To measure the extent to which individual glottal features obtained via inverse filtering can be used to discriminate between the speech of different speakers, we test a set of 16 glottal parameters on a 50-speaker subset of the TIMIT corpus in a frame-level pairwise speaker discrimination task. In addition, we compare a vector of glottal parameters to a set of spectral envelope features in the same task. The results showed higher discrimination ability between pairs of male speakers, and an overall inconsistency in feature rankings between genders. Glottal feature vectors were able to discriminate speakers at a much higher rate, and showed some ability to complement spectral envelope features.

Index Terms: glottal waveform, voice source, speaker identification, feature extraction