Abstract: A method for speech conversion includes obtaining a speech conversion model configured to convert input utterances of human speech directly into corresponding output utterances of synthesized speech. The method further includes receiving a speech conversion request including input audio data corresponding to an utterance spoken by a target speaker associated with atypical speech and a speaker identifier uniquely identifying the target speaker. The method includes activating, using the speaker identifier, a particular sub-model for biasing the speech conversion model to recognize a type of the atypical speech associated with the target speaker identified by the speaker identifier.

Abstract: A method for turn detection in a speech-to-speech model includes receiving, as input to the speech-to-speech (S2S) model, a sequence of acoustic frames corresponding to an utterance. The method further includes, at each of a plurality of output steps, generating, by an audio encoder of the S2S model, a higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames, and determining, by a turn detector of the S2S model, based on the higher order feature representation generated by the audio encoder at the corresponding output step, whether the utterance is at a breakpoint at the corresponding output step. When the turn detector determines that the utterance is at the breakpoint, the method includes synthesizing a sequence of output audio frames output by a speech decoder of the S2S model into a time-domain audio waveform of synthesized speech representing the utterance spoken by the user.

Abstract: A method includes receiving a current spectrogram frame and reconstructing a phase of the current spectrogram frame by, for each corresponding committed spectrogram frame in a sequence of M number of committed spectrogram frames preceding the current spectrogram frame, obtaining a value of a committed phase of the corresponding committed spectrogram frame and estimating the phase of the current spectrogram frame based on a magnitude of the current spectrogram frame and the value of the committed phase of each corresponding committed spectrogram frame in the sequence of M number of committed spectrogram frames preceding the current spectrogram frame. The method also includes synthesizing, for the current spectrogram frame, a new time-domain audio waveform frame based on the estimated phase of the current spectrogram frame.

Abstract: A method for speech conversion includes obtaining a speech conversion model configured to convert input utterances of human speech directly into corresponding output utterances of synthesized speech. The method further includes receiving a speech conversion request including input audio data corresponding to an utterance spoken by a target speaker associated with atypical speech and a speaker identifier uniquely identifying the target speaker. The method includes activating, using the speaker identifier, a particular sub-model for biasing the speech conversion model to recognize a type of the atypical speech associated with the target speaker identified by the speaker identifier.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for performing speech-to-speech translation, including real-time speech-to-speech translation.

Abstract: Implementations are described herein for processing a stream of time-varying input data to generate/predict a stream of time-varying output data in real-time or near-real time. In various implementations, while a stream of input frames, such as a stream of audio input frames, is received, audio input frames received up to a current time step may be tokenized (e.g., midstream) to generate a stream of audio input tokens. A Transformer-based causal attention model may be used to predict a stream of audio output tokens, e.g., by iteratively applying the Transformer-based causal attention model to: at least some of the audio input tokens tokenized up to the current time step, and at least some of the audio output tokens predicted up to the current time step. The stream of audio output tokens may be detokenized to generate a stream of audio output frames.

Abstract: A method includes obtaining a plurality of training samples that each include audio data characterizing a corresponding speech utterance and a transcription of the corresponding speech utterance. The method also includes training an automatic speech recognition (ASR) model on the plurality of training samples, the ASR model having a recurrent neural network-transducer (RNN-T) architecture. The method also includes quantizing the trained ASR model to an integer target fixed-bit width and providing the quantized trained ASR model to a user device.

Abstract: A method for training a model includes obtaining a plurality of training samples. Each respective training sample of the plurality of training samples includes a respective speech utterance and a respective textual utterance representing a transcription of the respective speech utterance. The method includes training, using quantization aware training with native integer operations, an automatic speech recognition (ASR) model on the plurality of training samples. The method also includes quantizing the trained ASR model to an integer target fixed-bit width. The quantized trained ASR model includes a plurality of weights. Each weight of the plurality of weights includes an integer with the target fixed-bit width. The method includes providing the quantized trained ASR model to a user device.

Abstract: A method includes obtaining a plurality of training samples that each include a respective speech utterance and a respective textual utterance representing a transcription of the respective speech utterance. The method also includes fine-tuning, using quantization and sparsity aware training with native integer operations, a pre-trained automatic speech recognition (ASR) model on the plurality of training samples. Here, the pre-trained ASR model includes a plurality of weights and the fine-tuning includes pruning one or more weights of the plurality of weights using a sparsity mask and quantizing each weight of the plurality of weights based on an integer with a fixed-bit width. The method also includes providing the fine-tuned ASR model to a user device.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for processing an input audio waveform using a generator neural network to generate an output audio waveform. In one aspect, a method comprises: receiving an input audio waveform; processing the input audio waveform using an encoder neural network to generate a set of feature vectors representing the input audio waveform; and processing the set of feature vectors representing the input audio waveform using a decoder neural network to generate an output audio waveform that comprises a respective output audio sample for each of a plurality of output time steps.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for processing an input audio waveform using a generator neural network to generate an output audio waveform. In one aspect, a method comprises: receiving an input audio waveform; processing the input audio waveform using an encoder neural network to generate a set of feature vectors representing the input audio waveform; and processing the set of feature vectors representing the input audio waveform using a decoder neural network to generate an output audio waveform that comprises a respective output audio sample for each of a plurality of output time steps.

