NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·2h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
Subtitle-Aligned Fine-Tuning of Whisper for Swiss German ASR: Benchmark Contamination, Convention Mismatch, and an Honest Baseline at 25.6% WER (13.8% cWER)
TL;DR
Swiss German automatic speech recognition (ASR) faced challenges due to a lack of suitable training data. Fine-tuning OpenAI's Whisper large-v3 model using 1,367 hours of broadcast speech with Standard German subtitles was conducted.
✦ Why It Matters
Engineers can leverage these findings to improve ASR systems for low-resource languages by focusing on data quality and training methods.
Key Takeaways
- Utilize the released Whisper models for Swiss German ASR tasks.
- Be aware of benchmark contamination affecting ASR performance metrics.
- Explore both LoRA and full fine-tuning methods for model training.
- Consider content WER for a more accurate assessment of ASR quality.
How It Works
The fine-tuning process involved using a large dataset of broadcast speech and aligning it with Standard German subtitles. This weak supervision allowed the model to learn from imperfect data while focusing on dialectal nuances.
The study employed advanced training techniques on a high-performance computing system to optimize the model's performance.