Quran Lab
Playground

Models

Free and public, never sold. Each model states its licence. Speech recognition comes first; restoration, alignment and data serve it.

Quran-Lab on Hugging Face

Speech recognition

One model line, built to hear tajweed. It writes phonemes, not words, and has no language model, so a reciter's mistake stays visible.

Speech recognition · NPL-1.2

TajweedFormer V4

A 65M-parameter streaming model that hears Quran recitation sound by sound, inside an app, on the learner's device: each sound with its timing and confidence, and separate outputs for lengthening, gemination, ghunna and qalqalah. It replaces Zipformer Quran 3.

How to use it

Through the free SDKs for Android, iOS and macOS, the web, Windows and Linux. Sign in at sdk.quranlab.ai with Hugging Face, copy one prompt into your coding assistant, and it sets the SDK up in your app. There are no weight files to download; if no SDK can serve what you build, ask for the raw weights there.

For developers sdk.quranlab.ai Docs Hugging Face
1.23%
PER on correct recitation by 16 reciters it never heard, 1,780 clips
65M
parameters
38
tokens, one per sound; tajweed in its own outputs
57 MB
about the size of the first download, once per device

Speech recognition · NPL-1.2 · the earlier model

Zipformer Quran 3.2

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.

TajweedFormer V4 has replaced it for apps: Quran Lab's SDK no longer serves it, and its files stay on Hugging Face for reference.

v3.2 is the corrected export of v3.1: the same weights, computing the network exactly as it was trained. It reports far more real recitation mistakes at the same accuracy, and it is what this site runs.

How to use it

Request access on Hugging Face, then download the INT8 ONNX file and tokens.txt. It is a standard sherpa-onnx streaming CTC model: feed it 16 kHz mono audio as 80-dim kaldi fbank features. On an iPhone, the Core ML package runs on the Neural Engine.

Model page Hugging Face Benchmark
4.13%
phoneme error, all 600 benchmark clips
65.5M
parameters
251
phonetic symbols, tajweed included
63 MB
iOS package, 8-bit weights

Speech restoration

SAWT restores damaged recordings to 48 kHz in the same voice, words and melody. Four versions are public. V6 Pro is the one we recommend, and V6 Flash does the same job in four steps.

SAWT V6 in detail

SAWT V6 Pro

Recommended

A 351M flow-matching transformer that reads the recording through omniASR and restores it in 16 steps. Silent pauses, the melody kept. New weights, V6.2, since 29 October 2026.

351M
generator
16
steps
16×
real time
Model pageHugging Face

SAWT V6 Flash

For large archives

A 105M student distilled from Pro's own restorations, in four steps. A full mushaf in well under an hour on one RTX 4090.

105M
generator
4
steps
74×
real time
Model pageHugging Face

SAWT V5

Earlier

Reads the recording with a Whisper large-v3 encoder and renders with a 372M generator, trained on speech and singing in more than 30 languages. V6 Pro has replaced it; it stays available.

372M
generator
64
steps
Hugging Face

SAWT V4

First release

The first public SAWT: a w2v-BERT 2.0 anchor and a 372M generator that checks every window it renders against the recording. Evaluated under a pre-registered protocol, with a preprint.

372M
generator
64
steps
Research noteHugging Face

On 120 archive recitations

RecordingSAWT V5 ProSAWT V6 FlashSAWT V6 Pro
DistillMOS, higher is better2.783.473.683.89
Pause floor, median-36 dBFS-53 dBFS-70 dBFS-79 dBFS
Speed on one RTX 4090, long filesn/a7.8×74×16×

Ten-second excerpts of real archive recitation, through the release code with default settings. DistillMOS is a learned quality predictor; the pause floor is the median level in the recording's pauses. From the SAWT V6 Flash model card.

Forced alignment

A time for every phone of an ayah, so madd, ghunnah and qalqalah can be measured, not only word boundaries.

Forced alignment · Apache-2.0

MFA Quran Hafs

A Montreal Forced Aligner model for recitation in the riwaya of Hafs. Its dictionary is built from whole ayat, so idgham, ikhfa and iqlab across word boundaries are already in place, and every sound that carries a tajweed rule is a single phone.

How to use it

Download it from Hugging Face without a request. Give it one 16 kHz wav and one text file with the Uthmani ayah per clip, then run mfa align with the dictionary and the acoustic model.

Model page Hugging Face
10 to 20 ms
median phone-boundary error
67
phones, one per articulation
20,967
dictionary entries, from whole ayat
93,430
tajweed rule positions indexed

Datasets

What the models are built and measured on.

  • quran-tajweed-phonetics

    Tajweed phonetics

    Every ayah of the Quran as phones, each tied to the tajweed rule that produced it and to its source in the classical books.

    6,236
    ayat
    522,475
    phones
    208
    rulings reviewed by experts
    NPL-1.2 Open download Hugging Face The story

  • quranic-asr-benchmark

    Held-out test sets

    600 clips the models never trained on: studio reciters, a professional reciter and real phone recordings, 200 of each.

    600
    clips
    3
    sets of 200
    v1.1
    corrected references
    NPL-1.2 Access on request Hugging Face Benchmark

  • QuranTTS

    Speech corpus

    Ear-verified recitation for speech synthesis and restoration: 16 reciters, Hafs murattal, 48 kHz masters at 24-bit.

    301.4 h
    of audio
    64,721
    clips
    6,120
    ayat
    Non-commercial 2.0 Access on request Hugging Face

  • sawt-v6-restored-recitations

    Restored recitation

    Restored by SAWT V6 Pro with its default settings, nothing edited by hand: a complete mushaf from an 11 kHz copy, and one surah from each of 200 reciters.

    47.6 h
    of audio
    114
    surahs of one mushaf
    200
    reciters
    SAWT Non-Commercial 2.1 Access on request Hugging Face

  • quranic-asr-cloud-rawdata

    Benchmark transcripts

    Transcripts, raw responses and scores from six runs of cloud speech recognition services on the 600 benchmark clips, with the scripts that produced them. The audio stays in the benchmark.

    3,600
    transcripts
    6
    service runs
    NPL-1.2 Access on request Hugging Face

Try them online

Nothing to install. The playground on the home page also runs recognition in your browser.

Try the playground
Quranic ASR leaderboard Word and character error of Arabic speech recognition models, all on the same 600 held-out clips.SAWT V6 gallery Twelve well-known recitations, before and after SAWT V6 Pro.Listening page Ten archive clips side by side through SAWT V6 Pro, V6 Flash and V5 Pro, loudness-matched.

Access and licences

Everything is free. Most repositories are gated: ask for access on Hugging Face, accept the licence, and access is granted by hand. MFA Quran Hafs and quran-tajweed-phonetics download straight away.

LicenceCovers
NPL-1.2, the Quran-Lab No-Profit LicenseTajweedFormer V4 (through the SDKs), Zipformer Quran 3 and 3.1, quran-tajweed-phonetics, quranic-asr-benchmark, quranic-asr-cloud-rawdata
SAWT License 2.1, non-commercialSAWT V4, V5, V6 Pro and V6 Flash, sawt-v6-restored-recitations
Quran-Lab No-Profit License 2.0, non-commercialQuranTTS
Apache-2.0MFA Quran Hafs

The full text of each licence is in its repository on Hugging Face.