Speech command recognition using self-supervised asr in Tamil

Abstract

ASR technology has made significant advancements, approaching performance levels comparable to humans. However, developing ASR models for all languages, especially Low Resource Languages (LRLs), presents challenges due to limited resources. Recent studies on intent classification in LRLs employ transfer learning techniques, leveraging state-of-the-art English ASR models to improve performance in these domains. Additionally, the emerging trend of self-supervised learning has proven advantageous in developing sophisticated ASR models, requiring less labeled data to achieve high performance. In our research, we utilized a self-supervised ASR model to classify intent in an LRL, specifically the Tamil language. We compared two methods for the same ASR framework: a fine-tuning approach and a transfer learning approach with a limited amount of labeled data in Tamil. Our findings indicate that the fine-tuning method outperforms the transfer learning technique. Moreover, the model exhibited a noteworthy increase in accuracy compared to the established phoneme-based speech intent classification methodology in Tamil. This study represents a significant step forward in enhancing speech recognition capabilities for LRLs.