Automatic speech recognition for Telugu: a comparative analysis of Wav2Vec 2.0 model variants and hyperparameter tuning

Automatic speech recognition (ASR) systems for the Telugu language can significantly enhance speech-based interaction and digital accessibility. Although Telugu is among the top languages in the Dravidian family in terms of the number of speakers, there is a notable absence of standardized, noise-robust resources for Telugu ASR development. The existing systems are not specifically designed for low-resource languages like Telugu. The datasets available for Telugu are restricted to controlled environments and do not include noise interferences. Most existing ASR models are optimized for high-resource languages such as English, leading to poor generalization when applied to Telugu. In this work, we fine-tuned three pre-trained variants of the Wav2Vec 2.0 model: XLS-R 300M, XLSR-53, and XLS-R-1B on a curated 48.4-h Telugu speech corpus compiled from Mozilla Common Voice, OpenSLR, and the Hugging Face dataset. This study utilized a comprehensive ablation study on four hyperparameters to evaluate their recognition performance using word error rate (WER) and character error rate (CER) as evaluation metrics. Among the three models, the fine-tuned XLS-R-1B performed the best, achieving a WER of 36.23% and CER of 15.44%, followed by the XLSR-53 (WER 37.78%, CER 15.81%) and XLS-R-300M (WER 37.87%, CER 15.81%). An important consideration of the research was finding the right trade-off between a model's complexity and the accuracy of its performance. The results showed that the size of the model is important, and hyperparameter selection significantly influences the development of an effective ASR model in low-resource languages such as Telugu.