Breaking a Paradox: Applying VTLN to Residuals

Abstract:
Vocal tract length normalization (VTLN) is a speaker normalization technique that tries to compensate for the effect of speaker-dependent vocal tract lengths. According to the source-filter model of speech production, the vocal tract can be represented by features as linear predictive or cepstral coefficients, whereas the residual is an estimate of the underlying excitation, i.e. the voice source. In speech recognition, VTLN is performed before deriving the features. Interestingly, recent studies on VTLN-based voice conversion suggest that the application of VTLN to the residuals, that should hardly contain vocal tract information, may result in a voice that is very similar to a voice generated in the conventional way, i.e., by applying VTLN to the original speech frames. In this paper, we discuss an objective experiment that is to confirm this observation.