Google Inc.

International Business Machines Corporation

Fadi Biadsy

Pedro Jose Moreno Mengibar

Mark Edward Epstein

Bhuvana Ramabhadran

Diego Melendo Casado

Andrew M Rosenberg

Diamantino Antonio Caseiro

Lidia Mangu

Daniel Martin Bikel

Hagen Soltau

Petar Aleksic

Ron J Weiss

Kaisuke Nakajima

Martin Jansche

Pedro J Mengibar

Diamantino Antionio Caseiro

Johan Schalkwyk

Yonghui Wu

Yu Zhang

Zhifeng Chen

Aleksandar Kracun

Ye Jia

Fernando Pereira

Wolfgang Macherey

Dirk Padfield

Zhehuai Chen

Jason William Pelecanos

Brian E Roark

Victoria Zayats

Gary Wang

Liyang Jiang

Kapil R Thadini

Maria Shugrina

Melvin Johnson

Pedro J Moreno Mengiba

A method for training a sub-model for contextual biasing for speech recognition includes obtaining a base speech recognition model trained on non-biased data. The method includes obtaining a set of training utterances representative of a particular domain, each training utterance in the set of training utterances including audio data characterizing the training utterances and a ground truth transcription of the training utterance. The method further includes, for each corresponding training utterance in the set of training utterances, determining, using an embedding encoder, a corresponding document embedding from the ground truth transcription of the corresponding training utterance. The method includes training, using the corresponding document embeddings determined from the ground truth transcriptions of the set of training utterances, a sub-model to bias the base speech recognition model to recognize speech in the particular domain.

Sub-models for neural contextual biasing with attention and embedding space

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for end to end speech conversion are disclosed. In one aspect, a method includes the actions of receiving first audio data of a first utterance of one or more first terms spoken by a user. The actions further include providing the first audio data as an input to a model that is configured to receive first given audio data in a first voice and output second given audio data in a synthesized voice without performing speech recognition on the first given audio data. The actions further include receiving second audio data of a second utterance of the one or more first terms spoken in the synthesized voice. The actions further include providing, for output, the second audio data of the second utterance of the one or more first terms spoken in the synthesized voice.

End-to-end speech conversion

A method for speech conversion includes receiving, as input to an encoder of a speech conversion model, an input spectrogram corresponding to an utterance, the encoder including a stack of self-attention blocks. The method further includes generating, as output from the encoder, an encoded spectrogram and receiving, as input to a spectrogram decoder of the speech conversion model, the encoded spectrogram generated as output from the encoder. The method further includes generating, as output from the spectrogram decoder, an output spectrogram corresponding to a synthesized speech representation of the utterance.

Conformer-based speech conversion model

A method for contextual biasing for speech recognition includes obtaining a base automatic speech recognition (ASR) model trained on non-biased data and a sub-model trained on biased data representative of a particular domain. The method includes receiving a speech recognition request including audio data characterizing an utterance captured in streaming audio. The method further includes determining whether the speech recognition request includes a contextual indicator indicating the particular domain. When the speech recognition request does not include the contextual indicator, the method includes generating, using the base ASR model, a first speech recognition result of the utterance by processing the audio data. When the speech recognition request includes the contextual indicator the method includes biasing, using the sub-model, the base ASR model toward the particular domain and generating, using the biased base ASR model, a second speech recognition result of the utterance by processing the audio data.

Sub-models for neural contextual biasing

Implementations disclosed herein are directed to techniques for selectively enabling and/or disabling non-transient storage of one or more instances of assistant interaction data for turn(s) of a dialog between a user and an automated assistant. Implementations are additionally or alternatively directed to techniques for retroactive wiping of non-transiently stored assistant interaction data from previous assistant interaction(s).

Preventing non-transient storage of assistant interaction data and/or wiping of stored assistant interaction data

A method includes receiving a set of training utterances each including a non-synthetic speech representation of a corresponding utterance, and for each training utterance, generating a corresponding synthetic speech representation by using a voice conversion model. The non-synthetic speech representation and the synthetic speech representation form a corresponding training utterance pair. At each of a plurality of output steps for each training utterance pair, the method also includes generating, for output by a speech recognition model, a first probability distribution over possible non-synthetic speech recognition hypotheses for the non-synthetic speech representation and a second probability distribution over possible synthetic speech recognition hypotheses for the synthetic speech representation. The method also includes determining a consistent loss term for the corresponding training utterance pair based on the first and second probability distributions and updating parameters of the speech recognition model based on the consistent loss term.

