Sinusoidal model based speaker identification using VQ and DHMM

In this work, we propose a new set of features, sinusoidal model features, for speaker identification. The performance of the features is evaluated using vector quantization (VQ) and discrete hidden Markov model (DHMM) for a speaker database with 20 speakers. Fifty speech utterances of duration 2 seconds each are recorded from each speaker for design and testing. Eighty percent of the speech data is used for training and twenty percent of the speech data issued for testing purpose. The speaker identification using sinusoidal model feature (amplitude) presents 98% speaker recognition for the test set by the vector quantization classifier. The frequency, phase features presents maximum of 79% and 32% recognition accuracy respectively.