"Mushaf al-Muraja'a"
As-salamu alaykum wa rahmatullahi wa barakatuh, brother Muno459,
I read your reply in discussion #7. Since my app has a paid subscription, I understand I need a separate written arrangement to use your models, and I would like to discuss one with you.
Jazak Allah khayran for fastconformer-quran-streaming and for documenting it so carefully, especially the real phone-audio results and the pronunciation head.
About the app: it is called "Mushaf al-Muraja'a" (Smart Quran), a revision app for huffaz who have already memorized and want to maintain their hifz. It uses the Madinah Mushaf layout (QCF) with visual revision tools such as progressive hiding, page-position recall, mutashabihat, and verse-start drills.
Voice recitation (tasmee') is our main obstacle. Chrome's built-in speech recognition on Android has not been reliable for Quran recitation in our tests, and the Whisper-based Quran model we tried works well on full recordings but is not suited to live tracking. Your streaming model fits exactly what we need: live word-level tracking (missed, wrong and extra words), and possibly the pronunciation head later.
Our planned use:
- Server-side inference over a WebSocket audio stream, with a VAD gate in front of the model, as you recommended.
- No audio is stored; it is processed and discarded.
- Clear attribution to Muno459 / Quran-Lab and NVIDIA FastConformer inside the app.
- We are happy to share back any improvements, bug fixes or evaluation results.
I am also building a small labeled test set of correct recitations and recitations with deliberate errors, and I can share the results with you if they are useful.
I am open to whatever terms you prefer, whether a fixed fee, a revenue share, or another arrangement. Since this discussion is public, could you share a private email or another way to contact you so we can discuss the details?
May Allah reward you for this work and make it a means of khayr for the Ummah. I look forward to hearing from you.
Adnan - Dubai