Transformer-Based Grapheme-to-Phoneme Conversion With Knowledge Transfer for Isarn

Past research has demonstrated the significant advantages of the transformer model in grapheme-to-phoneme (G2P) conversion. Transformers typically require a large amount of training data to generalize well, and this poses a challenge in the context of dialects for which limited data are available. In this paper, we propose a transformer-based G2P model for Isarn, which is a dialect of Thai. The proposed model is initially trained on a Thai language dataset to acquire pronunciation-related knowledge, and is then transferred and adapted to the Isarn dialect by using the appropriate representations of graphemes, including syllable boundaries, phonemes, and tones. The proposed transfer learning-based model achieved a word error rate (WER) of 15.63% and phoneme error rate (PER) of 3.98%, surpassing the transformer-based model trained from scratch by using only Isarn data that obtained a WER of 22.96% and a PER of 8.64%, as well as multilingual models such as Multilingual ByT5 G2P (17.86% WER, 5.06% PER). These evaluations, which were carried out on unseen Isarn words and loan words borrowed from Thai, offer compelling evidence that the proposed transfer learning-based model excels not only in comparison with a transformer-based model trained from scratch by using solely Isarn data, but also against other baseline models.