International Business Machines Corporation

Google Inc.

Nuance Communications, Inc.

Bhuvana Ramabhadran

Dimitri Kanevsky

Andrew M Rosenberg

Abhinav Sethy

Pedro Jose Moreno Mengibar

Sara H Basson

Edward Emile Kelley

Tara N Sainath

Yu Zhang

Zhehuai Chen

David Nahamoo

Stanley S Chen

Takashi Fukuda

Gakuto Kurata

Samuel Jonathan Thomas

Masayuki A Suzuki

Michael Alan Picheny

Osamu Ichikawa

Kenneth W Church

Vaibhava Goel

Etienne Marcheret

Tongzhou Chen

Irina Rish

Jiri Navratil

Paul Joseph Vozila

Lidia Mangu

Neeraj Gaur

Leonid Rachevsky

Laurence P Sansone

Yonghui Wu

Ryuki Tachibana

Ron J Weiss

Fadi Biadsy

Byungha Chun

Kartik Audhkhasi

Ehsan Variani

Ruhi Sarikaya

Parisa Haghani

Dan Jiang

Vijayaditya Peddinti

Jesse Emond

Ebru Arisoy

Clifford A Pickover

Guillermo A Cecchi

Zhifeng Chen

Yuqing Gao

Brian E D Kingsbury

Ron Hoory

Andrew Aaron

Ye Jia

Russell John Wyatt Skerry-Ryan

Jonathan Mamou

Nathan M Bodenstab

Brian Reginald Heasman

Yinghui Huang

Brian E Roark

Rick Hamilton

Katya Scheinberg

Mateusz Golebiewski

Yuan Wang

Rick A Hamilton, Ii

Chandrasekhar Narayanaswami

Mahesh Viswanathan

Rohit Prakash Prabhavalkar

Raimo Bakis

Trevor D Strohman

Chengjun Julian Chen

Abraham P Ittycheriah

Salim Estephan Roukos

Osamuyimen Stewart

I-Hsin Chung

Paul Gregory Crumley

Stephen Mingyu Chu

Peter V De Souza

Roberto Sicconi

Sameer Maskey

Real Tremblay

Narges Bani Asadi

Ankur Bapna

David Rybach

Petr Fousek

Peter Thomas Brunet

Weiran Wang

Maximilian Bisani

Yangzhang He

Hakan Erdogan

Yuging Gao

Charles Caleb Peyser

Ronny Huang

Rèal Tremblay

Brian Kingsbury

Yi Su

Min Ma

Arindrima Datta

Shay Maymon

Sepand Mavandadi

David Eubensky

Abihnav Sethy

Daniel Povey

A method of guided data selection for masked speech modeling includes obtaining a sequence of encoded representations corresponding to an utterance. For each respective encoded representation, the method includes processing the respective encoded representation to generate a corresponding probability distribution over possible speech recognition hypotheses and assigning, to the respective encode representation, a confidence score as a highest probability from the corresponding probability distribution over possible speech recognition hypotheses. The method also includes selecting a set of unmasked encoded representations to mask based on the confidence scores assigned to the sequence of encoded representations. The method also includes generating a set of masked encoded representations by masking the selected set of unmasked encoded representations. Here, each masked encoded representation in the set of masked encoded representations corresponds to a respective one of the unmasked encoded representations in the selected set of unmasked encoded representations.

Guided data selection for masked speech modeling based on an average score assigned to encoded representations of an utterance

A method for instantaneous learning in text-to-speech (TTS) during dialog includes receiving a user pronunciation of a particular word present in a query spoken by a user. The method also includes receiving a TTS pronunciation of the same particular word that is present in a TTS input where the TTS pronunciation of the particular word is different than the user pronunciation of the particular word. The method also includes obtaining user pronunciation-related features and TTS pronunciation related features associated with the particular word. The method also includes generating a pronunciation decision selecting one of the user pronunciation or the TTS pronunciation of the particular word that is associated with a highest confidence. The method also include providing the TTS audio that includes a synthesized speech representation of the response to the query using the user pronunciation or the TTS pronunciation for the particular word.

Instantaneous learning in text-to-speech during dialog

A method for distilling one or more trained teacher automatic speech recognition (ASR) models into a multilingual student model includes receiving a plurality of teacher training examples and a plurality of student training examples. The method also includes training one or more teacher automatic speech recognition (ASR) models using the plurality of teacher training examples. Each teacher ASR model is configured to output a respective textual representation of a respective audio input. The method further includes generating a multi-lingual student ASR model by training the multi-lingual student ASR model using the plurality of student training examples and distilling the trained one or more teacher ASR models into the multilingual student ASR model using a tunable distillation loss weight. Each student ASR model is configured to receive an audio input and output a corresponding textual representation of the received audio input.

