Discord ships automatic audio attenuation on Windows and macOS, but not on Linux. This reproduces it on PipeWire: when someone talks in a Discord call, your game/music/browser audio drops in real time and comes back when they stop.
It is signal-triggered, not call-state-triggered. Audio ducks while people are actually speaking, not for the whole duration of the call. Discord's own audio is never ducked by its own voice, and that is guaranteed structurally rather than by tuning.
Built and running on CachyOS with PipeWire 1.6.8 / WirePlumber 0.5.15.
Discord ──(in-app output select)──► [ DiscordSink ] (null sink)
│
├─► loopback ─────────────────► default sink (unity, never ducked)
│
└─► monitor ──► sidechain ┐
▼
Everything else ──► [ easyeffects_sink ] ──► compressor ───────────► default sink (ducked)
Discord is routed into a dedicated null sink. That sink's monitor does two jobs: it feeds a loopback straight to your speakers at full volume, and it drives the external sidechain of a compressor that every other application passes through.
Two properties make this work:
- Discord bypasses the compressor entirely, so it structurally cannot duck itself. There is no signal path from Discord into its own attenuator.
- The sidechain taps the monitor upstream of the loopback, so the duck engages slightly ahead of the audible voice rather than lagging it.
Both output paths follow whatever your current default sink is, and re-link automatically when
you switch devices. Verified across Bluetooth → HDMI → S/PDIF. Swap headsets freely; nothing needs
editing. This is why the loopback deliberately omits target.object and node.dont-reconnect.
- PipeWire + WirePlumber (
pipewire-pulsefor the PulseAudio layer) easyeffects(8.x) andlsp-plugins-lv2python-numpyis optional, only fordiscord-ducking test(the two-tone proof). Threshold calibration needs nothing beyond the Python standard library.- A Discord build whose Voice & Video → Output Device selector lists PipeWire sinks (the native Arch package does; Flatpak may differ)
sudo pacman -S --needed easyeffects lsp-plugins-lv2
# optional, only for `discord-ducking test`:
sudo pacman -S --needed python-numpygit clone https://github.com/K-Sugar/discord-ducking
cd discord-ducking
makepkg -si # builds from this tree; no remote or tarball needed
discord-ducking installThis installs the tooling to /usr/share/discord-ducking with a discord-ducking
command, and pulls in easyeffects / lsp-plugins-lv2 as real package dependencies.
Configuration lives in your home directory, which pacman must not write to, so
discord-ducking install is the separate per-user step. Everything is available as
subcommands:
discord-ducking install deploy config into ~/.config, enable the service
discord-ducking verify health check (exit 0 = all good)
discord-ducking measure measure speech level, recommend a threshold
discord-ducking test two-tone end-to-end proof
discord-ducking bt-mic on|off|status Bluetooth mic vs playback quality
discord-ducking uninstall full teardown
git clone https://github.com/K-Sugar/discord-ducking
cd discord-ducking
./scripts/install.shThe installer backs up anything it touches to ~/.local/share/discord-ducking/backup-<timestamp>/,
and merges into your EasyEffects config rather than overwriting it, so an existing effects chain
and blocklist are preserved.
These apply to both install methods. The package form is shown first, the from-the-repo form second.
1. Point Discord at the sink
Discord → Settings → Voice & Video → Output Device → Discord (ducking source)
If Discord was already running during the install, restart it first. It caches its device list at startup and will not show the new sink otherwise.
2. Check it
discord-ducking verify # or: ./scripts/verify.sh3. Calibrate the threshold, and do not skip this
discord-ducking measure # or: ./scripts/measure-sidechain.shRun it while someone is actually talking. This is the step people get wrong. See below.
The correct compressor threshold depends on your Discord volume, your output device and the speaker's mic. It is not portable between systems.
On the reference machine, measured Discord speech looked like this:
| Metric | Level |
|---|---|
| Median (silence / noise floor) | -59.1 dBFS |
| 95th percentile | -41.8 dBFS |
| Peak over 16 s of conversation | -34.8 dBFS |
The threshold value commonly suggested for this kind of setup is -30 dB. Real speech crossed it 0.0 % of the time. The compressor was wired perfectly and simply never triggered. Symptom: "everything looks right but I hear no difference."
discord-ducking measure records the real sidechain signal, prints the level distribution, and
recommends a threshold midway between your noise floor and your speech level. Apply it with:
systemctl --user stop easyeffects
sed -i 's/^threshold=.*/threshold=-50/' ~/.config/easyeffects/db/compressorrc
systemctl --user start easyeffectsOr just type it into the Threshold box in the EasyEffects window, with no stopping and no restarting. See Tuning for what every other control on that page does.
