HMM-based Tianjin Dialect speech synthesis using bilateral question Set

The main difference between Tianjin dialect and Mandarin is the tone change. However, because there is no pronunciation dictionary for Tianjin dialect and not all the tones follow the sandhi rules in a continuous utterance, we focus on improving the tone accuracy by using bilateral question set based on the traditional Mandarin synthesis system. By considering more contextual features in one question, we can obtain better splitting steps during clustering so that the models could be more precise. Through experiments, we select appropriate bilateral question set for both pith models and spectrum models. ABX test also shows that we could get more natural and accurate tones by using the bilateral question set than the ones using summarized rules for Tianjin Dialect.