Microsoft Technology Licensing, LLC

Takuya Yoshioka

Zhuo Chen

Dimitrios Basile Dimitriadis

Xuedong D Huang

Nanshan Zeng

Andreas Stolcke

Lijuan Qin

William Isaac Hinthorn

Xiong Xiao

Xiaofei Wang

Zhong Meng

Yashesh Gaur

Naoyuki Kanda

Huaming Wang

Hakan Erdogan

Xuankai Chang

Jinyu Li

Changliang Liu

Sefik Emre Eskimez

Yifan Gong

Fileno A Alleva

Zhenghao Wang

Min Tang

Hassan Taherian

Hemin Yang

Zirun Zhu

A hypothesis stitcher for speech recognition of long-form audio provides superior performance, such as higher accuracy and reduced computational cost. An example disclosed operation includes: segmenting the audio stream into a plurality of audio segments; identifying a plurality of speakers within each of the plurality of audio segments; performing automatic speech recognition (ASR) on each of the plurality of audio segments to generate a plurality of short-segment hypotheses; merging at least a portion of the short-segment hypotheses into a first merged hypothesis set; inserting stitching symbols into the first merged hypothesis set, the stitching symbols including a window change (WC) symbol; and consolidating, with a network-based hypothesis stitcher, the first merged hypothesis set into a first consolidated hypothesis. Multiple variations are disclosed, including alignment-based stitchers and serialized stitchers, which may operate as speaker-specific stitchers or multi-speaker stitchers, and may further support multiple options for differing hypothesis configurations.

Hypothesis stitcher for speech recognition of long-form audio

A computer implemented method processes audio streams recorded during a meeting by a plurality of distributed devices. Operations include performing speech recognition on each audio stream by a corresponding speech recognition system to generate utterance-level posterior probabilities as hypotheses for each audio stream, aligning the hypotheses and formatting them as word confusion networks with associated word-level posteriors probabilities, performing speaker recognition on each audio stream by a speaker identification algorithm that generates a stream of speaker-attributed word hypotheses, formatting speaker hypotheses with associated speaker label posterior probabilities and speaker-attributed hypotheses for each audio stream as a speaker confusion network, aligning the word and speaker confusion networks from all audio streams to each other to merge the posterior probabilities and align word and speaker labels, and creating a best speaker-attributed word transcript by selecting the sequence of word and speaker labels with the highest posterior probabilities.

Speaker attributed transcript generation

Examples of array geometry agnostic multi-channel personalized speech enhancement (PSE) extract speaker embeddings, which represent acoustic characteristics of one or more target speakers, from target speaker enrollment data. Spatial features (e.g., inter-channel phase difference) are extracted from input audio captured by a microphone array. The input audio includes a mixture of speech data of the target speaker(s) and one or more interfering speaker(s). The input audio, the extracted speaker embeddings, and the extracted spatial features are provided to a trained geometry-agnostic PSE model. Output data is produced, which comprises estimated clean speech data of the target speaker(s) that has a reduction (or elimination) of speech data of the interfering speaker(s), without the trained PSE model requiring geometry information for the microphone array.

Array geometry agnostic multi-channel personalized speech enhancement

A computer implemented method includes receiving audio signals representative of speech via multiple audio streams transmitted from corresponding multiple distributed devices, performing, via a neural network model, continuous speech separation for one or more of the received audio signals having overlapped speech, and providing the separated speech on a fixed number of separate output audio channels.

Processing overlapping speech from distributed devices

The disclosure herein describes using a transcript generation model for generating a transcript from a multi-speaker audio stream. Audio data including overlapping speech of a plurality of speakers is obtained and a set of frame embeddings are generated from audio data frames of the obtained audio data using an audio data encoder. A set of words and channel change (CC) symbols are generated from the set of frame embeddings using a transcript generation model. The CC symbols are included between pairs of adjacent words that are spoken by different people at the same time. The set of words and CC symbols are transformed into a plurality of transcript lines, wherein words of the set of words are sorted into transcript lines based on the CC symbols, and a multi-speaker transcript is generated based on the plurality of transcript lines. The inclusion of CC symbols by the model enables efficient, accurate multi-speaker transcription.

Training and using a transcript generation model on a multi-speaker audio stream

Systems and methods are provided for generating and operating a speech enhancement model optimized for generating noise-suppressed speech outputs for improved human listening and live captioning. A computing system obtains a speech enhancement model trained on a first training dataset to generate noise-suppressed speech outputs and an automatic speech recognition model trained on a second training dataset to generate transcription labels for spoken language utterances. A third training dataset comprising a set of spoken language utterances is applied to the speech enhancement model to obtain a first noise-suppressed speech output which is applied to the automatic speech recognition model to generate a noise-suppressed transcription output for the set of spoken language utterances. Speech enhancement model parameters are updated to optimize the speech enhancement model to generate optimized noise-suppressed speech outputs based on a comparison of the noise-suppressed transcription output and ground truth transcription labels.

Systems and methods for human listening and live captioning

A computer implemented method includes receiving information streams on a meeting server from a set of multiple distributed devices included in a meeting, receiving audio signals representative of speech by at least two users in at least two of the information streams, receiving at least one video signal of at least one user in the information streams, associating a specific user with speech in the received audio signals as a function of the received audio and video signals, and generating a transcript of the meeting with an indication of the specific user associated with the speech.

Audio-visual diarization to identify meeting attendees

A computer implemented method includes receiving audio streams at a meeting server from two distributed devices that are streaming audio captured during an ad-hoc meeting between at least two users, comparing the received audio streams to determine that the received audio streams are representative of sound from the ad-hoc meeting, generating a meeting instance to process the audio streams in response to the comparing determining that the audio streams are representative of sound from the ad-hoc meeting, and processing the received audio streams to generate a transcript of the ad-hoc meeting.

Distributed device meeting initiation

A system and method include reception of a first plurality of audio signals, generation of a second plurality of beamformed audio signals based on the first plurality of audio signals, each of the second plurality of beamformed audio signals associated with a respective one of a second plurality of beamformer directions, generation of a first TF mask for a first output channel based on the first plurality of audio signals, determination of a first beamformer direction associated with a first target sound source based on the first TF mask, generation of first features based on the first beamformer direction and the first plurality of audio signals, determination of a second TF mask based on the first features, and application of the second TF mask to one of the second plurality of beamformed audio signals associated with the first beamformer direction.

Low-latency speech separation

Systems and methods for providing customized output based on a user preference in a distributed system are provided. In example embodiments, a meeting server or system receives audio streams from a plurality of distributed devices involved in an intelligent meeting. The meeting system identifies a user corresponding to a distributed device of the plurality of distributed devices and determines a preferred language of the user. A transcript from the received audio streams is generated. The meeting system translates the transcript into the preferred language of the user to form a translated transcript. The translated transcript is provided to the distributed device of the user.

Customized output to optimize for user preference in a distributed system

This document relates to separation of audio signals into speaker-specific signals. One example obtains features reflecting mixed speech signals captured by multiple microphones. The features can be input a neural network and masks can be obtained from the neural network. The masks can be applied one or more of the mixed speech signals captured by one or more of the microphones to obtain two or more separate speaker-specific speech signals, which can then be output.

Growing community of inventors

Bellevue, WA, United States of America

Takuya Yoshioka