NEW APPROACHES TO VOICE CONVERSION USING STATISTICAL MAPPING FUNCTIONS

VOICE conversion (VC) is the process whereby the speech signal of one speaker (source) is transformed into the the voice of another speaker (target). Voice con- version can be used in many applications, example of which includes text to speech; speaker recognition; noise reduction in speech; neutral speech to emotional speech conversion; movie, animation, and music industry applications. The features trans- formed in VC systems are typically the parameters characterizing the speech and speaker individuality, including the fundamental frequency, spectral envelope, ape- riodicity, and phoneme duration. Among these, the spectral envelope is one of the most significant characteristics of the speaker identity. In this thesis, we propose four new approaches for spectral conversion: Mixture Density Network (MDN); Dynamic Multi-band Random Forest (DMRF); State Space Model (SSM) employing the Gaus- sian Mixture Model (GMM) for state-vector sequence conversion (SSM-GMM); and Sub-band Deep Gaussian Processes (SDGP). These new conversion methods were developed for both speech and singing applications. Experimental results show that the new methods have performance advantages over the conventional methods both subjectively and objectively.