Ffmpeg Audio Fx

ffmpeg Audio FX

Context: $ARGUMENTS

Creative and restoration audio effects. -af chain syntax. Every recipe below is verified against the FFmpeg filters doc. For basic gain / EQ / loudnorm / tempo see the sibling ffmpeg-audio-filter skill — this one is modulation, delay, restoration, true-peak limiting, dynamic EQ, spectral domain, and plugin hosts.

Quick start

• Add echo — Step 2 (Delay group), recipe aecho.
• Chorus / flanger / phaser / tremolo / vibrato — Step 2 (Modulation group).
• Normalize speech for a podcast — Step 2 (Dynamics group), recipe speechnorm.
• Rescue vinyl rip — Step 2 (Restoration group), chain adeclick → adeclip.
• Broadcast-safe true-peak limit — apsyclip (Dynamics group).
• Host a LADSPA or LV2 plugin — Step 2 (Plugin group).
• Shift frequencies (not pitch) — afreqshift (Filter-curve group).

When to use

• Music production: add modulation (chorus/flanger/phaser), widen a mono source, add shimmer (crystalizer), analog-warm limiting (asoftclip).
• Post / dialogue: enhance voice in a 5.1 mix (dialoguenhance), level an uneven interview (speechnorm), duck resonant frequencies (adynamicequalizer).
• Restoration: vinyl / shellac / cassette rips (adeclick + adeclip + asupercut rumble filter), decorrelate stereo (adecorrelate).
• Broadcast / streaming: psychoacoustic true-peak limiter (apsyclip) as the final stage before dithering to s16.
• DSP / research: arbitrary biquads, spectral-domain manipulation (afftfilt), host third-party plugins via LADSPA / LV2.

Step 1 — Classify the effect

Group the request into one of five buckets before picking a filter:

| Group | Filters | Use when | |---|---|---| | Modulation | chorus, flanger, aphaser, tremolo, vibrato | Add movement / wobble / thickness. | | Delay | aecho | Echo / slap-back / doubler. | | Dynamics | speechnorm, dialoguenhance, adynamicequalizer, adynamicsmooth, asoftclip, apsyclip, crystalizer | Level, duck, saturate, limit, sharpen transients. | | Restoration | adeclick, adeclip, adecorrelate, asupercut | Vinyl pops, clipped peaks, stereo smear, subsonic rumble. | | Filter-curve / spectral | bandpass, bandreject, biquad, allpass, asuperpass, asuperstop, atilt, afreqshift, aphaseshift, afftfilt | Surgical EQ, high-order filters, frequency shifting (not pitch), arbitrary spectral math. | | Plugin host | ladspa, lv2 | Load an external DSP plugin. Requires build flags. |

Step 2 — Build the filter chain

Copy the exact recipe. Parameters that look arbitrary (e.g. 0.8:0.9:1000:0.3) are positional and mismatched counts produce cryptic errors.

Modulation

# Echo (doubler / slap-back) — delays are MILLISECONDS
ffmpeg -i in.wav -af "aecho=0.8:0.9:1000:0.3" out.wav
# Chorus (6 positional args: in-gain:out-gain:delays-ms:decays:speeds:depths)
ffmpeg -i in.wav -af "chorus=0.7:0.9:55:0.4:0.25:2" out.wav
# Flanger (jet-swoosh; depth is ms range of the LFO)
ffmpeg -i in.wav -af "flanger=delay=5:depth=2:speed=0.5" out.wav
# Phaser (type=t triangle, s sine, n analog noise)
ffmpeg -i in.wav -af "aphaser=type=t:speed=2" out.wav
# Tremolo (volume modulation)
ffmpeg -i in.wav -af "tremolo=f=5:d=0.5" out.wav
# Vibrato (pitch modulation via delay)
ffmpeg -i in.wav -af "vibrato=f=7:d=0.5" out.wav

Dynamics

# Intelligent speech leveler — better than dynaudnorm for voice-only
ffmpeg -i podcast.wav -af "speechnorm=e=12.5:r=0.0001:l=1" out.wav
# Dialogue enhance (boosts the center channel voice band, REQUIRES 5.1+)
ffmpeg -i film.wav -af "dialoguenhance" out.wav
# Dynamic EQ — duck resonant 1 kHz band by 5 dB when it gets loud
ffmpeg -i in.wav -af "adynamicequalizer=bandwidth=300:dfrequency=1000:direction=below:mode=cut:makeup=5" out.wav
# Crystalizer (add harmonics / transient sharpening) — intensity > 3 = artificial
ffmpeg -i in.wav -af "crystalizer=i=2" out.wav
# Soft-clip (warm limiter; type= hard / tanh / atan / cubic / exp / alg / quintic / sin / erf)
ffmpeg -i in.wav -af "asoftclip=type=tanh" out.wav
# Psychoacoustic true-peak limiter (THE recommended broadcast limiter)
ffmpeg -i in.wav -af "apsyclip=level_in=1:level_out=1:clip=0.9" out.wav

