On the use of spectral peak parameters in voice conversion
Résumé
This paper addresses the problem of low transformed data variance, or "over-smoothing," in spectral transformation for Voice Conversion. In examining a classic GMM-based transformation with cepstral coefficients, we show that this problem lies, not only in the transformation model (as commonly assumed), but also in the choice of spectral parameterization. Consequently, we propose an alternative method for spectral transformation using spectral peaks and an HMM with Gaussian state distributions. The spectral peaks are shown to offer higher inter-speaker feature correlation and yield higher transformed data variance than their cepstral coefficient counterparts. Additionally, the accuracy of the transformed envelopes is examined.
Origine | Fichiers produits par l'(les) auteur(s) |
---|
Loading...