Mandarin-Tibetan Bilingual Cross-language Voice Conversion Based on Semi-hidden Markov Model

Abstract

In recent years, deep learning based voice conversion (VC) has significantly improved the performance of the conversion system. However, such systems generally require a large amount of parallel corpus from the source speakers and the target speaker, and the parallel corpus of Mandarin-Tibetan bilingual is difficult to obtain. In order to solve this problem, we propose a method based on the semi-hidden Markov model (HSMM) using Mandarin-Tibetan bilingual average voice model and speaker adaptive technology for VC. This method does not require parallel corpus. Firstly, it obtains the Mandarin-Tibetan bilingual average voice model using mixed language multi-speaker corpus training, then uses a small number of source speaker corpora to adaptively convert the average voice model to obtain the speaker-related acoustic model. Finally, the text corresponding to the source speaker's speech is translated, and the context-dependent labels of the translated text is input into the speaker-related acoustic model, the converted speech is output to realize cross-language VC. The experimental results show the effectiveness of this method, the converted speech MOS score: 3.37 points; DMOS score: 3.00 points.