Restoration

# Vinyl de-click (arorder 6-8 for vinyl; method a=autoregressive)
ffmpeg -i vinyl.wav -af "adeclick=window=55:overlap=75:arorder=8:threshold=2:burst=2" out.wav
# De-clip (threshold 10+ for obvious clipping; hsize=1000 history size)
ffmpeg -i clipped.wav -af "adeclip=window=55:overlap=75:arorder=8:threshold=10:hsize=1000:method=a" out.wav
# Subsonic rumble kill (sharp 20 Hz high-pass, 20th order Butterworth)
ffmpeg -i in.wav -af "asupercut=cutoff=20:order=20" out.wav
# Stereo decorrelation (adds a tiny inter-channel delay)
ffmpeg -i mono_ish.wav -af "adecorrelate" out.wav

Filter-curve / spectral

# Band-pass for voice intelligibility (Q=2 around 1 kHz)
ffmpeg -i in.wav -af "bandpass=f=1000:width_type=q:w=2" out.wav
# Frequency shift 50 Hz (NOT pitch shift — inharmonic, metallic)
ffmpeg -i in.wav -af "afreqshift=shift=50" out.wav
# Spectral tilt (lean EQ, +/- dB per octave)
ffmpeg -i in.wav -af "atilt=freq=1000:slope=-3" out.wav
# Arbitrary biquad IIR (coefficients b0,b1,b2,a0,a1,a2)
ffmpeg -i in.wav -af "biquad=b0=0.5:b1=0:b2=0.5:a0=1:a1=0:a2=0" out.wav
# FFT filter — real/imag expressions in spectral domain (null example)
ffmpeg -i in.wav -af "afftfilt=real='hypot(re,im)*sin(0)':imag='hypot(re,im)*cos(0)'" out.wav

Plugin host

# LADSPA: file=libname (no .so), plugin=label, c=key=val,...
ffmpeg -i in.wav -af "ladspa=file=cmt:plugin=amp_stereo:c=gain=10" out.wav
# LV2: p=plugin URI, c=key=val,...
ffmpeg -i in.wav -af "lv2=p=http://calf.sourceforge.net/plugins/Reverb" out.wav

Step 3 — Run

Encode to the target codec in the same invocation:

ffmpeg -i in.wav -af "speechnorm=e=12.5:r=0.0001:l=1,apsyclip=clip=0.95" \
       -c:a libopus -b:a 96k out.opus

For video input, pass the video through untouched with -c:v copy:

ffmpeg -i in.mp4 -af "dialoguenhance,speechnorm=e=6.25" \
       -c:v copy -c:a aac -b:a 192k out.mp4

Step 4 — A/B verify

Always listen to the result against the source. Use the ffmpeg-playback skill's ffplay recipes or diff loudness:

ffmpeg -i in.wav -af "ebur128=peak=true" -f null - 2>&1 | tail -20
ffmpeg -i out.wav -af "ebur128=peak=true" -f null - 2>&1 | tail -20

Target: integrated loudness inside spec (podcast -16 LUFS, streaming -14 LUFS, broadcast -23 LUFS), true peak ≤ -1 dBTP.

Available scripts

• scripts/afx.py — stdlib argparse front-end with subcommands for every recipe above. --dry-run prints the ffmpeg command without running it. --verbose shows the full argv.

uv run ${CLAUDE_SKILL_DIR}/scripts/afx.py echo --input in.wav --output out.wav --delay-ms 600 --decay 0.4
uv run ${CLAUDE_SKILL_DIR}/scripts/afx.py deckick --input vinyl.wav --output clean.wav
uv run ${CLAUDE_SKILL_DIR}/scripts/afx.py limit --input master.wav --output final.wav --method psy

Reference docs

Read [references/filters.md](references/filters.md) for per-filter option tables, effect-type taxonomy, LADSPA/LV2 plugin discovery, and production recipe book (vinyl chain, voice chain, radio master chain).

