Verification

Verification and Measurement

Quesynth treats the original Synth1 plugin as a behavioral reference. The goal is not merely to produce a similar patch in isolation, but to identify parameter laws and signal structures that remain valid across notes, levels, sample rates, and control settings.

Reference-comparison workflow

tools/s1probe hosts the reference VST and Quesynth, supplies identical patch and event data, and records their outputs. A typical investigation follows this order:

1. Isolate one parameter or signal stage.
2. Hold unrelated parameters at controlled values.
3. Render a sweep over the relevant range.
4. Align latency, note timing, level, and polarity.
5. Compare waveform, spectrum, envelope, stereo behavior, and residual.
6. Fit the simplest model that explains held-out measurements.
7. Add an independent regression test before changing the engine.

Principal metrics

Metric	Purpose	Common failure mode
Sample null	Measures total sample-domain residual	Timing or phase error dominates the result
Harmonic spectrum	Identifies waveform, nonlinear, and filter differences	Windows miss detuned or non-bin-centered components
RMS/peak level	Measures gain behavior and headroom	Silence or release tails bias whole-render RMS
Envelope contour	Separates timing from level behavior	Threshold choice changes reported segment time
Stereo correlation/width	Evaluates pan, phase, and unison structure	Mono energy is mistaken for independent stereo layers

For aligned signals r[n] and q[n], the residual e[n] = r[n] - q[n] may be reported relative to the reference:

null_dB = 20 log10(RMS(e)/RMS(r))

A more negative result is better, but it is not a diagnosis. Spectral and temporal measurements are required to determine why a null is limited.

Independent test inputs

A test must not generate its input through the same representation that it is intended to validate.

This rule applies to binary event layouts, patch serialization, display-keyed states, and host bridges. If producer and consumer share the same defect, their agreement provides false confidence. Fixture bytes, external host captures, and hand-verified parameter cases are used where the representation itself is under test.

Measurement validity

Several plausible metrics can produce systematically incorrect conclusions:

• Short-window autocorrelation may classify low fundamentals as less tonal because too few cycles fit in the analysis window.
• Whole-render RMS understates percussive sounds because the window contains a large silent region.
• Fixed-width harmonic windows measured in hertz become progressively narrower in cents at higher harmonics and can reject detuned energy.
• A single operating point cannot distinguish two nonlinear transfer functions that happen to share one harmonic ratio.

Every fitted curve should therefore include validation points that were not used for fitting, and each metric should be tested with signals whose expected answer is known analytically.

Reproducibility standard

An audible compatibility claim should record:

• reference plugin version and host configuration;
• patch or raw stored parameter values;
• note, velocity, sample rate, block size, and transport tempo;
• render duration and analysis window;
• alignment and normalization steps;
• metric definition and numerical result;
• command or test that reproduces the result.

The detailed campaign, current match quality, and known outliers are maintained in docs/null-test.md.

---

← Architecture · Home →