Blog · · 4 min read

DRAFT: not reviewed yet. Hidden from search engines and the sitemap until draft: false.

Automatic tajweed checking: what AI can and can't detect today (shadda, ikhfa, iqlab, madd)

"Does it check tajweed?" is the first question people ask about any Quran app that listens. The honest answer for every system we know, ours included, is: some rules, with measurable reliability, and not others. This post lays out exactly which, with the numbers, so you can judge for yourself. Our implementation is the Naml tajweed correction module, coming soon as a product.

How the check works

The speech model, Naml Tajweed, outputs tajweed phoneme units rather than letters (what those units are). For each word you finish, the analyser:

  1. takes the model's output for that stretch of audio, with the neighbouring words as context;
  2. force-aligns it to the units the word should have (for example a doubled ل for a shadda, or the nasal ں for an ikhfa);
  3. for each expected unit, aligns the likely mistake over the same audio: the consonant without its shadda, a clear noon instead of the ikhfa, ت instead of ط, the vowel missing;
  4. compares the two scores. If the mistake fits the audio clearly better than the correct reading, it is reported. A strong difference is an error, a weaker one uncertain.

It runs a second or two behind you, once per word, and costs about 0.3 to 0.6 ms per word on a server CPU. Findings go to a Mistake History with the ayah, the word, what was expected, what was heard, and your own audio of that moment to replay.

What it detects

These categories are on by default. The detection rates come from an "altered text" test: we take correct recitations and change the expected spelling, so the correct audio now contains the "mistake". At the default threshold:

Rule Detected
Shadda (doubled consonant read single) 92%
Ikhfa (hidden noon read clearly) 94%
Iqlab (noon not turned to meem before ب) 97%
Haraka (a short vowel read as another) 79-94%
ط read as ت 90%
ع read as ء 96%
ح read as ه 88%
ص read as س 60%

Also on: idgham, a dropped letter or vowel, and whole-word mistakes (missing, incorrect or extra words).

These rates are optimistic. In the altered-text test the surrounding words are recited correctly, which gives the analyser good context. Real mistakes often come with other hesitations around them.

How often it is wrong on correct recitation

A checker that complains about correct recitation is worse than none: the student learns to ignore it. We measured false alarms on recordings of professional reciters, where every flag is by definition false:

That is about one false alarm every five to six minutes of correct recitation. We think that is usable for practice, with the "uncertain" level shown separately, but it is not zero, and a teacher should still have the last word.

What it does not detect (yet)

Madd length and ghunna duration. This is the most requested rule and we do not claim it. The reason is instructive: the model's units are fixed by the text. A 4-count madd is written as one unit, "ا4", however long the reciter holds it. When we shortened or lengthened madds in the audio itself and re-ran the model, the analysis caught 0-1% of them. The same was true for a shortened ghunna. So these checks are off. Detecting them needs a measurement of duration, not a choice between units; it is on our list.

Qalqala and some letter pairs (ض/د, ق/ك, ظ/ذ/ز, ث/س) are also off by default, because they either raised too many false alarms or were not reliably separated.

Who it is for

It is not a replacement for a qualified teacher, and we would not present it as one.

See it

The tajweed demo replays a phone recording with the findings appearing as cards in sync with the audio. They are the analyser's own output, not reviewed by a teacher, which is the point: you see what it really does. Common confusions it targets are covered in this post on recitation mistakes.