To prove the whole chain end-to-end without needing a second person:
discord-ducking testIt injects a 440 Hz "music" tone and a 1000 Hz "voice" tone, captures the output, and reports the per-frequency change. Two distinct frequencies let the ducked signal and the trigger be measured independently in a single capture.
Everything below lives in one compressor. You can change it two ways: in the EasyEffects window, or by editing the config file. The window is easier and safer, so start there.
Launch easyeffects, stay on the Output tab, and click Compressor in the effects list.
Changes apply to live audio immediately and are saved for you. No restart, no file editing.
The Gain reduction meter on that page is the fastest tuning tool you have: get someone to
talk (or run discord-ducking test in another terminal) and watch it. If it never moves,
your threshold is too high. If it never returns to 0, it is too low.
| Control | Config key | What it does |
|---|---|---|
| Threshold | threshold |
How loud the Discord voice must be before ducking starts. The one you must calibrate. Never triggers → lower it. Ducks on keyboard noise → raise it. |
| Ratio | ratio |
How hard it ducks once triggered. High values give a near fixed-depth duck, which is what Windows does. Lower it for a gentler dip. |
| Attack | attack |
How fast the duck engages, in ms. Too slow and the first word gets through at full volume. |
| Release | release |
How fast the volume comes back, in ms. The setting most worth your time. Below ~400 ms it pumps audibly between words; too high and music stays quiet long after the talking stops. |
| Knee | knee |
How abruptly ducking starts around the threshold. A softer knee makes the onset less noticeable. |
| Makeup | makeup |
Adds gain back after compression. Leave at 0, because you want the level drop. |
| Compression mode | mode |
Must stay Downward. Upward or Boosting inverts the effect. |
| Sidechain → Type | sidechainType |
Must stay External. This is what makes the compressor listen to Discord instead of to the music. Changing it breaks the whole design. |
| Sidechain → Input device | sidechainInputDevice |
Must stay DiscordSink. This is where the trigger signal comes from. |
| Sidechain → Mode | sidechainMode |
RMS averages over a short window and tracks perceived loudness; Peak reacts to instantaneous spikes and is twitchier on speech. |
| Sidechain → Reactivity | sidechainReactivity |
Size of that RMS averaging window, in ms. Raise it if brief consonants cause flutter. |
| Sidechain → Preamp | sidechainPreamp |
Gain applied to the trigger signal only. An alternative to lowering the threshold if a quiet talker never triggers ducking. |
| Sidechain → Lookahead | sidechainLookahead |
Delays the audio so the duck can start before the trigger. Not needed here, because the sidechain already taps upstream of the loopback, so the duck leads the voice slightly. |
| Sidechain → Listen | sidechainListen |
Diagnostic toggle: replaces the output with the trigger signal itself, so you hear what the compressor is reacting to. Turn it on and you should hear Discord voice and nothing else. If you hear your own microphone or your music, the sidechain is wired to the wrong source. Turn it back off when done. |
The two rows marked must stay are structural. Everything else is taste.
Reference values (~/.config/easyeffects/db/compressorrc), for scripting or for copying a known-good
setup to another machine:
| Setting | Value | Notes |
|---|---|---|
threshold |
-50 |
calibrate this yourself |
ratio |
24 |
high ratio ⇒ near fixed-depth duck, like Windows |
attack |
5 |
ms |
release |
700 |
ms; below ~400 pumps between words |
knee |
-6 |
dB |
makeup |
0 |
dB |
mode |
0 |
Downward |
sidechainType |
2 |
External |
sidechainInputDevice |
DiscordSink |
the trigger source |
sidechainMode |
1 |
RMS |
sidechainReactivity |
10 |
ms |
Stop the service before hand-editing these files. A running EasyEffects holds settings in memory and overwrites the file on exit, so your edit silently disappears. This is the reason to prefer the window: it has no such trap.
systemctl --user stop easyeffects
sed -i 's/^release=.*/release=700/' ~/.config/easyeffects/db/compressorrc
systemctl --user start easyeffects- Quiet talkers do not trigger it → lower the threshold (or raise sidechain preamp)
- Background noise / keyboard ducks constantly → raise it
- Audible pumping between words → lengthen release
- First word slips through at full volume → shorten attack
- Duck is too aggressive → lower the ratio
Discord's join/leave notification pings will also trigger ducking. That is inherent to signal-triggered attenuation and matches Windows. Not a bug.
discord-ducking uninstall # or: ./scripts/uninstall.shRemoves the drop-ins and the unit, turns the catch-all off, clears the blocklist entries and restarts the audio stack. Afterwards, set Discord's Output Device back to your real device, since it will otherwise point at a sink that no longer exists.
If you installed the package, run discord-ducking uninstall before pacman -R discord-ducking,
because pacman will not remove per-user config.
