Google Inc.

Trevor D Strohman

Tara N Sainath

Ruoming Pang

Rohit Prakash Prabhavalkar

Yanzhang He

David Rybach

Arun Narayanan

Ehsan Variani

Qiao Liang

Weiran Wang

Rami Botros

Chung-Cheng Chiu

Jaclyn Konzelmann

Bo Li

Francoise Beaufays

Brian Strope

Michael Andrew Goodman

Françoise Beaufays

Valerie Nygaard

Shuo-Yiin Chang

Jiahui Yu

Ke Hu

Basilio Garcia Castillo

Cyril Allauzen

Pedro Jose Moreno Mengibar

Jun Xu

Bhuvana Ramabhadran

Paul Haahr

Johan Schalkwyk

Yonghui Wu

Ron J Weiss

Fuchun Peng

Giovanni Motta

Joseph Robert Smarr

Khe Chai Sim

Martin Baeuml

Gabriel Schine

Wei Li

Xin Lei

Ian C McGraw

Rajiv Mathews

Neeraj Gaur

Shankar Kumar

Parisa Haghani

Li Wei

Ronny Huang

Yangzhang He

Dan Rasin

Rafael Goldfarb

Dekel Auster

Lior Alon

Sepand Mavandadi

Nino Tasca

Guru Prakash Arumugam

Zhiyun Lu

Tongzhou Chen

Krzysztof Marcin Choromanski

Anmol Gulati

Adam Sadovsky

Jay Shrauner

Mirkó Visontai

Chao Zhang

Ian Mcgraw

Manaal Faruqui

Liangliang Cao

Qiumin Xu

Charles Caleb Peyser

Per Bjornsson

Zelin Wu

Shaojin Ding

Shuo-yiin Chang

Oren Litvin

Lillian Zhou

Mingqing Chen

Jonathan Bloom

Ding Zhao

Thibault Doutre

Adam Stambler

Shyam Upadhyay

Yanwei Pan

Vince Chen

Heguang Liu

Sameer Bidichandani

Dhruv Guliani

Harry Zhang

Alexander Bailey

Florent D'Halluin

Jonas Bragagnolo

Xin Wang

Dongseong Hwang

Robert David

Rina Panigrahy

Basi Garcia

Gleb Skobeltsyn

Tara C Sainath

Ian C Mcgraw

Basilio Castillo Garcia

Brian Patrick Strope

A method includes obtaining a set of training samples, wherein each training sample includes a corresponding sequence of speech segments corresponding to a training utterance and a corresponding sequence of ground-truth transcriptions for the sequence of speech segments, and wherein each ground-truth transcription includes a start time and an end time of a corresponding speech segment. For each training sample in the set of training samples, the method includes processing, using a speech recognition model, the corresponding sequence of speech segments to obtain one or more speech recognition hypotheses for the training utterance; and, for each speech recognition hypothesis obtained for the training utterance, identifying a respective number of word errors relative to the corresponding sequence of ground-truth transcriptions. The method trains the speech recognition model to minimize word error rate based on the respective number of word errors identified for each speech recognition hypothesis obtained for the training utterance.

Training for long-form speech recognition

Two-pass automatic speech recognition (ASR) models can be used to perform streaming on-device ASR to generate a text representation of an utterance captured in audio data. Various implementations include a first-pass portion of the ASR model used to generate streaming candidate recognition(s) of an utterance captured in audio data. For example, the first-pass portion can include a recurrent neural network transformer (RNN-T) decoder. Various implementations include a second-pass portion of the ASR model used to revise the streaming candidate recognition(s) of the utterance and generate a text representation of the utterance. For example, the second-pass portion can include a listen attend spell (LAS) decoder. Various implementations include a shared encoder shared between the RNN-T decoder and the LAS decoder.

Two-pass end to end speech recognition

An automated speech recognition (ASR) model includes a first encoder, a first encoder, a second encoder, and a second decoder. The first encoder receives, as input, a sequence of acoustic frames, and generates, at each of a plurality of output steps, a first higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames. The first decoder receives, as input, the first higher order feature representation generated by the first encoder, and generates a first probability distribution over possible speech recognition hypotheses. The second encoder receives, as input, the first higher order feature representation generated by the first encoder, and generates a second higher order feature representation for a corresponding first higher order feature frame. The second decoder receives, as input, the second higher order feature representation generated by the second encoder, and generates a second probability distribution over possible speech recognition hypotheses.

