Voice conversion method based on State-Space Model

This paper presents a novel method based on State Space Model (SSM) for voice conversion. The SSM has the ability to model the spectral parameter trajectory of speech which is desirable in voice conversion. Traditional training method for SSM is EM (Expectation Maximum) algorithm. Due to its low convergence speed, the training time is very long. Moreover increasing the spectral parameter order which helps to improve the speech quality will also cause SSM training difficulties. These problems are the underlying motivations of our work. In this paper we develop an efficient and robust training procedure based on LSF (Line Spectral Frequencies) segmentation method and subspace training algorithm. The objective and subjective experiments demonstrate the superior performance of the training approach as compared with EM method.