Vocalic system modeling: a VQ approach

This paper presents our work on vowel system identification as part of a project of automatic language identification using phonological typologies. The phonological study of vowel systems shows that a typology of the languages may be issued from the description of their vowel system (vowel number and acoustic space topology). We exploit this knowledge, to improve the phone recognizer currently used as front end of the automatic language identification systems. In a previous work, we have developed a vowel detection algorithm based on a spectral analysis of the acoustic signal and requiring no learning stage. Currently, we have modeled the vocalic space as a Gaussian mixture and we have compared two algorithms, the classical LBG algorithm and the LBG-Rissanen (1983) algorithm to estimate it. We use the Oregon Graduate Institute Multi-Language Telephone Speech Corpus (OGI-TS) to assess our approaches. We limit our experiments to five languages (French, Japanese, Korean, Spanish and Vietnamese). As we do not use any global acoustic model and any phone language modeling stage, the correct identification rate (about 75%) is a promising result and offers many perspectives.