On the Development of HindiEnglish Code Switching Speech Recognition Systems and Corpus

Abstract

quotCode-switching refers to the alternate use of two or more languages (or dialects) during the conversation. This phenomenon has been observed in many multilingual communities across the globe. Therefore, handling code-switching by the spoken input systems is very much required for e cient human-machine interaction. However, due to the lack of domain-speci c resources, the research in this domain is somewhat limited compared to the monolingual case. This thesis aims to address the acoustic and language modeling challenges in code-switching automatic speech newlinerecognition (ASR) tasks. In addition to that, a Hindi-English code-switching corpus has been created towards addressing the data scarcity issue. newlineThe early works on code-switching ASR happen to employ the hybrid framework typically developed for the monolingual case. The created Hindi-English code-switching corpus is rst evaluated in the hybrid framework. The hybrid framework comprises of three sub-modules, namely, a pronunciation model, an acoustic model, and a language model. The end-to-end (E2E) framework has recently emerged as a viable alternative to the hybrid systems in the ASR domain. Unlike the hybrid framework, the E2E framework does not require the phonetically labeled training data, and also does not include any explicit pronunciation model. In the case of code-switching ASR, for multiple languages being involved, these attributes become more attractive. Motivated by that, in this thesis, the E2E framework has been explored for developing the code-switching ASR systems.quot