Synth model internals

instrument/sei/: modelled on the Roland Juno-106 family. Marbble isn’t affiliated with Roland. Not compared with the hardware, nor heard: what follows is from documentation and measured on its own terms.

From what’s documented: one DCO per voice whose saw, pulse and sub square come from one clock, so they stay phase-locked and don’t drift (their phases are locked here, with no drift at all); PWM from the one LFO or by hand; 16’, 8’ and 4’; six voices; noise; a 24 dB OTA filter with resonance to self-oscillation, envelope polarity, LFO and keyboard; one ADSR, the VCA on it or on the gate; the four-position high-pass after the voices; one triangle LFO with a delay; Poly 1 (the next voice, so tails ring), Poly 2 (the lowest free) and Unison with no detune. Above all the chorus (chorus.rs): two bucket-brigade lines swept in opposite directions, the dry signal to both sides, and the modes I, II and I+II at 0.513 Hz and 0.863 Hz over 1.66 to 5.35 ms and 9.75 Hz over 3.28 to 3.82 ms. Those figures are the ones widely repeated for the earlier sibling’s board; they come from memory of those sources, weren’t measured here, and the 106’s own board may differ a little.

Guesses, as named constants: the filter’s chip constants; the high-pass corners (240 Hz and 720 Hz, and a shelf for position 0); the knob reaches (vibrato ±3 semitones, filter LFO ±30, envelope 7 octaves); the envelope’s linear attack and exponential decay (the usual reading of its CPU-made envelope, not at its real control rate); the widest pulse (5 %); the chorus’s filters (9 kHz) and saturation; the hiss level (-65 dBFS peak at Chorus Noise 100 %); the size of Voice Variation. High-pass after the voices is from memory of the block diagram.

Differences on purpose: the hiss comes with the voices and fades 0.25 s after the last, so a silent track stays silent and costs nothing; the hardware hisses always. Clock bleed isn’t modelled, being above the audio band. Added: Unison Detune, Velocity, Voice Variation, Chorus Noise; off or modest by default.

Marbble Cinque

instrument/cinque/: modelled on the Sequential Prophet-5. Marbble isn’t affiliated with Sequential. Not compared with the hardware, nor heard.

From what’s documented: two VCOs per voice (A saw and pulse, with sync to B; B saw, triangle and pulse, Lo Freq and a keyboard switch, any waves together); five voices; mixer with noise; a 24 dB filter with keyboard off, half or full; two ADSRs and the release switch; one LFO, saw, triangle and square in any mix; Wheel Mod blending it with noise to five switches; Poly Mod: the filter envelope and Osc B, per voice at the audio rate, to Osc A’s pitch, Osc A’s pulse width and the filter. B keeps running with its mixer knob down. Unison on the lowest key, with Retrigger. Drift as Slop.

Guesses: Poly Mod’s reach (60 semitones at the pitch, ±0.48 of the width, 72 at the filter), the source being the sum of B’s waves that are on; the wheel’s reach; the mixer’s drive; the two filter chips’ constants, Early (SSM-like) and Late (Curtis-like); B’s Lo range (7 octaves down). Filter tuning isn’t trimmed (within 23 cents, notes 36 to 84).

Added: Velocity (off), Mod Amount (where the wheel rests), Unison Detune. Exponential FM from Osc B moves the mean pitch, as exponential FM does.

Marbble Uno

instrument/uno/: modelled on the Roland TB-303. Marbble isn’t affiliated with Roland. The reference is what’s published on the circuit, Robin Whittle’s Devil Fish notes and Tim Stinchcombe’s filter analysis above all, from memory of them. Not compared with the hardware, nor heard.

From what’s published: one oscillator, saw or a square shaped from it; the diode ladder with its smaller first capacitor; no keyboard tracking of the filter; the MEG (3 ms attack, 200 ms to 2 s decay, always decaying whether the key’s up or not; Env Mod’s other job of pulling the base cutoff down); the VEG (a long decay of about 3.2 s with the key down, a fast close with it up); accent forcing the MEG to 200 ms, lifting the VCA and, through the capacitor whose charging time the Resonance pot’s other half sets, the “sweep” that lags and stacks on the cutoff; slide as a lag of 60 ms.

