Amazon Technologies, Inc.

Katrin Kirchhoff

Sravan Babu Bodapati

Monica Lakshmi Sunkara

Ashish Vishwanath Shenoy

Srikanth Ronanki

John Baker

Varun Sembium Varadarajan

Pu Paul Zhao

Anubhav Mishra

Jeffrey John Farris

Peng Zhang

Mohit Gupta

Harshal Pimpalkhute

Vivek Govindan

Arshit Gupta

Deepthi Devaiah Devanira

Dhanush Bekal Kannangola

Julian E S Salazar

Nilaksh Das

Yi Zhang

Rahul Gupta

Rashmi Gangadharaiah

Hassan Sawaf

Ankur Gandhe

Ganesh Kumar Gella

Rishita Rajal Anubhai

Roger Scott Jenke

Srikanth Vishnubhotla

Kalpesh N Sutaria

Asem Rustum

Yue Yang

Hadis Nosrati

Ravi Teja Gadde

Garima Lalwani

Bangrui Liu

Christopher Michael Hittner

Yilai Su

Anuj Gautam Surana

Yide Zou

Brandon Louis Mendez

Chengshun Zhang

Jinglun Cai

Silviu Mihai Burz

Nicholas G Aldridge

Goeric Sydney Huybrechts

Xilai Li

Chaitanya Shivade

Devang Kulshreshtha

Saket Dingliwal

Adel A Youssef

Mona Diab

Calvin Xuan Li

K P N V D S Siva Rama

Yibin Wang

Liangwei Guo

Akshay Nalwaya

Orestis Lykouropoulos

Abhinav Kumar

Sai Madhuri Bhavirisetty

Himalay Mohanlal Joriwal

Christian Egon Berkhoff Dossow

Techniques for customizable latency, from the customer's side, for automatic speech recognition (ASR) are described. In particular, the customer may specify a parameter that controls how fast or how slow the customer's media content will be streamed or processed. Slower processing means higher accuracy, with near real-time latency, while faster processing means lower accuracy, but offers much lower latency (e.g., less than 600 ms). Enabling tuning of the latency-versus-accuracy tradeoff of the ASR system offers customers the flexibility to meet varying needs for different ASR applications.

Customizable latency for automatic speech recognition

Language models may be dynamically updated for trending entities based on tuning data for particular users. A user may provide specific tuning data associated with trending entities within a class to generate a weight map for a language model. A class based model may be trained using the weight map specific for the user for the trending entities. Additionally, weights may be further boosted using a boosting language model to emphasize the trending entities.

Dynamic language model updates with boosting

Systems and methods are described for converting speech to text. In one aspect, a method for converting speech to text includes generating an initial transcript of an audio or video file. At least one named entity may be extracted from the initial transcript using entity recognition. A subset of nodes of a knowledge graph that include the at least one named entity may be selected, where the subset of nodes of the knowledge graph correspond to a set of named entities. The method may further include encoding the set of named entities to generate a set of entity embeddings. The speech in the audio or video file may then be decoded using the set of entity embeddings to produce a final transcript of the audio file.

Infusing knowledge graphs into automatic speech recognition

Bias may be removed from automatic speech recognition model predictions using internal language model estimates. Audio data may be received for speech recognition. The audio data may be processed both through an automatic speech recognition model to produce original word token predictions and masked in different portions of the audio data to produce other word token predictions for the masked audio. A comparison of the original word token predictions and the other word token predictions may provide an estimate of an internal language model for the automatic speech recognition model. This estimate can be used to modify the original word token predictions to remove the lexical bias and produce a speech prediction.

Removing bias from automatic speech recognition models using internal language model estimates

Techniques for at least the generation of a chatbot built from a custom vocabulary and to use runtime hints during inference are described. In some examples, the generation of the chatbot includes receiving a request to build a chatbot using a bot definition and a custom vocabulary, wherein the chatbot is to use runtime hints during usage; building the chatbot from the bot definition and custom vocabulary by at least: generating automatic speech recognition (ASR) artifacts to be used in decoding audio input into the chatbot into text for at least one other component of the chatbot to use in determining a next act to be performed, the ASR artifacts including artifacts that use the custom vocabulary and artifacts that do not use the custom vocabulary, and storing the ASR artifacts.

Dynamically selectable automated speech recognition using a custom vocabulary

Techniques for managing access to sensitive data in transcriptions are described. A method for managing access to sensitive data in transcriptions may include receiving a request to generate a redacted transcript of content, obtaining a transcript of the content, sending at least a portion of the transcript to a model endpoint to identify sensitive entities in the transcript, receiving an inference response identifying one or more sensitive entities in the transcript, and generating the redacted transcript based at least one the transcript and the inference response.

Managing access to sensitive data in transcriptions

A set of alternative vocal input styles for specifying a parameter of a dialog-driven application is determined. During execution of the application, an audio prompt requesting input in one of the styles is presented. A value of the parameter is determined by applying a collection of analysis tools to vocal input obtained after the prompt is presented. A task of the application is initiated using the value.

