Quranic ASR but using Turkish Transcription
Brothers Selamun Aleykum,
I had an idea which I had to park aside for now after I discovered the Zipformer model.
My app runs fot the moment with zipformer But I Ithink this is still worth a shot.
The problem:
Non-Arabic speaking muslims do not understand the recitation of the quran. They can not know which Ayah is being recited in a mosque, on tv or youtube. Best model yet Zipformer has flaws if sound quality is bad.
so we do not understand quran.
Possible solution:
Recitation-->package recording-->Transliteration as if heard text is Turkish-->compare the text output with Turkish Transcript of the Quran>Use ayah matcher to find the surah/ayah number. display quran and translation on the screen.
Advantage/Why it may work:
Turkish is written as it is read. If we write down the text of the recitation in Turkish we do not have the problem of Harekeh, Harfi med, Tajweed, Maqam and etc.
you literally write down the sounds that you hear.
I made my own model to do exactly this procedure.
The proof of concept was succesfull, I built a working android app.
It worked for the reciters which I used for the training but not so succesfull with unknown reciters.
The dataset was small. Training was made on an office laptop in many days. Maybe if we had bigger dataset or maybe if we started from scratch with the right programming tools this might be promising.
Anybody can help me develop this idea?
I do not have GPUs or supercomputers to chase the idea. My AI knowledge is also limited In fact I am a mechanical engineer :-)
Best Regards,
Kasım Yazan from Istanbul.
Wa alaykum as-salam brother Kasım,
JazakAllahu khayran for sharing this with us. May Allah bless your effort, put benefit in it, and reward you for trying to make the Quran easier to follow and understand.
Your approach is especially relevant for real-world situations where the recitation may come from a mosque, TV, YouTube, or a distant microphone. We are also actively working on improving the model’s robustness in these kinds of conditions, in Syaa Allah.
We would be very happy to receive more Quran audio, whether clean recordings or more challenging recordings, especially cases where the model still struggles to recognize the recitation well.
If you have more audio like this, please feel free to send it to us by email. We will review it, in Syaa Allah. May Allah make whatever benefit comes from it a source of ongoing reward for you.
We are also currently collecting Quran audio of any quality through https://contribute.quranlab.ai. Clean audio, noisy recordings, reverberant mosque recordings, and other real-world conditions are all valuable to us.
BarakAllahu feek, Akhi 🤍
We’re also very open to discussing this further by email, Akhi. If you have any thoughts or ideas, please feel free to share them with us!