Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 15·5 min read
Comparative Analysis of Multilingual Pre-trained Models for Nepali Automatic Speech Recognition
TL;DR
A benchmark comparing six multilingual pre-trained models for Nepali Automatic Speech Recognition (ASR) reveals that Whisper-Large-v3-Turbo and IndicWav2Vec achieve similar performance despite significant differences in size and training data. CTC decoders demonstrate superior speed, making them preferable for real-time applications.
✦ Why It Matters
Choose CTC decoders for faster Nepali ASR applications to enhance real-time performance.
Key Takeaways
- Consider Whisper-Large-v3-Turbo for high accuracy in Nepali ASR.
- Use CTC decoders for faster processing in real-time applications.
- Evaluate model performance using standardized benchmarks for better comparisons.
- Leverage multilingual pretraining to improve robustness in ASR tasks.
How It Works
The study fine-tunes various ASR models using a consistent protocol, including identical preprocessing and learning-rate schedules, to ensure fair comparisons. The models utilize different architectures, such as CTC self-supervised and autoregressive encoder-decoder, to process audio data effectively.