Flexible Harmonic/stochastic Modeling for Hmm-based Speech Synthesis

In this paper the preliminary results, of a new approach on speech modeling for statistical parametric HMM-based speech synthesis are presented. The proposed system is based on a flexible pitch-asynchronous harmonic/stochastic model (HSM) [1]. The speech is modeled as the superposition of two components: a harmonic component and a stochastic or aperiodic component. The fact that the synthesis model is pitch-asynchronous allows the direct integration to a HMM-based synthesis system. HTS [2], an open source software toolkit that provides HMM-based speech synthesis was used. The proposed HSM method was compared to the HTS baseline system with the same configurations and database. A number of different experiments were conducted. Results show that high quality of synthesized utterances is reached. A small perceptual test was carried out comparing the two systems on quality of the synthetic voice and similarity to the original voice. HSM outperforms the HTS baseline system in the quality test: HSM 53%, HTS 35,3%, and undecided 11,7%. Concerning similarity to the original voice, HSM-performed slightly better than HTS: HSM 35,3%, HTS 29,4%, and undecided 35,3%.