Abstract
Voice conversion based on State Space Model (SSM) has been recently proposed to address the discontinuity problem in the traditional frame-based voice conversion by considering the spectral envelope evolutions. However, the results are over-smoothed. To resolve this problem, in this paper we propose a new procedure for integrating the global variance constraint into the SSM-based voice conversion. Moreover, unlike the SSM-based method, we allow the state-vector order to be higher than the feature-vector order. Experimental results verify that the proposed method significantly improves the performance of the SSM-based voice conversion in terms of speaker individuality and speech quality. Our experiments also show that the proposed method outperforms the well-known Maximum Likelihood estimation method that considers the Global Variance in terms of speech quality.