers, normalization layers, and residual connections.

[0086] In these examples, by using the reprogrammed audio and video feature vectors instead of the original audio and video feature vectors, the system optimizes the extraction of valuable information via cross attention, by preprocessing the audio and visual feature vectors before they enter the cross attention stage, i.e., by enriching the audio and video feature vectors with additional information that enhances the performance of the crossattention operations.

[0087] For each updated reprogrammed video feature, the system generates a combined video feature by combining the updated reprogrammed video feature and the updated audio feature that corresponds to the same time step as the updated reprogrammed audio feature (step 506). For example, the system can concatenate or sum the updated video feature and the corresponding updated audio feature, i.e., the audio feature vector that corresponds to the same time step as the video from which the video feature vector was extracted, to generate the combined video feature.

[0088] The system processes the combined video features to generate the active speaker detection output (step 508).

[0089] As a particular example, the system can process the combined video features using one or more self-attention layer blocks to update the combined video features. That is, the system can process the combined video features through a sequence of one or more selfattention layer blocks that each update the combined video features.

[0090] Each self-attention layer blocks includes one or more self-attention heads. Each selfattention head generates queries, keys, and values from the combined video features and then applies an attention mechanism to the queries, keys, and values to generate the output of the attention head. When there are multiple attention heads, the self-attention layer block can then combine the outputs of the attention heads to generate a combined output, e.g., by concatenating each of the output vectors and then optionally applying a linear transformation to the concatenation. Optionally, the layer block can also include other operations, e.g., one or more of position- wise feed-forward layers, normalization layers, and residual connections.

[0091] After processing the combined video features using the one or more self-attention layer blocks to update the combined video features, the system can, for each combined video feature, process the combined video feature using an output neural network head to generate a score that indicates a likelihood that the respective speaker corresponding to the combined video feature is speaking at the time step corresponding to the combined video feature.

[0092] The output neural network head can generally have any appropriate architecture that maps a feature vector to a score. For example, the output neural network head can be a single linear layer, a single linear layer followed by a non-linear activation function, or a multi-layer perceptron (MLP).

[0093] Prior to using the active speaker detection neural network to generate speaker detection outputs, the system trains the speaker detection neural network on a set of training data that includes a set of training examples, where each training example includes a training input that, in turn, includes a video and corresponding audio and a respective ground truth speaker detection output that identifies the active speaker in some or all of the video frames in the video in the training example. For example, the system can train the active speaker detection neural network on an objective that measures, for a given training example, an error between (i) the active speaker detection output generated by the active speaker detection neural network for the training input in the training example and (ii) the ground truth output in the training example. For example, the objective can be a cross-entropy loss function between the active speaker detection output and the ground truth output.

[0094] As part of this training, the system also learns the parameters of the learned reprogramming transformation, e.g., by backpropagating gradients of the objective through the active speaker detection neural network.

[0095] Similarly, the system can also train the audio and video feature extraction neural networks, e.g., e.g., by backpropagating gradients of the objective through the active speaker detection neural network and through the cross-signal reprogramming operations.

[0096] FIG. 6 shows an example 600 of the performance of the described techniques relative to existing approaches.

[0097] In particular, the example 600 shows the performance on AVA- ActiveSpeaker dataset in terms of the mean average precision (mAP) metric of the described techniques (CSR-ASD) when the detection neural network uses the Transformer-based architecture described above with reference to FIG. 5. As can be seen from FIG. 5, the described techniques perform comparably to or outperform a variety of previously state-of-the-art techniques that use a variety of different neural network architectures. This is despite the fact that the described techniques are significantly more computationally efficient than many of the approaches. For example, compared to the ASDNet technique shown in example 600, the described techniques use a model that size ~3Xsmaller than that of ASDNet, with sizes of ~15 MB and ~48 MB. As another example, the described techniques demonstrate swift convergence during training, requiring 36% fewer epochs to attain a similar level of accuracy compared to the existing state-of-the-art transformer-based approaches shown in FIG. 6.

[0098] This specification uses the term “configured” in connection with systems and computer program components. For a system of one or more computers to be configured to perform particular operations or actions means that the system has installed on it software, firmware, hardware, or a combination of them that in operation cause the system to perform the operations or actions. For one or more computer programs to be configured to perform particular operations or actions means that the one or more programs include instructions that, when executed by data processing apparatus, cause the apparatus to perform the operations or actions.

