Google Inc.

Ankur Bapna

Orhan Firat

Yonghui Wu

Zhifeng Chen

Niki J Parmar

Wolfgang Macherey

Melvin Jose Johnson Premkumar

Michael Schuster

Llion Owen Jones

Macduff Richard Hughes

George Foster

Xu Chen

Tara N Sainath

Yu Zhang

Bhuvana Ramabhadran

Larry Paul Heck

Bo Li

Andrew M Rosenberg

Khe Chai Sim

Thang Minh Luong

Zhehuai Chen

Anjuli Patricia Kannan

Yanping Huang

Maxim Krikun

Ye Tian

Arindrima Datta

Junwen Bai

Nikhil Siddhartha

Sneha Kudugunta

Dmitry Lepikhin

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for machine translation using neural networks. In some implementations, a text in one language is translated into a second language using a neural network model. The model can include an encoder neural network comprising a plurality of bidirectional recurrent neural network layers. The encoding vectors are processed using a multi-headed attention module configured to generate multiple attention context vectors for each encoding vector. A decoder neural network generates a sequence of decoder output vectors using the attention context vectors. The decoder output vectors can represent distributions over various language elements of the second language, allowing a translation of the text into the second language to be determined based on the sequence of decoder output vectors.

Machine translation using neural network models

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 receiving audio features and generating a latent speech representation based on the audio features. The method also includes generating a target quantized vector token and a target token index for a corresponding latent speech representation. The method also includes generating a contrastive context vector for a corresponding unmasked or masked latent speech representation and deriving a contrastive self-supervised loss based on the corresponding contrastive context vector and the corresponding target quantized vector token. The method also include generating a high-level context vector based on the contrastive context vector and, for each high-level context vector, learning to predict the target token index at the corresponding time step using a cross-entropy loss based on the target token index. The method also includes predicting speech recognition hypotheses for the utterance and training a multilingual automatic speech recognition (ASR) model using an unsupervised loss and a supervised loss.

Joint unsupervised and supervised training for multilingual ASR

Systems and methods for routing in mixture-of-expert models. In some aspects of the technology, a transformer may have at least one Mixture-of-Experts ('MoE') layer in each of its encoder and decoder, with the at least one MoE layer of the encoder having a learned gating function configured to route each token of a task to two or more selected expert feed-forward networks, and the at least one MoE layer of the decoder having a learned gating function configured to route each task to two or more selected expert feed-forward networks.

Systems and methods for routing within multitask mixture-of-experts models

Determining slot value(s) based on received natural language input and based on descriptor(s) for the slot(s). In some implementations, natural language input is received as part of human-to-automated assistant dialog. A natural language input embedding is generated based on token(s) of the natural language input. Further, descriptor embedding(s) are generated (or received), where each of the descriptor embeddings is generated based on descriptor(s) for a corresponding slot that is assigned to a domain indicated by the dialog. The natural language input embedding and the descriptor embedding(s) are applied to layer(s) of a neural network model to determine, for each of the slot(s), which token(s) of the natural language input correspond to the slot. A command is generated that includes slot value(s) for slot(s), where the slot value(s) for one or more of slot(s) are determined based on the token(s) determined to correspond to the slot(s).

Processing natural language using machine learning to determine slot values based on slot descriptors

A method of transcribing speech using a multilingual end-to-end (E2E) speech recognition model includes receiving audio data for an utterance spoken in a particular native language, obtaining a language vector identifying the particular language, and processing, using the multilingual E2E speech recognition model, the language vector and acoustic features derived from the audio data to generate a transcription for the utterance. The multilingual E2E speech recognition model includes a plurality of language-specific adaptor modules that include one or more adaptor modules specific to the particular native language and one or more other adaptor modules specific to at least one other native language different than the particular native language. The method also includes providing the transcription for output.

Large-scale multilingual speech recognition with a streaming end-to-end model

Adapters for neural machine translation systems. A method includes determining a set of similar n-grams that are similar to a source n-gram, and each similar n-gram and the source n-gram is in a first language; determining, for each n-gram in the set of similar n-grams, a target n-gram is a translation of the similar n-gram in the first language to the target n-gram in the second language; generating a source encoding of the source n-gram, and, for each target n-gram determined from the set of similar n-grams determined for the source n-gram, a target encoding of the target n-gram and a conditional source target memory that is an encoding of each of the target encodings; providing, as input to a first prediction model, the source encoding and the condition source target memory; and generating a predicted translation of the source n-gram from the first language to the second language.

Neural machine translation adaptation

Growing community of inventors

Sunnyvale, CA, United States of America

Ankur Bapna