EvenAudio

A small Linux desktop audio compressor: seven bands, instant attack, adjustable release, stereo-linked compression, and optional automatic leveling for uneven speakers. C handles the sample loop; Python provides a small Tk interface and audio routing. No pip packages, lookahead buffer, or separate peak limiter.

The interface starts in dark mode. Use the Dark mode switch to select light mode; this preference is remembered for the next launch. Light-mode preview.

Install from the M45 APT repository

For Ubuntu 24.04 (amd64), add the signed repository once:

curl -fLO https://m45software.com/apt/m45-software-archive-keyring_1.0_all.deb
sudo apt install ./m45-software-archive-keyring_1.0_all.deb
sudo apt update
sudo apt install evenaudio

Launch EvenAudio from the application menu or run evenaudio. The package includes the compiled audio engine and installs its runtime dependencies and recommended tray support. It uses your existing PulseAudio or PipeWire audio service. Updates arrive through APT. Remove it with sudo apt remove evenaudio; per-user preferences stay in ~/.config/evenaudio.

Run from source

On Debian/Ubuntu, install the build and runtime dependencies:

sudo apt install build-essential python3 python3-tk pulseaudio-utils libpulse0
make
python3 evenaudio.py

Run from this directory. Requires a running PulseAudio server or PipeWire with its pipewire-pulse compatibility service. Use your existing desktop audio stack; you do not need to replace it or install a second sound server.

For an application-menu launcher and desktop icon:

make install-desktop

This installs only for your user, without sudo. The launcher builds the local audio engine when needed and opens the GUI. Keep the checkout in place; rerun the installer if you move it. Some desktops require right-click → Allow Launching the first time you use the desktop shortcut.

Select your speakers or headphones and click Start processing. The app creates a temporary “EvenAudio” output, makes it the default, and moves existing playback on the selected output into it. Its own processed output is explicitly pinned to your speakers/headphones to prevent feedback. Playback on other outputs is left alone. Some applications pin their own output: select EvenAudio in the desktop sound mixer if needed. Stop processing or choose Quit to restore playback and remove the temporary sink. A GUI bypass control lets you compare with the original signal.

Keep running when window is closed is enabled by default. Closing the window keeps processing active and hides the controls to the tray. If no tray is connected, the window minimizes to the taskbar instead. A one-time notice explains this behavior. Disable the checkbox to make closing the window quit. The preference is remembered in $XDG_CONFIG_HOME/evenaudio/window.json (normally ~/.config/evenaudio/window.json). The window's Quit button always exits and restores playback.

The optional tray menu offers Show EvenAudio, Start/Stop processing, Enable/Disable bypass, and Quit. On Debian/Ubuntu, enable tray support with:

sudo apt install python3-gi gir1.2-gtk-3.0 gir1.2-ayatanaappindicator3-0.1

KDE Plasma supports these indicators. GNOME needs the AppIndicator extension (some distributions include it). Legacy AppIndicator3 is also supported. These packages are optional; the Tk interface works without them. If the tray host or helper disappears while the window is hidden, EvenAudio shows the window again.

Without a GUI:

python3 evenaudio.py --list-outputs
python3 evenaudio.py --headless
python3 evenaudio.py --headless --preset gentle-music
python3 evenaudio.py --headless --preset uneven-voices

Use --output SINK_NAME to choose an output. Broadband automatic leveling is off by default for music; use --auto-level for uneven speech, or enable it in the GUI. Without --release, each band uses the default below; an explicit --release 100 overrides all seven bands to 100 ms. Ctrl+C and SIGTERM restore the routing. Audio settings are session-only; appearance is saved in $XDG_CONFIG_HOME/evenaudio/appearance.json (normally ~/.config/evenaudio/appearance.json).

Presets and voice leveling

The preset menu replaces all audio controls with editable starting values. You can switch while processing; manual adjustments show Custom. Bypass stays as selected. The application always opens on General purpose unless you pass --preset; explicit CLI options override that preset.

Ratio (:1, all bands) is one shared control, starting at 4:1. Thresholds and releases remain independently adjustable for each band. The shared ratio sets compression strength; each band still measures and reduces its own signal. Use --ratio 4 to set it from the command line; fractional values such as 2.5 are supported too.