Self-adaptive distillation

A method includes receiving training data that includes unspoken textual utterances in a target language. Each unspoken textual utterance not paired with any corresponding spoken utterance of non-synthetic speech. The method also includes generating a corresponding alignment output for each unspoken textual utterance using an alignment model trained on transcribed speech utterance in one or more training languages each different than the target language. The method also includes generating a corresponding encoded textual representation for each alignment output using a text encoder and training a speech recognition model on the encoded textual representations generated for the alignment outputs. Training the speech recognition model teaches the speech recognition model to learn how to recognize speech in the target language.

Using aligned text and speech representations to train automatic speech recognition models without transcribed speech data

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

A method includes receiving training data that includes unspoken text utterances, un-transcribed non-synthetic speech utterances, and transcribed non-synthetic speech utterances. Each unspoken text utterance is not paired with any corresponding spoken utterance of non-synthetic speech. Each un-transcribed non-synthetic speech utterance is not paired with a corresponding transcription. Each transcribed non-synthetic speech utterance is paired with a corresponding transcription. The method also includes generating a corresponding synthetic speech representation for each unspoken textual utterance of the received training data using a text-to-speech model. The method also includes pre-training an audio encoder on the synthetic speech representations generated for the unspoken textual utterances, the un-transcribed non-synthetic speech utterances, and the transcribed non-synthetic speech utterances to teach the audio encoder to jointly learn shared speech and text representations.

Advancing the use of text and speech in ASR pretraining with consistency and contrastive losses

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 includes receiving a sequence of acoustic frames extracted from audio data corresponding to an utterance. During a first pass, the method includes processing the sequence of acoustic frames to generate N candidate hypotheses for the utterance. During a second pass, and for each candidate hypothesis, the method includes: generating a respective un-normalized likelihood score; generating a respective external language model score; generating a standalone score that models prior statistics of the corresponding candidate hypothesis; and generating a respective overall score for the candidate hypothesis based on the un-normalized likelihood score, the external language model score, and the standalone score. The method also includes selecting the candidate hypothesis having the highest respective overall score from among the N candidate hypotheses as a final transcription of the utterance.

Multilingual re-scoring models for automatic speech recognition

A method includes receiving audio data corresponding to an utterance and generating a pair of positive audio data examples. Here, each positive audio data example includes a respective augmented copy of the received audio data. For each respective positive audio data example, the method includes generating a respective sequence of encoder outputs and projecting the respective sequence of encoder outputs for the positive data example into a contrastive loss space. The method also includes determining a L2 distance between each corresponding encoder output in the projected sequences of encoder outputs for the positive audio data examples and determining a per-utterance consistency loss by averaging the L2 distances. The method also includes generating corresponding speech recognition results for each respective positive audio data example. The method also includes updating parameters of the speech recognition model based on a respective supervised loss term and the per-utterance consistency loss.

Supervised and unsupervised training with contrastive loss over sequences

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 training data that includes unspoken text utterances and un-transcribed non-synthetic speech utterances. Each unspoken text utterance is not paired with any corresponding spoken utterance of non-synthetic speech. Each un-transcribed non-synthetic speech utterance is not paired with a corresponding transcription. The method also includes generating a corresponding synthetic speech representation for each unspoken textual utterance of the received training data using a text-to-speech model. The method also includes pre-training an audio encoder on the synthetic speech representations generated for the unspoken textual utterances and the un-transcribed non-synthetic speech utterances to teach the audio encoder to jointly learn shared speech and text representations.

Injecting text in self-supervised speech pre-training

A method for an automated speech recognition (ASR) model for unifying streaming and non-streaming speech recognition including receiving a sequence of acoustic frames. The method includes generating, using an audio encoder of an automatic speech recognition (ASR) model, a higher order feature representation for a corresponding acoustic frame in the sequence of acoustic frames. The method further includes generating, using a joint encoder of the ASR model, a probability distribution over possible speech recognition hypothesis at the corresponding time step based on the higher order feature representation generated by the audio encoder at the corresponding time step. The audio encoder comprises a neural network that applies mixture model (MiMo) attention to compute an attention probability distribution function (PDF) using a set of mixture components of softmaxes over a context window.

Mixture model attention for flexible streaming and non-streaming automatic speech recognition

A method includes receiving an input text sequence to be synthesized into speech in a first language and obtaining a speaker embedding, the speaker embedding specifying specific voice characteristics of a target speaker for synthesizing the input text sequence into speech that clones a voice of the target speaker. The target speaker includes a native speaker of a second language different than the first language. The method also includes generating, using a text-to-speech (TTS) model, an output audio feature representation of the input text by processing the input text sequence and the speaker embedding. The output audio feature representation includes the voice characteristics of the target speaker specified by the speaker embedding.

Multilingual speech synthesis and cross-language voice cloning

A method for subword segmentation includes receiving an input word to be segmented into a plurality of subword units. The method also includes executing a subword segmentation routine to segment the input word into a plurality of subword units by accessing a trained vocabulary set of subword units and selecting the plurality of subword units from the input word by greedily finding a longest subword unit from the input word that is present in the trained vocabulary set until an end of the input word is reached.

Regularizing word segmentation

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

Growing community of inventors

Mount Kisco, NY, United States of America

Bhuvana Ramabhadran