A hybrid statistical/RNN approach to prosody synthesis for taiwanese TTS

m is a spoken dialect widely used in the south-eastern China and Taiwan. Just like Mandarin, Min-Nan speech is a monosyllabic and tonal language. Each character is pronounced as a syllable carrying a lexical tone. There are only 877 base-syllables and 8 tones including a degenerated one which is not used in modern Taiwanese. These 877 basesyllables also have the same initial-final structure like Mandarin base-syllables. There are 18 initials and 82 finals. Although tonal syllables are the basic pronunciation units, Min-Nan speech is colloquial and does not have a standard written form. There are two popular written forms used in Taiwan. One is the Romanization form which uses Roman alphabets to spell each base-syllable and uses a number to specify its tone. The other is a hybrid one which uses Chinese characters to represent ordinary words and represents some extraordinary syllables in the Romanization form. It is noted that for the hybrid written form the system to represent words in Chinese characters is still not standardized nowadays in Taiwan. This makes the text analysis very difficult for Min-Nan language. In the past, there are very few studies in Min-Nan TTS. There are only a few preliminary studies in Taiwan in recent years [1-3]. This is mainly due in part to the lack of a standardized written-form representation as mentioned above and in part to the difficulty of collecting a large speech database. In this paper, a hybrid statistical/RNN approach to prosodic information synthesis for Min-Nan TTS is proposed. It tries to first model some primary factors affecting the generation of prosodic information by the statistical-based method and then uses RNN to take care of all other affecting factors. Advantages of the new approach as compared with the conventional RNNbased approach are two-folded: (1) It provides more consistent training data to train the RNN via suppressing some interference factors such as speaking rate variation; and (2) It removes some loads of the RNN prosodic modeling. The paper is organized as follows. Section 2 presents the proposed hybrid prosody synthesis method. Two schemes of the proposed method are discussed. Performance of these two schemes are evaluated by experiments discussed in Section 3. Some conclusions are given in the last section.