Text-Independent Voice Conversion

This thesis deals with text-independent solutions for voice conversion. It first introduces the use of vocal tract length normalization (VTLN) for voice conversion. The presented variants of VTLN allow for easily changing speaker characteristics by means of a few trainable parameters. Furthermore, it is shown how VTLN can be expressed in time domain strongly reducing the computational costs while keeping a high speech quality. The second text-independent voice conversion paradigm is residual prediction. In par-ticular, two proposed techniques, residual smoothing and the application of unit selection, result in essential improvement of both speech quality and voice similarity. In order to apply the well-studied linear transformation paradigm to text-independent voice conversion, two text-independent speech alignment techniques are introduced. One is based on automatic segmentation and mapping of artificial phonetic classes and the other is a completely data-driven approach with unit selection. The latter achieves a performance very similar to the conventional text-dependent approach in terms of speech quality and similarity. It is also successfully applied to cross-language voice conversion