A GMM-STRAIGHT Approach to Voice Conversion

This paper explores the topic of voice conversion as explored in a joint project with Percy Liang (EECS, Berkeley). For our purposes, voice conversion is the process of modifying the speech signal of one speaker (source) such that it sounds as though it had been pronounced by a different speaker (target). Following the Source-Filter model of speech production, we begin by assuming that most of a speaker’s characteristics can be summarized in the spectral envelope as represented by a set of Linear Predictive Coefficients. By using a Gaussian mixture model (GMM) to model the features of the source speaker, we can then learn a mapping of features from the source to the target, and then resynthesize via various methods. In this paper, we explore different approaches to model the features and describe our results from experimenting with the resynthesis process, including the integration of a STRAIGHT vocoder system, which provides an advanced model of the excitation signal. Further discussion includes ways to immediately improve the system and how we would like to proceed in the future.