Dialog-driven applications supporting alternative vocal input styles

Techniques for personalized batch and streaming speech-to-text transcription of audio reduce the error rate of automatic speech recognition (ASR) systems in transcribing rare and out-of-vocabulary words. The techniques achieve personalization of connectionist temporal classification (CT) models by using adaptive boosting to perform biasing at the level of sub-words. In addition to boosting, the techniques encompass a phone alignment network to bias sub-word predictions towards rare long-tail words and out-of-vocabulary words. A technical benefit of the techniques is that the accuracy of speech-to-text transcription of rare and out-of-vocabulary words in a custom vocabulary by automatic speech recognition (ASR) system can be improved without having to train the ASR system on the custom vocabulary. Instead, the techniques allow the same ASR system trained on a base vocabulary to realize the accuracy improvements for different custom vocabularies spanning different domains.

Personalized batch and streaming speech-to-text transcription of audio

Portions of text data generated from inverse text normalization may be redacted. Text data for redaction may be obtained. One or more inverse text normalization models may be applied to the text data to generate normalized text data. A machine learning model, trained to recognize text for redaction, may be applied to identify portions of the normalized text data for redaction. The identified portions may be redacted and the redacted normalized text provided to a destination.

Redacting portions of text transcriptions generated from inverse text normalization

Domain-specific parameters may be used for tuning speech processing. A pre-trained transformer-based language model may train domain-specific parameters using domain-specific unlabeled text data. This domain-specific parameters can then be appended to candidate texts produced by a speech model on received speech data and input to the transformer-based language model to score the candidate texts. The scores of the candidate texts determined using the pre-trained transformer-based language model can then be used to select a candidate text for further speech processing.

Domain-specific parameter pre-fixes for tuning automatic speech recognition

Techniques for implementing a chatbot that utilizes context embeddings are described. An exemplary method includes determining a next turn by: applying a language model to the utterance to determine a probability of a sequence of words, generating a context embedding for the utterance based at least on one or more of: a dialog act as defined by a chatbot definition of the chatbot, a topic vector identifying a domain of the chatbot, a previous chatbot response, and one or more slot options; performing neural language model rescoring using the determined probability of a sequence of words as a word embedding and the generated context embedding to predict an hypothesis; determining at least a name of a slot and type to be fulfilled based at least in part on the hypothesis and the chatbot definition; and determining a next turn based at least in part on the chatbot definition, any previous state, and the name of the slot and type to be fulfilled.

Contextual biasing of neural language models using metadata from a natural language understanding component and embedded recent history

Techniques are described for a contextual natural language understanding (cNLU) framework that is able to incorporate contextual signals of variable history length to perform joint intent classification (IC) and slot labeling (SL) tasks. A user utterance provided by a user within a multi-turn chat dialog between the user and a conversational agent is received. The user utterance and contextual information associated with one or more previous turns of the multi-turn chat dialog is provided to a machine learning (ML) model. An intent classification and one or more slot labels for the user utterance are then obtained from the ML model. The cNLU framework described herein thus uses, in addition to a current utterance itself, various contextual signals as input to a model to generate IC and SL predictions for each utterance of a multi-turn chat dialog.

Contextual natural language understanding for conversational agents

Techniques for predicting punctuation and casing using multimodal fusion are described. An exemplary method includes processing generated text by: tokenizing the generated text into sub-words, and generating a sequence of lexical features for the sub-words using a pre-trained lexical encoder; processing audio of the audio by: generating a sequence of frame level acoustic embeddings using a pre-trained acoustic encoder on the audio, and generating task specific embeddings from the frame level acoustic embeddings; performing multimodal fusion of the sub-word level acoustic embeddings and the sequence of lexical features by: aligning the task specific embeddings to the sequence of lexical features, and combining the sequence of lexical features and aligned acoustic sequence; predicting punctuation and casing from the combined sequence of lexical features and aligned acoustic sequence; concatenating the sub-words of the text, and applying the predicted punctuation and casing; and outputting text having the predicted punctuation and casing.

Multimodal based punctuation and/or casing prediction

A spoken language understanding (SLU) system may include an automatic speech recognizer (ASR), an audio feature extractor, an optional synchronizer and a language understanding module. The ASR may produce a first set of input data representing transcripts of utterances. The audio feature extractor may produce a second set of input data representing audio features of the utterances, in particular, non-transcript specific characteristics of the speaker in one or more portions the utterances. The two sets of input data may be provided for the language understanding module to predict intents and slot labels for the utterances. The SLU system may use the optional synchronizer to align the two sets of input data before providing them to the language understanding module.

Multi-modal spoken language understanding systems

A transcription service may receive a request from a developer to build a custom speech-to-text model for a specific domain of speech. The custom speech-to-text model for the specific domain may replace a general speech-to-text model or add to a set of one or more speech-to-text models available for transcribing speech. The transcription service may receive a training data and instructions representing tasks. The transcription service may determine respective schedules for executing the instructions based at least in part on dependencies between the tasks. The transcription service may execute the instructions according to the respective schedules to train a speech-to-text model for a specific domain using the training data set. The transcription service may deploy the trained speech-to-text model as part of a network-accessible service for an end user to convert audio in the specific domain into texts.

Model training system for custom speech-to-text models

Growing community of inventors

Seattle, WA, United States of America

Katrin Kirchhoff