Microsoft Technology Licensing, LLC

Nanshan Zeng

Xuedong D Huang

Dimitrios Basile Dimitriadis

William Isaac Hinthorn

Andreas Stolcke

Zhuo Chen

Takuya Yoshioka

Lijuan Qin

Chenguang Zhu

Christopher Lovett

Yu Shi

Greg S Friedman

Jianfeng Gao

Liyang Lu

Kenichi Kumatani

Furu Wei

Shujie Liu

Jinchao Li

Xiujun Li

Nithyalakshmi Sampathkumar

Yu Wu

Chengyi Wang

Baolin Peng

Xuedong David Huang

Daniel Mikusik

Chunyuan Li

Ruochen Xu

Yao Qian

Dong Li

Jun Zhang

Ling Tony Chen

Wei Xiong

Meng Ye

Pengcheng He

Yashesh Gaur

Jinghua Chen

Xie Chen

Yining Chen

Lingfeng Wu

Dong Han Wang

Masaki Itagaki

Junwei Xu

Honghua Feng

Robert Peter Gmyr

Yu-chao Cao

Xiaofei Huang

Shayin Jing

Lin Song

Yingjun Yu

Systems and methods are provided for training and using a novel unified language foundation model. An encoder-decoder natural language model is obtained and various training data is obtained and used for training. The training process integrates a combination of replaced token detection, corrupted span reconstruction, and disentangled attention methodologies to produce a unified encoder-decoder model. The trained model is trained for performing both natural language understanding (NLU) tasks and natural language generation (NLG) tasks. Attention applied to the model is applied discretely to segmented chunks of encoded data during processing to improve the efficiency of applying attention by the model.

Pre-training a unified natural language model with corrupted span and replaced token detection

A speech module is joint trained with a knowledge module by transforming a first knowledge graph into an acoustic knowledge graph. The knowledge module is trained on the acoustic knowledge graph. Then, the knowledge module is integrated with the speech module to generate an integrated knowledge-speech module. In some instances, the speech module included in the integrated knowledge-speech module is aligned with a language module to generate an optimized speech model configured to leverage acoustic information and acoustic-based knowledge information, along with language information.

Generation of optimized spoken language understanding model through joint training with integrated acoustic knowledge-speech module

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

A system obtains a first training data set comprising labeled speech data or both labeled and unlabeled data corresponding to a high-resource data set as well as latent speech representations based on the first training data set. The system trains a machine learning model on the first training data set to learn phonetically aware speech representations corresponding to the first training data set. The system applies the latent speech representations to a transformer context network to generate contextual representations. The system aligns each of the contextual representations with a phoneme label to generate phonetically-aware contextual representations. The system causes a refinement engine to further refine the machine learning model.

Unified speech representation learning

The disclosure herein describes training a global model based on a plurality of data sets. The global model is applied to each data set of the plurality of data sets and a plurality of gradients is generated based on that application. At least one gradient quality metric is determined for each gradient of the plurality of gradients. Based on the determined gradient quality metrics of the plurality of gradients, a plurality of weight factors is calculated. The plurality of gradients is transformed into a plurality of weighted gradients based on the calculated plurality of weight factors and a global gradient is generated based on the plurality of weighted gradients. The global model is updated based on the global gradient, wherein the updated global model, when applied to a data set, performs a task based on the data set and provides model output based on performing the task.

Dynamic gradient aggregation for training neural networks

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

A transcription of audio speech included in electronic content associated with a meeting is created by an ASR model trained on speech-to-text data. The transcription is post-processed by modifying text included in the transcription, for example, by modifying punctuation, grammar, or formatting introduced by the ASR model and by changing or omitting one or more words that were included in both the audio speech and the transcription. After the transcription is post-processed, output based on the post-processed transcription is generated in the form of a meeting summary and/or template.

Automated meeting minutes generator

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

This document relates to machine learning. One example includes a method or technique that can be performed on a computing device. The method or technique can include obtaining a task-semantically-conditioned generative model that has been pretrained based at least on a first training data set having unlabeled training examples and semantically conditioned based at least on a second training data set having dialog act-labeled utterances. The method or technique can also include inputting dialog acts into the semantically-conditioned generative model and obtaining synthetic utterances that are output by the semantically-conditioned generative model. The method or technique can also include outputting the synthetic utterances.

Synthetic data generation for training of natural language understanding models

A language module is joint trained with a knowledge module for natural language understanding by aligning a first knowledge graph with a second knowledge graph. The knowledge module is trained on the aligned knowledge graphs. Then, the knowledge module is integrated with the language module to generate an integrated knowledge-language module.

Generation of optimized knowledge-based language model through knowledge graph multi-alignment

Systems and methods are provided for training a machine learning model to learn speech representations. Labeled speech data or both labeled and unlabeled data sets is applied to a feature extractor of a machine learning model to generate latent speech representations. The latent speech representations are applied to a quantizer to generate quantized latent speech representations and to a transformer context network to generate contextual representations. Each contextual representation included in the contextual representations is aligned with a phoneme label to generate phonetically-aware contextual representations. Quantized latent representations are aligned with phoneme labels to generate phonetically aware latent speech representations. Systems and methods also include randomly replacing a sub-set of the contextual representations with quantized latent speech representations during their alignments to phoneme labels and aligning the phonetically aware latent speech representations to the contextual representations using supervised learning.

Attributes of electronic content from a meeting are identified and evaluated to determine whether sub-portions of the electronic content should or should not be attributed to a user profile. Upon determining that the sub-portion should be attributed to a user profile, attributes of the sub-portion of electronic content are compared to attributes of stored user profiles. A probability that the sub-portion corresponds to at least one stored user profile is calculated. Based on the calculated probability, the sub-portion is attributed to a stored user profile or a guest user profile.

Automated meeting minutes generation service

This document relates to machine learning. One example includes a method or technique that can be performed on a computing device. The method or technique can include obtaining a task-adapted generative model that has been tuned using one or more task-specific seed examples. The method or technique can also include inputting dialog acts into the task-adapted generative model and obtaining synthetic utterances that are output by the task-adapted generative model. The method or technique can also include populating a synthetic training corpus with synthetic training examples that include the synthetic utterances. The synthetic training corpus may be suitable for training a natural language understanding model.

The present disclosure provides method and apparatus for processing a message. A statement sentence message and a message processing parameter associated with a user's session message are obtained. One or more first statement sentence nodes that are semantic-matched with the statement sentence message are determined in the knowledge map. One or more second statement sentence nodes corresponding to the message processing parameters are obtained from the knowledge map, based on the node relationship properties of the first statement sentence nodes. A response is generated based at least in part on statement sentences of the one or more second statement sentence nodes. The generated response is provided to the user.

Session message processing with generating responses based on node relationships within knowledge graphs

Growing community of inventors

Bellevue, WA, United States of America

Nanshan Zeng