Ambi Encoder Pulsar

s3g Ambi Encoder Pulsar is a zero-input pulsar and neural-synthesis instrument with direct first-through-seventh-order ACN/SN3D output. It distributes three-layer sound objects across a field of four to thirty-two encoded points.

Pulsar synthesis repeats short, windowed waveforms called pulsarets. Slow emission rates produce separate events and rhythms; faster rates fuse into beating textures and continuous pitch. The A, B, and C layers provide independent waveforms, tuning, overlap, and synchronized FM, while a selectable sixteen-, thirty-two-, or sixty-four-node recurrent ecology can sound directly or supply captured pulsaret, envelope, and FM shapes. Four or eight internal Ambisonic ears can excite, imprint, or rebalance the emitted organism as well as return the spatial field to its neural circuit.

Workflow

1. Insert s3g Ambi Encoder Pulsar on a 64-channel REAPER track, choose a factory state or load a .s3gpulsar preset, and select the ambisonic ORDER.
2. Follow it with an ambisonic decoder.
3. Use SAVE when a state should remain portable, then set EMISSION for the overall rhythm-to-tone timescale.
4. Set each lane's carrier TUNING, CARRIER HZ or CLOCK RATIO, OVERLAP, LEVEL, OFFSET, and PULSARET.
5. Use FM RATIO and FM INDEX for clock-synchronous phase modulation, WIN SKEW to move the window peak, and TRIGGER to choose reset, continuous-phase, or idle-only behavior.
6. Choose four to thirty-two POINTS, then combine probability, burst/rest, sieve, and point-randomness controls.
7. Raise DIRECT to hear the neural circuit, then shape its signed ring and matrix feedback. After the capture meter arms, press CAPTURE to snapshot new pulsaret, envelope, and FM tables.
8. Open LISTEN, choose a TETRA 4 or CUBE 8 auditory body and a listening topology, then enable listening. Choose RESONATE, IMPRINT, or BALANCE as its response law and raise ECOLOGY; RETURN BYPASS controls only the deeper neural-audio feedback loop.
9. Set the field center, width, scatter, motion shape, orbit rate/depth, air, and Doppler behavior.

User presets are stored by default in ~/Music/s3g/Presets/Ambi Pulsar Encoder. A preset contains every editable synthesis, event, lane, neural, spatial, motion, order, and output parameter. Captured neural tables are not stored in preset files: loading a preset preserves the captured table already in memory while restoring its pulsaret, envelope, and FM mix settings. The custom preset name is retained in normal CLAP project state.

Synthesis Model

Pulsar synthesis separates the repetition rate from the short waveform, or pulsaret, heard inside each event. At slow emission rates the events form rhythms; as the rate rises, their repetition becomes a continuous tone whose spectrum depends on the pulsaret, carrier, window, and overlap.

Each emission launches the A, B, and C lane pulsarets as one spatial object. OVERLAP controls whether events remain separate or layer into denser textures, and automation applies to newly launched events without reshaping those already sounding.

Pulsaret Lanes

The three lanes are parallel components of one spatial object. Every accepted emission selects one point and normally launches A, B, and C there; OFFSET may displace each chain by a fraction of the emission period while retaining its point assignment. OVERLAP expresses window duration as a multiple of the current emission period, allowing isolated classical pulsars below one and dense concurrent pulsarets above one.

Pulsaret choices are Sine, Triangle, Saw, Square, Overtone, Fold, Impulse, and Noise. Saw and Square use polynomial bandlimited edge correction, Overtone removes partials above its safe band, and Impulse uses a finite bandlimited impulse train. High and Ultra render both nonlinear Fold and active phase-FM paths at two or four times the host rate and reconstruct them through cascaded low-pass stages.

Neural Synthesis

The neural section is a Tudor-inspired recurrent sound circuit rather than a literal emulation of a specific neural chip. One, two, or four independently seeded lobes provide sixteen, thirty-two, or sixty-four leaky tanh nodes. Every lobe contains four clusters running at progressively faster nominal R-C times of 80, 32, 9.5, and 2 ms.

Each cluster is a four-stage signed ring. A sparse signed matrix mixes the four clusters inside every lobe, while a second signed matrix exchanges bounded activity between lobes. Slower clusters gate the faster rings, and SLOW > FAST varies their effective R-C time. First-order allpass memories put phase-bearing paths inside the feedback web. Population changes use smoothed lobe activation, so adding or retiring neural material does not hard-switch a completed audio signal.

Ambisonic Self-Listening

AUDIO FB returns a conditioned mono sum of the Pulsar field to every neural lobe. Field listening provides a separate path that retains direction. The complete, post-depth point field is encoded internally to the currently selected HOA order, then decoded through four tetrahedral or eight cube-corner ears. Each ear tracks a fast and slow envelope, directional energy, onset novelty, slow habituation, and a consumable charge reservoir. Those features drive the selected response law and can also be distributed among the neural lobes before the next sample is evaluated. Motion, width, source direction, air, distance, and HOA order can therefore alter both what Pulsar hears and what it emits next.

