What the model hears, sound by sound, then the ayah it comes from, with every word checked against the text. Recite into the microphone or open a recording. The recognizer, zipformer v3.1, runs on the exact engine compiled to WebAssembly inside this tab, so your audio stays on your device.
The model writes in the Quran phonetic script: one letter per sound, a doubled letter for shadda, and the vowel letter repeated once per count of madd. The Latin keeps every count. Consonants follow the usual scheme (ḥ ṣ ḍ ṭ ẓ for the emphatics, ʿ for ʿayn, ʾ for hamza); the rest is below. Underlines are the model's confidence; amber marks a phoneme under 50%.