Unified cascaded encoder ASR model for dynamic model sizes

A computer-implemented method includes receiving audio data that corresponds to an utterance spoken by a user and captured by a user device. The method also includes processing the audio data to determine a candidate transcription that includes a sequence of tokens for the spoken utterance. Tor each token in the sequence of tokens, the method includes determining a token embedding for corresponding token, determining a n-gram token embedding for a previous sequence of n-gram tokens, and concatenating the token embedding and the n-gram token embedding to generate a concatenated output for the corresponding token. The method also includes rescoring the candidate transcription for the spoken utterance by processing the concatenated output generated for each corresponding token in the sequence of tokens.

Lookup-table recurrent language model

Implementations relate to dialog management of a large language model (LLM) utilized in generating natural language (NL) output during an ongoing dialog. Processor(s) of a system can: receive NL based input as part of the ongoing dialog, generate NL based output utilizing the LLM, and cause the NL based output to be rendered. Further, the processor(s) can receive subsequent NL based input as part of the ongoing dialog. In some implementations, the processor(s) can determine whether to modify a corresponding dialog context in generating subsequent NL based output, and modify the corresponding dialog context accordingly. For example, the processor(s) can restrict the corresponding dialog context, or supplant the corresponding dialog context with a corresponding curated dialog context. In additional or alternative implementations, the processor(s) can modify a corresponding NL based output threshold utilized in generating the subsequent NL based response to ensure the resulting NL based output is desirable.

Dialog management for large language model- based (LLM-based) dialogs

A method includes receiving a training example that includes audio data representing a spoken utterance and a ground truth transcription. For each word in the spoken utterance, the method also includes inserting a placeholder symbol before the respective word identifying a respective ground truth alignment for a beginning and an end of the respective word, determining a beginning word piece and an ending word piece, and generating a first constrained alignment for the beginning word piece and a second constrained alignment for the ending word piece. The first constrained alignment is aligned with the ground truth alignment for the beginning of the respective word and the second constrained alignment is aligned with the ground truth alignment for the ending of the respective word. The method also includes constraining an attention head of a second pass decoder by applying the first and second constrained alignments.

Emitting word timings with end-to-end models

A method of text-only and semi-supervised training for deliberation includes receiving training data including unspoken textual utterances that are each not paired with any corresponding spoken utterance of non-synthetic speech, and training a deliberation model that includes a text encoder and a deliberation decoder on the unspoken textual utterances. The method also includes receiving, at the trained deliberation model, first-pass hypotheses and non-causal acoustic embeddings. The first-pass hypotheses is generated by a recurrent neural network-transducer (RNN-T) decoder for the non-causal acoustic embeddings encoded by a non-causal encoder. The method also includes encoding, using the text encoder, the first-pass hypotheses generated by the RNN-T decoder, and generating, using the deliberation decoder attending to both the first-pass hypotheses and the non-causal acoustic embeddings, second-pass hypotheses.

Deliberation by text-only and semi-supervised training

A method includes generating, using an audio encoder, a higher-order feature representation for each acoustic frame in a sequence of acoustic frames; generating, using a decoder, based on the higher-order feature representation, a plurality of speech recognition hypotheses, each hypotheses corresponding to a candidate transcription of an utterance and having an associated first likelihood score; generating, using an external language model, for each speech recognition hypothesis, a second likelihood score; determining, using a learnable fusion module, for each speech recognition hypothesis, a set of fusion weights based on the higher-order feature representation and the speech recognition hypothesis; and generating, using the learnable fusion module, for each speech recognition hypothesis, a third likelihood score based on the first likelihood score, the second likelihood score, and the set of fusion weights, the audio encoder and decoder trained using minimum additive error rate training in the presence of the external language model.

Rare word recognition with LM-aware MWER training

Implementations described herein identify and correct automatic speech recognition (ASR) misrecognitions. For example, on-device processor(s) of a client device may generate a predicted textual segment that is predicted to correspond to spoken utterance of a user of the client device, and may receive further input that modifies the predicted textual segment to an alternate textual segment. Further, the on-device processor(s) may store these textual segments in on-device storage as a candidate correction pair, and transmit the candidate correction pair to a remote system. Moreover, remote processor(s) of the remote system may determine that the candidate correction pair is an actual correction pair, and may cause client devices to generate updates for a global ASR model for the candidate correction pair. Additionally, the remote processor(s) may distribute the global ASR model to the client devices and/or additional client devices.

