Automatic voice-source parameterization of natural speech

Abstract

We present here our work in automatic parameterization of natural speech by means of a pitch synchronous source-filter decomposition algorithm. The derivative glottal source is modelled using the Liljencrants-Fant (LF) model. The model parameters are obtained simultaneously with the coefficients of an all-pole filter representing the vocal tract response by means of a quadratic programming algorithm. Synthetic data has been created and analyzed in order to show the appropriate function of the estimation method. The parameterization results in high quality synthesized speech for voiced frames. Voice quality extraction is performed on basis to the LF source representation. The inherent modelling of the voice source makes it suitable for voice modification tasks. Work is in progress to add this speech representation to emotional speech synthesis and voice conversion algorithms. 1.