metadata
title: BuzzASR
emoji: 🐝
colorFrom: yellow
colorTo: red
sdk: static
pinned: false
license: mit
🐝 BuzzASR — A Swarm of 100+ Monolingual Speech Recognition Models
One specialist model per language. BuzzASR is a suite of 102 monolingual ASR models, each a fine-tune of Whisper-large-v3 specialized to a single language.
The work has been accepted at EMNLP 2026 · built by Lemn Lab in collaboration with EleutherAI.
The idea
Big multilingual models spread themselves thin across hundreds of languages. We asked whether a single specialist, the same size as Whisper, could beat the giant generalists. It can to a considerable extent :)
Highlights
- ✅ Beats Whisper-large-v3 on 89 of 102 languages (FLEURS)
- 🏆 Open-source state-of-the-art (beats Whisper, Omnilingual 1B/7B, MMS, Qwen3-ASR, Cohere) on 27 languages
- 📉 Cuts character error rate ~3x on average, and 3.45x across Whisper's 51 worst languages
- 🧩 Each model is the same 1.55B parameters as Whisper — about 4.5x smaller than the strongest baseline (Omni-7B)
A few standouts (combined FLEURS + Common Voice, normalized CER %)
| Language | BuzzASR CER | Whisper zero-shot | Note |
|---|---|---|---|
| Cantonese | 13.0 | 36.4 | SOTA — 2x better than the next-best system |
| Punjabi | 8.8 | 42.3 | SOTA |
| Mongolian | 5.2 | 38.2 | SOTA |
| Korean | 4.6 | 5.7 | SOTA |
| Amharic | 9.1 | 191 | >20x reduction over Whisper |
Find your language
Browse all 102 models in the Models tab above, or go to huggingface.co/BuzzASR/<language>.
Usage
from transformers import WhisperForConditionalGeneration, WhisperProcessor
model = WhisperForConditionalGeneration.from_pretrained("BuzzASR/mongolian")
proc = WhisperProcessor.from_pretrained("BuzzASR/mongolian")
# the language/task prompt is baked in — just call model.generate(input_features)
Links
- 📄 Paper — Findings of EMNLP 2026 (https://arxiv.org/abs/2609.09554)
- 🌐 Project & docs — https://lemn-lab.github.io/buzz-asr/