A

Ayah Finder

A NivareQ™ project
Starting…
Private · local-first

Recite what you remember.

Ayah Finder uses local speech recognition only to generate a rough search query. Canonical Quran text always comes from the bundled verified corpus.

Preparing local AI…
Checking WebGPU and model cache

Type or paste a remembered phrase

Arabic keyboard dictation also works. Exact spelling is not required.

Local AI

Engine: automaticModel: Base Q8

The default iPhone path is WebGPU when available, with CPU/WASM fallback.

Quran

Read locally, with selectable script and translation.

Saved

Bookmarks stay on this device.

Guide

Simple ways to get good results.

1. Recite a short fragment

About 4–12 seconds is ideal. You do not need to remember the whole ayah.

2. Let the matcher do the forgiving

The transcript can be wrong. Ayah Finder compares the rough Arabic against 1–3 ayah windows locally.

3. Treat similarity as similarity

Scores are not probabilities. If confidence is modest, compare the top few candidates.

4. Offline use

Open the app once online and tap “Prepare AI for offline use.” Quran reader data is bundled with the app.

5. Desktop / NVIDIA

In Settings → Local AI, Advanced Local Model can load your own Quran-Turbo Q4/Q5/Q8/F16 GGML file directly from disk.

6. Safety invariant

AI output may identify Quran text; AI output may never become Quran text. Displayed Arabic comes only from the local corpus.

Settings

Display, reader, and local AI.

Appearance

Reader

Arabic scriptChoose Uthmani or IndoPak display tradition.
TranslationFive English and two Bengali options are bundled.
Show translation footnotesWhere the selected source provides them.

Local AI

Inference modeAutomatic prefers WebGPU. CPU fallback uses the proven no-pthread runtime.
Model modeBase Q8 is the phone-safe default. Advanced lets desktop GPUs load a larger local model file.

Diagnostics

About

Product identity, privacy model, sources, and credits.

About

Ayah Finder · Product Candidate 1

Ayah Finder
A NivareQ™ project

Ayah Finder is a local-first Quran verse finder and reader. Speech recognition produces approximate search input only. Canonical Quran text is always loaded from the bundled corpus.

Privacy

Speech inference runs on your device through WebGPU or the CPU/WASM fallback. Advanced model files are read locally from your device. Audio is not uploaded for transcription.

Quran & translation sources

Quran display text: Tanzil Uthmani, distributed verbatim under CC BY 3.0 with attribution; changing the Quran text is not permitted. IndoPak display text: DigitalKhatt dataset. Translations: QuranEnc sources where noted, plus public-domain Pickthall and Yusuf Ali. Full source/version details are bundled in ATTRIBUTION.md and the licensing review.

Local speech model

Default: pinned Quran-specialized Whisper Base Q8, SHA-256 72194195f7d280adebec57acf2c6e01e209484322ec1cec21c275a3c0b1e3d77. The model is verified before initialization. Desktop users can optionally load local Quran-Turbo GGML models from Settings.