How MIDI maps: a note is a step. Overlapping notes slide: the second comes in by the lag, without restarting either envelope. Any gap restarts them and jumps the pitch. A velocity at or above Accent at Velocity (110 by default, above the panel keyboard’s and a new note’s 100) accents; a slid-to note can be accented without a restart. The last key held sounds.

Guesses: Env Mod’s reach (40 semitones) and its pull (14); accent’s reach (36 semitones and +110 % on the VCA) and the sweep lag (47 to 147 ms, from a reading of a 1 µF and 47 k + 100 k); the square’s droop and slew; Resonance stopping at 0.965 of self-oscillation; the feedback high-pass (150 Hz); the input drive. Added: Slide Time, Accent at Velocity, Drift, a bend range.

Marbble Sedici

instrument/sedici/: modelled on the Yamaha DX7. Marbble isn’t affiliated with Yamaha. Its scales are from the published reverse engineering of the chips (Ken Shirriff’s die photographs, the fits in Raph Levien’s MSFA engine that Dexed grew from), from memory of them, not checked against a source here. Not compared with the hardware, nor heard.

From what’s documented: six sine operators, the 32 algorithms from the panel’s diagrams (algorithms.rs), one feedback loop (the average of two samples; 4 into 6 on algorithm 4, 5 into 6 on 6, the others on themselves); four rates and four levels, moving in a log scale, with the attack’s jump and curve; output level scale, keyboard level scaling with four curves, rate scaling; ratio or fixed frequency; detune; operator velocity sensitivity and amp mod sensitivity; pitch envelope; one LFO with six waves, delay, key sync; 16 voices; Osc Key Sync; mono with portamento. It runs at the instrument’s own 49.096 kHz at any session rate, and FM’s partials past that rate’s Nyquist fold, as they do on the instrument; one 48-tap Kaiser sinc brings it to the session’s rate.

The grit: with Converter Grit on: an operator’s output in 14 bits, a sine table of 4096 points read without interpolation, a 12-bit mantissa with an 8-step exponent at the output. Those are what’s published. How it sits in the real instrument’s analog stages isn’t documented well enough to model, so the output filter is the resampler’s alone. Its clean side costs little.

Approximations, not the tables: velocity (a power-law fit to the published curve’s shape), the pitch envelope’s level and rate (fits to their tables), LFO delay. Guesses: vibrato’s reach at full (±12 semitones), amp mod’s (about 48 dB), portamento’s scale, the converter’s clipping point (1.5 times a full carrier).

Differences on purpose: a voice stolen with key sync fades for 2 ms first (the hardware clicks; the contract says it mustn’t). A released voice at a fourth level above zero drones, as on the hardware, until the transport stops, when it fades. An operator’s level or envelope moved under a held note is heard on that note, so a lane works (the hardware’s edit buffer acts from the next key). Added: Volume, Touch and the wheel’s routings as knobs rather than function settings. A seed only changes anything with Osc Key Sync off or a random LFO.

Marbble Onda

instrument/onda/: a modern wavetable synth of the kind Serum made the standard. It isn’t modelled on one instrument’s panel and isn’t affiliated with Xfer Records or anyone else. Its group is Modern Synths. Its tables (tables.rs) are generated in code: Basic Shapes, Harmonic Rise, Vowels, Pulse Sweep, Sync Sweep, FM Sweep, Stepped Digital, Drawbars, Wavefolder, Comb Sweep.

Each has 32 frames, each with 10 levels halving the harmonics from 512, made by FFT and shared read-only between instances (12 MB, built the first time one is made, off the audio thread). Playback crossfades two frames at the level whose top harmonic stays under the voice’s Nyquist, read by a Catmull-Rom cubic, at 2x with the shared decimator. Up to 7 unison copies per oscillator; sub, noise, FM from B to A; Low 12 and 24, Band, High, Notch, Ladder and Diode filters (stereo in every voice). Env 1 is the VCA; Env 2 and 3, two LFOs (free or synced, shared or per voice), velocity, key, wheel and aftertouch reach eleven destinations through eight slots. The matrix runs every 16 ticks and ramps between. Choices that are judgement, not modelling: the reaches, the unison spread, the table contents. A tempo-synced LFO reads the song position per block: its output can differ by rounding across buffer sizes (under 1e-5 measured).