The matched factory states Ecology A: Listen Off, Ecology B: Ears Only, and Ecology C: Closed Loop hold the underlying synthesis settings constant and use RESONATE. Step through them in order to hear the unlistening reference, the ear-driven response ecology without neural return, and the complete self-listening feedback ecology. Because the response laws act directly on event and grain synthesis, listening remains audible even when neural DIRECT, captured-shape mixes, and FIELD RETURN are all zero.

Both enable and bypass are smoothed. Every neural return is high-pass conditioned inside the neurons, low-pass conditioned according to propagation distance, normalized during HOA decoding, and bounded before re-entry. Grain imprint is latched at trigger time, so an active pulse does not jump when a moving listener crosses an ear boundary. Output trim is outside the internal return, so changing final gain does not alter the ecology. CLASSIC is the default response law.

Neural Capture Utility

The utility samples the running network at 4096 Hz into a half-second rolling window. With 32 or 64 nodes, corresponding node and cluster positions are averaged across smoothly active lobes before capture. CAPTURE derives three different 2048-point tables: fast bipolar activity becomes the pulsaret, rectified slow activity becomes the envelope, and a zero-mean mid/fast mixture becomes the FM contour. An initial snapshot is taken automatically once the window is full.

PULSARET and ENVELOPE blend those captured tables with the selected analytic waveform and envelope. FM DEPTH reads the captured FM contour across each grain and expresses its depth in semitones. Pulsaret and FM captures receive eight progressively filtered mip levels; playback selects and interpolates levels from the instantaneous phase increment. Each capture crossfades over 80 ms so active grains never see a one-sample table swap.

The capture strip distinguishes NO TABLE, STORED / MIX OFF, and ACTIVE. Its PUL, ENV, and FM flags show exactly which captured paths are audible. Before the first table exists, enabled paths are marked ARM; after capture, an active table uses the bright inverted strip.

Composable Masks

The event masks are applied together. PROB supplies stochastic density. BURST ON and REST form a repeating gate. SIEVE MOD and RESIDUE retain one residue class of the integer event sequence; modulo one disables the sieve. POINT RAND moves from even round-robin point assignment toward stochastic point choice. Each accepted event still carries all three formant chains.

Spatial Model

The point count is constrained to four through thirty-two. Grains retain their assigned point for their complete lifetime; the three lane sums at each point are combined, conditioned, and multiplied by that point's complete spherical-harmonic basis vector. Point 1 starts at the field origin, while the surrounding points use polyhedral or Fibonacci-sphere anchors. MOTION: FREE applies Ambi Encoder Stochastic's bounded three-axis random walk: Cauchy, logistic, and Gaussian perturbations drive reflected velocities and positions without collecting at an edge. The local 3D cloud is rotated around the chosen center as a complete sphere, rather than compressing its azimuth and elevation into a center sector; SPATIAL INERTIA affects only trajectory smoothing in this mode. FORSY uses the family topology shared by Stochastic and Wrangler: a three-axis, interlaced sine structure under their FREE topology modulation, with the bounded secondary random walk underneath. Its main trajectory uses a reflected 0-to-1-to-0 phase, retracing like a palindrome instead of wrapping to a new position; a bounded angular slew also protects points at topology-basis crossings. ORBIT, constant-speed triangular SWAY, and two-axis FIGURE 8 provide deterministic alternatives. ORBIT RATE and ORBIT DEPTH control motion speed and excursion. Individual HOA channels are never moved or scaled independently. With width and motion depth at zero, every point collapses exactly to the selected center direction.

The upper-left visual viewport follows the Encoder-family layout used by Water, Wind, Path, and the other spatial instruments. Its FIELD page provides the projected sphere, numbered square markers, ring links, event and energy halos, selectable points, zoom, and TOP, SIDE, and 3/4 camera presets. PULSARETS shows the window, waveform, synchronized FM, and current mode of all three lanes; selecting a trace also selects that lane in the compact editor below. NEURAL shows all four lobes and marks inactive populations as standby. LISTEN shows the tetrahedral/cube auditory body, live ear energy, ear-to-lobe relationships, return state, and the listening controls.

The complete control surface retains its native 1160-by-858 layout and control sizes. On a smaller display the plug-in opens inside a screen-fitted native scroll viewport, with horizontal and vertical access as needed. Hosts that support CLAP GUI resizing may reduce or restore the viewport freely; moving the window to a smaller screen requests a fitting viewport automatically instead of leaving controls beyond the visible desktop.

AIR applies distance-dependent high-frequency loss before encoding. DOPPLER adds a smoothed variable delay for radial movement. Both paths remain internally warm and crossfade into use, so automating either amount does not expose stale filter or empty delay state. Leave Doppler at zero when the pulsar clock itself should remain the only timing structure.

Order and Routing

The output bus is always 64 channels wide. The selected order activates (order + 1)² ACN channels and clears the remainder.