Grapheme-to-Phoneme Conversion for Thai using Neural Regression Models

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 4251 - 4255 July 10-15, 2022 ©2022 Association for Computational Linguistics Grapheme-to-Phoneme Conversion for Thai using Neural Regression Models Tomohiro Yamasaki Toshiba Corporation, Research and Development Center, Japan tomohiro2.yamasaki@toshiba.co.jp Abstract We propose a novel Thai grapheme-to- phoneme conversion method based on a neu- ral regression model that is trained using neu- ral networks to predict the similarity between a candidate and the correct pronunciation. Af- ter generating a set of candidates for an input word or phrase using the orthography rules, this model selects the best-similarity pronunci- ation from the candidates. This method can be applied to languages other than Thai simply by preparing enough orthography rules, and can reduce the mistakes that neural network mod- els often make. We show that the accuracy of the proposed method is .931, which is com- parable to that of encoder-decoder sequence models. We also demonstrate that the pro- posed method is superior in terms of the differ- ence between correct and predicted pronunci- ations because incorrect, strange output some- times occurs when using encoder-decoder se- quence models but the error is within the ex- pected range when using the proposed method. 1 Introduction Grapheme-to-phoneme conversion (G2P) is the task of converting grapheme sequences into cor- responding phoneme sequences. Many languages have the difficulty that some grapheme sequences correspond to more than one different phoneme se- quence depending on the context. G2P plays a key role in speech and text process- ing systems, especially in text-to-speech (TTS) systems. These systems have to produce speech sounds for every word or phrase, even those not contained in a dictionary. In low-resource lan- guages, it is fundamentally difficult to obtain large vocabulary dictionaries with pronunciations. Therefore, pronunciations need to be predicted from character sequences. In many languages, each word is composed of syllables and each syllable is composed of char- acters following the orthography rules of that lan- guage. This means that G2P for languages with syllabic orthography rules can be formulated as the task of selecting the best path in a lattice gen- erated for a given input word or phrase if we pre- pare enough orthography rules to make sure that any lattice generated almost certainly includes the path for the correct pronunciation. ��� �� ��� �� ���� �� ���� �� ���� � ��� � ��� ��� Figure 1: G2P can be formulated as the task of se- lecting the best path in a lattice. In English, because “ea” can be pronounced as /e/ and /i:/ and “ow” can be pronounced as /oU/ and /aU/, the word “meadow” has 4 pronunciation candidates, excluding stress posi- tion estimation. As the result of some effort, we prepared Thai orthography rules. Almost all possible paths in a lattice can be generated from these, and each path needs to be evaluated using a phonological lan- guage model to select the best path. With this in mind, we propose …

Content truncated to 500 words