Mandarin-Tibetan Cross-Lingual Voice Conversion System Based on Deep Neural Network

This paper realizes a Mandarin-Tibetan cross-lingual voice conversion system to solve the communication problem between the Mandarin speaker and the Tibetan speaker. Mandarin speech recognition and Tibetan speech synthesis techniques based on deep neural network(DNN) are adopted to convert Mandarin to Tibetan. In this way, we can effectively avoid the problem of building large parallel corpus and complex conversion rules. Meanwhile, we modify the converted Tibetan speech features so that it is perceived as a sentence uttered by the Mandarin speaker. The experimental results show that Mean Opinion Score (MOS) is 3.26 points and the degradation mean opinion score (DMOS) of the timbre similarity between the converted Tibetan speech and the Mandarin speech is 3.07 points.