We benchmarked all 19 WAXAL ASR languages — here's what we found
#22
by Professor - opened
👋 Hi WAXAL community!
We've just released the first systematic ASR benchmark across all 19 WAXAL languages — evaluating Whisper Large-v3, MMS-1B, and Omnilingual-1B (zero-shot) against fine-tuned edge models (Whisper Tiny, Whisper Small, MMS-300M).
🔗 Project: https://waxalnet.vercel.app/
📄 Paper: https://arxiv.org/abs/2606.02375
🤗 Models: https://huggingface.co/waxal-benchmarking
Key findings
- Fine-tuned edge models (39M–300M params) achieve 38.0% macro-avg WER vs 64.9% for the best zero-shot baseline — using models 3–40× smaller
- Whisper Large-v3 natively supports only 4 of the 19 languages
- Architecture follows language family: MMS-300M wins on all Bantu languages; Whisper Small wins on Afro-Asiatic
- WER alone misrepresents performance for Ge'ez-script languages (Amharic, Tigrinya) — CER tells a substantially different story
All 57 fine-tuned checkpoints, evaluation code, and a cleaned test subset are released under CC-BY 4.0. Built with 32 contributors and native speakers across all 19 language communities.
Happy to answer any questions! 🙌