Voice conversion based on formant frequency shift technique

Abstract

In this thesis, two new speech-related algorithms are proposed. One is a fully automated systematic formant estimation algorithm while the other is a voice conversion one based on the formant shift concept practically implemented by shifting LSFs (Line Spectrum Frequency). The former algorithm can be considered as a modified version of the peak-picking one. The modification comes from the basic idea that the closer the distance between an LSF pair, the higher the probability that a formant exists between them. This formant information is directly utilized in the voice conversion algorithm. For voice conversion, two methods are investigated. The first is based on the formant shift and can change the formant frequencies and their bandwidths. The second transforms an input speech uttered by a specific speaker into another target speaker``s voice by applying the formant shift technique. Our formant shift algorithm is tested and shows that perceptually different voice color can be produced. The XAB and the MOS (Mean Opinion Score) tests are performed for voice conversion with target voice and the results show 95% of correct response and 4.2, respectively. As a conclusion, it can be said that, by using the proposed algorithms, an automatic formant estimation and voice conversion can be achieved with relatively successful results.