You can deploy this model and start earning money today!

dev-ahmedhany/
whisper-large-v3-arabic-ft-v3

Name: whisper-large-v3-arabic-ft-v3
Rating: 2.5 (5 reviews)

transformers

safetensors

whisper

automatic-speech-recognition

arabic

multi-dialect

openai/whisper-large-v3

finetune

apache-2.0

model-index

Views

No views yet

Whisper-large-v3 — Arabic 4-dialect FT v3 (merged HF safetensors)

The float32 PyTorch safetensors version of openai/whisper-large-v3 (1.55 B base) fine-tuned with QLoRA on 38 hours of dialect-balanced Arabic. The "v3" of the project — bigger base + bigger data than the v2 turbo variant.

This is the transformers-ready merged model. For production CPU inference, prefer the int8 CT2 build at whisper-large-v3-arabic-ft-v3-ct2-int8 (1.6 GB, real-time on commodity CPU). For further fine-tuning or to inspect the training trajectory, use the LoRA adapter repo which has every save during training as a Git revision.

Headline WER (n=100/dialect, eval at int8/beam=2/threads=8)

50% Casablanca + 50% broadcast for Egyptian/Levantine; 100% Casablanca UAE for Gulf; 100% FLEURS for MSA. Same exact recordings + decoding config used for both rows.

Dialect	Test composition	Zero-shot Whisper-large-v3	This model (v3-ft)	Δ
MSA	FLEURS broadcast	8.51%	10.52%	+2.01 pp
Egyptian	50 Casablanca + 50 MGB-3	38.48%	23.90%	−14.58 pp ✅
Levantine	50 Casablanca JO + 50 MASC	37.70%	30.63%	−7.07 pp ✅
Gulf	Casablanca UAE	52.72%	41.46%	−11.26 pp ✅
avg-4		34.35%	26.63%	−7.72 pp ✅

v3-ft beats zero-shot Whisper-large-v3 by 7.72 pp average WER, with double-digit gains on Egyptian and Gulf. MSA loses 2.01 pp — known dialect-vs-MSA tradeoff that v4 (planned) addresses.

Quickstart with transformers

python

from transformers import pipeline

asr = pipeline("automatic-speech-recognition",
               model="dev-ahmedhany/whisper-large-v3-arabic-ft-v3",
               torch_dtype="bfloat16", device_map="cuda")
out = asr("audio.wav", generate_kwargs={"language": "ar", "task": "transcribe"})
print(out["text"])

For CPU production deployment, use the CT2 int8 build instead — same weights, ~5× faster on CPU.

Training recipe

Base: openai/whisper-large-v3 (1.55 B params)
QLoRA: NF4 + bf16, r=8, α=16, dropout 0.05, target modules q/v/k/out_proj + fc1/fc2
Optimizer: paged_adamw_8bit, lr=1e-4, warmup 0.1
Effective batch 16 (per-device 4 × grad-accum 4), gradient_checkpointing
Best checkpoint: step 4750 of 10000 (early-stop after plateau across ckpts 5000-6000)
Training data: ~38 h, 26,817 train / 924 val rows
- MSA (9.6 h): Common Voice 18 Arabic (capped 15 h)
- Egyptian (18.6 h): Casablanca Egypt + MGB-3 + cleaned MGB-3
- Levantine (9.8 h): Casablanca Jordan + Palestine + MASC
- Gulf (1.9 h): Casablanca UAE + Yemen
Maghrebi excluded (84.7% zero-shot WER too far gone for QLoRA budget)
Compute: ~25-28 h on a single GCP L4

Limitations

Maghrebi out of scope — zero-shot-quality (≥80% WER) on Algerian/Moroccan/Tunisian dialects.
MSA: zero-shot Whisper-large-v3 is slightly better. If your traffic is overwhelmingly broadcast Arabic, plain zero-shot is fine.
Gulf test is Casablanca-only — no public broadcast Gulf corpus available.

Citation

bibtex

@misc{hany2026whisperarabic,
  title        = {Production-Aware Fine-Tuning of Whisper Variants for Multi-Dialect
                  Arabic ASR: A Cross-Platform CPU Inference Study},
  author       = {Hany, Ahmed},
  year         = {2026},
  howpublished = {Preprint, arXiv (in preparation)},
  url          = {https://github.com/dev-ahmedhany/whisper-arabic-dialects},
}

License

Apache-2.0 (inherits from openai/whisper-large-v3).