Abstract: The present disclosure relates to a streaming speech-to-speech conversion model, where an encoder runs in real time while a user is speaking, then after the speaking stops, a decoder generates output audio in real time. A streaming-based approach produces an acceptable delay with minimal loss in conversion quality when compared to other non-streaming server-based models. A hybrid model approach for combines look-ahead in the encoder and a non-causal stacker with non-causal self-attention.

Abstract: Apparatus and methods related to enhancement of audio content are provided. An example method includes receiving, by a computing device and via a communications network interface, a compressed audio data frame, wherein the compressed audio data frame is received after transmission over a communications network, The method further includes decompressing the compressed audio data frame to extract an audio waveform. The method also includes predicting, by applying a neural network to the audio waveform, an enhanced version of the audio waveform, wherein the neural network has been trained on (i) a ground truth sample comprising unencoded audio waveforms prior to compression by an audio encoder, and (ii) a training dataset comprising decoded audio waveforms after compression of the unencoded audio waveforms by the audio encoder. The method additionally includes providing, by an audio output component of the computing device, the enhanced version of the audio waveform.

Abstract: The present disclosure relates to a streaming speech-to-speech conversion model, where an encoder runs in real time while a user is speaking, then after the speaking stops, a decoder generates output audio in real time. A streaming-based approach produces an acceptable delay with minimal loss in conversion quality when compared to other non-streaming server-based models. A hybrid model approach for combines look-ahead in the encoder and a non-causal stacker with non-causal self-attention.

Abstract: Exemplary embodiments provide distributed parallel training of a machine learning model. Multiple processors may be used to train a machine learning model to reduce training time. To synchronize trained model data between the processors, data is communicated between the processors after some number of training cycles. To improve the communication efficiency, exemplary embodiments synchronize data among a set of processors after a predetermined number of training cycles, and synchronize data between one or more processors of each set of the processors after a predetermined number of training cycles. During the first synchronization among a set of processors, compressed model gradient data generated after performing the training cycles may be communicated. During the second synchronization between the set of processors, trained models or full model gradient data generated after performing the training cycles may be communicated.

Abstract: A method for turn detection in a speech-to-speech model includes receiving, as input to the speech-to-speech (S2S) model, a sequence of acoustic frames corresponding to an utterance. The method further includes, at each of a plurality of output steps, generating, by an audio encoder of the S2S model, a higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames, and determining, by a turn detector of the S2S model, based on the higher order feature representation generated by the audio encoder at the corresponding output step, whether the utterance is at a breakpoint at the corresponding output step. When the turn detector determines that the utterance is at the breakpoint, the method includes synthesizing a sequence of output audio frames output by a speech decoder of the S2S model into a time-domain audio waveform of synthesized speech representing the utterance spoken by the user.

Abstract: A method for training a model includes obtaining a plurality of training samples. Each respective training sample of the plurality of training samples includes a respective speech utterance and a respective textual utterance representing a transcription of the respective speech utterance. The method includes training, using quantization aware training with native integer operations, an automatic speech recognition (ASR) model on the plurality of training samples. The method also includes quantizing the trained ASR model to an integer target fixed-bit width. The quantized trained ASR model includes a plurality of weights. Each weight of the plurality of weights includes an integer with the target fixed-bit width. The method includes providing the quantized trained ASR model to a user device.

Abstract: A method for speech conversion includes obtaining a speech conversion model configured to convert input utterances of human speech directly into corresponding output utterances of synthesized speech. The method further includes receiving a speech conversion request including input audio data corresponding to an utterance spoken by a target speaker associated with atypical speech and a speaker identifier uniquely identifying the target speaker. The method includes activating, using the speaker identifier, a particular sub-model for biasing the speech conversion model to recognize a type of the atypical speech associated with the target speaker identified by the speaker identifier.

Abstract: A method includes receiving a current spectrogram frame and reconstructing a phase of the current spectrogram frame by, for each corresponding committed spectrogram frame in a sequence of M number of committed spectrogram frames preceding the current spectrogram frame, obtaining a value of a committed phase of the corresponding committed spectrogram frame and estimating the phase of the current spectrogram frame based on a magnitude of the current spectrogram frame and the value of the committed phase of each corresponding committed spectrogram frame in the sequence of M number of committed spectrogram frames preceding the current spectrogram frame. The method also includes synthesizing, for the current spectrogram frame, a new time-domain audio waveform frame based on the estimated phase of the current spectrogram frame.

Abstract: Real-time evaluation and enhancement of image quality prior to capturing an image of a document on a mobile device is provided. An image capture process is initiated on a mobile device during which a user of the mobile device prepares to capture the image of the document, utilizing hardware and software on the mobile device to measure and achieve optimal parameters for image capture. Feedback may be provided to a user of the mobile device to instruct the user on how to manually optimize certain parameters relating to image quality, such as the angle, motion and distance of the mobile device from the document. When the optimal parameters for image capture of the document are achieved, at least one image of the document is automatically captured by the mobile device.