Investigating Automatic Language Discrimination via Vowel System andConsonantal System Modeling
Résumé
This paper presents an approach to Automatic Language Identification (ALI) based on a differentiated modeling of vowel and consonantal systems. The objective is to consider phonetic and phonological features that are not taken into account in the standard phonotactical approach. For each language, two Gaussian Mixture Models (GMM) are trained respectively with automatically detected vowel and non-vowel segments. Since this vocalic detection is unsupervised and language independent, no labeled data are required. GMMs are initialized using a datadriven variant of the LBG vector quantization algorithm: the LBG-Rissanen algorithm. Experiments show that this algorithm behaves efficiently to take the vowel system structure into account. With 5 languages from the OGI MLTS corpus and in a close set identification task, we reach 85 % of correct identification for the 45 second duration utterances, considering the male speakers.