Myanmar Text-to-Speech System based on Tacotron-2

Myanmar is one of the developing countries situated in South-East Asia, and there are still many areas that have been under-developed with respect to advanced natural language processing technologies, where text-to-speech is one of them. The main motivation of this paper is to improve the naturalness of Myanmar text-to-speech system that is able to generate human-like speech. In this paper, we apply the neural network architecture based on Tacotron-2 that generates a mel spectrogram for speech synthesis directly from the sequence of text. Our proposed method is composed of three steps. In the first step, we create a speech corpus of 5k sentences of text and audio pair of Myanmar text from a large set of news articles, novel books, daily usages and travel-related expressions. We segment the Myanmar text into a sequence of characters by using a syllable segmenter and text normalizer. In the second step, we utilize the recurrent sequence-to-sequence feature prediction network that maps character embedding to mel-scale spectrograms. In the final step, we use Griffin-Lim algorithm to convert the corresponding text into generate Myanmar speech output. We compare our proposed method with an end-to-end generative model based on Tacotron. Furthermore, we investigate the subjective evaluation for both methods in speech synthesis by using mean opinion score (MOS). The experimental results show that our proposed method obtains an improvement over Tacotron based speech synthesis in terms of naturalness and intelligibility.