Voice source waveforms for utterance level speaker identification using support vector machines

The voice source waveform generated by the periodic motion of the vocal folds during voiced speech remains to be fully utilised in automatic speaker recognition systems. We perform closed-set speaker identification experiments on the YOHO speech corpus with the aim of continuing our investigation into the level of speaker discriminatory information present in a data driven parameterisation of the voice-source waveform obtained by closed-phase inverse filtering. Discriminatory modelling using support-vector-machines resulted in utterance level correct identification rates of 85.3% when using a multi-class model, and 72.5% when using a binary, one-against-all regression model, each on cohorts of 20 speakers respectively. These results compare well with other speaker identification experiments in the literature employing features derived from the voice source waveform, and are positive when observed under the hypothesis that they should be complementary to the common magnitude spectral parameters (mel-cepstra).