BuzzASR
100+ monolingual speech recognizers, one per language¶
We fine-tune Whisper-large-v3 into a separate model for each of 102 FLEURS languages, with two recipes: plain fine-tuning (SFT) and full fine-tuning with a native tokenizer (FFT). This site is a short technical overview of the project: the two recipes, why the tokenizer matters, and what the models achieve.
How to read the SOTA counts
"Lowest CER among open systems" means the best of SFT or FFT (picked per language on validation, never on test) beats Omni-1B/7B, MMS-1B, Qwen3-ASR, and Cohere on CER. We report it on FLEURS-test (31) and on the combined FLEURS + CommonVoice set (41).
The two recipes¶
FFT's advantage comes from its tokenizer. Whisper's tokenizer was built for English-heavy data, so it splits other languages into many small pieces. FFT trains a native tokenizer that keeps whole morphemes. That one change makes FFT more accurate, and it also makes decoding faster.
Where to start¶
-
Technical overview
What BuzzASR is, and how the two recipes (SFT and FFT) differ.
-
The tokenizer
Why a native tokenizer helps, and a bug worth knowing about.
-
Results
How accurate the models are, and why FFT decodes faster.
-
Models
Per-language weights on Hugging Face, released as they are finalized.
Model weights are being released on Hugging Face as the per-language models are finalized. See Models.
Acknowledgments¶
Developed in collaboration with EleutherAI, whose compute and support made these training runs possible.