Using speaker adaptive training to realize Mandarin-Tibetan cross-lingual speech synthesis

Abstract

This paper presents a method to realize the hidden Markov model (HMM)-based Mandarin-Tibetan cross-lingual statistical speech synthesis using speaker adaptive training. A set of Speech Assessment Methods Phonetic Alphabet (SAMPA) is designed to label the pronunciation of the initial and the final of Mandarin and Tibetan syllables according to the similarities in pronunciation between Mandarin and Tibetan. A grapheme-to-phoneme conversion method is realized to convert Chinese or Tibetan sentences to SAMPA-based Pinyin sequences. A Mandarin statistical speech synthesis framework is employed to realize Mandarin-Tibetan cross-lingual speech synthesis. A set of context-dependent label format is designed to label the context information of Mandarin and Tibetan sentences. A question set is also realized for context dependent decision tree clustering. The initial and the finalare used as the synthesis units with training using a set of average mixed-lingual models from a large Mandarin multi-speaker-based corpus and a small Tibetan one-speaker-based corpus using speaker adaptive training (SAT). Then, the speaker adaptation transformation is applied to the speaker dependent (SD) training data to obtain a set of speaker dependent Mandarin or Tibetan models from the average mixed-lingual models. The Mandarin speech or Tibetan speech is then synthesized from the speaker dependent Mandarin or Tibetan models. Tests show that this method outperforms the method using only Tibetan SD models when only a small number of Tibetan training utterances are available. When the number of training Tibetan utterances is increased, the performances of the two methods tend to be the same. Mixed Tibetan training sentences have a small effect on the quality of synthesized Mandarin speech.