Studying the Use of Phonetic Language Script in a Limited Training Environment for Text-to-Speech

Abstract

This study looks into a variety of characteristics of text-to-speech (TTS) model training and transfer learning with a phonetic language script. When compared to training using a non-phonetic script, training with a phonetic language script appears to speed up the process. In both Hindi and Indian English evaluation data, transfer learning with phonetic writing (in Devanagari script) leads to remarkable improvements in the number of pronunciation errors (87.5% and 16.67% relative). As a result, a case is made for publishing models pre-trained on phonetic language transcripts rather than the models that have been pretrained on non-phonetic transcripts for the purpose of transfer learning. Going further, a mix-lingual training approach is also explored to solve the problem of transforming code-switched text to speech. It is emphasized that even with commercial text-to-speech solutions, the pronunciation of Indian proper nouns contained in English text is a difficulty, which has been addressed in the current framework.