Restoration for archive recitation. SAWT V6 brings tape, hall recordings and low-bitrate uploads back to 48 kHz. The voice, the words and the melody stay as they were recorded, and the pauses fall silent. Pro gives the best result; Flash does the same job in four steps, for whole archives.
| SAWT V6 Pro | SAWT V6 Flash | |
|---|---|---|
| Task | Speech restoration to 48 kHz mono WAV, with a JSON sidecar of the settings for every file | |
| Generator | 351M flow-matching transformer | 105M student, distilled from Pro |
| Reads the recording with | omniASR CTC 1B | omniASR CTC 300M |
| Sampling steps | 16 | 4 |
| Speed on one RTX 4090, long files | 16× real time | 74× real time |
| Download to restore | 4.6 GB | 1.8 GB |
| Licence | SAWT License 2.1, non-commercial | |
| Latest update | V6.2 weights, 29 October 2026 | V6.1 restorer, 21 October 2026 |
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
Ten-second excerpts of archive uploads in three input-quality tiers, through the release code with the V6 weights and default settings. V5 Pro keeps the speaker a little better; V6 Pro is ahead on every other measure.
| Recording | SAWT V5 Pro | SAWT V6 Pro | |
|---|---|---|---|
| DistillMOS | 2.74 | 3.48 | 3.89 |
| Audiobox PQ | 6.76 | 7.55 | 7.60 |
| Pause floor, median | -36 dBFS | -56 dBFS | -74 dBFS |
| Speech to background | n/a | 29.4 dB | 43.6 dB |
| Speaker consistency | n/a | 0.957 | 0.947 |
DistillMOS and Audiobox PQ are learned quality predictors. The pause floor is the median level of the restored audio in the recording's pauses. Next to five other open restorers on 120 archive clips, V6 Pro has the highest DistillMOS; that table is in the story. Read the story
Two updates in the first ten days. The original V6 weights stay in the repository; the tables above and the restored recitations were made with them.
V6.1 · 21 October 2026 · restorer, Pro and Flash
The pause gate no longer fires inside speech on crushed or limited uploads: 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%. Stretches of digital silence in the input now stay silent.
V6.2 · 29 October 2026 · new Pro weights
The V6 weights, fine-tuned with preference optimisation on restorations of real recitation and of general speech, ranked by automatic measures of intelligibility, quality and speaker identity. Phoneme error falls most on the hardest recordings.
| 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 |
Ask for access on Hugging Face and accept the licence, then download only the files that restore audio. Every setting below is the evaluated default.
pip install huggingface_hub
python - <<'EOF'
from huggingface_hub import snapshot_download
snapshot_download("Quran-Lab/sawt-v6-pro", local_dir="sawt-v6-pro", allow_patterns=[
"config.json", "sawt2/*", "requirements.txt",
"generator.pt", "omniasr-ctc-1b/*", "vae48_dacvae_d2.pt", "dacvae_stats.pt",
])
EOF
cd sawt-v6-pro
pip install -r requirements.txt
python sawt2/restore.py --in recordings/ --out restored/A CUDA GPU is required. Inputs can be WAV or FLAC at any sample rate; outputs are 48 kHz mono WAV, each with a JSON sidecar of the settings.
For Flash, use the repository Quran-Lab/sawt-v6-flash and the folder omniasr-ctc-300m in place of omniasr-ctc-1b. The same restorer reads Flash's recipe from its generator.
--steps 8faster, with a little less detail--pause-gate 0keep the recording's room in the pauses--lanes 4more files at once, on a GPU with 40 GB or more