Telugu Language Low-resource ASR Fine-Tuning with Wav2Vec2-XLS-R-300 and Interface Development

The development of Automatic Speech Recognition (ASR) systems for low-resource languages is a challenging task because of the limited amount of annotated speech data and linguistic variation. This work is concerned with the fine-tuning and development of a Telugu language ASR system based on Wav2Vec2-XLS-R-300, a self-supervised multilingual model developed for low-resource languages. Because of Telugu’s phonetic richness and dense morphological structure, training ASR models poses the need for efficient management of acoustic and linguistic variations. This project utilizes transfer learning by fine-tuning Wav2Vec2-XLS-R-300 on the Microsoft Speech Corpus for Telugu, tuning model parameters to maximize recognition accuracy. The training pipeline involves dataset preprocessing, feature extraction, augmentation, and hyperparameter tuning to enhance generalization. Moreover, an easy-to-use web-based interface is implemented to support real-time speech-to-text conversion, allowing for easy user interaction. The framework is developed to support varied speech inputs with accent, dialect, and background noise variations, rendering it applicable in educational, accessible, and online communication contexts. Experimental experiments show performance improvement in Word Error Rate (WER) and Character Error Rate (CER) over baseline models available, illustrating the potential of self-supervised learning in low-resource ASR development. This work adds to the general cause of improvement of multilingual speech recognition and reducing the digital divide for Indian languages.