How Tehseen AI checks your Quran recitation
You recite an ayah into your phone or computer. Tehseen AI compares what it hears with the exact phonetic form of that ayah and marks every word correct, a mistake, or skipped, about a second behind your voice.
Last updated 7 October 2026
1. The ayah becomes a list of sounds
Before you start, every ayah has already been turned into its phonetic form in the Hafs ʿan ʿĀṣim reading: each letter, vowel, shadda and madd, including how words join when read without stopping and how the last word changes when you pause. That is the reference your recitation is checked against, so you are not compared with one particular qari's voice.
2. A speech model trained on Quran recitation listens
Your audio streams to our servers and runs through a speech model trained on hundreds of hours of expert Quran recitation, which hears sounds rather than guessing likely words. That matters: general speech recognition “autocorrects” a learner's mistake into the word it expects, which is the one thing a recitation checker must not do. Tehseen AI is built on muaalem, an open-source Quran recitation model (MIT licence).
Results arrive word by word each time you pause for breath, so you see where you are while you recite. The full summary appears when you finish the ayah.
3. Every word gets a verdict
Each word is marked correct, a mistake or skipped. A mistake names what went wrong:
| Mistake | What it means | Example |
|---|---|---|
| Wrong word | A different word, or one not in the ayah | Reading نَسْتَعِينُ where نَعْبُدُ belongs |
| Similar letter | A letter swapped for its close neighbour | ظ for ض, س for ص, ت for ط, ه for ح |
| Articulation (makhraj) | A different letter, made from the wrong place | ل where ر belongs |
| Vowel (haraka) | The wrong short vowel on a letter | A fatha read as a kasra |
| Shadda | A doubled letter said once | ٱلرَّحْمَٰنِ without the doubled ر |
| Madd length | A long vowel cut short | مَٰلِكِ said as مَلِكِ |
| Skipped | A word left out | Jumping from رَبِّ to ٱلرَّحْمَٰنِ |
Each mistake comes with a plain-language explanation first and the tajweed term second. You can play the qari, play yourself, and retry that one word on its own.
What it doesn't grade yet
Tajweed rules such as ghunnah, ikhfa', idgham and qalqalah are colour-coded in the reader. In Strict mode some of them are checked as a beta; by default they are not graded. When the model is unsure about a sound, it gives you the benefit of the doubt rather than flag a mistake you didn't make.
How we measure accuracy
Every change is tested on recordings of professional qaris (where every word should be marked correct) and on recordings with deliberate mistakes spliced in (where each one should be caught). A test set of learners, labelled word by word by a qualified teacher, is being collected now; we will publish results once it is complete. If a result looks wrong, tap “Not right? Report it” and we review it.
What happens to your audio
Your recitation is analysed in memory on our own servers (Hetzner, Germany) and then discarded. It is not sent to a third-party AI provider. A clip is kept only if you choose to attach it to a report about a result. See the privacy policy.
A practice aid, not a teacher
Tehseen AI helps you practise between lessons. A qualified teacher (ustadh or ustadha with ijazah) should still guide your recitation.