SAWT V6 Pro and V6 Flash were released on 19 October 2026. Two updates have followed: a change to the restorer on 21 October, and new weights for Pro on 29 October. Both are in the repositories now, and the restorer reads its settings from the checkpoint, so a fresh download is all it takes.
V6.1: the pause gate stays out of speech
SAWT V6 hides the damaged recording from the generator wherever the reciter is silent, so the old room does not come back in the pauses. The gate decided what counts as a pause by comparing loudness with the file's own floor.
On crushed, limited or noisy uploads, speech and pauses are almost equally loud. There the gate also fired during speech, and the generator rebuilt those stretches from its anchor alone: a flatter, less faithful voice. On two loudness-maxed uploads from 2004 it gated 80 to 88% of the file.
Since V6.1, frames within 20 dB of a file's speech level are never treated as pauses. On 15 real archive and live-mosque recordings, pitch frames more than 50 cents off the recording fell from 11.9% to 9.2%, and from 23% to 7% on the most crushed file, with phoneme error slightly lower. Clean studio input is unaffected; on noisy live recordings the pauses keep 1 to 3.5 dB more background.
V6.1 also keeps digital silence silent. Some uploads hold stretches of pure digital silence, left by an uploader's noise gate, an edit or dead air. Given nothing there, the generator used to invent sound: one radio-era upload is 105 seconds of digital silence out of 273, and V6.0 filled those gaps at -15 to -20 dB. Stretches of 0.3 seconds or longer now stay silent.
V6.1 changes the restorer, not the weights, and applies to Pro and Flash alike.
V6.2: new weights for Pro
V6.2 is the V6 Pro generator fine-tuned with preference optimisation (Diffusion-DPO) on restorations of real recitation and of general speech. Restorations were ranked by automatic measures of intelligibility, quality and speaker identity, and the model learned to prefer the better ones.
| V6.1 | V6.2 | |
|---|---|---|
| Phoneme error, 15 real archive and live-mosque recordings | 0.230 | 0.210 |
| Phoneme error, very reverberant recordings (126 files) | 0.269 | 0.202 |
| Phoneme error, 120 archive clips | 0.082 | 0.077 |
| Phoneme error, a second reciter 3 dB below the first | 0.067 | 0.065 |
| DistillMOS, 120 archive clips | 3.89 | 3.89 |
| NISQA, 120 archive clips | 3.57 | 3.54 |
The gain is largest where restoration is hardest: on very reverberant recordings, phoneme error falls from 0.269 to 0.202. Predicted quality holds: DistillMOS stays at 3.89, and NISQA moves from 3.57 to 3.54.
On general speech, the URGENT 2025 non-blind test of 1,000 clips in five languages, V6.2 is better than V6.1 on 12 of 13 measures, among them character accuracy (81.9% to 84.3%), speaker similarity (0.684 to 0.690) and UTMOS (2.84 to 2.99). SDR is lower, 2.09 to 1.73 dB.
The original weights stay in the repository as variants/sawt_v6_generator.pt. Pass that file as --ckpt and the restorer uses its own defaults, so earlier output is unchanged. The evaluation tables on the model card and the 47.6 hours of restored recitation were made with those V6 weights.
Which one to download
For a recording that matters most, SAWT V6 Pro. For whole archives, SAWT V6 Flash, which keeps its weights and runs with the V6.1 restorer. Both are free for non-commercial use under the SAWT License 2.1.
All posts