Myanmar text-to-speech system with rule-based tone synthesis

Abstract

We have introduced a novel Myanmar text to speech (MyanmarTTS) system with rule-based tone synthesis. Myanmar is a tonal language that possesses unique characteristics compared with other tonal languages such as Chinese, Vietnamese and Thai. Such languages have complicated fundamental-frequency (F₀) patterns of tones, and F₀ is of foremost importance. Myanmar tones are unique in their simplistic pattern related not only to F₀ but also, more specifically to duration. Myanmar tones have different durations between short-tone and long-tone groups. In accordance, we defined a tone rule employing two parameters F₀ at the center of the syllable and the syllable’s duration. The rule is implemented with a linear F₀ pattern. Large variability exists in the F₀ and duration uttered by different speakers of different syllables. Hence, for tone synthesis, normalization of the F₀ and duration is important and necessary to discriminate tones. We proposed a normalization method and the effectiveness of this method was confirmed in the distribution of the F₀ and duration. The intelligibility of the synthesized tone was confirmed through listening tests with correct rates of 95.6% for male and 97.8% for female speech. As a result, we showed that the linear pattern is sufficient for Myanmar tone synthesis.