Create a Production FX track from location recordings: retain clothing movement, footsteps, handling and props while reducing dialogue, breathing, mouth noises and unwanted ambience.
The Windows frontend transfers files through Google Drive. Audio processing runs in Google Colab; the final download is a Pro Tools session with its audio files.
Both backends are supported and use the same workflow. Choose by listening to the results on your recordings.
| Bandit V4.2 | MDX23C + BS-RoFormer V3.5.2 | |
|---|---|---|
| Separation | One model trained for film dialogue/music/effects separation; uses its SFX output | Two models whose instrumental outputs are blended |
| Scene handling | Each WAV processed independently | Aligns BOOM/LAV when paired; temporarily groups short clips, then splits them back |
| Bodytalk gain | Up to +4 dB, with local peak headroom | Up to +3.5 dB, with peak headroom for the processed group |
| When to choose | For speed and processing each scene independently | An alternative when its human-sound removal sounds better on your footage |
Both default to 40% adaptive pre-cleaning and cumulative YAMNet removal: detection on the raw recording first, then residual detection on the cleaned output. The residual detector uses temporary gain up to +18 dB; this does not raise the output volume. Bodytalk gain targets clothing-movement detections, not footsteps.
Neither backend removes every human sound. Strong removal can also reduce simultaneous Foley; listening and manual finishing remain necessary.
Bandit retains V4.1's priority removal of shouting and screaming at 95%, amplified residual detection, and inhibition of bodytalk gain where detected. This can further reduce simultaneous Foley. V4.2 changes dependency checks only: it accepts the documented TorchAudio 2.11 / PyTorch ≥2.11 pairing, checks CUDA build tags and loaded/installed versions, and keeps pip from replacing either package during auxiliary dependency installation. V4.1 and V4 are retained; audio code and settings are unchanged.
Requires Windows, Python 3.10+, Git, Google Drive and a GPU-enabled Colab runtime. A T4 is the usual processing configuration.
- Download or clone this repository.
- Place Google OAuth Desktop app credentials at the root as
credentials.json, and supply a compatibletemplate.ptx. The template is proprietary and not distributed here; its profile is documented in architecture.md. - Run
start.bat. It creates the Python environment and installs dependencies on first launch. - Open localhost:7862 and authorize Drive when prompted.
To update an existing installation, run this in the project folder, then restart the frontend:
.\.venv\Scripts\python.exe -m pip install -r requirements.txtPro Tools export requires pt_api 1.6.2 or later. Dependencies install v1.6.2; older or unidentified loaded versions are refused.
- Add mono 48 kHz BWF WAV files with their original timestamps. One or two clearly named sets are accepted:
BOOM/LAV, or a stable prefix followed by a clip number (for examplePr_Rep_1-Gain_01.wavandPr_Rep_2-Gain_01.wav). Two sets must have equal file counts; their clip numbers need not match. Generic names do not imply BOOM/LAV roles or trigger microphone alignment. - Click Upload vers Google Drive. The progress bar shows the current file, its transfer percentage and the number of files sent.
- Open either backend in Colab and run its cells in order. You can save its URL in Paramètres avancés for the Ouvrir Google Colab button.
- Once processing finishes, click Télécharger les fichiers traités en ZIP.
- Extract the ZIP and open the
.ptxsession. It includesAudio Files/, with clips placed at their BWF timestamps on up to two PFX tracks, at 48 kHz / 23.976 fps.
Switching backends requires no manual Drive changes. Previous output generations are retained; incomplete or inconsistent deliveries are refused. Run only one notebook at a time.
Keep the ZIPs you need before using Effacer la cache: after confirmation, it permanently removes active local working files/exports, files in the Drive input/output folders, and generated WAVs/manifests in their run_… generations. Empty generation folders are removed too. Progress and the final summary report deleted files, folders and errors. Model caches, notebooks and unrelated folders are preserved. A running or unreadable manifest blocks cleanup; unidentified nonempty generations are preserved and reported. Do not start Colab during cleanup. Restart the frontend after updating it (interface v3.0.1).
Audio stays in your Colab/Drive workflow; no recording is sent to an external separation service. Credentials, tokens, audio caches and the proprietary template are excluded from Git.
If download or export fails, inspect the status and work/app.log. Check that Colab completed, the WAV format is supported, and the local template and API version are compatible. Downloaded WAVs remain available for diagnosis after an export failure.
Technical details: architecture.md. Version history: changelog.md.
Historical notebooks V3.5.1 and Test V3 are local-only archives, excluded from Git. The two supported backends are linked above.
Created by Sébastien Bédard / sebedard-creator. Use, modification and redistribution are governed by the Attribution Assurance License.
Redistribution must retain the license. Redistributed executables and dependent programs must display the required attribution at launch:
sebedard-creator — PFX Extractor developer — https://github.com/sebedard-creator/PFX-Extractor