Técnicas para conversão de orador em sinais de voz

Abstract

Presents a voice conversion system, a system that transforms a voice signal spoken by some speaker into a signal that sounds like it was spoken by another speaker, without changing the textual content of the speech or changing information like emotion or emphasis. The main objective of this work is to compare the conversion as done by different methods. To accomplish this, a unified voice conversion system containing the analysis, conversion and synthesis steps necessary to transform the speaker was implemented. Four voice conversion techniques, three from the literature, based on Gaussian mixture models, hidden Markov models and feed forward neural networks, and one novel based on recurrent neural networks, were evaluated. Two methods to generate the excitation used in the synthesis step were also implemented, one utilizing a parametric pulse trained on the speech signals, and one utilizing the PSOLA algorithm. On this system a couple of experiments were conducted to assess the conversion quality of each method: one measuring the distance between the cepstra of the signals, and the other employing a speaker recognition system. In these experiments the conversion based on Gaussian mixture models yielded the best results, but all techniques were relatively close in terms of performance.