autocut

Cut the waffle. Keep the signal. Drop a talk, demo or screen recording in — get a tighter, louder-sounding edit out. One command.

autocut transcribes your video with word-level timestamps, kills silence, fillers and restated sentences, optionally lets an LLM flag redundant passages, and renders a clean result with a broadcast-standard audio chain (EBU R128, −16 LUFS). No Premiere. No DaVinci. No manual scrubbing.

autocut talk.mp4 --speed 1.25 --protect 303:415 --llm
# → output/talk/talk_cut.mp4 · 12:14 → 6:27 · clean audio · usable immediately

Why autocut

Typical screen recordings are 30–50 % pauses, "ums", and restatements. Manual editing is tedious; AI "magic editors" are opaque black boxes.

autocut is the opposite:

• Deterministic — every cut is in cut_plan.json, you can inspect it.
• Honest defaults — the talk preset is tuned on real lecture recordings; nothing touches your video until you like the plan (--dry-run).
• Local-first — Parakeet ASR runs on your machine (CoreML on macOS, CUDA/CPU elsewhere). An LLM pass is optional and works with any OpenAI-compatible endpoint, including Ollama, LM Studio, or your own gateway.
• Zero lock-in — MIT, zero Python runtime deps. pip install trsdn-autocut and go.

Use cases

Install

pip install trsdn-autocut       # or: uv tool install trsdn-autocut
autocut --help

# no install at all:
uvx --from trsdn-autocut autocut talk.mp4

Why two names? The PyPI distribution name is trsdn-autocut because the plain autocut name on PyPI belongs to an unrelated project. The command you run and the Python import package are still autocut — only the pip install name differs.

From source:

git clone https://github.com/trsdn/autocut.git
cd autocut
pip install -e .       # or: uv pip install -e .

That's it on macOS once ffmpeg is installed (brew install ffmpeg). The FluidAudio Parakeet CLI auto-builds on first run.

External dependencies

Parakeet-TDT v3 is the same multilingual model across all three backends.

Quick start

Drop a video into input/ and run:

autocut                       # picks newest video from ./input/
autocut my_video.mp4          # bare filename resolves against ./input/
autocut /abs/path/clip.mov    # absolute paths work too

Output lands under output/<stem>/<stem>_cut.mp4. Intermediates are cached in output/<stem>/.work/ so repeat runs (different settings, same source) skip re-transcription:

output/talk/
├── talk_cut.mp4
└── .work/
    ├── audio_16k.wav
    ├── transcript.json      ← word-level timings
    └── cut_plan.json        ← every kept/removed range, inspectable

Common recipes

# Preserve a negotiation block exactly, speed up the rest
autocut talk.mp4 --speed 1.25 --protect 303:415

# Preview the plan without rendering
autocut talk.mp4 --dry-run

# Use an LLM to flag restated sentences (optional, any OpenAI-compatible API)
export AUTOCUT_LLM_API_KEY=sk-...
autocut talk.mp4 --llm

# Meeting preset: keep turn-taking pauses, don't touch fillers
autocut kickoff.mp4 --preset meeting

# Cross-platform (Windows / Linux)
pip install 'trsdn-autocut[nemo]'
autocut talk.mp4 --asr-backend nemo

How it works

 source.mp4
     │
     ▼
 ┌────────────────────┐      ┌──────────────────────────┐
 │ 1. extract 16k wav │ ───▶ │ 2. ASR (word timestamps) │
 └────────────────────┘      │  FluidAudio / NeMo /      │
                             │  whisper.cpp              │
                             └────────────┬──────────────┘
                                          │
              ┌───────────────────────────┼───────────────────────────┐
              ▼                           ▼                           ▼
     ┌────────────────┐         ┌──────────────────┐         ┌────────────────┐
     │ 3. silences    │         │ 4. filler regex  │         │ 5. LLM pass    │
     │ −35 / −40 dB   │         │ um / äh / ähm    │         │ (optional)     │
     └───────┬────────┘         └─────────┬────────┘         └───────┬────────┘
             └──────────────┬─────────────┴──────────────────────────┘
                            ▼
                 ┌──────────────────────┐
                 │ 6. cut_plan.json     │  ← every cut is auditable
                 │    keep + remove     │
                 └──────────┬───────────┘
                            ▼
                 ┌──────────────────────────┐
                 │ 7. ffmpeg render         │
                 │    split/trim/atrim ·    │
                 │    40 ms fades · speed · │
                 │    highpass · comp ·     │
                 │    de-ess · loudnorm     │
                 └──────────┬───────────────┘
                            ▼
                      output.mp4

Every step is a small, independent module. Swap the ASR, tweak the audio chain, bring your own cut plan.

