BuzzASR — Urdu

A monolingual automatic speech recognition model for Urdu, fine-tuned from openai/whisper-large-v3. Part of BuzzASR, a suite of 102 language-specialized ASR models (Findings of EMNLP 2026).

This model uses simple fine-tuning (Whisper's tokenizer, ASR fine-tuning only).

Results (normalized CER / WER, %)

Test set	CER	WER	Whisper-large-v3 (zero-shot) CER
FLEURS	7.5	21.8	8.08
Combined	9.24	26.77	8.08

~0.9x CER reduction over Whisper zero-shot on the combined test set.

Usage

import torch, torchaudio
from transformers import WhisperForConditionalGeneration, WhisperProcessor

model = WhisperForConditionalGeneration.from_pretrained("BuzzASR/urdu", torch_dtype=torch.float16).to("cuda").eval()
proc  = WhisperProcessor.from_pretrained("BuzzASR/urdu")

wav, sr = torchaudio.load("audio.wav")           # 16 kHz mono
feats = proc(wav[0], sampling_rate=16000, return_tensors="pt").input_features.to("cuda").half()
ids = model.generate(feats, num_beams=1, no_repeat_ngram_size=3, repetition_penalty=1.2)
print(proc.batch_decode(ids, skip_special_tokens=True)[0])

The language/task prompt is baked into the generation config, so no language= argument is needed.

Training data

FLEURS, capped per the paper. Text-only data from the Goldfish corpus (Chang et al., 2026).

Limitations

Monolingual (Urdu only). Evaluated on FLEURS / Common Voice test splits; other domains or dialects may differ.

Citation

Project page: https://lemn-lab.github.io/buzzasr-docs/