International Business Machines Corporation

Ohio State Innovation Foundation

George Andrei Saon

Brian E D Kingsbury

Samuel Jonathan Thomas

Michael Alan Picheny

Hong-Kwang Jeff Kuo

Dimitrios B Dimitriadis

Gakuto Kurata

Mukund Padmanabhan

David C Haws

Zoltan Tueske

Tara N Sainath

Ramesh Ambat Gopinath

Xiaodong Cui

Lidia Mangu

Kenneth W Church

Hagen Soltau

Petr Fousek

Miroslav Novak

Geoffrey G Zweig

Dimitri Kanevsky

Erik Rueger

David Nahamoo

Sachindra Joshi

Sasha P Caskey

Sreeram Viswanath Balakrishnan

Stanley S Chen

Kartik Audhkhasi

Jatin Ganhotra

Benoit Maison

Nicolò Sgobba

Daniel Bolanos

Yuging Gao

Vishal Sunder

Eric Fosler-Lussier

Brian Kingsbury

Scott Elliot Axelrod

Antonello Izzi

One or more systems, devices, computer program products and/or computer-implemented methods of use provided herein relate to length perturbation techniques for improving generalization of DNN acoustic models. A computer-implemented system can comprise a memory that can store computer executable components. The computer-implemented system can further comprise a processor that can execute the computer executable components stored in the memory, wherein the computer executable components can comprise a frame skipping component that can remove one or more frames from an acoustic utterance via frame skipping. The computer executable components can further comprise a frame insertion component that can insert one or more replacement frames into the acoustic utterance via frame insertion to replace the one or more frames with the one or more replacement frames to enable length perturbation of the acoustic utterance.

Length perturbation techniques for improving generalization of deep neural network acoustic models

Systems, computer-implemented methods, and computer program products to facilitate fine-grained textual knowledge transfer to improve speech recognition and understanding are provided. According to an embodiment, a system can comprise a processor that executes components stored in memory. The computer executable components comprise deriving component that can derive one or more speech-based embeddings from an utterance via a speech encoder. The computer executable components can comprise a cross-attention component that can align, at a token level, one or more large language model (LLM) based sentence embeddings with the one or more speech-based embeddings. The computer executable components can comprise a loss component that can combine an alignment loss and an automatic speech recognition (ASR) loss.

Textual knowledge transfer for improved speech recognition and understanding

Features of two or more single speaker utterances are concatenated together and corresponding labels of the two or more single speaker utterances are concatenated together. Single speaker acoustic embeddings for each of the single speaker utterances of the concatenated single speaker utterances are generated using a single speaker teacher encoder network. An enhanced model is trained on the concatenated single speaker utterances using a classification loss Land a representation similarity loss L, the representation similarity loss Ldefined to influence an embedding derived from the concatenated single speaker utterances, the influence being based on the single speaker acoustic embeddings derived from the single speaker teacher encoder network.

Multi-speaker data augmentation for improved end-to-end automatic speech recognition

A computer-implemented method is provided for model training. The method includes training a second end-to-end neural speech recognition model that has a bidirectional encoder to output same symbols from an output probability lattice of the second end-to-end neural speech recognition model as from an output probability lattice of a trained first end-to-end neural speech recognition model having a unidirectional encoder. The method also includes building a third end-to-end neural speech recognition model that has a unidirectional encoder by training the third end-to-end neural speech recognition model as a student by using the trained second end-to-end neural speech recognition model as a teacher in a knowledge distillation method.

Accuracy of streaming RNN transducer

Mechanisms are provided for performing machine learning training of a computer model. A perturbation generator generates a modified training data comprising perturbations injected into original training data, where the perturbations cause a data corruption of the original training data. The modified training data is input into a prediction network of the computer model and processing the modified training data through the prediction network to generate a prediction output. Machine learning training is executed of the prediction network based on the prediction output and the original training data to generate a trained prediction network of a trained computer model. The trained computer model is deployed to an artificial intelligence computing system for performance of an inference operation.

Reducing exposure bias in machine learning training of sequence-to-sequence transducers

Audio signals representing a current utterance in a conversation and a dialog history including at least information associated with past utterances corresponding to the current utterance in the conversation can be received. The dialog history can be encoded into an embedding. A spoken language understanding neural network model can be trained to perform a spoken language understanding task based on input features including at least speech features associated with the received audio signals and the embedding. An encoder can also be trained to encode a given dialog history into an embedding. The spoken language understanding task can include predicting a dialog action of an utterance. The spoken language understanding task can include predicting a dialog intent or overall topic of the conversation.

