Most systems of Automatic Language Identification are based on phonotactic approaches. However, it is more and more evident that taking other features (phonetic, phonological, prosodic, etc.) into account will improve performances. This paper presents an unsupervised phonetic approach that aims to consider phonological cues related to the structure of vocalic and consonantal systems.
In this approach, unsupervised vowel/non vowel detection is used to model separately vocalic and consonantal systems. These Gaussian Mixture Models are initialized with a data-driven variant of the LBG algorithm: the LBG-Rissanen algorithm.
With 5 languages from the OGI MLTS corpus and in a closed set identification task, the system reaches 85 % of correct identification using 45-second duration utterances for male speakers. Using the vowel system modeling as a complement to an unsupervised phonetic modeling increases this performance up to 91 % while still requiring no labeled data.
Cite as: Pellegrino, F., Farinas, J., André-Obrecht, R. (1999) Vowel system modeling: A complement to phoneticmodeling in language identification. Proc. Multi-Lingual Interoperability in Speech Technology, 73-78
@inproceedings{pellegrino99_mist,
author={François Pellegrino and Jérôme Farinas and Régine André-Obrecht},
title={{Vowel system modeling: A complement to phoneticmodeling in language identification}},
year=1999,
booktitle={Proc. Multi-Lingual Interoperability in Speech Technology},
pages={73--78}
}