Panaudium

An open-source, sample-free layered synthesizer plugin (VST3 + standalone, Windows) inspired by Spectrasonics Omnisphere. Every sound is generated in code by physical models, formant, additive, granular and analog-style synthesis. There are no recorded samples or impulse-response files anywhere, and the whole plugin is about 5 MB.

Install

Download the latest installer from Releases:

• Panaudium-<version>-Windows-x64-Setup.exe installs the VST3 into C:\Program Files\Common Files\VST3 (the folder every DAW scans) and, optionally, the standalone app. Then rescan plugins in your DAW. In FL Studio, use Options → Manage plugins → Find plugins. FL only scans that standard folder for VST3s; custom search paths are treated as VST2 paths.
• Panaudium-<version>-Windows-x64.zip contains the same files for a manual install: copy Panaudium.vst3 into the VST3 folder.

The binaries link the C++ runtime statically, so no Visual C++ redistributable is needed.

What's inside

• 2 layers × 16 voices, each layer with one of 9 synthetic engines:

  • Analog/FM/sync
  • Spectral (additive)
  • Granular
  • Karplus-Strong plucked string
  • Modal resonator bank (morphs string → bar → bell → membrane)
  • Glottal-pulse + formant voice
  • STK-style flute
  • STK-style bowed string
  • Atmos (noise through a tuned comb)

• Per layer: unison with analog drift, drive, a filter (SVF LP/HP/BP/notch, formant, comb), amp and filter envelopes, and 2 LFOs.

• Omnisphere-style FX rack: 3 slots per layer plus 4 common slots, with 17 units and 84 models:

  • Stompbox (Tube Screamer, Fuzz Face, Rat, Big Muff, ...)
  • Analog Phaser, Chorus/Ensemble, Flanger, Tremolo/Gate, Rotary (Leslie)
  • Delay, Pro-Verb (8×8 FDN)
  • Innerspace: convolution with procedurally synthesised impulse responses (Coin Dropper, Spring Tank, Metal Box, ...)
  • Shimmer, Studio EQ, Compressor (incl. Pump), Lo-Fi, Ring/Freq-shift, Wah/Env filter, Tape, Amp sim (synthetic cabinets)

• 115 factory presets in 18 categories:

  • Pad, Atmos, Drone, Texture, Cinematic, Pulse, FX
  • Keys, Bass, Lead, Pluck, Mallet, World, Brass, Percussion, Strings, Wind, Vocal

  They are loudness-matched to −18 LUFS (standard deviation 0.2 LU).

• Custom GUI: layer pages with engine-specific knob labels, and the FX rack with per-unit themes and model menus. All parameters show real units, in the GUI and in host automation.

Measured quality (reproducible with the scripts below)

Check	Result
Pitch (pYIN) for all 9 engines, A2–A5	in tune; bowed string C2–E6 and flute C2–A6 over full parameter grids
Preset loudness (BS.1770)	mean −18.1 LUFS, σ 0.20 LU (was 27 LU spread before matching)
Timbral diversity (MFCC/spectral/temporal features)	original 35 presets: mean pairwise 7.02, coverage 3.74 → all 115: 8.03 / 4.13; 64 of 80 new presets sit outside the original sound space (report)
FX torture test (17 units × 84 models × 4 knob extremes)	no NaN/Inf, bounded peaks, ≥29× realtime, default wet level 0.5–1.6×
Headless host test (VST3 hosting like a DAW, all 115 programs)	pass: programs, automation fuzz, 1–1024-sample blocks, 44.1–192 kHz, bit-exact state restore, editor open/close, ≥5× worst-block real-time headroom
pluginval (strictness 8) / Steinberg VST3 validator	pass / 47 of 47

Build from source

cmake -S . -B build -G "Visual Studio 17 2022" -A x64
cmake --build build --config Release --target Panaudium_VST3 Panaudium_Standalone

JUCE 8.0.8 is fetched automatically, or a local ./JUCE checkout is used if one exists. -DPANAUDIUM_BUILD_PLUGIN=OFF builds only the JUCE-free DSP core tools.

Tools (all in build/):

• panaudium_render <dir> renders every preset to WAV. It also takes --tuning (pitch probes), --probe <engine> <note> <timbre> <shape> <char>, and --fxtest (the FX torture test plus a WAV per FX model).
• panaudium_hosttest <Panaudium.vst3> <outdir> [--all-programs] is the headless DAW-style test.
• panaudium_snapshot <outdir> [program] renders the GUI pages to PNG.

Analysis (Python: numpy, scipy, librosa, soundfile, matplotlib, pyloudnorm, scikit-learn):

• analysis/loudness.py measures loudness, and write_trims.py folds the trims into src/dsp/presets/Trims.h.
• analysis/diversity.py and compare_banks.py measure timbral diversity.
• analysis/fx_models.py checks that the FX models sound distinct.
• analysis/tuning.py, pitchsweep.py and sweep.py check engine pitch.

Every push of a v* tag builds, tests and publishes a release through GitHub Actions (.github/workflows/release.yml).

How it relates to Omnisphere

The architecture comes from a reverse-engineering pass over the shipped Omnisphere.vst3, done with Ghidra via REA plus string and RTTI analysis. It looked at the layered RhLayeredSynthEngine, the per-layer parameter block (oscillator type, Timbre/Shape/Symmetry, FM/ring, unison+drift, dual filter, envelopes, LFOs), and the FX rack of 100+ Eff* units. The decompiled VowelFilter uses 128-step formant tables, a size scale and 5 two-pole resonators, and Panaudium's Vocal engine follows that structure. No Spectrasonics code, samples or data are used; Panaudium re-creates the kinds of sounds with synthesis instead of a sample library.

How it was built

The DSP core, FX-rack architecture, integration and all verification were written by Claude (Anthropic). Self-contained pieces went to Google's Antigravity agent (Gemini) through written briefs, kept in docs/agent-briefs/. Those pieces were:

• the FX unit implementations;
• the GUI;
• preset banks 2 and 3;
• the loudness and diversity scripts.

Each delegated piece was then compiled, measured and corrected (see the commit history).

License

AGPL-3.0 (see LICENSE), compatible with the JUCE dependency.