GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
Dr-AliGomaa
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Dr-AliGomaa
Model Tree
openai/whisper-large-v3
Input Modalities
Output Modalities
Supported Functionality
Normalized WER/CER (%) on the frozen validation splits. Every system is scored with one
identical Arabic normalizer applied to hypotheses and references —
eval/asr_score.py.
| System | Quran WER | Quran CER | Hadith WER | Hadith CER | MGB-2 WER | MGB-2 CER |
|---|---|---|---|---|---|---|
| whisper-large-v3 (zero-shot) | 3.50 | 0.81 | 7.49 | 2.22 | 10.85 | 4.76 |
| Cohere Transcribe Arabic | 3.96 | 2.73 | 3.99 | 1.46 | 6.77 | 3.60 |
| this model | 0.33 | 0.10 | 3.60 | 0.99 | 9.52 | 2.92 |
Split sizes: Quran 7,280 clips · Hadith 4,752 · MGB-2 494.
On sacred text this is the state of the art: 0.33 vs 3.96 WER on Quran against the strongest open Arabic model — a ≈12× relative reduction — and 3.60 vs 3.99 on Hadith. Relative to its own zero-shot base: −91 % Quran, −52 % Hadith, −12 % MGB-2.
On general broadcast MSA we are competitive, not leading. Cohere reaches 6.77 WER to our 9.52 — but on CER we are ahead, 2.92 vs 3.60, meaning our errors are shallow near-misses rather than dropped or wrong words. MGB-2 is strictly zero-shot for this model and may be in-domain for the others.
python
import torchfrom transformers import pipelinepipe = pipeline("automatic-speech-recognition",model="Dr-AliGomaa/whisper-large-v3-ar",torch_dtype=torch.float16,device="cuda:0",chunk_length_s=30,)# The decoding config the reported numbers use -- fixed, not tuned per dataset.gen = {"language": "arabic","task": "transcribe","num_beams": 5,"temperature": (0.2),"condition_on_prev_tokens": False,"compression_ratio_threshold": 1.35,"logprob_threshold": -1.0,"return_timestamps": False,"max_new_tokens": 444,}print(pipe("audio.mp3", generate_kwargs=gen)["text"])
This model does not emit timestamps. For time-aligned output use
whisper-large-v3-ar-eg-timestamps.
1. Run the audio pipeline first. The model was trained on optimized audio: silence-trimmed,
loudness-normalized to −16 LUFS, edge-padded 100 ms, gain-corrected, segmented ≤ 30 s. Applying
the same chain at inference is required to reproduce these numbers — it is the single most
common reason people can't. Scripts:
pipeline/.
2. Score with eval/asr_score.py. Arabic WER moves materially with the normalizer, so a
number computed under a different one is not comparable to the table above.
| Base | openai/whisper-large-v3 |
| Trained on | Quran + Hadith + Common Voice 17 + FLEURS + MediaSpeech. No Egyptian, no MGB-2. |
| Learning rate | 5e-6, linear decay with warmup |
| Batch | 4 per device × 8 GPUs = 32 effective |
| Precision / distributed | bf16 + tf32, DeepSpeed ZeRO |
| Weight decay | 0.05 (no dropout, no label smoothing) |
| Augmentation | waveform (MUSAN + ESC-50 + Gaussian, band-pass) via audiomentations; Quran/Hadith oversampled through independently-augmented passes |
Full recipe: training/training.py.
bibtex
@misc{kotb2026quranhadith,title = {A Quran and Hadith Speech Resource and Benchmark for Arabic ASR,with Professional-Reciter Training and Validation},author = {Mohamed Kotb},year = {2026},publisher = {Zenodo},doi = {10.5281/zenodo.21927416},url = {https://doi.org/10.5281/zenodo.21927416},note = {Preprint}}
| Frameworks | Transformers 4.51.3 · PyTorch 2.5.1+cu121 · Datasets 3.6.0 |