Omanoise

Procedural concentration soundscapes for the Omarchy shell, in the spirit of Endel. A small native engine plays sampled acoustic instruments through libfluidsynth, lays them over wavetable pads it generates at startup with Paul Nasca's PADsynth algorithm, and mixes the lot with a noise bed and a bank of six environmental sounds straight into PipeWire — forever and never the same twice. A bar widget drives it.

Sound engine v5. v2 synthesized its musical layer by hand (detuned supersaw pads, Risset and FM bells, Karplus-Strong plucks, a shimmer bus); it was, in the author's words, faintly anxiety-inducing. v3 replaced that layer entirely and removed every source of periodic modulation from the audio path. v4 keeps Focus exactly as it was, folds the old Sleep mode into Relax, and turns the old Ocean mode into Environment: a bank of blendable environmental sounds over a tonal layer that is now only a faint background. v5 admits what two attempts at a synthesized fireplace proved — a convincing fire is not worth synthesizing — and replaces the fire and wind generators with public-domain and CC0 recordings, adding a stream and a dawn chorus alongside them. Ocean and rain stay synthesized. The reasoning is in docs/research-opensource.md, the specifications in docs/PLAN-sound-v3.md, docs/PLAN-sound-v4.md and docs/PLAN-sound-v5.md, and the measurements in docs/ACCEPTANCE.md.

Install

omarchy pkg add fluidsynth soundfont-fluid      # sampled instruments (LGPL / MIT)
omarchy plugin add https://github.com/pepsoren/omanoise --enable

Then open the panel from the bar icon and press Build engine (or run build.sh in the plugin directory). That compiles the native engine with the system gcc and PipeWire/FluidSynth headers and converts the bundled environment recordings to WAV; nothing is downloaded and no sudo is used. Without the SoundFont the plugin still runs with pads and environments only.

Remove with omarchy plugin remove io.github.pepsoren.omanoise; the only file left behind is the settings file ~/.local/state/omanoise.json, which you can delete.

Optional keybinding for ~/.config/hypr/bindings.lua:

o.bind("SUPER + SHIFT + N", "Omanoise", "omanoise")

Dependencies

package	why
pipewire (+ headers)	audio output
fluidsynth	sampled instrument playback (LGPL-2.1)
soundfont-fluid	/usr/share/soundfonts/FluidR3_GM.sf2, the sample set (MIT)
gcc	./build.sh

Without soundfont-fluid the engine still runs: it reports "sf2":false, the panel says Install soundfont-fluid for piano and harp, and the pads, bass, noise bed and ocean play on their own. $OMANOISE_SF2 overrides the SoundFont path (used by the acceptance tests).

The recorded environments are files in sounds/, loaded at startup (about 58 MB of PCM). A missing or unreadable one disables just that slot; everything else runs. $OMANOISE_SOUNDS overrides the directory.

Layout

manifest.json            plugin manifest (bar-widget)
Service.qml              owns the engine process, mirrors its state
Panel.qml                bar icon + popup panel
engine/common.h          constants, RNG, small maths
engine/dsp.[ch]          Cytomic SVF, biquads, delay lines, value-noise drift
engine/padsynth.[ch]     radix-2 FFT + the PADsynth wavetable generator
engine/bank.[ch]         the startup bank: pad tables, bass, pulse one-shots
engine/sf2.[ch]          libfluidsynth: presets, forced attacks, note playback
engine/brain.[ch]        mode profiles, key, chord pool, Markov, voice leading
engine/wav.[ch]          WAV reader + BS.1770 loudness, for the recorded sounds
engine/env.[ch]          the rain generator and the random-segment sample player
engine/layers.c          voices, nature layers, buses, the realtime render loop
engine/fx.[ch]           the Dattorro plate
engine/master.[ch]       HP/AM/width/comp/LP/LUFS-trim/limiter
engine/main.c            PipeWire, stdin protocol, state file, offline renders
build.sh                 one gcc invocation -> bin/omanoise-engine
bin/omanoise-engine      built binary (not tracked; run build.sh)
sounds/*.wav             the recorded environments (see sounds/LICENSES.md)
sounds/src/*.ogg         the originals they were converted from
docs/                    the plans, the research, acceptance runs

The engine is started by the plugin and lives as long as the shell does. It appears in the audio mixer as its own app ("Omanoise"), so the normal per-app volume works too. Settings persist in ~/.local/state/omanoise.json.

