The models
- All three models, weights and wasm core
- Unit lists, model card, how to run it
- CC BY 4.0, credit Fanah Solutions and NVIDIA
Naml follows a recitation word by word. Three small streaming models, about 17 MB of weights each, run in any modern browser on our WebAssembly core. Take the model for free, buy the tracker that runs recite.nabilmh.com, or put the Tarawih on your masjid's screen.
every demo below is the real engine's output, replayed
Press play. Every 0.32 s each model reads the newest audio (with 0.16 s of look-ahead) and writes what it heard. The timings are what each step took on one CPU core of our server.
Cached streaming FastConformer, 8 or 10 blocks, ternary weights (−1 / 0 / +1 per 128-input block, 2 bits each). Base model: NVIDIA NeMo FastConformer, CC BY 4.0.
Loading the demo…
Run Naml Tajweed on your own microphone, in this tab. It downloads about 23 MB once (weights, front end, the wasm core). Nothing is uploaded.
The tracker turns the model's stream into a place in the mushaf: the word you are reciting, highlighted on the Madani page, the page turned when you reach the end. Start anywhere; it finds you.
Loading the demo…
Each completed word is force-aligned on Naml Tajweed's frames and compared with the likely mistake (a missing shadda, a clear noon for ikhfa, ط read as ت). Strong findings are errors, weaker ones "uncertain". They land in a Mistake History with your own audio to replay.
| False alarms on correct professional recitation | 0.16 / min (EveryAyah, 73 min) · 0.20 / min (long QUD recordings, 196 min) |
|---|---|
| Detected, altered-text test | shadda 92% · ikhfa 94% · iqlab 97% · haraka 79–94% · ط as ت 90% · ع as ء 96% · ح as ه 88% · ص as س 60% |
| Cost | ~0.3–0.6 ms per word on a server CPU |
Altered-text figures are optimistic (the text context helps). Madd length and ghunna duration are not detected yet: they are off, and we do not claim them.
Loading the demo…
One email when it is ready. Nothing else.
Put a laptop or a TV by the musalla: the ayah the imam is reciting, the word he is on and its English, page and ayah progress. It listens in Sync mode from Al-Fatiha, as the app's Prayer view does, and moves with him from surah to surah.
Loading the demo…
Also coming: tajweed correction ($1,000 planned). Get notified.
Checkout coming soon. Write to us and we will send an invoice.
The model runs in the browser tab, on the phone's own CPU. There is no speech server, so no audio is uploaded, and there is no per-minute bill.
After the first download the model needs no network. The mushaf pages and translation are cached the same way.
Read the model card, run your own tests, keep your own copy. Nothing phones home.
The tajweed Mistake History lives in the browser's storage on that device. Clear it and it is gone.
The three Naml models: weights, the WebAssembly core that runs them, the unit lists and a model card, under CC BY 4.0 (credit Fanah Solutions and NVIDIA's NeMo FastConformer, which the models derive from). The tracker code and the masjid kit are paid; tajweed correction is coming.
Any browser with WebAssembly SIMD: current Chrome, Edge, Firefox and Safari. Measured per 0.32 s step: about 18 ms on a budget Android phone (4 GB, Chrome), about 15 ms on an iPhone in Safari, about 15 ms on one server CPU core. Older browsers without SIMD fall back to a slower path.
On 31 reviewed clips (about 1,240 ayahs), Naml Letters puts both the start and the end of an ayah within 250 ms for 51% of ayahs and within 500 ms for 68%. wav2vec2 with Viterbi alignment: 43% and 62%.
A download of the recite app's source (TypeScript, React, the Rust source of the wasm core), its README and a licence to use it in your products. Not included: test recordings, our development plugins, third-party models. See the licence draft.
The recite app's source set up as a Tarawih screen: the Prayer view (it listens from Al-Fatiha, follows the imam into the passage and shows each ayah with its English) and the large single-ayah view, plus a setup guide for one laptop, a TV or projector and an audio feed from the mixing desk. It runs offline. If you would rather not set it up yourself, choose "Built for you".
On by default: shadda, ikhfa, iqlab, idgham, confused letters (ص/س, ط/ت, ح/ه, ع/ء), short vowels, a dropped letter or vowel, and missing, incorrect or extra words. Not detected yet: how long a madd or a ghunna is held. On correct professional recitation it raises about one false alarm every 5 to 6 minutes.
See the refund policy draft. If the code does not run as described, write to us.
No hosted API: everything runs on the user's device. If you want it built into your app or a masjid screen, choose "Built for you".
What a Quran recitation tracker does between your voice and the highlighted word, why it can run on a phone with no server, and what it still gets wrong.
2026-11-19 · draftHow a masjid can show the ayah the imam is reciting, the current word and its English on a screen during Tarawih, with no internet and no operator.
2026-11-19 · draftA memorization test that asks a random ayah from a range, lets the student continue, and checks the continuation automatically, with the adult still deciding.
Tell me what you are making: an app, a hifz class, a masjid screen. I will say whether the code already does it, or quote the work.