Muno459

/

fastconformer-quran

like 5

Wa alaikom assalam wa rahmatullah, brother. No headache at all, this is a pleasure <3

You are exactly right. The plain q8 (model.q8.onnx) only outputs the text logprobs, so it cannot feed the tajweed head. The pronunciation head needs the encoder features, which come from model_with_encoder. I went ahead and quantized that one for you and just uploaded it as onnx/model_with_encoder.q8.onnx (about 132 MB). It keeps both outputs, logprobs and encoder_output, so you can run the tajweed head straight from it without quantizing anything yourself.

Also, I just pushed a proper streaming version of the ASR under streaming_onnx/, so it can transcribe live during recitation instead of only at the end. That should help on the latency side too. Happy to help wire either one in, baraka Allah feek.

Small follow-up on the streaming model link. I moved it to its own repo to keep things clean:

https://huggingface.co/Muno459/fastconformer-quran-streaming

The README there has an offline-vs-streaming comparison (latency and accuracy) so you can pick per use case. The encoder q8 for the tajweed head stays in this repo at onnx/model_with_encoder.q8.onnx.

thanks my brother <3

I'll try the tajweed mode , I am forking an edit for sherpa to export the logprobs , encoder and load the head onnx instead of using the onnxruntime

and i'll try the streaming mode and send updates there <3

Thanks for your patience and careness elhamdle Allah <3

i was using the offline model as a sliding windows which gave me the a streaming stimulated but it used alot of cpu
ofcourse the streaming model is the better way , i was waiting for this <3

1️⃣ Step 1: Join the Early Access Group
https://groups.google.com/g/recite-quran

2️⃣ Step 2: Download the App from Google Play
https://play.google.com/store/apps/details?id=com.recitequran.app

you can try the sliding window using the offline model it was a real time masha'a Allah

One more upgrade for you, akhi. I retrained the offline ASR with the same canonical-label fix that helped the streaming model, and it cut the real-phone word error roughly in half (about 20% down to ~10% on our phone-recorded holdout; clean studio stays near-perfect).

I uploaded it as a MATCHED PAIR so the tajweed head keeps working with the new encoder:

• onnx/model_with_encoder_4h.q8.onnx (~126 MB, outputs both logprobs and encoder_output)
• head/pronunciation_head_v8.pt (the pronunciation head retrained on the 4h encoder features)

Use those two together (4h encoder + v8 head). Your current model_with_encoder.q8.onnx + head v7 still work fine, so there is no rush, but the 4h pair is noticeably better on real phone audio. Happy to help wire it in. baraka Allah feek.

Thank you my brother for your care <3
I'll try it and make another repo for the flutter app using tajweed insha'a Allah
Elhamdle Allah for this streaming and offline taweed model
I didn't think that i'll find that type of model <3
ربنا تقبل منا انك انت السميع العليم
Our Lord Accept from us <3