Sojo-JustHuman/ebabil-tecvid-onnx

Ebabil Tecvid (ONNX, 4-bit)

This repository holds a 4-bit quantized ONNX export of obadx/muaalem-model-v3_2 (the Quran Muaalem project, https://github.com/obadx/quran-muaalem). Only the phoneme head is exported. It runs fully on the user's device, in the browser, with onnxruntime-web.

It is used by Ebabil (https://ebabil.web.app), a free, ad-free, charitable Quran learning app, to give learners tajweed feedback. Audio never leaves the device.

What was changed

• Only the phonemes CTC head is kept; the sifat heads are dropped.
• The 1×1 pointwise convolutions are rewritten as equivalent linear layers.
• MatMul weights are quantized to 4 bits (block-wise, symmetric) with onnxruntime's MatMulNBitsQuantizer.
• The file is split into ~32 MB parts (tecvid.partNN). Concatenate them in order to get the .onnx file. manifest.json lists the parts and the accuracy comparison against the original PyTorch model.

Input / output

• Input: input_features float32 [1, T, 160]. These are SeamlessM4T/Wav2Vec2-BERT fbank features at 16 kHz with stride 2, as in preprocessor_config.json.
• Output: logits [1, T/2, 43]. Greedy CTC decoding with vocab.json["phonemes"] (0 = blank) gives the Quran phonetic script.

License and use conditions

This model is a derivative of Quran Muaalem and is distributed under the same Quran Muaalem License (QML) 1.0. See LICENSE and NOTICE. In short, on top of Apache-2.0, the license requires:

• no advertising in any product that uses it;
• a free access path for users;
• a clear notice that the AI may make mistakes and does not replace a qualified teacher;
• use in keeping with Islamic principles.

All credit for the original model and data goes to the Quran Muaalem team.

⚠️ AI advisory: automatic tajweed feedback can be wrong. It is a practice aid and does not replace reciting to a qualified teacher.