Sixth European Conference on Speech Communication and Technology

Budapest, Hungary
September 5-9, 1999

The Analysis of Speaker Individual Features Based on Autoregressive Hidden Markov Models

Evgeny I. Bovbel, Polina P. Tkachova, Igor E. Kheidorov

Dept. of Radiophysics, Belarusian State University, Minsk, Belarus

The speech-based analysis of speaker individual features has found wide application area. In order to analyse the speaker individual features it is necessary to use high frequencies and accurate spectrum estimation methods. It was found out that the best way to analyse the personal voice individuality is to use bark scaled spectrum estimation based on arithmetic Fourier transform. For each of 10 speakers the autoregressive hidden Markov model was trained. The experiments show that such models provide high accuracy of person identification based on bark- cepstrum analysis and high sampling frequency.

Full Paper (PDF)

Bibliographic reference.  Bovbel, Evgeny I. / Tkachova, Polina P. / Kheidorov, Igor E. (1999): "The analysis of speaker individual features based on autoregressive hidden Markov models", In EUROSPEECH'99, 1191-1194.