A hybrid LF-Rosenberg frequency-domain model of the glottal pulse

In this paper we describe innovative advances to the design of a new frequency-domain algorithm to glottal source estimation whose conceptual approach we have reported recently [1]. Those advances result from accurate sinusoidal/harmonic analysis and synthesis of two concomitant acoustic signals: the glottal source signal captured near the vocal folds, and the corresponding voiced signal captured outside the mouth. We describe the experimental procedure which was performed by an ORL specialist using a rigid video-laryngoscope and two tiny and high-quality microphones. Six subjects have participated in the tests and records were made for vowels /a/ and /i/. The data analysis allowed us to conclude on the magnitude and on the phase-related NRD features of the glottal source signal. In addition, a new frequency-domain glottal pulse model combining features of the Liljencrants-Fant and Rosenberg models has been devised that is a better match to the observed data. The derivatives of the three models are obtained using accurate frequency-domain processing. The paper concludes with next research steps.