NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·Jun 19·5 min read
A Comparative Study of Pretrained Transformer Models for Quranic ASR: Speech Representations, Label Formats, and Dataset Composition
TL;DR
Existing automatic speech recognition (ASR) systems for Quranic Arabic face challenges due to unique speech characteristics and limited datasets. This study evaluates various pretrained transformer models, including Wav2Vec 2.0 and HuBERT, to improve ASR performance.
✦ Why It Matters
Engineers can leverage pretrained transformer models to enhance ASR systems for low-resource languages like Quranic Arabic.
Key Takeaways
- Explore pretrained transformer models for ASR applications.
- Consider fine-tuning on domain-specific datasets for better accuracy.
- Evaluate different speech representations and label formats.
- Aim for significant reductions in word error rates in ASR.
Full Summary
Automatic speech recognition (ASR) for Quranic Arabic is hindered by the language's unique phonetic features and the scarcity of relevant training data. This study investigates the effectiveness of pretrained transformer models, specifically Wav2Vec 2.0 and HuBERT, in transcribing Quranic audio.
Researchers fine-tuned these models on a curated dataset of Quranic recitations, employing various speech representations and label formats. The evaluation metrics showed that fine-tuning led to a substantial improvement in transcription accuracy, with a word error rate reduction of up to 92%.
These findings suggest that leveraging advanced transformer architectures can significantly enhance ASR systems for low-resource languages. The implications extend to improving accessibility and usability of Quranic texts in digital formats.