A Great Reduction of WER by Syllable Toneme Prediction for Thai Grapheme to Phoneme Conversion

Thai toneme prediction has been one of the greatest difficulties for Thai grapheme to phoneme conversion (G2P). This paper presents an improvement in the prediction of linguistic features in terms of tone rules. Among these, there will always be exceptions, for example, the tones used in loan words and transliterated words, which are usually adopted from the original language. This paper does not concern itself with the transliteration problem, but aims to show the success of a method which uses an automatic toneme predictor based on the tone rules of Thai pronunciation for the development of a machine learning model. The proposed method attaches a predictor to the final stage of converting a grapheme to a phoneme. Furthermore, this work also explores end-to-end prediction using Long Short Term Memories (LSTM) that takes its input sequence from the National Electronic and Computer Technology Center's Pseudo-Syllable segmentation and alignment tool. An evaluation was conducted to show the success of the proposed system, and also to compare the results with our traditional end-to-end sequence-to-sequence G2P. A comparison of the results shows that sequence-to-sequence modeling obtains the lowest Word Error Rate at 1.6%, and the proposed system works well on a 2018 small device (Raspberry Pi).