[0099] Embodiments of the subject matter and the functional operations described in this specification can be implemented in digital electronic circuitry, in tangibly-embodied computer software or firmware, in computer hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, e.g., one or more modules of computer program instructions encoded on a tangible non transitory storage medium for execution by, or to control the operation of, data processing apparatus. The computer storage medium can be a machine- readable storage device, a machine-readable storage substrate, a random or serial access memory device, or a combination of one or more of them. Alternatively or in addition, the program instructions can be encoded on an artificially generated propagated signal, e.g., a machine-generated electrical, optical, or electromagnetic signal, that is generated to encode information for transmission to suitable receiver apparatus for execution by a data processing apparatus.

[0100] The term “data processing apparatus” refers to data processing hardware and encompasses all kinds of apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, or multiple processors or computers. The apparatus can also be, or further include, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application specific integrated circuit). The apparatus can optionally include, in addition to hardware, code that creates an execution environment for computer programs, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, or a combination of one or more of them.

[0101] A computer program, which may also be referred to or described as a program, software, a software application, an app, a module, a software module, a script, or code, can be written in any form of programming language, including compiled or interpreted languages, or declarative or procedural languages; and it can be deployed in any form, including as a stand alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment. A program may, but need not, correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data, e.g., one or more scripts stored in a markup language document, in a single file dedicated to the program in question, or in multiple coordinated files, e.g., files that store one or more modules, sub programs, or portions of code. A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a data communication network. In this specification, the term “database” is used broadly to refer to any collection of data: the data does not need to be structured in any particular way, or structured at all, and it can be stored on storage devices in one or more locations. Thus, for example, the index database can include multiple collections of data, each of which may be organized and accessed differently.

[0102] Similarly, in this specification the term “engine” is used broadly to refer to a software-based system, subsystem, or process that is programmed to perform one or more specific functions. Generally, an engine will be implemented as one or more software modules or components, installed on one or more computers in one or more locations. In some cases, one or more computers will be dedicated to a particular engine; in other cases, multiple engines can be installed and running on the same computer or computers.

[0103] The processes and logic flows described in this specification can be performed by one or more programmable computers executing one or more computer programs to perform functions by operating on input data and generating output. The processes and logic flows can also be performed by special purpose logic circuitry, e.g., an FPGA or an ASIC, or by a combination of special purpose logic circuitry and one or more programmed computers.

[0104] Computers suitable for the execution of a computer program can be based on general or special purpose microprocessors or both, or any other kind of central processing unit. Generally, a central processing unit will receive instructions and data from a read only memory or a random access memory or both. The essential elements of a computer are a central processing unit for performing or executing instructions and one or more memory devices for storing instructions and data. The central processing unit and the memory can be supplemented by, or incorporated in, special purpose logic circuitry. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto optical disks, or optical disks. However, a computer need not have such devices. Moreover, a computer can be embedded in another device, e.g., a mobile telephone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a Global Positioning System (GPS) receiver, or a portable storage device, e.g., a universal serial bus (USB) flash drive, to name just a few.

[0105] Computer readable media suitable for storing computer program instructions and data include all forms of non volatile memory, media and memory devices, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto optical disks; and CD ROM and DVD-ROM disks. To provide for interaction with a user, embodiments of the subject matter described in this specification can be implemented on a computer having a display device, e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, for displaying information to the user and a keyboard and a pointing device, e.g., a mouse or a trackball, by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, speech, or tactile input. In addition, a computer can interact with a user by sending documents to and receiving documents from a device that is used by the user; for example, by sending web pages to a web browser on a user’s device in response to requests received from the web browser. Also, a computer can interact with a user by sending text messages or other forms of message to a personal device, e.g., a smartphone that is running a messaging application, and receiving responsive messages from the user in return.

[0106] Data processing apparatus for implementing machine learning models can also include, for example, special-purpose hardware accelerator units for processing common and compute-intensive parts of machine learning training or production, e.g., inference, workloads.

[0107] Machine learning models can be implemented and deployed using a machine learning framework, .e.g., a TensorFlow framework or a Jax framework.

[0108] Embodiments of the subject matter described in this specification can be implemented in a computing system that includes a back end component, e.g., as a data server, or that includes a middleware component, e.g., an application server, or that includes a front end component, e.g., a client computer having a graphical user interface, a web browser, or an app through which a user can interact with an implementation of the subject matter described in this specification, or any combination of one or more such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication, e.g., a communication network. Examples of communication networks include a local area network (LAN) and a wide area network (WAN), e.g., the Internet.

[0109] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. In some embodiments, a server transmits data, e.g., an HTML page, to a user device, e.g., for purposes of displaying data to and receiving user input from a user interacting with the device, which acts as a client. Data generated at the user device, e.g., a result of the user interaction, can be received at the server from the device.

