A Systematic Approach to Derive a Refined Speech Corpus for Sinhala

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5107–5113 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5107 A Systematic Approach to Derive a Refined Speech Corpus for Sinhala Disura Warusawithana, Nilmani Kulaweera, Lakshan Weerasinghe, Buddhika Karunarathne Department of Computer Science and Engineering University of Moratuwa Sri Lanka {disurawaru.17, rukshilakulaweera.17, lakshan.17}@cse.mrt.ac.lk, buddhikak@uom.lk Abstract Speech Recognition is an active research area where advances of technology have continuously driven the development of research work. However, due to the lack of adequate resources, certain languages such as Sinhala, are left to underutilize the technology. With techniques such as crowdsourcing and web scraping, several Sinhala corpora have been created and made publicly available. Despite them being large and generic, the correctness and consistency in their text data remain questionable, especially due to the lack of uniformity in the language used in the different sources of web scraped text. Addressing that requires a thorough understanding of technical and linguistic particulars pertaining to the language, which often leaves the issue unattended. We have followed a systematic approach to derive a refined corpus using a publicly available corpus for Sinhala speech recognition. In particular, we standardized the transcriptions of the corpus by removing noise in the text. Further, we applied corrections based on Sinhala linguistics. A comparative experiment shows a promising effect of the linguistic corrections by having a relative reduction of the Word-Error-Rate by 15.9%. Keywords: speech recognition, Sinhala speech corpora, refined corpus 1. Introduction With the advent of artificial intelligence (AI) in computer science, the conventional techniques of human-computer interaction (HCI) have been redefined with more natural human behaviours. Among such techniques, conversational interfaces driven by speech recognition have gained great attention, since speech is the most natural way of human communication (Gunasekara and Meegama, 2015). Since the 1950s (Malik et al., 2021), research and development in this area has evolved over the years and has produced many applications powered by the speech recognition technology. Examples of such applications include virtual assistants such as Google Assistant, Apple Siri, and Amazon Alexa (Karunathilaka et al., 2020) which are widely used in everyday life. A typical automatic speech recognition (ASR) system comprises three main models, namely, acoustic model, language model, and pronunciation model. The models use a probabilistic approach to collaboratively decode a sequence of text that matches to a sequence of features extracted from the audio input. Having an abundance of resources, widely used languages such as English are leading in the development of speech recognition. On the contrary, languages used by relatively smaller populations remain as low-resourced in this field, making them still remain in their early stage of development. Sinhala, a language used by the majority of the population in Sri Lanka, is recognized as such a language (de Silva, 2021; Dilshani et al., 2018). Being a member of the Indo-Aryan language family, Sinhala is a phonetically rich language and has a large lexical diversity (Karunathilaka et al., 2020). Therefore, developing an ASR system …

Content truncated to 500 words