Loading...
Optimizing vowel recognition in Thai spoken language using reduced LPC spectrum and reduced feature set of critical band intensities
Author(s)
Suktangman, Nongnuch
Khanthavivone, Kham
Songwatana, Kraisin
Date Issued
December 1, 2006
Type
Conference Paper
Abstract
This paper presents an optimization for vowel recognition in Thai spoken language. Thai language consists of 18 unmixed vowels (a, aa, i, ii,ω,ωω, u, uu, e, ee,ε,εε, o, oo,γ,γγ,); and 6 mixed vowels (ua, u:a:, ia, i:a:,ωa,ω:a:). Previously we have proposed use a 3-stage decision-making process [1]: using voice energy and polynomial regression for short and long vowels classification; identifying voice segment (frame) into basic vowels; and using decision rule for recognition on unmixed and mixed vowels. The optimization is done by reducing the order of LPC and choosing a reduced set of Critical Band Intensities (CBI) [2]. With this, the percentage accuracy remains high while processing time is decreased by 85%. The experiment is carried out on voice samples from 3 male model, 3 female model, and 2 male and 2 female model, yielding an accuracy of 91.47, 94.98, and 91.72 percents respectively. © 2006 IEEE.
Citation
2006 International Symposium on Communications and Information Technologies Iscit, 128-132, 2006