ODYSSEY 2004 - The Speaker and Language Recognition Workshop

May 31 - June 3, 2004
Toledo, Spain

Voice Profile: A Structured Probability Model with Application to Voice Morphing

Dimitrios Rentzos (1), Saeed Vaseghi (1), Qin Yan (2)

(1) Department of Electronics and Computer Engineering, Brunel University, Middlesex, UK
(2) Fortune Institute of Technology, Kaohsiung, Taiwan

This paper presents the concept of a voice profile as a complete description of the distributions of the acoustic correlates and the speaking characteristics of a speaker. A voice profile can be considered as a unified speaker-dependent probability model of speech with applications in speaker identification, adaptive speech recognition, voice morphing and text to speech synthesis. The spectral and temporal parameters that define a voice profile are obtained from hidden Markov models (HMMs) of speech. The HMMs are trained on extended feature vectors that include features for recognition, synthesis and identification. A method of ranking the acoustic correlates of a speakerís voice is proposed based on an analysis of the relative distance of each voice correlate from that of the gender-dependent modal voice. The voice profile is used effectively for voice conversion. Experimental results of speaker profiling and its evaluation in voice morphing are presented.

Full Paper

Bibliographic reference.  Rentzos, Dimitrios / Vaseghi, Saeed / Yan, Qin (2004): "Voice profile: a structured probability model with application to voice morphing", In ODYS-2004, 193-198.