Muno459

/

fastconformer-quran

like 4

i tried to use the coreml streaming model but i didn't find a .nemo to export the streaming model to onnx
if you can share an onnx model for the streaming <3

Wa alaikum as-salam wa rahmatullahi wa barakatuh akhi

Jazak Allahu khayran, really glad it's useful! In sha Allah I'll export the streaming model to ONNX and upload it as soon as I can. I'll update you here in this thread once it's up.

I was using a fastconformer model that outputs arabc without tashkeel but when i Allah subhanu guided me to yours , i dropped the old fast conformer and started using yours masha'a Allah , it has higher accuracy and the tashkeel option is good
will be adding the tajweed option also using the pronunciation_head onnx insha'a Allah

i invited you to the private repo in github (but still in the repo it is using the old fast conformer (offline ) and muaalem api for tajweed ) i will drop this branch and add a new branch using your model insha'a Allah (https://github.com/Iam-Muslim/ReciteQuran) insha'a Allah i will make the repo public but after the app be stable

also don't forget to add q8 quantization to the streaming model <3

Your work is so good my brother, may Allah the great accept <3
ربنا تقبل منا انك انت السميع العليم

Wa ʿalaykum as-salām wa raḥmatullāhi wa barakātuh, my dear brother 🤍

Jazāk Allāhu khayran for your kind words and your duʿā, wallāhi they mean more to me than anything. All praise
belongs to Allah alone; whatever good is in this work is purely from His tawfīq, and I ask Him to make it ṣadaqah
jāriyah for us both.

It truly moved my heart to read that Allah guided you to the model and that it's now serving your ReciteQuran app,
that the accuracy and the tashkīl are helping after the old model that had none. This is the exact niyyah I made when
I began it: that it would serve the Book of Allah and whoever recites it. Al-ḥamdulillāh.

About the streaming, I didn't want to hand you something that only half-works, so I'm now training a proper
cache-aware streaming model from scratch, the right way, so it genuinely streams token-by-token instead of being a
patched offline model. It's in training, in shā' Allah, and I'll add q8 quantization to the streaming model so it
stays light and fast on-device.

And the tajweed option you're planning through the pronunciation head, that makes me genuinely happy; it's exactly
what that head was built for. If you ever want help wiring it up, I'm right here, in shā' Allah.

Thank you for inviting me to the repo 🤍 truly an honor. Take all the time you need to make it stable, when it goes
public, in shā' Allah it will be a means of khayr for the whole Ummah.

🤍 ميلعلا عيمسلا تنأ كنإ ،انّم لبّقت مهللا
Ameen to your duʿā, your brother in this, always.

insha'a Allah
Thanks my brother , and thanks for sharing your work

rarely i find a real believers like you <3
I love you my brother
when i started my app i made a rule that it will be only for the sake of Allah the great without any money or purchases (Only for Allah )
I prayed and made duaa Allah to find an offline model that is better than mine and Alhamdle Allah , he guided me to here 🤍