Biometry of Voice based on the Glottal-Source Spectral Profile

Through the present work a biometric pattern of a speaker's glottal source based on the power spectral density profile of the mucosal wave correlate residual is defined, after estimations derived from the removal of the vocal tract transfer function by inverse filtering. This pattern may be parameterized accordingly to its peak-trough profile, which may be shown to be related to the biomechanics of the vocal folds. Using Principal Component Analysis on the resulting observation parameters the set of speaker samples are mapped to a new manifold where using k-means clustering are blindly divided into two clusters. It may be shown that the two clusters reflect the distribution of speakers by gender in a natural way. The resulting clusters may be represented on the first three most relevant principal components, revealing the structure of the subclusters on the reduced 3D manifold. Results from a set of 100 normophonic (pathology-free) speakers are presented. The results reveal that some speaker's meta-features may be hidden in the parameterization proposed.