Multisource Speech Analysis for Speaker Recognition

Abstract

On a comprehensive speech database, speaker recognition characteristics are compared under the usage of various voice-source models. Inverse problems to find a source via vowel speech segments are solved on the base of a special speech-production model and voice-source models (A-source, piecewise-linear source, nonparametric source, and source found by means of the spectral relation method). In the first stage, we find the pulses such that the relative residuals of their segmented and their theoretical analogs computed by means of the speech-production model are less than 0.25. For the selected pulses, a posteriori estimates of the error of their determining are computed and the final selection of the source pulses is performed: for the recognition procedure, we leave only pulses with a posteriori estimates of the error less than the accepted level 0.3. In the space of parameters found for each source model, a statistical model is created for each speaker and the recognition is performed. For the speaker recognition with respect to one vowel, the mean error is approximately equal to 66% for the piecewise-linear source, 61% for the spectral relation method, and 33% for the A-source.