Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

Method for Making Confusion Matrix by DP Matching

IP.com Disclosure Number: IPCOM000060547D
Original Publication Date: 1986-Apr-01
Included in the Prior Art Database: 2005-Mar-08
Document File: 2 page(s) / 45K

Publishing Venue

IBM

Related People

Sugawara, K: AUTHOR

Abstract

The technique disclosed in this article is a method for making a label confusion matrix for smoothing the label output distribution obtained from small size training data. A biased distribution of labels degrades the accuracy of the speech recognition technique which uses a labeled sequence. A technique for smoothing label distribution using a label confusion matrix has been proposed to resolve the problem. This article shows the way to obtain the confusion matrix based on the mutual information among label prototypes computed through DP matching. The matrix is computed as follows (see figure). Select a pair of utterances of the same word in the training data. Convert them into sequences of labels and carry out the labeled DP matching based on a certain distance measure.

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 100% of the total text.

Page 1 of 2

Method for Making Confusion Matrix by DP Matching

The technique disclosed in this article is a method for making a label confusion matrix for smoothing the label output distribution obtained from small size training data. A biased distribution of labels degrades the accuracy of the speech recognition technique which uses a labeled sequence. A technique for smoothing label distribution using a label confusion matrix has been proposed to resolve the problem. This article shows the way to obtain the confusion matrix based on the mutual information among label prototypes computed through DP matching. The matrix is computed as follows (see figure). Select a pair of utterances of the same word in the training data. Convert them into sequences of labels and carry out the labeled DP matching based on a certain distance measure. Let (h, k) be the coordinate of a point on the optimal path. Increase the confusion count of the corresponding label pair (l(i, H), l(j, K)) by one. Iterate the procedure on all the points on the optimal path, on all the pairs of the utterances in the training data, and on all the words in the vocabulary. Normalize the confusion counts to make the confusion matrix. The confusion matrix obtained by the method explained in this article is used to smooth output probabilities as follows.

(Image Omitted)

1

Page 2 of 2

2

[This page contains 1 picture or other non-text object]