Using non-parallel voice conversion for speech conversion models

A method includes receiving a reference audio signal corresponding to reference speech spoken by a target speaker with atypical speech, and generating, by a speaker embedding network configured to receive the reference audio signal as input, a speaker embedding for the target speaker. The speaker embedding conveys speaker characteristics of the target speaker. The method also includes receiving a speech conversion request that includes input audio data corresponding to an utterance spoken by the target speaker associated with the atypical speech. The method also includes biasing, using the speaker embedding generated for the target speaker by the speaker embedding network, a speech conversion model to convert the input audio data corresponding to the utterance spoken by the target speaker associated with atypical speech into an output canonical representation of the utterance spoken by the target speaker.

Speaker embeddings for improved automatic speech recognition

A method () of training a text-to-speech (TTS) model () includes obtaining training data () including reference input text () that includes a sequence of characters, a sequence of reference audio features () representative of the sequence of characters, and a sequence of reference phone labels () representative of distinct speech sounds of the reference audio features. For each of a plurality of time steps, the method includes generating a corresponding predicted audio feature () based on a respective portion of the reference input text for the time step and generating, using a phone label mapping network (), a corresponding predicted phone label () associated with the predicted audio feature. The method also includes aligning the predicted phone label with the reference phone label to determine a corresponding predicted phone label loss () and updating the TTS model based on the corresponding predicted phone label loss.

Training speech synthesis to generate distinct speech sounds

The present disclosure provides systems and methods that train and use machine-learned models such as, for example, sequence-to-sequence models, to perform direct and text-free speech-to-speech translation. In particular, aspects of the present disclosure provide an attention-based sequence-to-sequence neural network which can directly translate speech from one language into speech in another language, without relying on an intermediate text representation.

Direct speech-to-speech translation via machine learning

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for language models using domain-specific model components. In some implementations, context data for an utterance is obtained. A domain-specific model component is selected from among multiple domain-specific model components of a language model based on the non-linguistic context of the utterance. A score for a candidate transcription for the utterance is generated using the selected domain-specific model component and a baseline model component of the language model that is domain-independent. A transcription for the utterance is determined using the score the transcription is provided as output of an automated speech recognition system.

Language models using domain-specific model components

A method includes receiving acoustic features of a first utterance spoken by a first user that speaks with typical speech and processing the acoustic features of the first utterance using a general speech recognizer to generate a first transcription of the first utterance. The operations also include analyzing the first transcription of the first utterance to identify one or more bias terms in the first transcription and biasing the alternative speech recognizer on the one or more bias terms identified in the first transcription. The operations also include receiving acoustic features of a second utterance spoken by a second user that speaks with atypical speech and processing, using the alternative speech recognizer biased on the one or more terms identified in the first transcription, the acoustic features of the second utterance to generate a second transcription of the second utterance.

Speech recognition

A method for training a speech conversion model personalized for a target speaker with atypical speech includes obtaining a plurality of transcriptions in a set of spoken training utterances and obtaining a plurality of unspoken training text utterances. Each spoken training utterance is spoken by a target speaker associated with atypical speech and includes a corresponding transcription paired with a corresponding non-synthetic speech representation. The method also includes adapting, using the set of spoken training utterances, a text-to-speech (TTS) model to synthesize speech in a voice of the target speaker and that captures the atypical speech. For each unspoken training text utterance, the method also includes generating, as output from the adapted TTS model, a synthetic speech representation that includes the voice of the target speaker and that captures the atypical speech. The method also includes training the speech conversion model based on the synthetic speech representations.

Synthesized data augmentation using voice conversion and speech recognition models

Methods and systems for automatic speech recognition and methods and systems for training acoustic language models are disclosed. In accordance with one automatic speech recognition method, an acoustic input data set is analyzed to identify portions of the input data set that conform to a general language and to identify portions of the input data set that conform to at least one dialect of the general language. In addition, a general language model and at least one dialect language model is applied to the input data set to perform speech recognition by dynamically selecting between the models in accordance with each of the identified portions. Further, speech recognition results obtained in accordance with the application of the models is output.

Growing community of inventors

New York, NY, United States of America

Fadi Biadsy