Study on attenuated tone for Mandarin Text-To-Speech

Tone plays an important role in Mandarin. Some syllables may be neutralized while their lexicon tones are not neutral. This neutralization is extremely complicated and often speaker-by-speaker or case-by-case. It challenges TTS system in tone annotation when developing speech corpus. This paper makes some attempts on this aspect. Firstly, a new tone type, attenuated tone, is defined to distinguish these neutralized tones from the lexicon neutral tones. Secondly, a method is proposed to automatically detect the attenuated tones in a speech corpus. Both the subjective and objective evaluations show that the introduction of attenuated tone improves the performance of prosody models and the annotation method is reliable.