Microsoft Technology Licensing, LLC

Mitsubishi Electric Research Laboratories, Inc.

Kaustubh Prakash Kalgaonkar

Yifan Gong

Rui Zhao

Hosam Adel Khalil

Bhiksha Ramakrishnan

Khuram Shahid

Emilian Y Stoimenov

Guoli Ye

Ivaylo Andreanov Enchev

Anthony Phillip Stark

Mahadevan Srinivasan

Alejandro Acero

Yan Huang

Michael Lewis Seltzer

Kshitiz Kumar

Chaojun Liu

Generally discussed herein are devices, systems, and methods for on-device detection of a wake word. A device can include a memory including model parameters that define a custom wake word detection model, the wake word detection model including a recurrent neural network transducer (RNNT) and a lookup table (LUT), the LUT indicating a hidden vector to be provided in response to a phoneme of a user-specified wake word, a microphone to capture audio, and processing circuitry to receive the audio from the microphone, determine, using the wake word detection model, whether the audio includes an utterance of the user-specified wake word, and wake up a personal assistant after determining the audio includes the utterance of the user-specified wake word.

On-device custom wake word detection

The technology described herein uses a modular model to process speech. A deep learning based acoustic model comprises a stack of different types of neural network layers. The sub-modules of a deep learning based acoustic model can be used to represent distinct non-phonetic acoustic factors, such as accent origins (e.g. native, non-native), speech channels (e.g. mobile, bluetooth, desktop etc.), speech application scenario (e.g. voice search, short message dictation etc.), and speaker variation (e.g. individual speakers or clustered speakers), etc. The technology described herein uses certain sub-modules in a first context and a second group of sub-modules in a second context.

Modular deep learning model

Described is a technology by which a speech recognizer is adapted to perform in noisy environments using linear spline interpolation to approximate the nonlinear relationship between clean speech, noise, and noisy speech. Linear spline parameters that minimize the error the between predicted noisy features and actual noisy features are learned from training data, along with variance data that reflect regression errors. Also described is compensating for linear channel distortion and updating noise and channel parameters during speech recognition decoding.

Acoustic model adaptation using splines

A method and system recognizes an unknown speaker by directing an ultrasonic signal at a face of the unknown speaker. A Doppler signal of the ultrasonic signal is acquired after reflection by the face, and Doppler features are extracted from the reflected Doppler signal. The Doppler features are classified using Doppler models storing the Doppler features and identities of known speakers to recognize and identify the unknown speaker.

Ultrasonic doppler sensor for speaker recognition

A method and system detect speech activity. An ultrasonic signal is directed at a face of a speaker over time. A Doppler signal of the ultrasonic signal is acquired after reflection by the face. Energy in the Doppler signal is measured over time. The energy over time is compared to a predetermined threshold to detect speech activity of the speaker in a concurrently acquired audio signal.

Ultrasonic Doppler sensor for speech-based user interface

Growing community of inventors

Atlanta, GA, United States of America

Kaustubh Prakash Kalgaonkar