Speaker Verification using HMM for Romanian Language

In this paper we analyse a speaker verification system with fixed uttered text for Romanian language. The feature extraction in the system is based on perceptual cepstral analysis, giving the melfrequency cepstral coefficients. The acoustical modeling of speech in the statistical framework is based on hidden Markov models. To assess performance of the system, the models were trained with each speaker and than tested with all speakers, on our own database, containing speech data from ten speakers. The results are encouraging, a false acceptance error of 4% being achieved. Informal tests show the possibility to enhance this performance in a future work