Real-time stereo-to-multichannel upmix audio plugin using first-order Ambisonics. Available as VST3, AU, and Standalone.

UpmixRT doesn't simply copy the stereo signal to every speaker. It analyzes the spatial properties of the input and reconstructs a 3D sound field that distributes audio intelligently across the target speaker layout.

The stereo input is split into 8 frequency bands. For each band, the plugin measures:

• Inter-channel correlation (ICC) — how similar L and R are. High correlation means centered content (vocals, bass); low correlation means wide, ambient content (reverb, room tone).
• Azimuth — where the sound sits in the stereo panorama.
• Diffuseness — derived from ICC. Ambient and reverberant content scores high.
• Elevation — inferred from the spectral energy distribution.

These per-band values are energy-weighted into a single set of spatial parameters per sample.

The spatial parameters drive a first-order Ambisonic (B-format) encoder that produces four channels:

W and Y preserve the original stereo image bit-exactly. X and Z add spatial depth without altering the L/R information.

A per-layout decoder matrix maps the 4 B-format channels to the target speaker configuration. Each speaker gets a unique blend weighted by its physical position — not a copy of the input.

The LFE channel receives a lowpass-filtered version of W at -10 dB.

• Main output 1-2 is always dry stereo passthrough.
• Upmix channels are routed only to multi-out aux outputs (labeled 1/2, 3/4, ...).
• For full 22.2 wet routing, enable aux outputs up to 23-24.
• Dry/Wet controls only the aux upmix level.

• Content-adaptive: correlated content (vocals, dialog) stays up front; diffuse content (reverb, ambience) spreads to surrounds and height channels.
• Mathematically reversible: decoder matrices are constrained so that an ITU-standard downmix of the output reconstructs the original stereo input within float precision.
• Zero reported latency: the main signal path (W, Y) is pure arithmetic. Only the X/Z decorrelators introduce a small delay (~4 ms) for spatial enrichment.
• Real-time safe: no allocations, locks, or system calls in the audio path. All processing is sample-by-sample.

Requires CMake 3.22+, a C++20 compiler, and Ninja.

# macOS: install Ninja if needed
brew install ninja

# Build
cmake --preset default && cmake --build build

# Run tests
ctest --preset default

JUCE and GoogleTest are fetched automatically via CPM.

MIT