Gotchas

• aecho delays are milliseconds in a pipe-separated string: "1000|1500|2000" means three echoes at 1 s / 1.5 s / 2 s. Decays list must match the delay count.
• chorus positional args are in-gain:out-gain:delays-ms:decays:speeds:depths. Mismatched list counts produce a silent error / empty output.
• flanger depth is the LFO oscillation range in milliseconds — typical 0.5–10. Higher sounds like pitch-bend seasickness.
• aphaser type=t triangle LFO, s sine, n analog noise — noise sounds vintage / broken.
• tremolo modulates volume, vibrato modulates pitch (via delay). They are easy to confuse; pick the one that matches what "tremolo" means on your instrument.
• adeclick arorder should be 6–8 for vinyl; higher = more CPU and diminishing returns. threshold 2 is a good starting point.
• adeclip needs threshold=10 or more for obviously clipped material. Lower thresholds miss hard clips entirely.
• crystalizer intensity > 3 gets fatiguing fast. Most mastering uses i=1 to i=2.
• apsyclip is the recommended true-peak limiter for broadcast — psychoacoustic masking hides artifacts better than plain alimiter. Set clip=0.9 (about -0.9 dBTP) to stay under -1 dBTP.
• speechnorm is better than dynaudnorm for voice-only material (podcasts, audiobooks). For music or mixed content use loudnorm instead.
• dialoguenhance operates on stereo per current docs, but the real-world gain is on 5.1/7.1 where it targets the center channel. On stereo it will run but the effect is subtle.
• adynamicequalizer supports both mode=cut (duck a resonance) and mode=boost (lift under a threshold). direction=below or above chooses whether it acts on quiet or loud material.
• biquad is the generic bottom-level IIR. Prefer highpass, lowpass, bandpass, allpass, equalizer, treble, bass unless you genuinely need raw coefficients.
• afftfilt takes two expressions (real and imag) — advanced spectral domain. re, im, bin_h, mag, pha are available symbols.
• afreqshift is frequency shift (single-sideband), not pitch shift. It destroys harmonic relationships. For pitch shift use rubberband or asetrate+atempo.
• asupercut / asuperpass / asuperstop are high-order Butterworth (up to 20th order) — use when a regular highpass/bandpass rolloff isn't steep enough.
• atilt is linear-slope spectral tilt in dB/octave. Useful for pink-noise-style mastering adjustments.
• LADSPA requires ffmpeg built with --enable-ladspa. Plugin .so files live in /usr/lib/ladspa (or set LADSPA_PATH). Invoke with file=cmt (not file=cmt.so) and plugin=.
• LV2 requires ffmpeg built with --enable-lv2 (lilv). Plugins live in /usr/lib/lv2 (or set LV2_PATH). Invoke with p=. Calf's massive library is discoverable with lv2ls | grep -i calf.
• After heavy processing always re-dither if going back to integer PCM: -af "...,aformat=sample_fmts=s16,aresample=44100:dither_method=triangular".
• Never chain two strong limiters back-to-back — the second pumps against the first's reduction and creates distortion artifacts.

Examples

Example 1: Podcast voice chain

Input: noisy lavalier recording, interview.wav, needs to land at -16 LUFS for podcast.

ffmpeg -i interview.wav \
  -af "asupercut=cutoff=80:order=10,\
speechnorm=e=12.5:r=0.0001:l=1,\
adynamicequalizer=dfrequency=5000:mode=cut:bandwidth=800:makeup=4,\
apsyclip=level_out=0.9" \
  -c:a libopus -b:a 96k episode.opus

High-pass kills rumble → speech leveler brings quiet/loud to parity → dynamic EQ tames sibilant 5 kHz resonance → psyclip catches stray peaks.

Example 2: Vinyl restoration

Input: lp_side_a.flac from a USB turntable, clicks and a couple clipped peaks.

ffmpeg -i lp_side_a.flac \
  -af "adeclick=arorder=8:threshold=2,\
adeclip=threshold=10:method=a,\
asupercut=cutoff=30:order=12,\
aformat=sample_fmts=s16,aresample=44100:dither_method=triangular" \
  restored.flac

Example 3: Shimmer guitar

Chorus + crystalizer for a clean arpeggio:

ffmpeg -i guitar.wav -af "chorus=0.6:0.9:40|60:0.3|0.4:0.2|0.25:1.5|2.0,crystalizer=i=1" shimmer.wav

Example 4: Host a Calf LV2 reverb

ffmpeg -i dry.wav -af "lv2=p=http://calf.sourceforge.net/plugins/Reverb:c=room_size=5|decay_time=3" wet.wav

Troubleshooting

Error: Option chorus not found

Cause: ffmpeg build missing the filter — rare, but Alpine / minimal builds sometimes strip creative filters. Solution: ffmpeg -filters | grep chorus. If absent, install the full ffmpeg package (Debian/Ubuntu) or rebuild with --enable-filter=chorus.

Error: Error while parsing the string for 'aecho'

Cause: mismatched delay/decay counts — e.g. two delays 1000|1500 with only one decay 0.3. Solution: counts must match. aecho=0.8:0.9:1000|1500:0.3|0.4.

Error: LADSPA/LV2 No such filter

Cause: ffmpeg was built without --enable-ladspa / --enable-lv2. Solution: ffmpeg -buildconf | grep -E 'ladspa|lv2'. On macOS via Homebrew: brew install ffmpeg ships with ladspa/lv2 when available; on minimal Docker images you usually must rebuild.

Error: plugin loads but parameters silently ignored

Cause: wrong parameter name for the plugin (LADSPA/LV2 parameter names are plugin-specific). Solution: listplugins for LADSPA, lv2ls + lv2info for LV2 — copy the exact control-port name.

Symptom: dialoguenhance has no audible effect

Cause: running on mono or stereo source. The algorithm targets a dialogue-band extraction that is most effective on multichannel content with real separation. Solution: for stereo, try bandpass=f=2500:w=1200 + a light boost (+2 dB) instead, or upmix first with surround.

Symptom: distortion after limiter

Cause: chaining two strong dynamics stages (e.g. speechnorm l=1 into apsyclip level_in=2) pumps against itself. Solution: drop limiter input gain to 1.0, or remove one stage. Rule: one make-up-gain stage, one peak-catching stage — never both loud.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: damionrashford
• Source: damionrashford/media-os
• License: MIT
• Homepage: https://damionrashford.github.io/media-os/

Install and usage instructions live in the source repository linked above.