Preset	Intended use	Shared ratio	Threshold	Bass → treble release	Output gain	Voice leveling
General purpose	Gentler everyday leveling	4:1	−18 dBFS	2,000 → 250 ms	0 dB	Off
Gentle music	Lighter compression, more dynamics	3:1	−18 dBFS	2,000 → 250 ms	+3 dB	Off
Steady music	Stronger leveling, longer recovery	10:1	−28 dBFS	3,000 → 500 ms	+8 dB	Off
Uneven voices	Podcasts/calls with differing voice levels	10:1	−24 dBFS	1,000 → 200 ms	+6 dB	On

These are proposed presets for listening and tuning, not professionally mastered profiles or a guarantee against pumping. General purpose retains the original logarithmic release curve; the other release curves also use geometric spacing.

Level uneven voices (podcasts/calls) raises quiet passages and lowers loud ones by following overall volume. It does not identify individual speakers or separate overlapping voices. It is useful when podcast or call voices differ in volume; because it affects the whole mix, it can also make music swell or duck. Usually leave it off for music. Its target, boost limit, and time controls are disabled while it is off. The seven-band compressor remains active with it off.

Starting settings

These are this prototype's proposed listening defaults, not universal mastering settings. In particular, threshold depends on the incoming audio level.

The shared ratio is 4:1, output gain is 0 dB, and voice leveling is off. Compared with the earlier 10:1 / −24 dBFS / +6 dB default, this starts compression at a higher level, retains more level variation above threshold, and avoids automatically boosting quiet audio. It is a starting point for natural sound; listening with your own material is still needed.

Band	Threshold	Attack	Release
Below 80 Hz	−18 dBFS	Instant	2,000 ms
80–200 Hz	−18 dBFS	Instant	1,414 ms
200–500 Hz	−18 dBFS	Instant	1,000 ms
500–1,250 Hz	−18 dBFS	Instant	707 ms
1,250–3,150 Hz	−18 dBFS	Instant	500 ms
3,150–8,000 Hz	−18 dBFS	Instant	354 ms
Above 8,000 Hz	−18 dBFS	Instant	250 ms

The release values form a geometric progression: 2000 × (1/8)^(band/6) ms, rounded to whole milliseconds. The slow bass recovery is intended to keep gain steadier across successive beats, rather than recover between every hit. The two endpoints are this project's chosen listening defaults, not a published seven-band preset or a guarantee against pumping. Releases can be set from 10 to 5,000 ms independently.

Research behind the design:

• Universal Audio explains how too-fast recovery can cause repeated bass-driven compression and audible breathing, and describes release ranges extending into seconds.
• Rane warns that release definitions differ between manufacturers. EvenAudio's values are exponential time constants applied to gain reduction in dB. With no further compression, 10 dB of reduction becomes about 3.68 dB after one time constant, and 1 dB after about 2.3 time constants. These are not fixed times to recover 10 dB.
• Waves C4's manual explains that splitting bands prevents bass from modulating the whole mix, and discusses both the artifacts of fast recovery and the drawbacks of holding gain down too long. It also uses adaptive release; EvenAudio currently does not.

The optional leveler applies one gain before the split and can therefore modulate all bands together. It is off by default to preserve independent music dynamics. Long release still has a tradeoff: audio after an isolated loud event can remain attenuated. Listening tests across different music remain necessary.

Output gain starts at 0 dB. When enabled, automatic leveling targets −18 dBFS short-term RMS, with up to +18 dB of boost and a 1.5-second adjustment time. It measures stereo power over roughly 250 ms before compression. Gain freezes during pauses below −50 dBFS; it does not identify speakers, separate overlapping voices, remove room noise, or guarantee a particular LUFS value. Disable it for music if it changes the intended dynamics too much. Quiet speech below the pause floor will not get additional automatic boost. The floor prevents gain from increasing in pauses; it does not remove noise that is already present.

Tune release by ear: shorten it if one bass note ducks the next; lengthen it if you hear modulation or distortion. Release is an exponential time constant in gain dB, not a promise that gain fully recovers in that interval. Ratios range from 1:1 to 1,000:1. A 6 dB soft knee smooths the transition into compression.

