CleanTake 0.1.0 — recover dialogue from your backup recordings
CleanTake recovers damaged dialogue from simultaneous recordings. This first release includes the local browser studio, CLI, saved projects, reversible repair decisions and portable exports.
Import your main microphone and its backups, align them, compare Original/Repair/Source at the same moment, and keep each source switch editable. Export float WAV or PCM24 FLAC, aligned stems, a Reaper session, a contributor source map or a portable project archive. Optional finishing preserves the unmastered mix and reports actual loudness, including a warning when its target is missed.
Install and try the sample
Install FFmpeg/FFprobe, then install the attached wheel with Python 3.12+:
uv tool install --python 3.12 ./cleantake-0.1.0-py3-none-any.whl
cleantake doctor
cleantake demo
cleantake studioThe wheel includes the studio and licensed offline sample. cleantake-demo.zip is the same sample with its sole automatic proposal explicitly accepted; import it through the studio's portable-project action to inspect that result. The default cleantake demo leaves the proposal pending for your own review.
The demo recordings are from the AMI Project Consortium, meeting ES2004a, and are licensed under CC BY 4.0. The headset includes an injected half-second gap; the accepted repair uses the simultaneous lapel recording. Keep the attribution and modification record with the demo when redistributing it.
Evidence and limits
Native release checks passed: 242 Python tests on Linux (3.12/3.13/3.14) and macOS, and 238 on Windows with four explicitly POSIX-only skips. All five jobs built and installed the wheel outside the checkout, then completed a real sample recovery/export/archive/studio check. All 24 browser workflows passed across Chromium, Firefox and WebKit, followed by eight additional bundled-studio workflows under production security headers. The published evaluation includes untuned AMI excerpts, missed proposals, exact contributor reconstruction, a 30-minute resource run and the actual same-input finishing baseline.
Automatic proposals need listening review and can miss damage. The output timeline is 48 kHz mono. There is no recorded replacement when every source lost a passage. Independent listening preference, organic-damage coverage and editor-time advantage have not been established.