# Audio Edit

> Clean up or mix a video's audio track — normalize loudness, reduce background noise/hiss, or add background music (optionally ducked under speech). Use whenever the user wants to fix quiet/loud/inconsistent audio, remove hiss/hum, or lay music under a talking-head/voiceover. For removing um/uh filler words use filler-removal instead; for replacing the audio entirely use plain ffmpeg map.

- Skill: `agamm/audio-edit` (Agent Skill)
- Install (CLI): `npx skillmds@latest add agamm/audio-edit`
- Raw SKILL.md: https://api.skillmd.com/api/skills/agamm/audio-edit/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: agamm (https://skillmd.com/u/agamm)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/agamm/audio-edit

---


# Audio editing (ffmpeg)

Loudness normalize, denoise, and background-music mixing. These are plain ffmpeg
one-liners — no CLI command wraps them. Every recipe **stream-copies the video**
(`-c:v copy`, fast + lossless) and re-encodes only the audio to `aac`.

Pick the operation:
- Audio too quiet / loud / inconsistent across clips → **normalize**.
- Background hiss / hum / fan noise → **denoise**.
- Lay a music bed under narration → **add music** (use `--duck`-style sidechain so music
  drops under speech).
- Cutting um/uh hesitations → use the `filler-removal` skill, not this one.

## Normalize loudness

```bash
ffmpeg -y -i in.mp4 -af loudnorm=I=-16:TP=-1.5:LRA=11 -c:v copy -c:a aac out.mp4
```
`-16` LUFS / `-1.5` dBTP is a good general/web target. The single pass above is fine for
most clips. For **precise** targets (or batch consistency), do two-pass: run once with
`loudnorm=...:print_format=json`, read the measured values, then pass them back as
`measured_I`/`measured_TP`/`measured_LRA`/`measured_thresh` on a second run.

## Even out inconsistent levels (loud speaker vs quiet audience)

The most common "fix the audio" complaint on talks/panels/interviews isn't noise — it's
**level swings**: the presenter is loud and near the mic, audience questions are faint, peaks
nearly clip. Compress to tighten the dynamic range, then normalize:

```bash
ffmpeg -y -i in.mp4 -af \
  "acompressor=threshold=-21dB:ratio=3:attack=15:release=250:makeup=3,\
   equalizer=f=2800:t=q:w=2:g=2,\
   loudnorm=I=-14:TP=-1.5:LRA=9" \
  -c:v copy -c:a aac out.mp4
```
- `acompressor` pulls the loud parts down; `makeup` lifts everything so the quiet parts come
  up — net result is consistent loudness. Lower `threshold` / higher `ratio` = more leveling.
- The gentle `equalizer` presence bump (~2.5–3.5 kHz) adds intelligibility.
- `loudnorm` with a **tight `LRA` (7–9)** finishes the leveling; check the output isn't pumping.
- Diagnose first with `astats` (`RMS level dB` / `Peak level dB`) — a wide RMS-to-peak gap or a
  peak near 0 dB confirms it needs compression, not denoise.

## Denoise (hiss / hum)

```bash
ffmpeg -y -i in.mp4 -af afftdn=nr=12 -c:v copy -c:a aac out.mp4
```
`afftdn` is an FFT denoiser — no extra deps. `nr` is noise reduction in dB (10–20 typical;
higher = more aggressive but more artifacts/underwater sound — start at 12 and listen).
Combine with normalize in one pass: `-af "afftdn=nr=12,loudnorm=I=-16:TP=-1.5:LRA=11"`.

## Add background music

**If the edit is an EDL, put the bed in the EDL instead** — `"music": {"src": …, "gain_db":
-18, "duck": true, "fade_out": 2.0}` mixes it in the same render pass (looped to cover the
whole cut), so you don't pay for a second re-encode. The recipes below are for a video that's
already finished, or for a one-off mix outside an EDL.

Plain mix (music sits at a fixed level under the existing audio):
```bash
ffmpeg -y -i in.mp4 -i bg.mp3 -filter_complex \
  "[1:a]volume=-18dB[bg];[0:a][bg]amix=inputs=2:duration=first[a]" \
  -map 0:v -map "[a]" -c:v copy -c:a aac -shortest out.mp4
```

Music **ducked under speech** (music automatically drops when someone talks) — usually what
you want for narration:
```bash
ffmpeg -y -i in.mp4 -i bg.mp3 -filter_complex \
  "[1:a]volume=-18dB[bg];\
   [bg][0:a]sidechaincompress=threshold=0.03:ratio=8:attack=20:release=300[bgd];\
   [0:a][bgd]amix=inputs=2:duration=first[a]" \
  -map 0:v -map "[a]" -c:v copy -c:a aac -shortest out.mp4
```
The sidechain compresses the music (`[bg]`) using the original speech (`[0:a]`) as the
trigger: louder speech → more music attenuation. Tune `ratio` (duck depth), `attack`
(how fast it ducks, ms), `release` (how fast music returns, ms).

## Gotchas

- **Always `-shortest`** when adding music, or a longer music track extends the clip past
  the video and everything downstream desyncs.
- **`amix` can lower overall level** — if the result sounds quiet, append
  `,loudnorm=I=-16:TP=-1.5:LRA=11` after the `amix` output, or raise the speech with
  `volume` before mixing.
- Re-encode audio to **aac** (`-c:a aac`); don't `-c:a copy` after a filter.
- `duration=first` ties output length to the first `amix` input (the speech) — keep speech
  first in the `amix` chain.