DOSSIER Methods, measurements, results
These are the methods and measurements behind the current engine, with the audible reason each one matters.
The resonators and filters are TPT state-variable structures solved without unit delays. A separately implemented analogue reference checks the core to 0.02 dB at resonance and 0.15 dB across the tested band. Each voice holds its pitch, level and decay budgets from 44.1 to 192 kHz.
A trapezoidal companion model represents the pulse-shaper capacitor. The diode equation has a closed-form Wright omega solution. Accent drives that network harder, so the click changes shape in the same circuit that generates it.
The 6-oscillator metallic bank runs polyBLEP-corrected squares at measured frequencies with a 47.98% duty cycle. That asymmetry is part of the spectrum. Substance uses polyBLEP saw and pulse oscillators with hard sync.
Pitch-moving voices update filter coefficients through an approximation tested below 1 part per million of relative error. Expensive math is prepared before the audio callback. The bound applies at every supported sample rate.
Discrete white noise spreads fixed power across sample rate; a naive port therefore produces quieter snares at 96 kHz. Voltage 80 makes its shared noise bus power-spectral-density invariant. A 24 dB/oct band limit follows the hardware audio bandwidth before a voice draws from the bus.
Integer sample rounding can shift that pulse area by up to 0.25% across rates. The engine corrects the area and preserves the edge, which holds the attack specification at every supported rate.
Substance uses a 4-pole transistor-ladder model with zero-delay feedback. Drive and feedback paths saturate inside the model. One core produces low-pass / band-pass / high-pass output at 12 or 24 dB per octave, and the full-resonance state remains bounded.
Page-referenced service manuals establish the circuit. Published analyses answer documented gaps. Hardware measurements decide disagreements; unresolved values do not ship as fact.
Reference recordings calibrate the kick fundamental, decay and tone laws. The snare shell pair has a separate sweep; so does its snappy taper. Cowbell uses a measured pair. Six tom/conga curves pin their tuning, and the open-hat decay closes the set.
Zero-crossing analysis gives pitch trajectories. Log-domain RMS slopes measure decay laws; spectral peaks locate oscillator pairs. Accent tapers are checked step by step. Several widely repeated numbers did not survive those tests.
Automated checks exercise each machine during processing. A single observed allocation fails the gate.
It takes 0 locks. It starts 0 background threads. Panics are forbidden in the processing path, and runtime work has a declared bound.
The test renders one performance across changing buffer sizes and event boundaries. All sample values must remain bit-identical.
Each machine owns one seeded noise source. Two identical renders null to digital silence. The source still free-runs between hits, so repeated notes sample different points in a reproducible sequence.
Design math and filter state use 64-bit precision. Coefficients are range-checked; sub-audible tails follow an explicit policy.
The current engine synthesizes each hit at trigger time. Its recorded one-shot count is 0.
The synthesis engine has no runtime dependencies. All 3 plug-in formats wrap that core. The callback contract therefore reaches the same implementation in each format.
Voltage 80 and Voltage 78 synthesize each drum voice from measured circuit behaviour, so the circuit-modelling claim covers the entire instrument.
A machine must trace to circuit sources. Hardware calibration is separate. Sample-rate tests cover the complete supported matrix, and real DAW buffer changes exercise the render before release.
05 / 05 Generated proof
Plots regenerate from engine output and recorded hardware references. The plotting step preserves the measured samples.
Kick pitch is plotted at 44.1 through 192 kHz. The shaded area is the cents budget.
Two fixed-seed renders null to negative infinity dBFS.
The reference sweep covers the kick Decay control. Its measured T60 values appear against the model law. Each point is generated by the same analysis.
The 2 spectra are level-normalized before comparison. Their remaining difference comes from nonlinear circuit drive.