quran-asr-small

A 33M-parameter Quran recitation recognizer that runs offline on a phone (36 MB int8 ONNX, sherpa-onnx). It writes the Arabic letters of what is recited, for professional reciters and ordinary people, Hafs and Warsh, slow tartil to fast hadr. Big brother of quran-asr-tiny.

4.49% WER · 2.17% CER on the Quran-Lab benchmark, #2 of 33 on the leaderboard at submission. None of the benchmark's reciters or recordings were used for training.

Results

Quran-Lab benchmark, official score.py, 600 clips. WER / CER in %.

Model	Params	Overall	Phone	EveryAyah	QUL
zipformer_p-arabic-v3 (retrieval)	66M	3.46 / 1.84	7.56 / 1.79	0.00 / 0.00	4.40 / 3.77
quran-asr-small (int8)	33M	4.49 / 2.17	10.45 / 2.84	0.69 / 0.20	4.61 / 3.80
fastconformer-quran	115M	5.41 / 2.96	8.88 / 2.66	0.93 / 0.26	7.78 / 5.91
NVIDIA FastConformer	115M	7.44 / 3.08	16.77 / 5.61	1.50 / 0.49	7.61 / 4.20
Whisper large-v3	1.5B	10.10 / 4.26	26.23 / 12.39	5.48 / 1.79	4.69 / 1.83

Reciters never seen in training (our test set, 1,000 clips): 0.30% CER, 0.46% WER.

Notes

• The model writes Uthmani spelling. For the benchmark, output words are rewritten in standard spelling with a fixed table (spelling.json) that never changes a misrecognized word. The phone references are themselves Uthmani, where raw output scores 3.29 / 1.16.
• One phone clip of the benchmark (36:71) turned out to be the same recording as one in our training data, uploaded twice under different ids; it is excluded from future training.
• Greedy CTC, no language model, no lexicon, no hints.

Usage

pip install sherpa-onnx soundfile numpy scipy
python transcribe.py recitation.wav              # Uthmani letters
python transcribe.py recitation.wav --standard   # standard spelling

On Android / iOS the model loads as a sherpa-onnx zipformer CTC model (model.int8.onnx + tokens.txt).

Model and training

• Conformer-CTC: 9 blocks, d=384, 6 heads, ×4 subsampling, 31 output tokens (Arabic letters + word separator). InterCTC after block 5 during training.
• Data (1,532 h): 36 EveryAyah Hafs reciters (1,132 h), 3 Warsh reciters (96 h), tarteel-ai/tlog phone recordings (283 h, 200 shards, labels checked against the audio), Iqra'Eval Quranic sentences (21 h). Ordinary people are 30% of each epoch.
• 343k steps, 20 epochs, ~33,000 h of audio seen, ~37 h on one laptop GPU (RTX 5070 Ti Laptop). AdamW 4e-4, cosine, bf16; speed, tempo (×0.75 to ×2.5), noise, reverb, low-pass and SpecAugment.
• Released checkpoint: step 310k (lowest validation error, 0.84% CER).

Training code: code/ (config/base.yaml).

License

MIT. Training data keeps its own terms (tlog: CC-BY-4.0).