Spectral engine

Blending as spectra

With Blend Domain = Waveform (default) the echo adds up the traces’ audio. That’s exact, but two memories that aren’t aligned in time can cancel or comb-filter each other; the lower the Activation Power, the more memories blend and the more it happens.

With Blend Domain = Spectral the echo is built from short-time spectra instead: for every ~43 ms frame, each active memory’s magnitude spectrum is weighted by its activation and summed, and every frequency takes the phase of the memory that is loudest there. Memories blend without cancelling (two opposite-phase copies of a tone, which cancel completely as waveforms, add up spectrally). With a single memory the result is identical to the waveform echo.

Spectral Voices (1–32): how many of the most active memories are analysed per frame. Fewer is cheaper and clearer.
Frames are about 43 ms long, so changes in the echo fade in over about that time. There is no added latency.
In spectral mode, voice detune, wow/flutter and Loop are ignored.
It costs more CPU: roughly 3–4 % of one core per instance, against well under 1 % for waveform blending.

Spectral freeze

Spectral Freeze holds the echo’s current spectrum as a drone. Its partials keep turning at their measured frequencies (with a little jitter so it doesn’t sound static), so it keeps its pitch even after memory is cleared. Hold MIDI note base + 2 to freeze while the note is down.

Stretch

Length Mismatch = Stretch plays traces recorded at another length through a phase vocoder, so they fit the new length without changing pitch. It uses the spectral engine even when Blend Domain is Waveform.

Grain memories

Set Sync to Free and Trace Length (Free) to 15–50 ms, raise Memory Capacity into the thousands and turn Self Match off: every grain is answered by the most similar earlier grains, a granular cloud that follows what you play. Max Active Traces keeps it affordable. Try the 20 Grain Cloud preset.