Using it

• Bar icon 󰋋 — left: panel · right: play/pause · middle: next mode · scroll: volume. The icon is accent-coloured while playing and glows with the output level.

• Panel — play/pause, three modes, an Environment section with one slider per environmental sound (shown in Environment mode), Volume / Intensity / Brightness / Tonal sliders, neural modulation, binaural beats, time-of-day adaptation, focus pulse, and a stop timer.

• CLI — omanoise (in ~/.local/bin) wraps the shell IPC:

  omanoise              toggle play/pause
  omanoise next         cycle modes
  omanoise mode relax   focus | relax | environment
  omanoise env rain 0.6   ocean | rain | fire | wind | stream | birds, 0-1 each
  omanoise volume 0.5
  omanoise timer 45     stop after 45 minutes
  omanoise panel        open/close the popup
  

  A keybinding is one line in ~/.config/hypr/bindings.lua, e.g.

  o.bind("SUPER + SHIFT + N", "Omanoise", "omanoise")

Modes

All three are in major pentatonic (v2's minor Relax is gone), all events have attacks of 200 ms or more, and no mode has a shimmer, a chorus, a delay or a granular layer any more.

Mode	Key	Character	Target loudness
Focus	C major pentatonic	Soft grand piano and vibraphone, ~6 events/min in phrases of 3–6, over a warm PADsynth pad and a quiet pink-brown bed. Unchanged from v3.	−21 LUFS
Relax	A major pentatonic	Orchestral harp and low piano, single notes and dyads every 7–16 s, a very soft music box now and then, a pad that is now half glass, distant waves, and a broader, darker bed. This is v3's Relax and Sleep merged into one.	−25 LUFS
Environment	C major pentatonic	Ocean, rain, fire, wind, stream and birds, each on its own 0–1 slider and all six able to play at once, over a tonal layer 18 dB down — a faint background, not a mode of its own. No melody.	−23 LUFS

The Environment bank

Six sounds share one bus: two synthesized, four played from recordings in sounds/. Levels are perceptual (gain goes as level²) and each sound is calibrated so that it alone at 1.0 lands on the mode's loudness target; the bus sums them power-preservingly, so four sounds at 0.7 carry the same energy as one at 1.0. Defaults: ocean 0.8, everything else 0.

sound	how it is made
Ocean	Synthesized — the v3 ocean engine, unchanged: three wave engines (a noise rumble through an 85–350 Hz sweep driven by an asymmetric wave envelope, plus a "foam" band lagging 0.8 s behind) on 9–17 s periods that are re-drawn every cycle.
Rain	Synthesized, Farnell-style. A Poisson patter of 800–1600 impulses/s through a 2-pole band-pass whose centre wanders 350–800 Hz, then LP 3 kHz; and a hiss bed (HP 1.8 kHz, LP 6.5 kHz 4-pole) 16 dB under it with a slow ±3 dB shower envelope. No droplet chirps (they sounded like bubbles) and no resonant whistles.
Fire	sounds/fireplace.wav, 25.5 s, public domain.
Wind	sounds/wind.wav, 14.8 s, CC0.
Stream	sounds/stream.wav, 145.5 s, CC0.
Birds	sounds/birds.wav, 129.8 s, CC0 — a dawn chorus.

Sources and authors are in sounds/LICENSES.md; they come from the Blanket sound set, converted to 48 kHz 16-bit stereo WAV.

Why recordings. Two goes at a synthesized fireplace (v4 and v4.1) were rejected — the second sounded "like being on the phone with someone with bad internet in wind and rain". Fire is a famously hard thing to synthesize convincingly and a public-domain recording is simply the right tool. Wind went the same way; ocean and rain synthesize well and stayed.

How a recording is played — never a loop. Looping a 25 s fire would be recognisably periodic within a minute. Instead each recorded slot runs two read heads. A head plays a randomly chosen segment of the file at speed 1.0 (no pitch change: birds and water at altered speed sound wrong); when it has one crossfade-length left, the other head starts a new random segment and the two cross-fade equal-power (sin/cos, so their powers sum to a constant). Segment lengths and positions are re-drawn every time, so nothing repeats:

file length	segment	crossfade
under 40 s	6–14 s	2.5 s
40 s and over	20–45 s	4 s

A segment never comes within 1 s of either end of the file, and it is capped at 1/1.7 of the usable span so that two consecutive segments can overlap by no more than 30 % of their length (for the 14.8 s wind file that caps segments at 7.6 s). Each slot then gets HP 40 Hz, a gentle low-pass at 9 kHz that Brightness moves ±0.7 octaves, and a ±1.5 dB level drift under 0.05 Hz so a long session never settles.

