| name | cleanvoice-audio |
| description | Improve short-form video voice audio with Cleanvoice. Use when the user wants studio-quality phone/selfie audio, before/after audio demos, or a timing-preserving voice replacement pass for Reels, TikToks, Shorts, or vertical video. |
Cleanvoice Audio
Use this skill to polish spoken audio without disturbing the locked picture edit.
Workflow
- Identify the locked input video and the desired output path. Use
.voice.mp4 for the enhanced version.
- Confirm
CLEANVOICE_API_KEY is available in the environment or .env.local.
- If this repo's pipeline is present, prefer:
python3 pipeline/enhance_audio.py INPUT.mp4 --provider cleanvoice --out OUTPUT.voice.mp4
- If the repo pipeline is not present, use the Cleanvoice REST flow:
- Request a signed upload URL with
POST https://api.cleanvoice.ai/v2/upload?filename=....
PUT the local audio or video file to the signed URL.- Submit
POST https://api.cleanvoice.ai/v2/edits with the uploaded file URL.
- Poll
GET https://api.cleanvoice.ai/v2/edits/{id} until SUCCESS.
- Use conservative enhancement defaults for iPhone Reel voice:
{
"remove_noise": true,
"studio_sound": true,
"normalize": true,
"target_lufs": -14,
"export_format": "wav"
}
- Mux the enhanced audio back into the existing video stream. Do not rerender video frames for audio-only changes.
ffmpeg -y -i LOCKED_VIDEO.mp4 -i CLEANED_AUDIO.wav \
-map 0:v:0 -map 1:a:0 \
-c:v copy -c:a aac -b:a 256k -shortest -movflags +faststart \
OUTPUT.voice.mp4
Taste Rules
- Keep a deliberate raw sample raw when the video is proving the before/after difference.
- Do not cut the end of words. Leave 0.15-0.30s after important comparison lines.
- If Cleanvoice creates doubled consonants or blurred speech, rebuild only that audio section from the raw source and re-mux.
- Preserve caption timing unless the spoken waveform clearly shifted.
- Do not put copyrighted music into the voice enhancement pass; process voice first, add music later.
Verification
- Compare a raw/clean voice sample around the hook and the CTA.
- Confirm the output duration still matches the locked video.
- Confirm the video stream was copied, not rerendered, when only audio changed.