Presets

Every preset knob is also exposed as a flag — override any single value without abandoning the preset.

Audio chain (talk preset)

Values come from iterating on real bilingual screen recordings — not from a textbook. Tweak freely via flags.

LLM config (optional)

The LLM pass flags semantic cuts — restated sentences, aborted restarts, redundant examples — that regex can't catch. Config precedence:

1. CLI flags: --llm-base-url, --llm-api-key, --llm-model, --llm-max-tokens, --llm-temperature
2. Env vars: AUTOCUT_LLM_BASE_URL, AUTOCUT_LLM_API_KEY, AUTOCUT_LLM_MODEL, AUTOCUT_LLM_MAX_TOKENS, AUTOCUT_LLM_TEMPERATURE
3. Config file: ~/.config/autocut/config.json (see config.example.json)

Works with anything that speaks /v1/chat/completions: OpenAI · Azure OpenAI · Ollama · LM Studio · together.ai · Groq · your own gateway.

Default: gpt-5.4-mini, 16 384 output tokens, temperature 0.1.

CLI reference

autocut [INPUT] [-o OUTPUT] [--root DIR]
        [--preset talk|meeting|vlog]
        [--speed FLOAT] [--fade-ms INT]
        [--protect START:END ...] [--no-fillers] [--language en|de]
        [--pause-above SEC] [--pause-keep SEC]
        [--silence-db DB] [--silence-db-fine DB]
        [--asr-backend fluidaudio|nemo|whisper-cpp]
        [--whisper-model PATH] [--nemo-model HF_ID] [--transcript JSON]
        [--llm] [--llm-base-url URL] [--llm-api-key KEY]
        [--llm-model NAME] [--llm-max-tokens INT] [--llm-temperature FLOAT]
        [--crf INT] [--preset-enc NAME]
        [--loudnorm-i LUFS] [--loudnorm-tp DB]
        [--compressor-ratio FLOAT] [--deess-db DB] [--highpass-hz HZ]
        [--workdir DIR] [--keep-intermediates] [--dry-run]

Run autocut --help for the full list with descriptions.

Platform notes

• macOS (recommended) — FluidAudio Parakeet-TDT runs on CoreML at ~150× realtime. First run builds the CLI (~1 min) and downloads the model (~600 MB).
• Windows / Linux — pip install 'trsdn-autocut[nemo]', then --asr-backend nemo. Parakeet-TDT v3 via NeMo with chunked (30 s) inference; uses CUDA if available, falls back to CPU. First run downloads the model (~2.5 GB).
• Minimal-deps fallback (any OS) — --asr-backend whisper-cpp --whisper-model /path/to/ggml-large-v3.bin.

Project layout

src/autocut/
├── cli.py           # argparse entry point, presets, orchestration
├── audio.py         # wav extraction / probe helpers
├── deps.py          # external tool detection + FluidAudio bootstrap
├── transcribe.py    # ASR backend dispatcher
├── nemo_backend.py  # cross-platform Parakeet via NeMo (optional extra)
├── silence.py       # silencedetect parsing + snap-to-boundary
├── plan.py          # cut plan: fillers, pauses, redundancies
├── render.py        # ffmpeg filter_complex render
└── llm.py           # OpenAI-compatible redundancy detection (stdlib urllib)

Every module is under 200 lines. Read it. Fork it. Make it yours.

Contributing

Issues and PRs welcome. This started as a tool to shorten a 12-minute negotiation demo; it generalised from there. If you hit an edge case — unusual language, weird microphone, exotic codec — open an issue with the source clip (or a redacted transcript) and I'll take a look.

License

MIT — see LICENSE.

Acknowledgements

• FluidAudio — CoreML Swift ASR runner
• NVIDIA Parakeet-TDT — the excellent multilingual ASR model
• NeMo toolkit — cross-platform inference
• whisper.cpp — portable fallback
• FFmpeg — without which none of this would exist