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

Area of Science:

• Computational Linguistics
• Speech Processing
• Artificial Intelligence

Background:

• Automatic Speech Recognition (ASR) systems are crucial for digital accessibility and interaction.
• Telugu, despite its speaker base, lacks standardized, noise-robust ASR resources, hindering development.
• Existing ASR models often perform poorly on low-resource languages like Telugu due to optimization for high-resource languages.

Purpose of the Study:

• To fine-tune pre-trained Wav2Vec 2.0 models for Telugu ASR.
• To evaluate the performance of different model variants (XLS-R 300M, XLSR-53, XLS-R-1B) on a Telugu speech corpus.
• To investigate the impact of hyperparameter tuning on ASR accuracy for low-resource languages.

Main Methods:

• Fine-tuning three Wav2Vec 2.0 variants (XLS-R 300M, XLSR-53, XLS-R-1B) on a 48.4-hour Telugu speech corpus.
• Utilizing a curated dataset from Mozilla Common Voice, OpenSLR, and Hugging Face.
• Conducting an ablation study on four hyperparameters and evaluating performance using Word Error Rate (WER) and Character Error Rate (CER).

Main Results:

• The fine-tuned XLS-R-1B model achieved the best performance with a WER of 36.23% and CER of 15.44%.
• XLSR-53 and XLS-R-300M also showed competitive results, with WERs of 37.78% and 37.87% respectively.
• Hyperparameter selection and model size were identified as significant factors influencing ASR performance in Telugu.

Conclusions:

• The XLS-R-1B model demonstrates superior effectiveness for Telugu ASR among the tested Wav2Vec 2.0 variants.
• Optimizing hyperparameters and considering model complexity are critical for developing accurate ASR systems for low-resource languages.
• This research contributes valuable insights and resources for advancing Telugu ASR technology.