Listen to an old recording of a beloved reciter and you hear two things at once: the recitation, and everything that happened to it. The hiss of a cassette copied many times. The hall the microphone stood in. The upload that kept only what lies below 8 kHz. Between the ayat, where the reciter breathes, the damage is all there is to hear.
SAWT exists to take that away and leave the recitation exactly as it was. Not a cleaner imitation of it: the same voice, the same words, the same melody, in full 48 kHz. SAWT V6 is the first version built around that job from the start, and it is the one we recommend.
Hear it first
Four ten-second archive excerpts, before and after, on one timeline. Switch while it plays to hear the same instant both ways. Headphones help, and the pauses are where to listen.
A good copy with a constant hiss under it. The pauses fall from -28 to -85 dBFS; quality score 2.71 in, 3.77 out.
Switch while it plays to hear the same instant, before and after. Headphones help. More on the listening page
Silence is part of the recitation
Earlier restorers, ours included, cleaned the voice well and then let the old room creep back into every pause. V6 adds a pause gate: wherever the reciter is silent, the generator stops looking at the damaged recording and draws only from what it knows clean recitation sounds like. The median background in the pauses falls from -36 dBFS in the recordings to -74 dBFS. SAWT V5 Pro reached -56.
A restorer that sounds beautiful but moves the melody has changed the recitation. We measure that, not only how clean it sounds.
The melody stays where the reciter put it
Some of the strongest open restorers rebuild speech from scratch out of what a speech model heard, and invent the pitch along the way. On recitation that is the maqam itself. With SAWT V6 Pro only 11% of voiced frames move more than half a semitone from the recording. Sidon, the best open restorer we tested, moves 20%.
The reason is the anchor. SAWT reads the damaged recording with Meta's omniASR, a recognition encoder trained on more than 1,600 languages. A model that knows what was said is exactly what a restorer must not overrule. A 351M flow-matching transformer then renders 48 kHz sound under that reading, in 56-second windows long enough to hold a whole ayah and the pauses around it.
Next to every open restorer we measured
The same 120 archive recitations through each system's published recipe, levelled to the input. SAWT V6 Pro has the highest predicted quality and among the quietest pauses.
Quality predictors miss things a listener hears at once, in both directions. The numbers point the way; the listening page is the real test.
Two models: Pro and Flash
SAWT V6 Pro
351M generator. The best quality and the quietest pauses. 16× real time on long recordings on one RTX 4090: a 3.5-minute recitation in 13 seconds.
For large archivesSAWT V6 Flash
105M student in four steps. DistillMOS 3.68, close to Pro, at 74× real time: a 3.5-minute recitation in under 3 seconds.
47 hours, already restored
Alongside the models we released 47 hours of restored recitation, including a complete mushaf restored from an 11 kHz copy. Everything is free for non-commercial use under the SAWT License 2.1: individuals, mosques, schools, charities, archives and researchers. Audio restored with SAWT should say so: "Restored with SAWT by Quran-Lab".
SAWT means voice. May it carry these recitations to many who have not yet heard them clearly.
All posts