Speech recognition In training
TajweedFormer
The next recognition model. It was going to be Zipformer Quran 4; it has its own name because the architecture is new.
Evaluation after training
Release forthcoming
الفاتحةThis is the voice spectrum of Al-Fatiha. Press play.
Open, on-device AI for Quran recitation.
Our models listen to a recitation and write it down phoneme by phoneme, tajweed included, on a phone with no internet. Free, never sold, built as sadaqah jariyah.
Zipformer Quran 3.1 0:00
سورة الفاتحة
1 / 7
Measured on recitations it never heard
Zipformer Quran 3.1 on all 600 clips of the held-out benchmark. Lower is better.
- Phoneme error, all 600 clips
- 4.13%
- 3 held-out sets, none trained on
- Held-out studio reciters
- 2.06%
- 200 clips
- Real phone recordings
- 5.96%
- 200 clips
- A professional reciter it never heard
- 5.09%
- 200 clips
Playground
Try each model on real recitations, or on your own voice.
Transcript
سورة الفاتحة
٣
The Beneficent, the Merciful.
ررَ rra حِ ḥi ۦۦ ii maddم m
0:14 / 0:46
Recite with it
ReadRead along from the mushaf page, ayah by ayah, or as a teleprompter. The text follows your voice.
MemoriseThe words stay hidden until you recite them. Stuck? A hint shows the next word, one per tap.
ShuffleRandom ayat from the surahs, juz or pages you choose. Recite each from memory, then see your score.
Models
Free and public, never sold. Each model states its licence. Speech recognition comes first; restoration, alignment and data serve it.
Speech recognition Released
Zipformer Quran 3.1
Writes a recitation down phoneme by phoneme, tajweed included, streaming on a phone with no internet. No language model, so a reciter's mistake is never quietly corrected.
4.13%
PER, full 600-clip benchmark
Hugging Face NPL-1.2
Speech restoration Released
SAWT v4
Restores archive recitations to 48 kHz studio sound in the same voice, without moving a phoneme. Each restored window is checked against the input before it is kept.
0.071
PER on 120 restored archive recitations
Hugging Face SAWT Non-Commercial 2.0
Forced alignment Released
MFA Quran Hafs
Aligns each ayah to its text phone by phone, so madd, ghunnah and qalqalah can be measured, not only word timestamps.
10–20 ms
median phone-boundary error
Hugging Face Apache-2.0
Speech synthesis Planned
Quranic TTS
A text-to-speech model for research and teaching. Planned, not in training yet. Its corpus, QuranTTS, is already public.
301 h
QuranTTS corpus, published
Corpus To be announced
Research
- Sep 2026 Proving the potential of ternary quantization in ASR An internal experiment: Zipformer Quran 3.1 at −1, 0 and +1 reaches 3.34% phoneme error, but it is overfit to the canonical text. Not for feedback yet; ready to build a canonicalizer on. ASR
- Sep 2026 The data behind the models: a 59,000-hour corpus, and what the models took from it Quran Lab holds more than 59,000 hours of recitation. No model trains on all of it; this is how the selections for Zipformer Quran 3 and 3.1 were made. Data
- Sep 2026 Zipformer Quran 3.1: it now writes what was said 4.13% phoneme error on all 600 benchmark clips, madd held as it was recited, and the Neural Engine. ASR
- Sep 2026 SAWT v4: 48 kHz restoration of archive recitations Bringing archive recitations back to studio quality with a 372M flow-matching generator that renders texture but never decides what was said. Restoration
- Sep 2026 Zipformer Quran 3: closing the emphatic-letter gap How ص ض ط ظ went from three times the error of plain letters to parity on every held-out set. ASR
Every voice helps the model hear better.
Recite an ayah, share a recitation, or check one. Recordings are encrypted on your device and never published. Open to everyone, from children to seniors.
