Automatic speaker identification using the magnitude and phase spectra of inverse - filtered voiced speech

A speaker identification experiment was performed with the 10 digits of the TI46 data corpus, spoken in isolation by 8 female and 8 male speakers. From each voiced frame of an utterance, a single fundamental period was extracted automatically and inverse-filtered to obtain an approximation of the glottal excitation function. Differently sized feature vectors were derived from the magnitude and phase spectra of the glottal excitation function. They were considered by themselves and in combination with the fundamental frequency. The best identification results from the excitation spectra alone were 28% for the 8 female speakers and 34% for the 8 male speakers. When the feature vectors were extended to also contain the fundamental frequency of the frame, the identification results for the 8 female speakers increased to 47% and for the 8 male speakers to 37%.