whisper-yemeni

Whisper-Yemeni — Yemeni Arabic ASR (AYDID)

openai/whisper-large-v2 fine-tuned on the AYDID corpus for Yemeni Arabic speech recognition across seven sub-dialects (Adeni, Badawi, Hadrami, Sana'ani, Ta'izzi, Tihami, Standard Yemeni).

• Base model: openai/whisper-large-v2
• Dataset: mansoorSaleh/AYDID
• Task: automatic speech recognition

Result

In-domain fine-tuning cuts normalized WER+ from the best zero-shot system's 43.92% to 19.14% overall — a 24.8-point reduction. Per sub-dialect:

WER+ = normalized (punctuation and diacritics removed). Hadrami is hardest, consistent with its distinctive /q/→/g/ realisation; Standard Yemeni is easiest, reflecting MSA proximity.

Usage

import librosa
from transformers import WhisperProcessor, WhisperForConditionalGeneration

model_id = "mansoorSaleh/whisper-yemeni"
processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(model_id).eval()

wav, _ = librosa.load("clip.wav", sr=16000)
inputs = processor(wav, sampling_rate=16000, return_tensors="pt")
ids = model.generate(inputs.input_features, language="ar", task="transcribe")
print(processor.batch_decode(ids, skip_special_tokens=True)[0])

Citation

@inproceedings{aydid2027,
  title     = {AYDID: A Sub-Dialectal Yemeni Arabic Corpus for Dialect
               Identification and Speech Recognition},
  author    = {Ba Mahel, Mansoor S. M. and Wei, Jianguo and Yue, Xianghu and
               Awn, Norah Saeed and Bamahel, Abdulaziz S.},
  booktitle = {Proc. IEEE ICASSP},
  year      = {2027}
}

License: CC BY-NC-SA 4.0 (research use).