AI Enabled Text-to-Speech Synthesis for Unicode Language

Objectives: To explore the advancements and challenges in the implementation of an AI-enabled Text-to-Speech (TTS) for the Assamese language (an Indo-Aryan language) by discussing technological approaches, linguistic considerations, and its potential applications. Methods: The developed system has been experimented with 40K (approx) collected words of five different categories and tested with different datasets for each model. Four prominent methods (Dictionary, HMM, CNN & G2P) are commonly found used in TTS. With the Grapheme- to- Phoneme (G2P) conversion technique, all phonemes are ordered sequentially to generate a continuous voice by using a good synthesizer to translate into words, letters, ligatures, and symbols of the language into sound. Findings: Compared to the other three approaches, the G2P approach shows higher performance (89%) in terms of accuracy. We experimented with our system using different sample sizes from three Indian languages (Hindi, Assamese, and Bengali) across four approaches. With a dataset of 1,000 words, the Dictionary-based approach resulted in 60% accuracy, HMM in 70%, CNN in 85%, and remarkably, the G2P approach reported 89% accuracy. Upon uploading the input file into the system and listening to the system read words, we manually calculated the number of words spelled correctly, incorrectly, or not producing any sound. Precision, F1-score, and recall were analyzed and found to achieve a 90% result, which may be considered a satisfactory conversion. Furthermore, by giving additional thought to pronunciation generation, the accuracy level might be improved. For experimental purposes, only female voices were used in this study. The smaller database size and minimal word decomposition algorithms utilized in this study account for the improved accuracy. Novelty: This effort offers a unique solution for the Assamese language. It has been designed with a pleasant voice and is in interactive mode which wais not found earlier. Keywords: AI, Assamese Language, NLP, G2P, Text-To-Speech