公开(公告)号：US10229356B1

公开(公告)日：2019-03-12

申请号：US14581969

申请日：2014-12-23

Applicant: Amazon Technologies, Inc.

Inventor： Baiyang Liu ,  Michael Reese Bastian ,  Bjorn Hoffmeister ,  Sankaran Panchapagesan ,  Ariya Rastrow

IPC: G06N3/08

Abstract: Features are disclosed for error tolerant model compression. Such features could be used to reduce the size of a deep neural network model including several hidden node layers. The size reduction in an error tolerant fashion ensures predictive applications relying on the model do not experience performance degradation due to model compression. Such predictive applications include automatic recognition of speech, image recognition, and recommendation engines. Partially quantized models are re-trained such that any degradation of accuracy is “trained out” of the model providing improved error tolerance with compression.

公开(公告)号：US10304440B1

公开(公告)日：2019-05-28

申请号：US15198578

申请日：2016-06-30

Applicant: Amazon Technologies, Inc.

Inventor： Sankaran Panchapagesan ,  Bjorn Hoffmeister ,  Arindam Mandal ,  Aparna Khare ,  Shiv Naga Prasad Vitaladevuni ,  Spyridon Matsoukas ,  Ming Sun

IPC: G10L15/06 ,  G10L15/08 ,  G10L15/14 ,  G10L15/16 ,  G10L15/28

Abstract: An approach to keyword spotting makes use of acoustic parameters that are trained on a keyword spotting task as well as on a second speech recognition task, for example, a large vocabulary continuous speech recognition task. The parameters may be optimized according to a weighted measure that weighs the keyword spotting task more highly than the other task, and that weighs utterances of a keyword more highly than utterances of other speech. In some applications, a keyword spotter configured with the acoustic parameters is used for trigger or wake word detection.

公开(公告)号：US10964315B1

公开(公告)日：2021-03-30

申请号：US15639330

申请日：2017-06-30

Applicant: Amazon Technologies, Inc.

Inventor： Minhua Wu ,  Sankaran Panchapagesan ,  Ming Sun ,  Shiv Naga Prasad Vitaladevuni ,  Bjorn Hoffmeister ,  Ryan Paul Thomas ,  Arindam Mandal

IPC: G10L15/22 ,  G10L15/14 ,  G10L15/02 ,  G10L15/16 ,  G10L25/30 ,  G10L15/08

Abstract: An approach to wakeword detection uses an explicit representation of non-wakeword speech in the form of subword (e.g., phonetic monophone) units that do not necessarily occur in the wakeword and that broadly represent general speech. These subword units are arranged in a “background” model, which at runtime essentially competes with the wakeword model such that a wakeword is less likely to be declare as occurring when the input matches that background model well. An HMM may be used with the model to locate possible occurrences of the wakeword. Features are determined from portions of the input corresponding to subword units of the wakeword detected using the HMM. A secondary classifier is then used to process the features to yield a decision of whether the wakeword occurred.

公开(公告)号：US10147442B1

公开(公告)日：2018-12-04

申请号：US14869803

申请日：2015-09-29

Applicant: Amazon Technologies, Inc.

Inventor： Sankaran Panchapagesan ,  Shiva Kumar Sundaram ,  Arindam Mandal

IPC: G10L25/30 ,  G10L15/02 ,  G10L15/06 ,  G10L25/12 ,  G10L15/20

Abstract: A neural network acoustic model is trained to be robust and produce accurate output when used to process speech signals having acoustic interference. The neural network acoustic model can be trained using a source-separation process by which, in addition to producing the main acoustic model output for a given input, the neural network generates predictions of the separate speech and interference portions of the input. The parameters of the neural network can be adjusted to jointly optimize all three outputs (e.g., the main acoustic model output, the speech signal prediction, and the interference signal prediction), rather than only optimizing the main acoustic model output. Once trained, output layers for the speech and interference signal predictions can be removed from the neural network or otherwise disabled.