The default 4:1 ratio is gentler than the 10:1 Steady music and Uneven voices presets. Instant attack can still change transients or produce audible distortion. The shared ratio remains adjustable; attack is fixed at zero. RØDE's compression guide describes the tradeoffs of high ratios and fast attack/release. FFmpeg's dynamic normalization documentation explains why a gain cap is useful when raising quiet sections. EvenAudio implements its own causal RMS gain control, not FFmpeg's normalization filter.

Signal path and limits

Desktop apps → virtual sink → optional RMS leveler → seven-band compressor → output gain → selected output.

• Fixed crossovers at 80, 200, 500, 1,250, 3,150, and 8,000 Hz, using 24 dB/octave Linkwitz–Riley splits with phase compensation. At unity band gains the combined magnitude response is flat; the filters do introduce phase shift.
• Each band's peak detector uses the louder stereo channel and applies additional gain reduction to the same sample. Release smooths recovery. Controls run outside the audio sample loop; output gain and bypass changes fade over 20 ms.
• No final limiter or hard clipper. Instant attack eliminates attack overshoot in each compressor, but finite ratios still permit peaks above threshold. Summing seven bands can also exceed 0 dBFS. The output meter warns when this happens; reduce output gain or lower thresholds. This is not a true-peak guarantee.
• Audio is stereo float32 at 48 kHz, in 10 ms chunks. Playback requests an 80 ms target buffer, a 160 ms maximum, and 20 ms minimum requests. Automatic prebuffering pauses and refills the stream after an underrun; capture buffering is capped at half a second to bound backlog after a long scheduling stall. Server, device, crossover, and buffering latency still apply. “Instant” describes gain attack, not end-to-end playback latency. Bluetooth adds its normal transport latency.
• Output hot-plugging/reconnection and autostart are not implemented. Stop, refresh outputs, and start again after changing devices. An audio error stops processing and restores routing.
• A forced kill or power loss cannot execute cleanup. On startup, the app attempts to restore playback and remove orphaned EvenAudio sinks. Recovery leaves sinks alone when an EvenAudio playback worker or their owner process is still alive. If manual recovery is needed, run pactl list short sinks, set your real output with pactl set-default-sink YOUR_SINK, inspect pactl list short modules, and unload only the module-null-sink whose arguments contain the matching sink_name=evenaudio_… using pactl unload-module MODULE_ID.

Audio access uses the PulseAudio simple API. Do not run the app as root.

Debian builds and releases

sudo apt install build-essential dpkg-dev
bash packaging/build-deb.sh
sudo apt install ./dist/evenaudio_0.1.1_amd64.deb

VERSION is the package version. CI uses Ubuntu 24.04, caches downloaded APT archives with the same setup action as the other M45 projects, and uses ccache for the native DSP. APT still refreshes indexes and verifies current dependency versions on cache hits. Pull requests and main pushes build and test packages; pushing a matching vVERSION tag also creates a GitHub release with SHA256SUMS and publishes it through a package-restricted SSH receiver. The archive signing key stays on the server. The repository secret M45_APT_DEPLOY_KEY holds the restricted deployment key; packaging/apt/ssh_known_hosts pins the server key.

Development checks

make test
python3 tests/integration_audio.py
# Optional GUI smoke check (Tk, Xvfb, and ImageMagick required):
xvfb-run -a python3 tests/integration_audio.py --gui
# Include window lifecycle tests with a virtual display:
xvfb-run -a make test
# Optional real tray protocol check (tray packages, D-Bus, and Xvfb required):
dbus-run-session -- xvfb-run -a python3 tests/integration_tray.py

The unit tests cover crossover flatness, stereo linking, shared 4:1 and 10:1 compression, same-sample attack, adjustable release, automatic gain, silence behavior, bypass, unboosted quiet audio at the default settings, live ratio edits and presets, restoration rules, saved window behavior, closing while processing, and recovery when the tray disappears. The integration script creates its own temporary PulseAudio server and null output, exercises real capture/playback and worker failure, and checks cleanup without touching desktop playback. It needs the pulseaudio executable and permission to create Unix sockets. Its GUI check also verifies that closing preserves live routing and Quit restores playback. The tray check runs on a private session bus and verifies actual AppIndicator registration, menu commands, state updates, and watcher loss/reconnection. Synthetic tests do not replace listening tests with real podcasts, music, speakers, and headphones.

License

MIT; see LICENSE. The application code and SVG icon are original project assets. Linked manuals are references and are not included in the license.