Browse Prior Art Database

Speech Recognition Method Using Multi-Labeling

IP.com Disclosure Number: IPCOM000038909D
Original Publication Date: 1987-Mar-01
Included in the Prior Art Database: 2005-Feb-01
Document File: 2 page(s) / 56K

Publishing Venue

IBM

Related People

Nishimura, M: AUTHOR [+2]

Abstract

This article describes a speech recognition method using multi-labeling, that is, assigning a plurality of labels for each frame, in order to reduce the deterioration by error in vector quantization of speech feature. The theory of vector quantization of speech feature (so-called labeling) is widely used for speech recognition methods such as HMM (Hidden Markov Model) or DTW (Dynamic Time Warping) type. Although vector quantization is very useful for speech recognition to decrease computation time and memory space, it causes recognition errors because of its quantization error. In the method proposed here, the multi- labeling makes speech recognition free from deterioration by the quantization error. Multi-labeling Let Dmin and Si be as follows. Dmin = min di ...(1) Si = di/Dmin ...

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 64% of the total text.

Page 1 of 2

Speech Recognition Method Using Multi-Labeling

This article describes a speech recognition method using multi-labeling, that is, assigning a plurality of labels for each frame, in order to reduce the deterioration by error in vector quantization of speech feature. The theory of vector quantization of speech feature (so-called labeling) is widely used for speech recognition methods such as HMM (Hidden Markov Model) or DTW (Dynamic Time Warping) type. Although vector quantization is very useful for speech recognition to decrease computation time and memory space, it causes recognition errors because of its quantization error. In the method proposed here, the multi- labeling makes speech recognition free from deterioration by the quantization error. Multi-labeling Let Dmin and Si be as follows.

Dmin = min di ...(1)

Si = di/Dmin ...(2) where di is a distance between a code-book entry i and a speech feature at certain time. All label code-book entries which satisfy the following condition are selected at each frame, as shown in Fig. 1. SiThd ...(3) where Thd is a constant value. Training and Decoding The Multi-labeling method can also be applied to both training and decoding, both HMM and DTW. When applied to HMM, this method requires less computation time and memory space than DTW. For estimating the parameters of HMM, the conventional Baum-Welch algorithm is still available. Every label generated at each frame is equally used for parameter estimation. Also, in decod...