First International Conference on Spoken Language Processing (ICSLP 90)

Kobe, Japan
November 18-22, 1990

Speaker Based on Multipulse Excitation and UPC Vocal-Tract Model

Seiichiro Hangai, Kazvhiro Miyauchi

Department of Electrical Engineering, Science University of Tokyo, Tokyo, Japan

In automatic speaker identification, the reduction of dimensions of template is a key to realize a quick identification and to save storage. In this study, we extract some glottal wave parameters which does not seem to be susceptible to mimicry and combine them with some LPC vocal tract parameters to make a smaller sized template. In the extraction of feature parameters of glottal wave, the multipulse excitation model is adopted under modifying the pulse search algorithm. According to experimental results obtained from 30 speakers' 5 vowels, 94% identification rate with a template of 10 feature parameters (8 vocal tract and 2 glottal parameters) and 99% with a template of 36 feature parameters (14 vocal tract and 22 glottal parameters) are obtained respectively. In comparison with the number of parameters to get same identification rate only with a formants' template, the number of feature parameters is reduced by 5 in case of getting 94% identification rate and 7 in case of getting 99% identification rate.

Full Paper

Bibliographic reference.  Hangai, Seiichiro / Miyauchi, Kazvhiro (1990): "Speaker based on multipulse excitation and UPC vocal-tract model", In ICSLP-1990, 1269-1272.