Parameter Reference¶

Every setting in the GUI maps 1:1 to a field on AutoCutParams (autocut.py). Presets are snapshots of these values.

Silence removal¶

remove_silence — default: on: The core feature. Detects below-threshold regions in each clip's source audio and plans a cut for each one.
auto_threshold — default: on: Pick a silence threshold per file by measuring its noise floor and speech level (see How It Works). Recommended: camera auto-gain keeps "silence" around −20 dB on real footage, where any sane fixed threshold finds nothing.
silence_threshold_db — default: −42.0: Fixed threshold in dBFS, used only when auto_threshold is off. Audio below this level counts as silence. Only reach for this when the auto threshold misjudges a specific file (e.g. music beds under speech).
min_silence_sec — default: 0.40: A quiet region shorter than this is not cut. Lower values (0.3) give a tighter, faster-paced edit; higher values (0.6+) preserve natural breathing pauses — the podcast template uses 0.6.
padding_frames — default: 6: Frames kept on each side of every detected silence, so words are never clipped mid-syllable. At 24 fps the default keeps a quarter second around each cut. If cuts feel like they swallow word endings, raise this before touching anything else.

Filler word removal¶

remove_fillers — default: off: Cut transcript words that match the filler list. Requires Whisper (a transcription run happens during Analyze).
filler_words — default: um, uh, uhm, uhh, er, erm, hmm: Comma-separated, case-insensitive; punctuation is stripped before matching, so "Um," matches "um". Each removed filler is cut with one frame of margin on each side.

J-cut¶

jcut_frames — default: 0 (off): How many frames the previous shot's video holds over the start of the next line at each cut. The audio is never moved — see How It Works for the exact hold math. 4–8 frames reads naturally; more starts to feel like a mistake. Holds only happen between segments of the same source clip.

Punch-in zoom¶

punch_in — default: off: Alternate every second clip on the built timeline to a zoomed-in framing — the standard trick for hiding jump cuts in a single-camera setup.
zoom_amount — default: 1.15: Zoom factor for the punched-in clips. 1.1–1.2 is subtle; beyond 1.3 you will see the crop on 1080p footage.

Captions¶

captions — default: off: Generate an SRT from the transcript, remapped onto the cut timeline so timings match the edit. Requires Whisper. The app tries to place it on a subtitle track; if your Resolve build refuses, it lands in the media pool for a manual drag. Thai text gets syllable-safe line breaking.

Chapters¶

chapters — default: off: Add a purple marker at every segment preceded by a long removed pause, titled with the first words spoken there. Requires Whisper. Useful as YouTube chapter starting points.
chapter_gap_sec — default: 2.0: Minimum removed source time before a segment for it to count as a chapter boundary. This measures the pause you actually took while recording, not time on the finished timeline.

Whisper¶

model_size — default: base

Which model to transcribe with. Accepts any faster-whisper size name or Hugging Face CTranslate2 repo id (CPU), any mlx:-prefixed repo (Apple GPU), or one of the built-in aliases:

Alias	What it is	Engine
`base`, `medium`, …	vanilla Whisper sizes	CPU
`large-v3-turbo-mlx`	fast multilingual	Apple GPU
`thai-distill-mlx`	distilled Thonburian Thai fine-tune	Apple GPU
`thai-large-v3-mlx`	full Thonburian Thai model — best Thai accuracy, needs a one-time local conversion (below)	Apple GPU
`thai-large-v3`	same weights on CPU, roughly 10× slower	CPU

Models download on first use and are cached in ~/.cache/huggingface. Everything runs locally — no API keys.

language — default: auto-detect

ISO 639-1 code (en, th, de, …). Setting it explicitly skips detection and helps short or noisy clips.

Converting the full Thai model¶

No published MLX build of the full Thonburian model exists, so it needs a one-time local conversion:

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

(The converter writes model.safetensors; mlx-whisper loads weights.safetensors — hence the rename.) If the conversion isn't present, thai-large-v3-mlx falls back to the distilled model with a warning.

Build¶

build_mode — default: stream: stream appends segments to the new timeline live inside Resolve. fcpxml writes the whole edit as an FCPXML file instead — for Premiere Pro / Final Cut interchange; the free Mac App Store Resolve imports FCPXML with offline media (see How It Works).

Tracks¶

video_track, audio_track — default: 1: Which timeline tracks to read. Only one video track is processed per run; clips on other tracks are ignored.

Presets¶

Presets live at ~/.autocut/presets.json — plain JSON, one object per preset, keys matching the parameter names above. Built-in templates are always listed; saving a user preset under a template's name overrides it, and deleting that preset restores the template.

Template	What it changes from the base defaults
Talking Head (J-cut)	J-cut 4f
Podcast / Interview	min silence 0.6 s, padding 8f, J-cut 8f, chapters on (3 s gap)
Tutorial (tight cut)	min silence 0.3 s, padding 4f, J-cut 4f, captions on
Social Clip (subtitles + zoom)	min silence 0.3 s, padding 4f, fillers removed, captions + punch-in on