Speaker's Gender Detection from Glottal Biometry

Abstract

Through the present work a biometric signature of a speaker’s voice is proposed for the detection of the speaker’s gender. The estimation method relies on the extraction of the glottal flow derivative from voice after removing the vocal tract transfer function by inverse filtering. This spectral density is related to the vocal fold cover biomechanics, and it is well known that certain speaker’s features as gender, age or pathologic condition are present in it. For such a database of 100 pathology-free speakers equally balanced in gender and age is used as an experimental framework to draft the results exposed in the work. As the estimated biometric parameters show a certain degree of crosscorrelation Principal Component Analysis (PCA) is used to reduce parameter dimension. The principal components are used in unsupervised k-means clustering of speakers (unsupervised gender detection). The outcome grouping shows an almost complete separation of speakers by gender in terms of the most relevant parameters derived from a statistical dispersion study. Possible applications of the study can be found in forensic acoustics as well as in speaker identification and verification tasks.