detection and ducking

the features

A feature a user cannot predict is worse than none: they patch it, get something confusing, and blame the plugin. So bambi offers only interpretable features:

feature	answers	character
level	how loud	continuous
attack	how sharp the onsets are	impulsive
tonal	how sustained or pitched	continuous
low, mid, high	how much there is in each band	continuous

attack and tonal pair well: in a struck note they come one after the other, strike then ring, so routed to different targets they give “the hit does this, the sustain does that” with no envelope at all.

Measures that are informative but unpredictable — spectral centroid, crest factor, flatness as a control, pitch, consonance — are not offered. They may be computed internally to build an interpretable feature; the rule is about what is exposed.

the contract

Every feature must:

1. read 0 in silence — silence must move nothing;
2. declare whether it follows loudness — only level does, and the bands weakly;
3. declare its character — impulsive features suit direct targets, continuous ones suit rates;
4. be calibrated on real material, not on test tones — a miscalibrated feature and a bad feature look the same downstream, so measure the raw value first;
5. look ahead by nothing — latency would break the feel of responding to what you hear, and impose delay compensation on plugins that otherwise need none.

what a field effect hears

The encoder hears its own input. A field effect never sees a source, only the field, so its features read the energy of the whole field, summed over every channel. The first channel alone is wrong: sounds in antiphase cancel in it, and a field with no omnidirectional part has none. The whole field cannot cancel, and for a single source it reads exactly what the first channel would, so the calibration carries over. The spectral features take their shape from the first channel and their level from the field.

Features are always computed locally. No plugin reads another’s detector: a bounce of one track alone would then differ from the mix. A sidechain is the way to listen to something else: each plugin has an optional aux input with its own six features on the matrix’s sidechain tab.

ducking is a row

Ducking is not a module and has no tab. It is level, the input’s or the sidechain’s, onto wet, with a negative depth. wet always has a row, so there is nothing to add.

What makes it musical is level’s release, set on its own page: how long the detector takes to fall, and so how long the effect takes to come back. The input and the sidechain each have one. It keeps falling into digital silence, so a hard stop releases as gently as a fade.