Toward an Automatic Fongbe Speech Recognition System: Hierarchical Mixtures of Algorithms for Phoneme Recognition

Abstract

In this paper, we have demonstrated the efficacy of an automatic continuous speech recognition system by mixing fuzzy and neuronal approaches and an acoustic analysis of the sounds of an under-resourced language. The system we propose integrates the modules such as extraction module, segmentation and phoneme recognition modules and whose the core is based on the phoneme detection in continuous speech. This work offers a complete recipe of algorithms to perform hierarchically the following tasks: speech segmentation - phoneme classification - phoneme recognition. The segmentation task provides as output phoneme segment which are subsequently classified according to their nature (consonant or vowel voiced or unvoiced etc.). The segmentation and classification are based exclusively on a fuzzy approach while the phoneme recognition task exploits the acoustic features such as the formants for vowels and the pitch and intensity for consonants. Experiments were per- formed on Fongbe language (an African tonal language spoken especially in Benin, Togo and Nigeria) and results of phoneme error rate are reported.