mutqin-asr

Mutqin ASR — sherpa-onnx export of Tarteel's Qur'an Whisper models

ONNX (int8) exports of Tarteel AI's Qur'an-fine-tuned Whisper checkpoints, in the layout sherpa-onnx expects, for on-device recitation checking in the Mutqin app. Converted with tools/asr/convert.py.

Attribution and licence

The models are derived from Tarteel AI's checkpoints, licensed Apache-2.0:

• tarteel-ai/whisper-tiny-ar-quran
• tarteel-ai/whisper-base-ar-quran

These exports are redistributed under the same Apache-2.0 licence. Weights are unchanged — only remapped to the openai-whisper layout, exported to ONNX and dynamically quantized to int8. All credit for the fine-tuning is Tarteel's.

Layout

models/quran-tiny/{encoder.int8.onnx,decoder.int8.onnx,tokens.txt}   # ~99 MB
models/quran-base/{encoder.int8.onnx,decoder.int8.onnx,tokens.txt}   # ~153 MB

The decoder dominates: sherpa's export quantizes MatMul only, so the 51865-row token embedding stays float32. Quantizing Gather too makes it larger — Whisper reuses that embedding for the output projection, so an int8 copy is added rather than substituted.

Accuracy

WER reported by Tarteel on clean audio: 7.05% (tiny), 5.75% (base). Not independently verified here, and int8 quantization is not free — measure on your own audio before trusting it.