The later models, measured

All at 48 kHz, through instrument/test_kit.rs and each model’s tests.rs. Aliasing is the largest thing in the audio band that isn’t a harmonic, against the strongest.

Model	Tuning at 44.1, 48, 96 kHz	Aliasing	Clicks: largest step against the steady note’s
Sei	under 0.01 cents	saw at note 96 -90 dB, pulse -85, all waves at note 84 -94	note-off 1.00, steal 0.59, gate VCA 1.0
Cinque	under 0.01 cents	saw at note 96 -88, pulse -88, sync -90, Poly Mod to pitch (synced so it can be measured) at note 72 -89	note-off 1.00, steal 0.46
Uno	under 0.01 cents	saw at note 84 -77, square -95, saw with resonance -65	note-off 0.94, new step 1.26, slide 0.79
Sedici	under 0.01 cents	sine at note 84 -95 clean, -72 with the grit. FM is meant to fold: level 90 into a carrier at note 96 is -77, level 99 at note 84 -92 (nothing reaches Nyquist yet) and at note 96 -6: that’s its character at full index, and every session rate gets the same	note-off 1.00, steal 0.90
Onda	under 0.01 cents	worst of saw, square, narrow pulse, sync, FM, stepped, folded and comb frames at notes 84, 96, 108: -68 dB (sync at note 84; a sync frame at note 108 has nothing in the band)	note-off 0.98, steal 1.17

CPU (instrument_bench, release, i7-13700KF with other work on the machine, mean of the buffer’s time, 512 samples and 128 are within a percent of each other):

Model, at full polyphony	Mean
Sei, 6 voices, all waves, chorus	4.1 %	Unison the same.
Cinque, 5 voices, Poly Mod on or off	4.7 %	Costs the same: B always runs.
Uno, its one voice	0.74 %
Sedici, 16 voices	1.4 to 2.3 %	The cheapest per voice: a sine per operator at about half the others’ rate.
Onda, 8 voices, 2 oscs, 1 copy each	5.6 %
Onda, 16 voices, 2 oscs, 1 copy each	11 %
Onda, 8 voices, 2 oscs, 7 copies each	15 %
Onda, 16 voices, 2 oscs, 7 copies each	30 %	The worst this instrument does, and no worse than one core’s third. Multi-core shares tracks out, not voices within one.

Worst single blocks were 2 to 6 times the mean at 128 samples for every model including Tre, on a busy machine: scheduling here, not a property of a model, and all under the budget. An instance at rest costs nothing (under 0.1 µs a block).

Determinism: the same MIDI gives the same bits at 1, 37, 128, 512 and 1024 frames, also across a note after a long rest; renders repeat after a reset; parallel_check plays each model on a track of its own, serial against 2, 3, 4 and 8 threads, identical; rt_alloc_check plays them through parameter changes, presets and Track Off with 0 allocations; instrument_check takes each through region rendering, edit, undo, redo, every factory preset, saving and loading a user preset, reopening, Freeze and Track Off.

Going to rest is by the sample (instrument/common.rs): a model that stopped at a block boundary would run its LFO and noise on differently at another buffer size.

To add a model

1. Write instrument/<name>.rs (or a folder, as tre/): pub const PARAMS, SECTIONS, PRESETS; a struct implementing Instrument; pub fn info() -> InstrumentInfo. Build voices from synth_dsp.
2. List info() in instrument::registry(). That is all: the menus, New Tracks, the Library, panel, presets, automation, MIDI learn, undo, saving, Freeze and copy and paste work from the entry. The instrument registry test checks the spec (unique ids, defaults in range, every parameter in a section, presets in range) and that it runs.
3. Use instrument/test_kit.rs for what yours must hold: tuning, no NaN, block-size independence, reset, clicks and presets; a model’s tests.rs shows how (tre/tests.rs is the older long form).
4. Naming: an original name, in the product and presets, with no maker’s trademark. The docs may say what it’s modelled on, and that it isn’t affiliated. tests/builtinInstruments.test.ts scans each model’s files for it; add yours to its list.
5. If it needs a custom layout, register a panel in FX_PANELS (src/fx.ts). Otherwise sections will do.