[0110] While this specification contains many specific implementation details, these should not be construed as limitations on the scope of any invention or on the scope of what may be claimed, but rather as descriptions of features that may be specific to particular embodiments of particular inventions. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially be claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.

[0111] Similarly, while operations are depicted in the drawings and recited in the claims in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system modules and components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0112] Particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. As one example, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some cases, multitasking and parallel processing may be advantageous.

Claims

CLAIMS1. A method performed by one or more computers, the method comprising: receiving anew input that comprises (i) a video comprising a respective video frame at each of a plurality of time steps and (ii) a corresponding audio signal; generating, from the video, a set of video feature vectors that each correspond to a respective face of a respective candidate speaker at a respective time step in the video, wherein each video feature vector includes a first number of features; generating, from the audio signal, a respective audio feature vector for each of the time steps in the video, wherein each of the audio feature vectors includes a second number of features; for each audio feature vector, generating a reprogrammed audio feature vector that includes a source number of features by applying, to the audio feature vector, a learned reprogramming transformation; for each video feature vector, generating a reprogrammed video feature vector that includes the source number of features by applying, to the video feature vector, the learned reprogramming transformation; and processing the reprogrammed audio feature vectors and the reprogrammed video feature vectors using an active speaker detection neural network to generate an active speaker detection output that indicates, for each video feature vector, whether the corresponding candidate speaker is speaking at the respective time step in the video.

2. The method of claim 1, wherein, for each audio feature vector, the respective reprogrammed audio feature vector comprises: for each feature in the source number of features that has a corresponding feature in the second number of features, the value of the corresponding feature from the audio feature vector; and for each feature in the source number of features that does not have a corresponding feature in the second number of features, a respective learned value for the feature.

3. The method of claim 1 or claim 2 , wherein, for each video feature vector, the respective reprogrammed video feature vector comprises: for each feature in the source number of features that has a corresponding feature in the first number of features, the value of the corresponding feature from the video feature vector; and for each feature in the source number of features that does not have a corresponding feature in the first number of features, a respective learned value for the feature.

4. The method of claim 2 or claim 3, wherein the respective learned values for the features are learned during training of the active speaker detection neural network.

5. The method of any preceding claim, wherein the source number of features is greater than the first number of features and the second number of features.

6. The method of any preceding claim, wherein the first number of features and the second number of features are equal.

7. The method of any preceding claim, wherein the active speaker detection neural network is configured to: process the reprogrammed video feature vectors conditioned on the reprogrammed audio feature vectors to update the reprogrammed video feature vectors; process the reprogrammed audio feature vectors conditioned on the reprogrammed video feature vectors to update the reprogrammed audio feature vectors; for each updated reprogrammed video feature, generate a combined video feature by combining the updated reprogrammed video feature and the updated audio feature that corresponds to a same time step as the updated reprogrammed audio feature; and process the combined video features to generate the active speaker detection output.

8. The method of claim 7, wherein processing the combined video features to generate the active speaker detection output comprises: processing the combined video features using one or more self-attention layer blocks to update the combined video features; and after processing the combined video features using one or more self-attention layer blocks to update the combined video features and for each combined video feature, processing the combined video feature using an output neural network head to generate a score that indicates a likelihood that the respective speaker corresponding to the combined video feature is speaking at the time step corresponding to the combined video feature.

9. The method of claim 7 or claim 8, wherein processing the reprogrammed video feature vectors conditioned on the reprogrammed audio feature vectors to update the reprogrammed video feature vectors comprises: performing cross-attention into the reprogrammed audio feature vectors.

10. The method of any one of claims 7-9, wherein processing the reprogrammed audio feature vectors conditioned on the reprogrammed video feature vectors to update the reprogrammed audio feature vectors comprises: performing cross-attention into the reprogrammed video feature vectors.

11. The method of any preceding claim, wherein, generating, from the video, a set of video feature vectors that each correspond to a respective face of a respective candidate speaker at a respective time step in the video comprises, for each video frame: applying a face detector to the video frame to identify one or more regions within the video frame that each depict a respective face of a respective candidate speaker; and generating, using a video feature extraction neural network, a respective video feature vector for each of the one or more regions.

12. The method of any preceding claim, wherein generating, from the audio signal, a respective audio feature vector for each of the time steps in the video that each include a second number of features comprises: processing a representation of the audio signal using an audio feature extraction neural network to generate the respective audio feature vectors.

13. The method of claim 12, wherein the representation of the audio signal is a spectrogram of the audio signal.

14. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one more computers to perform the operations of the respective method of any one of claims 1-13.

15. One or more computer storage media storing instructions that when executed by one or more computers cause the one more computers to perform the operations of the respective method of any one of claims 1-13.

Images