Semantifx JUCE plugin

This directory contains SemantiFX's first multi-effect slice: a six-band parametric EQ and a SemantiFX-owned distortion in either order. The product direction is prompt-driven optimization of general, supplied effect chains; this plugin proves the host, inference, and real-time DSP seam without implementing chain assembly. It builds a Standalone application plus AU and VST3 plug-ins on macOS. JUCE is fetched at the pinned 8.0.8 tag during configuration.

Build and run

Build all plug-in formats and run the lightweight DSP-math checks with:

Release artifacts are produced by configuring a separate build directory with -DCMAKE_BUILD_TYPE=Release. Plugin copying is deliberately disabled so local Audio Unit and VST3 directories are not mutated by an ordinary build.

Parameters

• Low and high shelf gain, frequency, and Q
• Four bell bands, each with gain, frequency, and Q
• Distortion drive, character, 3 kHz tone shelf, and parallel mix
• Selectable EQ → distortion or distortion → EQ order
• Output trim and wet/dry mix
• Smooth bypass
• Protocol-v2 EQ → distortion result import

The 18 EQ parameters and their ranges mirror dasp-pytorch's ParametricEQ. The four distortion parameters mirror SemantiFX's differentiable processor. All controls are host-automatable and serialized in plugin state. Filter coefficients are recalculated at a fixed control interval, while parameters, output gain, mix, and bypass transitions are smoothed. The audio callback does not allocate or lock. Effect-order changes take effect at the next audio block; there is no dual-path crossfade yet, so rapid order automation can click.

The editor's Import Analysis… button accepts a complete successful v2 result. It validates every name, unit, effect type, instance identity, and parameter position, clamps values to the plugin's published ranges, and notifies the host of each change. Import and parsing run on the UI thread, never the audio thread.

Development-only capture and analysis

The editor also contains a deliberately prominent DEVELOPMENT ONLY panel for testing the complete local interaction:

1. Enter a prompt and click Listen 10s while audio is passing through.
2. After capture reaches 100%, click Shape sound.
3. The plugin writes a temporary WAV, submits a 600-iteration request to a persistent local worker, validates its protocol-v2 JSON, and applies the complete effect-chain result, including its order.
4. Before the optimizer has a valid candidate, Cancel discards the request. After the first candidate, it becomes Use current and applies the latest visible curve after interrupting the optimizer. Escape always cancels and discards.

While inference runs, the status strip reports preparation, model loading, optimizer steps, elapsed time, and an estimated time remaining. A progress sidecar publishes intermediate parameter snapshots atomically; the graph draws them as a preview without changing audible or host-automated parameters until the final result has been validated.

Capture storage is allocated by prepareToPlay; the audio callback only makes bounded copies into that storage and publishes atomic state/progress. WAV I/O, worker launch, waiting, and JSON file reads happen on a dedicated background thread. Result application is posted back to JUCE's message thread. The worker stays alive until the plugin instance is destroyed, so later requests reuse its loaded CLAP model. Interruption is cooperative at optimizer progress callbacks. If model loading or another non-cooperative phase does not stop within three seconds, the plugin terminates that worker and lazily starts a new one for the next request; this intentionally sacrifices the cache to keep the UI responsive.

This is development infrastructure, not a distribution design. CMake embeds the local repository root as SEMANTIFX_DEVELOPMENT_REPO_PATH, so the resulting binary expects this checkout's scripts/worker to remain at that path. The script in turn expects the sibling Text2FX environment described in the root project documentation. Each plugin instance currently owns one worker; a release plugin must replace this checkout-specific arrangement with a packaged shared worker/service.

On macOS, the standalone target declares why it needs microphone access. The first launch should display a system permission prompt; choose Allow, then use Options to select the intended input and output devices. AU and VST3 instances receive audio from their host and do not request microphone access themselves.