Downspout Algorithms

This note summarizes the main generation and processing methods used by the current Downspout plugins.

MIDI generators

• BassGen: transport-synced monophonic bass generation with style and scale vocabularies, rhythm regeneration, color-driven harmonic tension, MIDI follow/dodge behavior, Jazz-specific ii-V-I / turnaround targeting, and Fugue subject/answer bass behavior.
• DrumGen: transport-aware drum pattern generation with fill logic, style modes, genre vocabularies, pattern mutation around the current cycle, and a sparse Fugue pulse mode. An optional Automaton layer treats the generated pattern as generation 0 and, on later loop passes, replaces the hits of chosen lanes (hats, percussion, toms) with the next row of a cellular automaton seeded from the lane (see “Cellular automata” below).
• MelGen: phrase-aware monophonic melody generation with contour, call and answer, structure, range, color, follow controls, and a high-structure Fugue-friendly subject/dominant-answer region. An optional Inertia control applies Narmour-style implication: after a leap the next move tends to turn back without overshooting, after a step it tends to continue.
• Ground: longer-horizon bass generation that plans phrase roles, movement, cadences, pedal phrases, color, and high-sequence subject/answer forms across sections rather than just local bars, with explicit Dub and Jazz styles and generated notes folded into a guarded bass register lane.
• Cadence: learned harmony generation from captured MIDI, with scale-aware voicing, chord-size modes, comp scheduling, variation, arpeggiated playback phrasing, Jazz ii-V-I bias, and high-color circle-of-fifths/suspended dominant support.
• Counterpointer: learns an incoming MIDI line and answers it with a monophonic complementary line that can lean into contrary motion, chromatic color, rhythmic response, and strict dominant-answer imitation. Its Inertia control adds a scoring term that favours turning back after a leap and continuing after a step (Counterpoint mode only).
• Polymeter: four lanes of different lengths on one transport, each lane either a Euclidean pattern or a one-dimensional cellular automaton. The automaton (Wolfram rule 1-255, wrapping edges) starts from a seeded row whose density is Pulses/Length and advances one generation per cycle of the lane, returning after 64 generations. All choices are pure functions of the step position and seed, so block size and loops do not change the output.
• Xoxolo: a step-grid drum sequencer for programmed patterns. Any lane can be marked to evolve: the programmed row is generation 0 and each later pass plays the next row of a cellular automaton seeded from it (rule chosen from ten classics, a generation held for 1 to 8 passes), while unmarked lanes keep their grid. The grid itself is never altered.
• Harmonic Atlas: seeded chord progressions from four movement tables. Gravity replaces chords with functional targets chosen from the previous chord (V goes home, IV goes to V, I moves to IV, V or vi, anything else drifts to I, IV or V). Voice-leading above 50% voices each chord by choosing the inversion and octave whose sorted voices move least from the previous chord, itself voiced six chords back so that any position renders the same.
• Sprout: an L-system sequencer. Seven grammars (Plant, Koch, Dragon, Sierpinski, Cantor, Levy, Tree) are rewritten for N generations, capped at 131,072 symbols (a longer expansion falls back to the deepest generation that fits), and read as a turtle: F/G play a note, f rests, + and - move the pitch by scale degrees, brackets branch and return, other letters only grow. Pitch folds back at a set range, branch depth lowers velocity, and Grow adds one generation every N bars, restarting the pattern at each change. The generation count is read once per bar, so any change lands on a bar line. With the optional MIDI input the pitch source can be the notes held (or latched) on the input instead of a scale: degree 0 is the lowest note and degrees are spread across the held tones at about seven per octave, so the grammar’s contour follows a chord. Two CC sets (the Drift and Conductor conventions) can steer probability, gate, range, generations, velocity and seed, and a Conductor reset restarts the pattern at the next bar line.
• Markov: a Markov-chain melody. The state is the pitch class above the root and a 12 x 12 matrix of weights (0-8) gives, for each state, how likely each state is to come next. Notes outside the chosen scale are zeroed before sampling, Chaos is an exponent 2^(1 - 2c) on the weights (so 50% is the matrix as drawn), and an empty row falls back to every allowed note. Each phrase restarts the walk from the tonic, seeded by (seed, phrase), so a note is a pure function of its position. The matrix can be filled from a style, edited, or learned: transitions between successive note-ons on the input are counted relative to the root (first order for every pair, second order for every triple), and a Learned mix blends them in row by row where a row has data. With Order 2 the next note is drawn from the learned second-order counts for the last two notes when there are any, and from the ordinary row otherwise. Each note takes the octave nearest the previous one inside the register.
• Lifeform: Conway’s Game of Life drives the MIDI output, advancing the grid one beat at a time and turning active cells into musical events.
• Luma: an 8x8 Launchpad grid drives a small set of musical agents where cells can act as bass, chord, melody, or drum sources.