Calibration. At load time the engine measures each file's loudness twice: the BS.1770 gated integrated value (the number ffmpeg -af ebur128 reports) and the mean of the short-term loudness through the slot's own filters and the master's stereo width — the statistic the loudness servo actually converges on. The second one sets the slot's gain, so any file lands near the mode's −23 LUFS target on its own and the master's auto-trim starts at zero.

Drop in your own. Put a 48 kHz 16-bit PCM WAV (mono or stereo) at sounds/<slot>.wav and it replaces that slot — fireplace.wav, wind.wav, stream.wav, birds.wav, and also ocean.wav or rain.wav, which take over from the synthesized generators when present. Anything else is rejected with a line on stderr and the slot reports "sounds":{"<name>":false}, which the panel shows as a dimmed row. $OMANOISE_SOUNDS overrides the directory.

Nothing in the bank is periodic: every envelope is value noise at 0.02–0.2 Hz, every event train is a Poisson process, and the recorded players schedule random segments of random length.

With Adapt to time of day on, evenings and nights darken the filters and drop the key by 2 / 5 semitones; mornings are a touch brighter.

The focus pulse and the 16 Hz neural modulation are now off by default (both are periodic modulation; see below). They are still toggles.

How the sound is built

Everything runs at 48 kHz, stereo.

1. Startup, in a worker thread (~100 ms for the tables, ~40 ms for the SoundFont, ~10 MB of tables):
   • Four PADsynth wavetables, 2^18 samples (5.46 s) each, tuned to 220 Hz: a warm profile (A_n = 1/n^1.6, odd harmonics favoured) and a glass profile (A_n = 1/n^2.4 with a bump on the 2nd and 3rd), two independent random-phase tables per profile. Each partial is a Gaussian band rather than a line, so the pad is lush without a single detuned oscillator anywhere — nothing beats. The left channel reads one table of a profile and the right the other: full stereo width, identical spectra, no pitch difference at all. Built with an FFT written for this engine (no FFTW).
   • FluidR3_GM loaded through libfluidsynth with reverb and chorus off, synth.threadsafe-api 0 (all synth calls happen on the audio thread) and synth.dynamic-sample-loading 1. Four presets are selected here, in the worker, because preset selection is what pages sample data in; after that fluid_synth_noteon never touches the disk.
   • A saw+sub bass and the two pulse one-shots, as before.
2. The brain (control rate, every 64 samples) — one key per mode, a pool of four wide add9/6/sus chords moving every 1–3 minutes through a Markov table biased toward common tones. Sustaining voices are snapped to just intervals above the key root, so where two chord tones share a harmonic it coincides exactly instead of beating at a few hertz, and below 250 Hz no two voices are allowed closer than an octave (Plomp & Levelt: a fifth down there sits inside one critical band). Melody notes come from a Markov chain over the five pentatonic degrees with steps strongly favoured over leaps.
3. Pads — four voices on incommensurate Eno periods (17.3 / 21.9 / 26.1 / 31.7 s, mode-scaled), 4–6 s raised-cosine attack and 8–12 s release, each re-voiced at the top of its own cycle and sometimes sitting a whole cycle out (never more than two of four). Per-voice keytracked low-pass at 1.6–2.5 × f0, drifted by value noise at about 0.05 Hz. No chorus, no waveshaper.
4. Melody — fluid_synth_noteon on four channels: Acoustic Grand Piano (0), Orchestral Harp (46), Vibraphone (11) and Music Box (10). Every channel gets a volume-envelope-attack generator that stretches its onset to 200–1160 ms measured (the startle threshold is 12 ms and 141–220 ms mitigates it fully), a filter-cutoff offset for darkness, and an attenuation offset that matches the four instruments to each other. Velocities are 28–62, never accented; events are at least 1.5 s apart. Melody events duck the pad bus by up to 2.5 dB.
5. Nature — a brown/pink bed low-passed at 0.9–2.8 kHz, plus the ocean engines described above: a quiet bed inside Relax, and in Environment the first of the six blendable sounds, where the whole bank also feeds the plate at −14 dB (−18 dB for stream and birds, whose recordings already carry their own room). Each wave keeps a floor under it so distant surf never stops, and the three periods (about 9, 12 and 15 s, re-drawn every cycle) never line up.
6. Space — one Dattorro plate (scaled ×1.61 to 48 kHz, 40 ms pre-delay, decay ≤ 0.90, damping 2.5–3.5 kHz), its input high-passed at 200 Hz. The tank allpasses are de-fluttered by two slow value-noise drifts instead of the paper's 0.7 Hz sine, because v3 allows no periodic modulation in the audio path.
7. Master — HP 30 Hz → optional 16 Hz amplitude modulation on the 200 Hz–1 kHz band (Focus only, off by default, depth capped at 0.35) → M/S width 1.3 with the side high-passed at 200 Hz → 1.5:1 bus compression → gentle 9 kHz roll-off → a BS.1770 K-weighted loudness meter driving a slow ±6 dB trim toward the mode's LUFS target → volume² → a 5 ms look-ahead soft limiter.

