audio-denoise · Audio Denoising
Remove background noise, hum, wind, and buzz from recordings. Three denoise tiers: basic FFT filtering, enhanced denoising, and RNN neural-network denoising (voice-optimized) — chosen automatically or manually by noise level.
Deterministic ffmpeg filter chains with transparent, reproducible commands; video inputs get audio-only processing with the picture untouched; original files are never deleted.
Example invocation: "Denoise this podcast recording — there's air-conditioner hum in the background."
Full brief
Positioning
audio-denoise is a dedicated denoising tool: it cleans background noise from recordings. It handles denoise tuning and execution only — not general audio editing, transcoding, or leveling.
Core capabilities
- Three denoise tiers: Tier 1 basic FFT filtering (light noise, pure ffmpeg); Tier 2 enhanced denoising (moderate/steady-state noise like AC and fans — the default); Tier 3 RNN neural-network denoising (voice-optimized for podcasts, interviews, complex noise).
- Deterministic filter chains: the script assembles the ffmpeg chain per tier and prints the exact command it runs — reproducible, never hand-built on the spot.
- Video-safe: video inputs get audio-only processing; the video track is copied untouched.
- Original-file protection: originals are never deleted; preservation is the default.
- Graceful downgrade: if the Tier 3 RNNoise model is missing, it falls back to Tier 2 with a download pointer.
Workflow
- Probe the input: report duration/bitrate/channels, determine audio vs. video
- Pick a tier (default Tier 2; Tier 3 for voice recordings — needs the RNNoise model)
- Run denoise into
outputs/<topic>/, with a report (before vs. after, tier and filter chain used, size delta) - If over-denoised, lower the tier or reduce
--mixand re-run
Inputs & outputs
| Input | Required | Notes |
|---|---|---|
| input_file | Yes | Audio or video path (wav/mp3/flac/aac/m4a/mp4/mkv/mov) |
| tier | No | 1 / 2 / 3, default 2 |
| mix | No | Denoise strength 0.0–1.0 (default 0.8, Tier 3 only) |
| output_file | No | Default outputs/<topic>/{filename}-clean.{ext} |
Output: denoised audio file + processing report.
Boundaries with adjacent skills
| Skill | Lane |
|---|---|
| audio-denoise (this) | Denoise tuning specialist |
| audio-editing | General audio ops (cut/transcode/level), with denoise fallback |
| auto-subtitle | Speech-to-subtitles; no denoising |
| audio-mix | Mixing/BGM; no denoising |
Fit
- Podcast and interview recordings with AC/fan/floor noise
- Hum and wind cleanup in talking-head recordings
- Meeting and course recordings with background interference
Before you start
- No API key; local ffmpeg processing, no metered cost.
- Tier 3 needs the RNNoise model
sh.rnnn(downloadable from GitHub); missing models degrade to Tier 2 automatically — never blocking. - Denoise strength trades against voice quality: over-denoising damages vocals — listen to the report output and fine-tune.
audio-denoise is part of the Aiglade Skill library. Invoke it from the Aiglade chat box in plain language.