Audio effects

• P-Mix: transport-aware probabilistic audio gating and blending with fades.
• E-Mix: Euclidean stereo gating that carves audio into repeating patterns using density, block, and fade controls.
• Rift: short-buffer capture and disruption with chop/stutter repeats, reverse, skip, smear, and pitch-slip actions.
• PaunchLad: a dub-style performance effect built around delay throws, sirens, spring splashes, dropouts, and rhythmic chops.

MIDI effects

• M-Mix: MIDI gating that combines transport-locked probability decisions with Euclidean block patterns.
• Gremlin Driver: modulation and action sequencing that emits CC movement, action notes, and patch-randomization bursts for Gremlin.

• Cadence: input-aware MIDI harmonizing and comping that learns a cycle, rebuilds voicings from captured material, and can thin or break chords with Spread and Arpeggio.

• Retune: Scala microtuning. A .scl file (cents or ratio lines) gives the scale; each note is rounded to the nearest 12-tone key and corrected by a 14-bit pitch bend sent on its own output channel (2-16) ahead of the note-on, with oldest-note stealing. CC and program change are broadcast to the pool, polyphonic aftertouch follows its note, incoming pitch bend and channel aftertouch are dropped. The receiving synth must bend per channel.

Instruments

Canticle, Moka, Floozy, Syrinx and Pratt apply pitch bend per MIDI channel (range 2 semitones by default, set per channel by RPN 0) and key their voices by channel and note, so one-note-per-channel input such as Retune’s bends each note on its own.

• Basilico: a monophonic bass synth with glide, accent, body, drive, tempo-aware wobble modulation for amplitude/filter/phase movement, and acid squelch across upright, electric, dub, acid, and industrial tones.
• DrumKit: a triggered drum synthesizer with a fixed MIDI map, voice controls, and mixer-style mute strips.
• Floozy: an 8-voice hybrid synth combining distortion, physical-model-style excitation, feedback, filtering, modulation, and reverb.
• Gremlin: a chaotic glitch synth with sound modes, scenes, fader macros, performance actions, hold pads, and MIDI LED feedback.
• Canticle: a 12-voice tonal synth for readable keys, reed, pad, pluck, and glass parts driven by melody, counterpoint, and harmony generators.

• Plank: eight strings, one per Launchpad grid column. Each column’s eight rows are one octave of the current scale. Two 32-bit phase oscillators are summed as a polyBLEP saw pair with the second inverted; a morph axis crossfades that into a band-limited wavetable pair read at a quarter-cycle offset, and the result feeds a two-pole resonant filter whose coefficient is modulated by a second envelope.

• Pratt: Pratt polynomials f_1 = 1, f_2 = x, f_p = 1 + f_{p-1} (odd prime p), f_n = prod f_p^v, so f_n(2) = n and f_mn = f_m f_n. The filter H_n(s) = n / f_n(2 + s/w0) is all-pole, unity at DC and strictly stable. Cascading two filters multiplies the indices, H_m H_n = H_mn. Voices sum partial k with weight k^-roll * H_n(i xi k) for n = (note + 1) x base into a band-limited table; the filter mode factors the roots into biquads and crossfades old and new cascades when index or cutoff changes. Based on the original pratt-synth.

Cellular automata

Polymeter, DrumGen and Xoxolo share one piece of arithmetic (include/downspout/cellular_automaton.hpp): a row of cells, each on or off, is advanced by a Wolfram elementary rule. A cell’s next state is bit (left << 2) | (self << 1) | right of the rule number, and the row wraps at its ends. Rule 90 turns a single hit into Sierpinski’s triangle (1, 101, 10001, 1010101); rule 30 is chaotic; 110 is busy but structured; 150 and 105 fill in densely. The generation to play is a pure function of the loop pass (the absolute step divided by the pattern length), so playback, restarts and offline renders agree, and after 64 generations the row returns to its start.

The three plugins differ in what the row means. Polymeter seeds each lane from a random row of density Pulses/Length. DrumGen and Xoxolo seed from the lane’s own hits, with the pattern as generation 0, so the first pass is always what was generated or programmed; a lane with no hits stays silent, and in DrumGen new hits take the lane’s average velocity while surviving hits keep theirs. Only what is played changes, never the stored pattern.

How the site treats these methods

The plugin cards on the GitHub Pages site stay short on purpose. This page is the place to read the implementation patterns in one place when comparing generators, effects, and instruments.