whisper-small-ardalani-sine
Whisper-small fine-tuned for Kurdish Ardalani, the Central Kurdish variety of Sanandaj, Iran. To our knowledge the first ASR model specialized for this dialect.
Training
Two stages. Stage 1: 8,000 steps on 104 hours of Central Kurdish (CORDI corpus, all varieties) with Sanandaj utterances oversampled 3x; lr 1e-5, effective batch 32, fp16. Stage 2: 2,000 further steps on Sanandaj-only data (15,305 utterances, 10.6 hours); lr 5e-6; checkpoint selection by Sanandaj-only dev WER. Whisper has no Kurdish token, so the Arabic language token is used as a stand-in at training and inference.
Evaluation
Frozen held-out benchmark: 1,758 Sanandaj utterances (1.12 hours), three whole episodes never seen in training. Scoring normalizes punctuation and Arabic yeh/kaf on both sides.
| Model | WER | CER |
|---|---|---|
| Whisper small, zero-shot | 373.2 | 423.0 |
| Whisper large-v3, zero-shot | 181.4 | 131.8 |
| Best public Sorani fine-tune | 133.0 | 103.6 |
| this model, plain decoding | 74.9 | 56.6 |
| this model, capped decoding | 59.6 | 31.8 |
Capped decoding limits generated tokens to 6 + 18 per second of audio (floor 10) and is the recommended deployment setting; it suppresses repetition loops on sub-2-second clips. Clips of 5 seconds and longer score 47.6 WER.
How to use
from transformers import pipeline
asr = pipeline(
"automatic-speech-recognition",
model="shahotoofani/whisper-small-ardalani-sine",
generate_kwargs={"language": "arabic", "task": "transcribe"},
)
print(asr("audio.wav")["text"])
For short clips, cap max_new_tokens at about 6 + 18 * seconds.
Limitations
- Training and test data come from one TV series; scores measure dialect adaptation, not speaker independence.
- Acted dialogue domain; spontaneous speech is harder.
- Sub-2-second clips remain the weakest case; a residual tail of repetition loops exists even with capped decoding.
- Verbatim-dialectal transcripts in Sorani script; the model preserves Ardalani forms rather than normalizing to standard Sorani.
Data and license
Trained on CORDI (CC BY-SA 4.0): Sina Ahmadi, Daban Jaff, Md Mahfuz Ibn Alam, Antonios Anastasopoulos, "Language and Speech Technology for Central Kurdish Varieties", LREC-COLING 2024. Model weights released under CC BY-SA 4.0.