Integrating dialog history into end-to-end spoken language understanding systems

Training data can be received, which can include pairs of speech and meaning representation associated with the speech as ground truth data. The meaning representation includes at least semantic entities associated with the speech, where the spoken order of the semantic entities is unknown. The semantic entities of the meaning representation in the training data can be reordered into spoken order of the associated speech using an alignment technique. A spoken language understanding machine learning model can be trained using the pairs of speech and meaning representation having the reordered semantic entities. The meaning representation, e.g., semantic entities, in the received training data can be perturbed to create random order sequence variations of the semantic entities associated with speech. Perturbed meaning representation with associated speech can augment the training data.

Training end-to-end spoken language understanding systems with unordered entities

A computer-implemented method is provided for improving accuracy recognition of digital speech. The method includes receiving the digital speech. The method further includes splitting the digital speech into overlapping chunks. The method also includes computing a bidirectional encoder embedding of each of the overlapping chunks to obtain bidirectional encoder embeddings. The method additionally includes combining the bidirectional encoder embeddings. The method further includes interpreting, by a speech recognition system, the digital speech using the combined bidirectional encoder embeddings.

Chunking and overlap decoding strategy for streaming RNN transducers for speech recognition

A computer-implemented method for customizing a recurrent neural network transducer (RNN-T) is provided. The computer implemented method includes synthesizing first domain audio data from first domain text data, and feeding the synthesized first domain audio data into a trained encoder of the recurrent neural network transducer (RNN-T) having an initial condition, wherein the encoder is updated using the synthesized first domain audio data and the first domain text data. The computer implemented method further includes synthesizing second domain audio data from second domain text data, and feeding the synthesized second domain audio data into the updated encoder of the recurrent neural network transducer (RNN-T), wherein the prediction network is updated using the synthesized second domain audio data and the second domain text data. The computer implemented method further includes restoring the updated encoder to the initial condition.

Customization of recurrent neural network transducers for speech recognition

A processor-implemented method trains an automatic speech recognition system using speech data and text data. A computer device receives speech data, and generates a spectrogram based on the speech data. The computing device receives text data associated with an entire corpus of text data, and generates a textogram based upon the text data. The computing device trains an automatic speech recognition system using the spectrogram and the textogram.

Integrating text inputs for training and adapting neural network transducer ASR models

Using an encoder neural network model, an encoder vector is computed, the encoder vector comprising a vector representation of a current portion of input data in an input sequence. Using a prediction neural network model, a prediction vector is predicted, the prediction performed using a previous prediction vector and a previous output symbol corresponding to a previous portion of input data in the input sequence. Using a joint neural network model, a joint vector corresponding to the encoder vector and the prediction vector is computed, the joint vector multiplicatively combining each element of the encoder vector with a corresponding element of the prediction vector. Using a softmax function, the joint vector is converted to a probability distribution comprising a probability that a current output symbol corresponds to the current portion of input data in the input sequence.

Multiplicative integration in neural network transducer models for end-to-end speech recognition

In an approach to soft-forgetting training, one or more computer processors train a first model utilizing one or more training batches wherein each training batch of the one or more training batches comprises one or more blocks of information. The one or more computer processors, responsive to a completion of the training of the first model, initiate a training of a second model utilizing the one or more training batches. The one or more computer processors jitter a random block size for each block of information for each of the one or more training batches for the second model. The one or more computer processors unroll the second model over one or more non-overlapping contiguous jittered blocks of information. The one or more computer processors, responsive to the unrolling of the second model, reduce overfitting for the second model by applying twin regularization.

Soft-forgetting for connectionist temporal classification based automatic speech recognition

Techniques for augmenting the output of generally available speech-to-text systems using local profiles are presented. An example method includes receiving an audio recording of a natural language command. The received audio recording of the natural language command is transmitted to a speech-to-text system, and a text string generated from the audio recording is received from the speech-to-text system. The text string is corrected based on a local profile mapping incorrectly transcribed words from the speech-to-text system to corrected words. A function in a software application is invoked based on the corrected text string.

Vocal recognition using generally available speech-to-text systems and user-defined vocal training

An approach is provided that receives an audio stream and utilizes a voice activation detection (VAD) process to create a digital audio stream of voices from at least two different speakers. An automatic speech recognition (ASR) process is applied to the digital stream with the ASR process resulting in the spoken words to which a speaker turn detection (STD) process is applied to identify a number of speaker segments with each speaker segment ending at a word boundary. A speaker clustering algorithm is then applied to the speaker segments to associate one of the speakers with each of the speaker segments.

Growing community of inventors

Stamford, CT, United States of America

George Andrei Saon