On Residual Prediction in Voice Conversion Task

Nowadays, voice conversion is a problem which is intensively analyzed by many researchers. A large group of existing voice conversion systems is based on RELP re-synthesis. Within these systems, the speech signal is pitchsynchronously segmented and described with LSF parameters. A transformation function is acquired by employing pairs of equal time-aligned utterances from source and target speaker. The conversion function for LSFs is often derived from probabilistic description of LSF pairs. The residual signal is also important for speech perception; it is transformed by so called residual prediction. Usually, a suitable residual signal is estimated from converted LSFs. We proposed two alternative approaches. First, we tried to predict the target speaker’s residual signal directly from source speaker’s LSFs. Then we combined both aforementioned approaches and used both source and converted LSFs for residual signal estimation. Moreover within each of these methods, the probabilistic and Euclidian-metrics-based descriptions of LSF parameter space were employed. Various pairs of source and target speaker were tested. Objective evaluation of converted speech using performance metrics and orientation listening tests was performed. The preliminary experiments revealed that no approach is generally better then others. However, in particular cases one method is usually preferable to others. Thus a universal voice conversion system should automatically decide which technique of residual prediction will be utilized.