Multi-language Southeast Asian speech2text development

Abstract

This project aims to improve on the current Code-Switch ASR technology for English- (Dravidian Language) further and help in the future development of an engine that allows speech to be detected in English and another Dravidian language. As a native speaker of English and Tamil, the two languages chosen for my project are English and Tamil. ASRs consist of the Acoustic Model, Language Model and Pronunciation Lexicon. This report will investigate how we can train better Language Models to improve the outputs. To solve this, we also aim to expand the text corpus by not only collecting data but also generating data for the text corpus