darten-zipformer-quran (mirror)

An un-gated mirror of Quran-Lab/zipformer_p-arabic-v3 zipformer_p_arabic_v3.int8.onnx, hosted so the Darten mobile app can download it at runtime. onnxruntime-react-native cannot present a Hugging Face token from a phone, which is the only reason this mirror exists.

The weights are unmodified. Nothing here is retrained, re-exported or re-quantised. For the model card, the benchmark numbers, the CoreML packages and everything else, go to the original.

Licence — read before using this

Released by Quran-Lab under the No-Profit License 1.2 (NPL-1.2); the full text is in LICENSE in this repo. In the authors' words: free to use, study, adapt and redistribute, but the model itself is never for sale, so you may not charge for it, for access to it, or for a feature it powers. Your own labour stays your own — a teacher may charge for teaching and a developer may draw a salary.

The upstream repository is gated, and accepting that gate is how you tell the authors you accept those terms. This mirror removes that step, so accepting the terms is on you. If you are here to put this in a paid product, the licence says no, and mirroring it does not change that.

Attribution

• Model: Quran-Lab/zipformer_p-arabic-v3 by Quran-Lab
• Fine-tuned from zipformer_p-arabic; architecture from k2-fsa/icefall
• Phonetic script from the muaalem line by obadx

Usage note

Streaming CTC, sherpa-onnx compatible. Feed 16 kHz mono, 80-dim kaldi fbank (povey window, low_freq=20, high_freq=-400); a slaney-mel front end roughly doubles the error rate. Decode with the upstream tokens.txt (blank is 250) and not phoneme_units.json (blank 0) — the two are offset by one.

onnxruntime 1.27.0 silently miscomputes this graph, returning a near-all-blank posterior with no error. 1.26 and below are correct; bisected against a clip whose decode is known exactly.