About

Quran Lab is a non-profit research lab, run as a waqf, a charitable endowment: what it builds is never sold. Our mission is to build the best tajweed AI and give it away free, forever. The work is done as sadaqah.

How we work

  • We write what was recited

    The recognition models have no language model. A model that knows the text quietly corrects a reciter's mistake, and the mistake is what a teacher needs to see.

  • One riwaya, done carefully

    Recognition, alignment and the datasets follow Hafs 'an 'Asim. Another riwaya changes exactly the details tajweed cares about, so they are not mixed.

  • Measured on recordings the model never heard

    Every model is scored on held-out sets. When a published number turns out to be wrong, the correction goes up in public, next to the number it replaces.

  • Tajweed from the books, checked by people

    The rules behind the phonetic layer come from the classical literature, each with its citation. All 208 rulings were reviewed by an ijazah-holding hafiz and by two professors of Qira'at, each working alone.

  • Free, and never sold

    Models and datasets are released at no cost, on Hugging Face or through the SDKs. None of them is sold or put behind a payment.

What has been released

Everything below is public: TajweedFormer V4 through the SDKs, the rest on Hugging Face. Newest first.

  1. Nov 2026 TajweedFormer V4 A streaming model that hears recitation sound by sound on the learner's device, with outputs for lengthening, gemination, ghunna and qalqalah. Free SDKs for Android, iOS and macOS, the web, Windows and Linux.
  2. Oct 2026 SAWT V6 SAWT V6 Pro and V6 Flash restore archive recitation to 48 kHz, released with 47.6 hours of restored recitation. New Pro weights, V6.2, followed on 29 October.
  3. Oct 2026 SAWT V5 A restorer that reads the recording with a Whisper encoder, trained on speech and singing in more than 30 languages.
  4. Sep 2026 SAWT V4 The first public SAWT, evaluated under a pre-registered protocol and published openly.
  5. Aug 2026 Zipformer Quran 3 · 3.1 Phoneme recognition that streams on a phone with no internet. Version 3.1, on 28 August, holds madd as it was recited.
  6. Aug 2026 quran-tajweed-phonetics All 6,236 ayat as phones, each tied to its tajweed rule and to the book it comes from.
  7. Aug 2026 QuranTTS · MFA Quran Hafs 301 hours of ear-verified recitation from 16 reciters, and an aligner that times every phone of an ayah.
  8. Jun 2026 quranic-asr-benchmark 600 held-out clips and a public leaderboard, so every model is scored the same way.
  9. Before 59,000+ hours collected A corpus of recitation from thousands of reciters, cleaned and checked, and our own forced aligner to cut it into ayat. Models train on selections from it.

الشعراء: ١٨٠

وَمَآ أَسْـَٔلُكُمْ عَلَيْهِ مِنْ أَجْرٍ ۖ إِنْ أَجْرِىَ إِلَّا عَلَىٰ رَبِّ ٱلْعَٰلَمِينَ

And I do not ask you for it any payment. My payment is only from the Lord of the worlds.

Ash-Shu'ara 26:180 · Saheeh International

Help arrived, in many forms

Experts reviewed 208 tajweed rulings, and volunteers test and build with us. May Allah reward them on behalf of the Quran and its people.

Take part

Working together, citing our work, or helping with data, review or time.