Write to us

Planning

Quran Audio Engine

Can AI listen to Quran recordings and say what was recited, without ever changing the text, sending a recording away, or passing off a guess as a fact?

Our research into a family of engines that listen to Quran recordings, for developers of Quran apps and websites and for readers of Warsh and Qalun in North and West Africa. The first engines already run inside our Mirtal app, and the wider project is still at the planning stage.

At a glance

Where we are

We are planning the project, and the engines built so far live inside our Mirtal app rather than as a separate engine.

What we are aiming for

  1. Engines that tell which ayat were recited in a recording, and where each one starts and ends.
  2. Tools that serve developers of Quran apps and websites, and readers of Warsh and Qalun in North and West Africa.
  3. The Quran text is never generated or changed by any engine.
  4. No engine ever sends a recording off the device it is on.
  5. An uncertain answer is never shown as an exact one, and when an engine is not sure it stays silent.

What we are doing

The Quran Audio Engine is a family of engines that listen to Quran recordings, to tell which ayat were recited and where each one starts and ends, and in time to recognise melodic recitation (مجوّد), the reading being recited (رواية) and the melodic mode (مقام). We are building it for developers of Quran apps and websites, and for readers of Warsh and Qalun in North and West Africa, who are poorly served by tools that only know Hafs.

The work began inside our Mirtal app, and it has grown into research of its own, larger than any one app.

Three promises

Every engine is held to three promises. The Quran text is never generated or changed by an engine, no engine ever sends a recording off the device it is on, and an uncertain answer is never shown as an exact one.

The last of these shapes how the engines behave, because an engine that stays silent does less harm than one that gives a wrong answer with confidence.

Where we are

The project is at the planning stage, and the working code lives inside Mirtal for now. The recognition engine, the audio toolkit, the names reader and the desktop analyser are built inside Mirtal, the engines for melodic recitation and for memorised study texts (متون) have been merged into Mirtal but are not yet released, the engine for the reading is still research, and the melodic mode work is research that we have parked.

The engines draw on open work by others, including whisper.cpp, the Tarteel speech model under the Apache 2.0 licence, ONNX Runtime, and the Quran text from the Tanzil Project under CC BY 3.0.

The engines and where each one stands

  • Recognition

    Built inside Mirtal

  • Audio toolkit

    Built inside Mirtal

  • Names reader

    Built inside Mirtal

  • Desktop analyser

    Built inside Mirtal

  • Melodic recitation (مجوّد)

    Merged into Mirtal, not yet released

  • Memorised study texts (متون)

    Merged into Mirtal, not yet released

  • Reading (رواية)

    Research

  • Melodic mode (مقام)

    Research, parked

What comes next

There are five decisions to make before the project can start in earnest, covering the licence, the wording we use to ask reciters for their consent, which texts to cover, the timing, and when the code is made public. We are also reviewing a list of open datasets, and nothing will be downloaded or used until that review is done.

What we will not claim

We are not publishing any figures on this page yet, because each one has to be checked before we stand behind it, and we will add them here once they are. The licence has not been chosen either, so we make no claim that the project is open source.