AutoCut for DaVinci Resolve

Automatic silence removal with real J-cuts — for the free version of DaVinci Resolve, which normally can't be scripted from outside at all.

Point it at your timeline and it builds a new one next to it: silences cut, and at every edit the previous shot holds over the start of your next line, so you hear the new sentence before the picture cuts. Your voice track is never moved or padded — the J-cut comes from trimming video, the way an editor would do it by hand.

Not affiliated with the commercial AutoCut plugin — this is an independent open-source tool.

Features

• Silence removal with per-file auto threshold (measures each clip's noise floor — works on camera audio where a fixed dB threshold finds nothing)
• Real J-cuts in the built timeline (configurable lead, off by default)
• Preview before cutting: analyze first, see every planned cut with its reason, toggle any of them off, then build
• Whisper AI features (all offline): filler-word removal, SRT captions remapped to the cut timeline, chapter markers on long pauses
• Apple GPU transcription via mlx-whisper, including Thai fine-tuned models; caption line-breaking is Thai-aware (never splits า/ำ from the word)
• Presets with built-in templates; progress bar and cancellable runs
• FCPXML export of the same edit for Premiere Pro / Final Cut interchange

How it works on the free version

Free Resolve blocks external scripting — but scripts run from its own menu are allowed. AutoCut installs a tiny bridge script into Resolve's Scripts menu; the desktop app talks to it through JSON files. One click on Workspace → Scripts → Utility → AutoCut Bridge starts the bridge and opens the app (macOS), which connects automatically.

There is no razor/trim API in any Resolve edition, so "editing" means building a new timeline from your current one — your original is never modified.

Requirements

• DaVinci Resolve 18+ (free or Studio). Tested on Resolve 21 free from the Mac App Store — the most restricted variant there is.
• Python 3.12+ managed by uv for the app
• For Resolve's script runner on macOS: Python from python.org (Homebrew/Xcode Pythons are invisible to Resolve)
• macOS is the primary platform (Apple Silicon gets GPU transcription). Windows/Linux bridge paths are implemented but untested — reports welcome.

Install

git clone <this repo> && cd autocut
./install.sh          # installs uv if needed, then dependencies

Then open DaVinci Resolve with a project, and run Workspace → Scripts → Utility → AutoCut Bridge (the app installs the script on first connect; run uv run python main.py once if the menu entry is missing). On macOS the app opens automatically from then on — one click is the whole startup.

Uninstall

uv run python uninstall.py            # removes the Resolve scripts, launch agent, bridge files
uv run python uninstall.py --purge    # also removes ~/.autocut (presets + converted models)

Add --dry-run to preview. The repo folder and the shared HuggingFace model cache are never touched.

Usage

1. Open the timeline you want to cut (it is read as-is; video/audio track 1 by default).
2. Pick a preset — Talking Head (J-cut) is the usual starting point.
3. Analyze (preview) — review the cut list on the Preview tab. Click a row's ✓ to keep that section instead of cutting it.
4. ▶ Build Timeline — a new timeline named <name> [AutoCut hh.mm.ss] appears, cuts landing clip by clip.

Or skip the preview with ▶ Run AutoCut.

Thai transcription (and other languages)

The model box accepts any faster-whisper model name or HuggingFace repo, plus mlx: prefixed repos for Apple-GPU inference. Included aliases:

Alias	What it is
base / medium / …	vanilla Whisper, CPU
large-v3-turbo-mlx	fast multilingual, Apple GPU
thai-distill-mlx	distilled Thonburian Whisper Thai fine-tune, Apple GPU
thai-large-v3-mlx	full Thonburian Thai model, Apple GPU — best Thai accuracy, needs a one-time local conversion (below)

Converting the full Thai model (no published MLX build exists):

curl -sLO https://raw.githubusercontent.com/ml-explore/mlx-examples/main/whisper/convert.py
uv run python convert.py --torch-name-or-path biodatlab/whisper-th-large-v3-combined \
    --mlx-path ~/.autocut/models/thai-large-v3-mlx --dtype float16
mv ~/.autocut/models/thai-large-v3-mlx/model.safetensors \
   ~/.autocut/models/thai-large-v3-mlx/weights.safetensors

If the conversion isn't present, thai-large-v3-mlx falls back to the distilled model automatically.

Limitations (honest ones)

• The built timeline is new; nothing is trimmed in place (no such API exists).
• On the free Mac App Store build, importing the FCPXML back into Resolve yields offline media — FCPXML mode is for Premiere/FCP; Resolve uses the live-append mode, which is the default.
• Captions: the app tries to place the SRT on a subtitle track; if your build refuses, the SRT lands in the media pool and one drag finishes the job.
• J-cut holds only happen between segments of the same source clip (that's where the removed silence provides the video to hold on).

Development

uv run python tests/test_autocut.py    # offline suite, no Resolve needed

Full documentation — pipeline internals, every parameter, the bridge protocol, and the Resolve API quirks this tool works around (sandboxed script processes, FCPXML importer behavior, Tk 9 threading) — lives in docs/, published via GitHub Pages. Preview locally with uv run --group docs mkdocs serve.

Credits

• Workflow inspired by YourAverageMo/auto-silence-cut
• Thai models by biodatlab (Thonburian Whisper)
• faster-whisper, mlx-whisper, librosa

MIT licensed — see LICENSE.