AI & ML interests

Automatic Speech Recognition, Supervised Fine-Tuning, 102 Language Support

Recent Activity

vamsin07  updated a Space 1 day ago
BuzzASR/README
vamsin07  updated a model 20 days ago
BuzzASR/korean
View all activity

Organization Card

🐝 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.

Accepted at EMNLP 2026 · Shivam Singh, Aditya Yadavalli, Catherine Arnett, Alex Warstadt · LeMN Lab, UC San Diego 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. In most languages, it can.

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
  • ⚡ Native per-language tokenizers: 2.38 → 5.04 characters per token, about 1.6x faster decoding
  • 🧩 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, independently re-scored
Amharic 9.1 191 >20x reduction over Whisper

Against a production API

We ran the full FLEURS test sets through Whisper-large-v3-turbo on Cloudflare Workers AI and through BuzzASR (CER %, lower is better):

Language Workers AI turbo BuzzASR
Amharic 97.9 11.5
Sorani Kurdish* 71.4 12.8
Maltese 21.0 5.1
Estonian 3.4 2.9
Hungarian 3.5 3.2

*Whisper has no Sorani option, so turbo had to detect the language.

In the wild

  • 39,000+ downloads since release
  • Community versions for phones and laptops: GGUF, MLX, Core ML, whisper.cpp
  • Independent evaluations from users in Korea and Hungary

Find your language

Browse all 102 models in the Models tab above, or go to huggingface.co/BuzzASR/<language>.

Usage

from transformers import pipeline

asr = pipeline("automatic-speech-recognition", model="BuzzASR/mongolian")
print(asr("audio.wav")["text"])

The language and task prompt is built into each model, so there's nothing else to set.

Links

datasets 0

None public yet