Denormals are flushed on the audio thread, all delay buffers are powers of two and the realtime thread never allocates, locks or prints (FluidSynth's global log handlers are replaced with no-ops for exactly that reason). Measured cost and memory are in docs/ACCEPTANCE.md (§v3-7 and the v4 section).

Engine protocol

Plain text on stdin, one command per line; every change is echoed as state {json} on stdout, plus meter <rms> ten times a second while playing. Unsolicited state lines are emitted when the wavetable bank finishes ("bank" → true) and when the SoundFont finishes ("sf2" → true).

play | pause | toggle | quit | state
mode focus|relax|environment        (sleep -> relax, ocean/env -> environment)
volume|intensity|brightness|tonal <0..1>
env ocean|rain|fire|wind|stream|birds <0..1>
binaural|adaptive|pulse|modulation 0|1

state JSON fields: version (5), playing, mode, volume, intensity, brightness, tonal, binaural, adaptive, pulse, modulation, env (an object with the six level keys), sounds (the same six keys, true when the slot can make a sound — the synthesized ones always, a recorded one once its WAV has loaded), daypart, bank, sf2, lufs (short-term K-weighted loudness).

Migrations are one-way and silent. A file written by v2 has no version key; modulation and pulse are forced off once, because both are periodic modulation. A v3 file names modes sleep and ocean; those resolve to relax and environment and the default env object is added. A v4 file gains stream and birds at 0. In every case the file is rewritten with "version":5. The old mode names keep working as commands for ever.

Offline, for testing without PipeWire:

omanoise-engine --render <mode> <seconds> <out.f32> [--tail <seconds>] [--stats] [--layers <mask>] [--env name=v,...]
omanoise-engine --render-bank <dir>

--tail keeps rendering with playback paused, so reverb tails can be checked for clean decay. --stats prints event counts, peak, DC, the largest sample-to-sample step, the segment log of every recorded slot that played (each entry is position+length in seconds, which is how the acceptance run shows the scheduling is random), and the roughness metric: the output envelope (|x| → four one-poles at 100 Hz → decimated to 400 Hz) is transformed with the PADsynth FFT and reported as FLUCT (energy in 1–8 Hz, where fluctuation strength peaks) and ROUGH (15–200 Hz, the roughness band), both in dB relative to the envelope's mean. OMANOISE_ENVPROF=1 adds the whole envelope spectrum in octave bands. --layers is a debug mask (1 = pads, bass and the noise bed; 2 = melody and pulse; 4 = the environment bank, the ocean bed and the 16 Hz modulation; 7 = everything). --env ocean=1,rain=0.5 overrides the persisted environment levels for one render. The plate and the master chain always run. --render-bank writes every wavetable and one soft note per instrument as raw mono f32 and prints each one's attack time, peak and spectral centroid.

Evidence notes

Neural modulation (16 Hz AM on the low-mids, Focus only) is the one "neuro" feature here with peer-reviewed support — Woods et al. 2024 found EEG phase-locking at 16 Hz and a behavioural benefit concentrated in listeners with high ADHD-symptom scores. It is nevertheless off by default: 16 Hz is inside the 15–300 Hz roughness band, and this build's whole premise is that periodic modulation in that range is what made v2 uncomfortable. Binaural beats are off by default too: the meta-analytic picture is weak and mixed. See docs/research-science.md and docs/research-opensource.md.