Myanmar Number Normalization for Text-to-Speech

--Text Normalization is an essential module for Text-to-Speech (TTS) system as TTS systems need to work on real text. This paper describes Myanmar number normalization designed for Myanmar Text-to-Speech system. Semiotic classes forMyanmar language are identified by the study of Myanmar text corpus and Weighted Finite State Transducers (WFST) based Myanmar number normalization is implemented. Number suffixes and prefixes are also applied for token classification and finally, postprocessing has been done for tokens that cannot be classified. This approach achieves average tag accuracy of 93.5% for classification phase and average Word Error Rate (WER) 0.95% for overall performance which is 5.65% lower than rule-based system. The results show that this approach can be used in Myanmar TTS system and to our knowledge, this is the first published work of Myanmar number normalization system designed for Myanmar TTS system. Keywords-Myanmar Number Normalization; Text Normalization; Weighted Finite State Transducer; Myanmar Text-to-Speech; Myanmar;