VITB-HEBiC: A bilingual corpus for evaluating ASR in diverse Indian code-switching scenarios

Abstract

Code-switching is a common form of communication in every bilingual or multilingual culture. It has become common in an ethnically diverse country like India, where multiple languages are spoken within the society. The need for automatic speech recognition (ASR) systems to handle code-switching has recently increased. However, there are currently relatively few code-switching resources accessible to train these kinds of systems. Therefore, it is highly essential to create code-switching resources. This paper explains the creation of a bilingual Hindi-English corpus named HEBiC corpus at VIT Bhopal University. This corpus has 7.5 hours of read speech data and has 58,245 words. The corpus is highly diverse and includes speakers from 27 states of India, each exhibiting distinct accents and dialects. The corpus also includes 600 sentences with a vocabulary size of 3575 words. The sources and methods used to gather the corpus and the statistical analysis are described in detail in this paper. We have evaluated the corpus performance with baseline and end-to-end models in terms of word error rate (WER).