Model shrinking for embedded keyword spotting

Patent metadata

Field	Value
Publication number	US-9600231-B1
Application number	US-201514751975-A
Country	US
Kind code	B1
Filing date	Jun 26, 2015
Priority date	Mar 13, 2015
Publication date	Mar 21, 2017
Grant date	Mar 21, 2017

Abstract

A revised support vector machine (SVM) classifier is offered to distinguish between true keywords and false positives based on output from a keyword spotting component of a speech recognition system. The SVM operates on a reduced set of feature dimensions, where the feature dimensions are selected based on their ability to distinguish between true keywords and false positives. Further, support vectors pairs are merged to create a reduced set of re-weighted support vectors. These techniques result in an SVM that may be operated using reduced computing resources, thus improving system performance.

First claim

What is claimed is:

1. A computer-implemented method for detecting a wakeword in a spoken utterance using a trained classifier, the method comprising:
determining, for each of a first plurality of audio feature dimensions, a respective utility of each audio feature dimension for classifying audio data as including or not including a wakeword;
determining a second plurality of audio feature dimensions using the respective utilities, wherein the second plurality of audio feature dimensions is a subset of the first plurality of audio feature dimensions;
configuring a first support vector machine (SVM) classifier to classify whether input audio data includes a wakeword, the first SVM classifier comprising a first plurality of support vectors, each of the first plurality of support vectors including a value for each of the second plurality of audio feature dimensions;
determining a positive pair of support vectors in the first plurality of support vectors, each support vector of the positive pair of support vectors being on a positive side of a binary classification of the first SVM classifier;
combining the positive pair of support vectors into a first combined support vector;
determining a negative pair of support vectors in the first plurality of support vectors, each support vector of the negative pair of support vectors being on a negative side of the binary classification of the first SVM classifier;
combining the negative pair of support vectors into a second combined support vector;
creating additional combined support vectors by alternating between combining a positive pair of support vectors and combining a negative pair of support vectors;
determining a second plurality of support vectors including at least the first combined support vector, the second combined support vector, and the additional combined support vectors;
configuring a second SVM classifier comprising the second plurality of support vectors; and
determining, using the second SVM classifier, that first audio data includes a representation of a first wakeword.

2. The computer-implemented method of claim 1 , wherein the determining the second plurality of audio feature dimensions comprises:
determining a plurality of feature vectors from a plurality of training audio data samples, wherein each feature vector includes a value for each of the first plurality of audio feature dimensions;
determining a threshold utility in identifying a wakeword in audio data;
determining, for a first audio feature dimension, a first utility metric for the audio feature dimension for classifying audio data as including or not including a wakeword;
determining that the first utility metric is above the threshold utility; and
including the first audio feature dimension in the second plurality of audio feature dimensions.

3. The computer-implemented method of claim 1 , wherein the determining the second plurality of support vectors comprises:
combining the first combined support vector and the second combined support vector into a third combined support vector by:
mapping the first combined support vector to high dimensional space to obtain a first mapped support vector;
mapping the second combined support vector to high dimensional space to obtain a second mapped support vector;
multiplying the first mapped support vector by a first weight to determine a first weighted mapped support vector;
multiplying the second mapped support vector by a second weight to determine a second weighted mapped support vector;
adding the first weighted mapped support vector to the second weighted mapped support vector to determine a merged support vector; and
mapping the merged support vector to the feature space to determine the third combined support vector.

4. The method of claim 3 , wherein the determining the second plurality of support vectors further comprises:
determining a third weight for the third combined support vector using the first weight and the second weight;
determining a target number of support vectors;
determining a number of merged support vectors that matches the target number;
after determining the number of merged support vectors that matches the target number, determining a revised weight for the third combined support vector;
determining a weighted third combined support vector using the third combined support vector and the revised weight; and
including the weighted third combined support vector into the second plurality of support vectors.

5. A computer-implemented method comprising:
determining a first plurality of feature dimensions, wherein the determining is based on a respective utility of each feature dimension for classifying whether audio data includes a keyword;
configuring a first support vector machine (SVM) classifier, the first SVM classifier comprising a first plurality of support vectors, each of the first plurality of support vectors including a value for each of the first plurality of feature dimensions;
combining a positive pair of support vectors in the first plurality of support vectors into a first combined support vector, each support vector of the positive pair of support vectors being on a positive side of a binary classification of the first SVM classifier;
combining a negative pair of support vectors in the first plurality of support vectors into a second combined support vector, each support vector of the negative pair of support vectors being on a negative side of the binary classification of the first SVM classifier;
creating additional combined support vectors by alternating between combining a positive pair of support vectors and combining a negative pair of support vectors;
determining a second plurality of support vectors including at least the first combined support vector, the second combined support vector, and the additional combined support vectors; and
configuring a second SVM classifier comprising the second plurality of support vectors.

6. The computer-implemented method of claim 5 , further comprising:
determining a first value of a first feature dimension, the first value associated with first audio data including a first keyword;
determining a second value of the first feature dimension, the second value associated with second audio data not including a keyword; and
determining a first utility metric for the first feature dimension using the first value and the second value.

7. The computer-implemented method of claim 6 , further comprising:
processing a plurality of training audio data samples to determine a plurality of feature vectors, wherein each feature vector includes a value for each of a second plurality of feature dimensions and wherein the first plurality of feature dimensions is a subset of the second plurality of feature dimensions and the second plurality of feature dimensions includes the first feature dimension;
determining a threshold utility;
determining the first utility metric is above the threshold utility; and
including the first feature dimension in the first plurality of feature dimensions.

8. The computer-implemented method of claim 5 , wherein configuring the first SVM classifier comprises:
determining a pair of support vectors from the second plurality of support vectors, wherein the pair comprises the first combined support vector and the second combined support vector and wherein the first plurality of support vectors is a subset of the second plurality of support vectors;
merging the first combined support vector and the second combined support vector into a third combined support vector; and
including the third combined support vector in the first plurality of support vectors.

9. The computer-implemented method of claim 5 , further comprising

Assignees

• Amazon Tech Inc

Inventors

• Sun Ming
• Hoffmeister Björn
• Vitaladevuni Shiv Naga Prasad
• Nagaraja Varun Kumar

Classifications

• G10L15/14

  using statistical models, e.g. Hidden Markov Models [HMMs] (G10L15/18 takes precedence) · CPC title

• G10L15/18

  using natural language modelling · CPC title

• G10L15/22

  Procedures used during a speech recognition process, e.g. man-machine dialogue · CPC title

• G10L2015/088

  Word spotting · CPC title

• G06F3/167 Primary

  Audio in a user interface, e.g. using voice commands for navigating, audio feedback · CPC title