Memory-based Data-driven Approach for Grapheme-to-Phoneme Conversion in Bengali Text-to-Speech Synthesis System

In this paper, we propose a memory-based data-driven model for grapheme-to-phoneme (G2P) conversion for Bengali text-to-speech synthesis (TTS) system. Previous studies have stated the significance of the linguistic and phonetic features for rule-based Bengali G2P conversion techniques. But due to the lack of proper morphological analyzer, the scope of rule-based approaches is bounded. The proposed method overcomes the limitation of rule-based methods by exploiting the variety of contexts present in the text corpus built in the current study. The model has been trained with a memory-base showing the relation between graphs and phones based on contexts. The model has been tested with 300 random words and it achieved accuracy of 79.33% at word-level and 96.28% at graph-level. This performance has been compared with a related rule-based approach to prove the effectiveness of a data-driven method. Furthermore, the model doesn't require any morphological knowledge of the words.