Amazigh speech recognition using triphone modeling and clustering tree decision

Abstract

The main objective of this paper is to develop an Amazigh automatic speech recognition system using a speech corpus composed on $187$ distinct Amazigh words. The speech corpus was recorded by $50$ ($25$ male and $25$ female) Amazigh-Tarifyt native speakers. The system was evaluated on a speaker-independent approach using Hidden Markov Models(HMMs). The tests were carried out basing essentially on the Gaussian mixture distributions(GMMs), tied states (senons), triphone modeling and clustering tree decision. The recognition rate increases significantly and reached $92,2\%$ which is a high and satisfactory recognition rate comparing to the systems developed for this language especially in relative to the size of the corpus used on our system.