The Quran Lab journal
Notes from the lab
What we are building for the Quran and its listeners, told slowly: the models, the recordings, and what we learned on the way.
NewRecFlow: 884,821 recitations from our own tajweed TTS
RecFlow recites the Quran from its text with the tajweed of Hafs, holds a madd for as many counts as it is asked to, and can change one sound and nothing else. It writes the data that teaches our recitation model to hear mistakes.
Read the storyEarlier
Every phone of the Quran, with the rule behind it
quran-tajweed-phonetics writes all 6,236 ayat as 522,475 phones and ties each one to the tajweed ruling that produced it, with its source in the classical books. It is open to download.
SAWT V6.1 and V6.2: what changed after release
Two updates in the first ten days. V6.1 keeps the pause gate out of speech and digital silence silent; V6.2 gives Pro new weights, tuned on real recitation, with lower phoneme error on the hardest recordings.
SAWT V6: the pauses fall silent
A century of recitation survives on tape, in echoing halls and as 32 kbps uploads. SAWT V6 brings it back at 48 kHz without changing a syllable, and lets the silence between ayat be silent again.
Zipformer Quran 3.1: it now writes what was said
4.13% phoneme error on all 600 benchmark clips, madd held as it was recited, and the Neural Engine.
The data behind the models
More than 59,000 hours of recitation, and how the selections for Zipformer Quran 3 and 3.1 were made.
Closing the emphatic-letter gap
How ص ض ط ظ went from three times the error of plain letters to parity on every held-out set.