clean-audio — voice cleanup

Take a locked master cut and remove its background noise, producing a cleaned master whose voice sounds natural and whose visuals are untouched. Runs early (once the cut is locked) so everything downstream — TSX bake, SFX mix, final assemble — sits on the clean voice. Work with the user; the final loudness/limiting is the final-mix step's job, this step is "denoise only, levels preserved."

The engine is tools/clean_voice.py; this skill is the judgment around it: diagnose → pick method → A/B → clean → rewire.

The two methods (pick by NOISE TYPE — this is the core decision)

Method	What it is	Use when	Cost
--method eleven	ElevenLabs Voice Isolator (cloud ML voice/noise separation)	Dynamic, broadband noise in the voice band — outdoor running water, wind, traffic, crowd, cafe. Local tools CANNOT remove these.	1000 credits/min ($1 for a 5.5-min video); needs ELEVENLABS_API_KEY
--method rnnoise --model sh (or cb)	Local RNNoise via ffmpeg arnndn (models in tools/models/rnnoise/)	Stationary / mild noise (steady hiss, fan, some room tone). Free/offline. Only PARTIALLY removes dynamic noise.	free

Proven on video-1 (shot outdoors with a stream): afftdn did ~nothing, RNNoise only partially darkened the water bed, ElevenLabs removed it near-completely (pauses to near-silence, voice + breaths intact). Rule of thumb: stationary noise → try local first; dynamic broadband (water/wind/traffic) → ElevenLabs.

Inputs (read/measure first, every time)

• The master — videos/video-N/reference/<cut>.mp4 (or the locked cut). Original is NEVER modified; output is a new -clean / -clean-<model> file.
• The composited preview (if it exists) — videos/video-N/output/video-N-preview.mp4, to make a clean in-context preview by swapping audio (its video is identical — no re-bake needed).
• videos/video-N/work/timeline.json — its master field; you rewire this to the clean master on approval.

Workflow

1. Diagnose the noise BEFORE choosing a method. Measure and look:
   • Levels: ffmpeg -i M -vn -af astats (RMS, peak, noise floor) + ebur128 (integrated LUFS, true peak).
   • Find speech-free gaps (grep edited-transcript.json for the biggest inter-word gaps) and measure the pure-noise RMS there vs speech RMS → the real SNR.
   • Spectrogram: ffmpeg -i M -vn -lavfi showspectrumpic=s=1500x600:legend=1:scale=log out.png and LOOK at it. Hum = steady horizontal lines (50/60Hz) → notch. Rumble = low band → high-pass. Broadband bed that fills the voice band and fluctuates = dynamic (water/wind) → ElevenLabs. HF hiss = bright top band.
   • Note if the export is already produced (compressed/normalized/peak-maxed) — it limits what's recoverable.
2. Decide the method with the user from the diagnosis (table above). If unsure, A/B both.
3. A/B on a short sample FIRST (prove before spending / committing): cut a ~15s pause-rich sample, run each candidate method, level-match them to each other, and compare — by ear (the real test) AND by spectrogram (pauses going dark = noise removed) and residual level. Let the user pick.
4. Clean the full master: python tools/clean_voice.py videos/video-N/reference/<cut>.mp4 --method <chosen> [--model sh] → <cut>-clean.mp4 (or -clean-<model>.mp4). Video stream COPIED (fast, non-destructive, keeps 4K60).
5. Levels are preserved by RMS-match, not LUFS (the tool does this). Never match integrated LUFS — it is gated and inflated by the removed noise, and over-boosts the voice into clipping. The clean file will read a lower integrated LUFS than the noisy original; that is expected (the noise was padding the number), the voice RMS is unchanged. Final loudness to -14 LUFS is the final-mix step's job.
6. Give the user an in-context preview (optional but recommended): swap the clean audio onto the composited preview — ffmpeg -i preview.mp4 -i <cut>-clean.mp4 -map 0:v -map 1:a -c:v copy -c:a aac -shortest preview-clean.mp4 (video identical, no re-bake). For a full A/B, also export FULL_*.mp3 scrub files.
7. On approval, rewire the pipeline: point timeline.json "master" at the clean file so every future bake/mix uses the clean voice; re-bake the preview if needed.

Decisions to surface to the user

• Method (from the diagnosis) — and A/B if unsure.
• Dead-silent gaps vs a faint ambience bed. Voice isolation removes ALL background; on an outdoor shot the dead-silent gaps can feel vacuum-sealed. Offer to add back a low-level neutral ambience if wanted.
• Cost for ElevenLabs (~1000 credits/min) — confirm before running on the full master.

Principles (the house style)

• Least processing that works. The goal is to remove distraction, not to make the voice sound processed. Prefer the gentlest method that clears the noise; don't over-strip a clean track.
• Diagnose, then choose. The right tool depends on the noise type — never crank a denoiser blind. Local spectral/RNNoise can't separate dynamic broadband noise; that's ElevenLabs' job.
• A/B before you commit (and before you spend). Prove on a sample; the user's ears decide.
• Preserve levels; loudness is the final-mix step's. RMS-match with a peak ceiling, no compression here.
• Non-destructive. Original master untouched; video stream copied; output is a new file.

Tooling quick reference

• Clean: python tools/clean_voice.py IN.mp4 [--method eleven|rnnoise] [--model sh|cb] [-o OUT.mp4] [--no-preserve-loudness] [--keep]
• Diagnose: ffmpeg -i M -vn -af astats -f null - · ffmpeg -i M -vn -lavfi showspectrumpic=... out.png (then Read the png).
• RNNoise models: tools/models/rnnoise/<model>.rnnn (sh, cb).
• Scratch samples/spectrograms go in the scratchpad, not the project.

Done = the noise is diagnosed, the method is chosen (A/B'd if needed), the full master is cleaned with levels preserved, the user has approved by ear, and — on approval — timeline.json points at the clean master. Update memory if a noise-type → method lesson emerges.