Identifying and correcting automatic speech recognition (ASR) misrecognitions in a decentralized manner

A method includes receiving, as input to a speech recognition model, audio data corresponding to a spoken utterance. The method also includes performing, using the speech recognition model, speech recognition on the audio data by, at each of a plurality of time steps, encoding, using an audio encoder, the audio data corresponding to the spoken utterance into a corresponding audio encoding, and decoding, using a speech recognition joint network, the corresponding audio encoding into a probability distribution over possible output labels. At each of the plurality of time steps, the method also includes determining, using an intended query (IQ) joint network configured to receive a label history representation associated with a sequence of non-blank symbols output by a final softmax layer, an intended query decision indicating whether or not the spoken utterance includes a query intended for a digital assistant.

Intended query detection using E2E modeling for continued conversation

A computer-implemented method includes receiving a sequence of acoustic frames as input to an automatic speech recognition (ASR) model. Here, the ASR model includes a causal encoder and a decoder. The method also includes generating, by the causal encoder, a first higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames. The method also includes generating, by the decoder, a first probability distribution over possible speech recognition hypotheses. Here, the causal encoder includes a stack of causal encoder layers each including a Recurrent Neural Network (RNN) Attention-Performer module that applies linear attention.

Optimizing inference performance for conformer

A method includes receiving a sequence of acoustic frames characterizing one or more utterances as input to a multilingual automated speech recognition (ASR) model. The method also includes generating a higher order feature representation for a corresponding acoustic frame. The method also includes generating a hidden representation based on a sequence of non-blank symbols output by a final softmax layer. The method also includes generating a probability distribution over possible speech recognition hypotheses based on the hidden representation generated by the prediction network at each of the plurality of output steps and the higher order feature representation generated by the encoder at each of the plurality of output steps. The method also includes predicting an end of utterance (EOU) token at an end of each utterance. The method also includes classifying each acoustic frame as either speech, initial silence, intermediate silence, or final silence.

Language agnostic multilingual end-to-end streaming on-device ASR system

An automated speech recognition (ASR) model includes a first encoder, a second encoder, and a decoder. The first encoder receives, as input, a sequence of acoustic frames, and generates, at each of a plurality of output steps, a first higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames. The second encoder receives, as input, the first higher order feature representation generated by the first encoder at each of the plurality of output steps, and generates, at each of the plurality of output steps, a second higher order feature representation for a corresponding first higher order feature frame. The decoder receives, as input, the second higher order feature representation generated by the second encoder at each of the plurality of output steps, and generates, at each of the plurality of time steps, a first probability distribution over possible speech recognition hypotheses.

Cascaded encoders for simplified streaming and non-streaming ASR

Implementations disclosed herein are directed to ephemeral learning of machine learning ('ML') model(s) based on gradient(s) generated at a remote system (e.g., remote server(s)). Processor(s) of the remote system can receive stream(s) of audio data capturing spoken utterance(s) from a client device of a user. A fulfillment pipeline can process the stream(s) of audio data to cause certain fulfillment(s) of the spoken utterance(s) to be performed. Meanwhile, a training pipeline can process the stream(s) of audio data to generate gradient(s) using unsupervised learning techniques. Subsequent to the processing by the fulfillment pipeline and/or the training pipeline, the stream(s) of audio data are discarded by the remote system. Accordingly, the ML model(s) can be trained at the remote system without storing or logging of the stream(s) of audio data by non-transient memory thereof, thereby providing more efficient training mechanisms for training the ML model(s) and also increasing security of user data.

Ephemeral learning of machine learning model(s)

A method includes receiving a sequence of acoustic frames and generating, by a first encoder, a first higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames. The method also includes generating, by a first pass transducer decoder, a first pass speech recognition hypothesis for a corresponding first higher order feature representation and generating, by a text encoder, a text encoding for a corresponding first pass speech recognition hypothesis. The method also includes generating, by a second encoder, a second higher order feature representation for a corresponding first higher order feature representation. The method also includes generating, by a second pass transducer decoder, a second pass speech recognition hypothesis using a corresponding second higher order feature representation and a corresponding text encoding.

Transducer-based streaming deliberation for cascaded encoders

Growing community of inventors

Sunnyvale, CA, United States of America

Trevor D Strohman