A Novel Approach for Bootstrapping and Automatic Transcription of Low Resourced Language Speech Corpus

Automatic Speech Recognition (ASR) systems have made significant advancements in the context of high-resource languages, primarily attributable to the abundant availability of extensive and diverse speech datasets. Nevertheless, the dearth of annotated data remains a substantial hurdle when it comes to low-resource languages. This study delves into the feasibility of development of an ASR system for low-resource languages by leveraging pre-trained models from other languages. The fine-tuned model is then deployed to transcribe speech segments from news bulletins and audio content found on the web. Subsequently, the generated transcript is heuristically aligned with existing news script. This newly aligned speech corpus is used incrementally to augment the existing corpus, and thus progressively bootstrapping the ASR models. The proposed work has been effectively carried out for Dogri, a low resource language of India. The proposed approach of incremental learning and data augmentation can be applied to other low resource languages as well, and thus would help in bridging the resource gap.