From samples to a playable field
Imported sounds are encoded with SAME-S and arranged as points through PCA projection. A movable Body becomes the performance control: its position determines a weighted blend of nearby sounds, so the geometry on screen corresponds to the timbre being heard.
One motion, two morphing layers
01 · Latent layer
Generate the in-between sound
A background MLX worker decodes an N-way weighted latent vector while the audio thread remains isolated from model inference.
02 · Grain layer
Hide generation latency musically
A six-voice granular engine moves to the decoded result across grain boundaries, turning model latency into a continuous transition rather than a playback pause.
Current state
The core loop runs as VST3, AU, and a standalone application: add sounds, map the field, move the Body, and play the result with MIDI. The project is still an active prototype; perceptual evaluation, DAW stress testing, and distribution packaging remain in progress.