Voice conversion: Wavelet based residual selection

Voice conversion has been studied over past few decades and yet no flawless system has been developed. Primary restriction in developing conversion systems is decayed output speech quality. Work presented here alleviates this problem by mapping higher order excitation features along with state of the art spectral parameters. Well known linear predictive analysis is used to extract shape of the vocal tract and corresponding residual signal. Higher feature dimensionality of the excitation signal is confronted using synchronous segmentation and windowing of the signal. Each of the resulting frames are wavelet analyzed to calculate normalized sub-band energy coefficients forming a codebook. Conversion is obtained by selecting target residual corresponding to minimized energy cost function. Primary advantage of this technique is reduced dimensionality with satisfactory conversion statistics. Proposed method is compared with baseline residual selection approach using various subjective and objective tests. Wavelet features provide better selection criteria with slight improvement in output speech individuality.