Pauses, classified seven ways
Every gap is labeled filler, hesitation, breath, emphasis, pre-laughter, transition or listening — each with its own behavior. Breaths get trimmed to 150 ms, not deleted. That's the line between natural and robotic.
AI semantic video editor · open source · 100% Free
Edit by listening,
not by scrubbing.
Cadence Lab drops your OBS recording into an AI pipeline: it transcribes the audio, asks Claude to judge every pause and filler word in context, and renders a tight, YouTube-ready cut. Then you just ask it to refine the edit.
-
100% FREENo subscription, ever
-
Open sourceMIT licensed on GitHub
-
Local-firstYour footage stays put
- Powered by
- Whisper large-v3
- /
- Claude Opus
- /
- FFmpeg
- /
- CLIP
- /
- DeepFilterNet
- /
- Tauri
Why it's different
Most “auto-edit” tools are just a regex over the waveform.
They cut anything below −30 dB and call it done — butchering intentional pauses, flattening breaths into something robotic, and keeping both takes when you flub a line. Cadence Lab reads the transcript, not just the volume, so it edits like a human who was actually paying attention.
Features
Editing decisions made by something that understands the words.
Context-aware filler removal
“Like” used as filler gets cut. “Nothing else like it” gets kept. The classifier reads the surrounding words — no blunt keyword deletion.
Retake detection
Flub a sentence and start over? Cadence spots the second attempt — or a “let me try that again” — and flags the worse take, the way an editor would.
Ask Cadence
“Cut every sniffle.” “Find when the walnut table is on screen.” “Pull a 60-second highlight.” Claude executes against your video and proposes edits — you accept each with one click.
Semantic visual search
CLIP frame embeddings indexed at 1 fps. Ask “find the part where the dog appears” and get ranked timestamps — without scrubbing the whole tape.
Sound-event cleanup
An AudioSet-trained model spots sniffles, coughs, throat-clears and sneezes. Pair it with “remove all sniffles” and Cadence proposes a precise cut for each.
Neural denoise
DeepFilterNet — trained on ~100k hours of speech — clears fan hum, keyboard clicks and room noise far better than a classical filter, near real-time on CPU.
Splicing timeline
Extract highlight clips, rearrange them, drop black between, and render an assembled cut — a dedicated path separate from the main pacing edit.
Hardware-fast render
Hardware H.264 by default — 5–15× faster than a CPU encode YouTube can't tell apart — with an opt-in archival libx264 mode when you want the master.
How it works
A typed pipeline, from raw recording to finished cut.
Each stage writes a structured file the next one reads. Stop anywhere, tweak, and resume.
1
Ingest & transcribe
Cadence extracts your mic track alone, then transcribes with Whisper large-v3 (~30× realtime via Groq, or fully local). Word-level timestamps; desktop audio never masks the speech.
2
Classify in context
A single Claude Opus call reads the whole transcript and judges every pause and filler candidate with schema-constrained output — plus detected retakes. No regex, no fragile parsing.
3
Plan the cuts
Pure interval algebra turns those decisions — plus any cuts you or Ask Cadence add — into clean keep-segments. Fully local, deterministic, with an audit log of the original intent.
4
Review & render
Listen to a 3-second clip around each cut, override with a click, re-plan instantly. Then FFmpeg renders a frame-aligned, loudness-normalized MP4 — ready to upload.
Ask Cadence
Talk to your timeline.
A dedicated Claude agent sits on top of your edit with read tools and action tools. Ask in plain English; it queries your transcript, audio events and visual index, then proposes typed edits you apply one at a time. Kick off a long scan and it picks the conversation back up the moment it finishes.
- remove the um at 1:23
- cut every sniffle in the intro
- find when the walnut table is on screen
- pull a 60-second highlight from the demo
Every action is a proposal. Nothing touches your video until you click Apply.
cut every sniffle in the intro
Ask CadenceScanning for non-speech events… found 4 sniffles between 0:00–2:10. Proposing a custom cut for each.
also find where I show the table
Ask CadenceIndexing frames… best match for “table on screen” is 14:02–14:48. Want a highlight clip from that range?Create clipNot now
Open source
Yours to run, read, fork and ship.
MIT licensed. Your media and API keys stay on your machine — there's no Cadence Lab cloud in the loop.
Local-first & private
Everything runs on your computer. Bring your own Groq and Anthropic keys, or go fully offline with local Whisper. Your footage never leaves the device unless you pick a cloud transcription backend.
Honest pricing
The app is free. You pay your own model costs — about $0.60–$2.00 for a 30-minute video, mostly transcription and one big classification call. No subscription, no markup.
A real reference build
Structured outputs, prompt caching, agentic tool use, a typed multi-stage data contract, a Tauri + React shell over a Python FastAPI sidecar. Read the code — it's a working example of all of it.
Get the source
Run it from source today.
Cadence Lab is in early open-source development. Clone the repo and run it from the command line — one-click installers are on the roadmap.
# 1 — prerequisites
brew install ffmpeg uv
# 2 — clone & run
git clone https://github.com/JosephLeon/Cadence-Lab
cd Cadence-Lab
uv sync && cp .env.example .env
uv run cadence-lab serverPre-release. macOS & Windows installers are on the roadmap (donations welcome). Until then, developers can clone the repo and run from source; non-developers, hang tight.
Full setup — API keys, the desktop app, all of it — is in the
install guide.
Requires ffmpeg and your own API keys.