Estimating the Stability and Dispersion of the Biometric Glottal Fingerprint in Continuous Speech

The speaker’s biometric voice fingerprint may be derived from voice as a whole, or from the vocal tract and glottal signals, after separation by inverse filtering. This last approach has been used by the authors in early work, where it has been shown that the biometric fingerprint obtained from the glottal source or related speech residuals gives a good description of the speaker’s identity and meta-information, as gender or age. In the present work a new technique is proposed based on the accurate estimation of the glottal residual by adaptive removal of the vocal tract, and the detection of the glottal spectral singularities in continuous speech. Results on a reduced database of speakers demonstrate that the biometric fingerprint estimation is robust, and shows low intra-speaker variability, which makes it a useful tool for speaker identification as well as for pathology detection, and other fields related with speech characterization.