Utterance copy through analysis-by-synthesis using genetic algorithm

Synthesizing artificial voices with the aim of performing an utterance copy is considered a difficult task even when using a synthesizer based on formants. That is justifiable for example, by the fact that one needs to adjust several input parameters values, with nonlinear relation to the output, to produce a short segment of voice. Furthermore, the naturalness and intelligibility of the artificial voice are aspects to be considered when producing and evaluating the synthesized voice. The main goal of this work is to present the current status of our system based on genetic algorithm to automatically estimate the input parameters of the Klatt synthesizer, and compare its results with the ones obtained withWinsnoori, the only available software that performs the same task. The proposed system results outperforms the baseline by a large margin. For example, in average, the mean square error is reduced to 9.45% of the values obtained with Winsnoori and the PESQ increased from 1.06 to 3.36.