It deliberately does not delete ~/.local/state/wireplumber/stream-properties, which holds
saved volume and routing for every application on the system.
Everything looks correct but nothing ducks.
Almost always the threshold. Run discord-ducking measure. See the calibration section.
Discord ducks itself / its audio sounds compressed.
The loopback got captured by EasyEffects. The blocklist must contain discord_direct_playback.
EasyEffects matches the blocklist against node.name, not application.name, so the node's
application.name (DiscordDirectOut) will not work. If you rename node.name in
20-discord-loopback.conf, you must update the blocklist to match or self-ducking returns silently.
Ducking stops after switching headphones.
Check useDefaultOutputDevice=true in easyeffectsrc. Without it EasyEffects pins one device.
DiscordSink is missing from Discord's dropdown.
Restart Discord first. Discord enumerates audio devices through libpulse at startup and caches
the list, so a sink created (or recreated) while it is running will not appear until it restarts.
This is the usual cause right after installing, and after any install/uninstall cycle, where
Discord can also end up holding a stale reference to the old sink and showing an incomplete device
list until restarted.
Only if it is still absent after a restart, set Discord's output to Default and install the
optional pulse.rules fallback described in docs/DESIGN.md §4.3.
Bluetooth audio goes mono / sounds bad during calls.
Check discord-ducking bt-mic status. If the active profile is headset-head-unit, the device is
in HSP/HFP: mono, ~24 kHz. Unrelated to ducking, but it degrades everything you hear for the whole
call.
Two different causes, with different fixes:
WirePlumber switched it. It flips Bluetooth devices to headset mode whenever any application
opens a capture stream, even when that app records from a different microphone. Switch back
with discord-ducking bt-mic off. To stop it happening at all, enable the optional drop-in:
mkdir -p ~/.config/wireplumber/wireplumber.conf.d
cp config/wireplumber/wireplumber.conf.d/51-bt-no-autoswitch.conf ~/.config/wireplumber/wireplumber.conf.d/
systemctl --user restart wireplumberTrade-off: your Bluetooth mic then stops being selected automatically. Use it deliberately, with
discord-ducking bt-mic on before you need it and off afterwards. This is the better setup if you
usually use a separate mic but occasionally want the headset one.
The device connected as a headset. If bt-mic status shows no A2DP profile at all, no
software switch can fix it, because the profile was never negotiated. This happens when the device
connects while an app is already recording. Disconnect and reconnect it with nothing recording.
No audio at all after installing.
A malformed drop-in can stop PipeWire from starting. Run discord-ducking uninstall, then
journalctl --user -u pipewire -n 50.
PKGBUILD builds from the working tree; makepkg -si
discord-ducking.install post-install notes shown by pacman
bin/discord-ducking CLI front-end (subcommand dispatcher)
config/
pipewire/pipewire.conf.d/10-discord-sink.conf null sink
pipewire/pipewire.conf.d/20-discord-loopback.conf loopback (follows default sink)
systemd/user/easyeffects.service graphical-session.target, not default.target
easyeffects/compressorrc compressor + sidechain parameters
easyeffects/easyeffectsrc.reference keys install.sh merges (reference only)
wireplumber/.../51-bt-no-autoswitch.conf OPTIONAL: stop BT dropping to headset mode
scripts/
install.sh idempotent installer, backs up and merges
uninstall.sh full teardown
verify.sh health check; exit 0 = all good
measure-sidechain.sh measures real speech level, recommends a threshold
test-ducking.sh two-tone end-to-end proof
bt-profile.sh Bluetooth A2DP <-> headset-mic toggle
docs/
DESIGN.md full spec, and every deviation with the evidence behind it
BUILD-LOG.md gated build checklist with results
docs/DESIGN.md §3 is the interesting part: every deviation from the original plan, each with the
measurement that forced it.
- Discord notification sounds trigger ducking (inherent; matches Windows).
- The threshold needs re-calibrating if you substantially change your Discord volume or output device chain.
easyeffects.servicehooksgraphical-session.target, which KDE Plasma activates (verified). Most desktop sessions do. If yours does not, the service simply will not autostart. Check withsystemctl --user is-active graphical-session.targetand fall back to EasyEffects' own "start service at login" toggle.- Discord screen-share-with-audio is unaffected by the routing change (verified): the capture taps application output nodes directly, upstream of the effects sink.
- Bluetooth devices that connect while an app is already recording may negotiate as a headset and
offer no A2DP profile at all.
discord-ducking bt-mic statusdetects this; only a reconnect fixes it.
Bug reports and PRs welcome. If something does not work on your setup, open an issue with the
output of discord-ducking verify.
MIT, see LICENSE.