Reply to thread

Message: [QUOTE="Bombastus, post: 91813, member: 2178"] A deterministic audio plugin fed the same samples from the same initial state should produce the same output on each render. Once a developer adds per-instance tolerances, randomized noise, drift, modulation, or another time-varying process, repeated passes can stop matching perfectly. Testing analog variation therefore starts with repeatability, not with deciding whether a residual sounds pleasantly vintage. Rocksolid Audio gives the idea a concrete form through [B][URL='https://goldmidi.com/community/threads/rocksolid-audio-introduced-the-circuit-modeled-teq-6p.77153/']instance-specific component variation in TEQ-6P[/URL][/B]. A useful test is not simply whether two instances sound different. You want to know what changes, whether the change remains stable after recall, how large the spread is, and whether it depends on signal level. Keep the first setup boring. Use one preset, one input file, fixed gain staging, and identical sample-rate and oversampling settings. Render several passes from the same instance before comparing different instances, because a plugin that changes between passes is demonstrating time variance or randomness rather than fixed manufacturing-style variation. [HEADING=2]Start with repeatability before chasing character[/HEADING] Render the same short test signal several times without touching the plugin. If every pass cancels to numerical silence, or to the practical noise floor of your setup, the processor is behaving deterministically under those conditions. A stable tolerance model should normally pass this test because its chosen component values remain fixed. Next, save the session, close it, reopen it, and repeat the render. This catches a different failure mode. A plugin can create unique values when an instance is first made but still needs to restore those values correctly if the supposed variation is meant to survive session recall. Now duplicate the plugin or create fresh instances with identical controls. Compare their frequency response, phase response, output level, and harmonic spectrum at the same input level. A genuine instance-variation system may show small but repeatable differences across copies, while every individual copy remains internally consistent from pass to pass. One comparison is not enough to describe the system. Test several instances and look at the spread rather than celebrating the largest difference you can find. Realistic variation should behave like a bounded population, not like a lottery where one dramatic outlier becomes the marketing demo. [HEADING=2]Different tests expose different kinds of variation[/HEADING] Use a low-level sine sweep or broadband test first to measure frequency response without driving nonlinear stages hard. Small shifts in filter center, Q, gain, or stereo balance can show up here. Repeat the measurement at higher input levels, and any new changes may point toward level-dependent behavior rather than static tolerance alone. Single sine tones are useful for harmonic distortion because the new frequencies are easy to identify. Two-tone tests add another layer by exposing intermodulation products, which can reveal nonlinear interaction that a simple harmonic display misses. Comparing both at several drive levels prevents one flattering screenshot from standing in for the whole processor. Time-varying behavior needs repeated measurements rather than a single sweep. Academic [B][URL='https://www.dafx.de/paper-archive/2015/DAFx-15_submission_6.pdf']measurement work on audio-effect time variance[/URL][/B] used repeated exponential sweeps with uneven pauses to help detect systems whose response changes over time. The uneven spacing matters because perfectly regular tests can accidentally synchronize with periodic modulation and hide part of what is happening. Stereo deserves its own pass. Feed identical material into left and right channels and compare magnitude, phase, harmonics, and residual level. Channel differences can come from modeled tolerances, but they can also come from unlinked random processes, noise generators, or deliberate stereo decorrelation, so the measurement tells you a difference exists without automatically telling you why. [HEADING=2]Null tests need careful interpretation[/HEADING] A null test is excellent for proving two signals are not identical, but it is poor at explaining the cause by itself. Tiny gain mismatches can leave a surprisingly large residual, as can phase shifts, timing offsets, different noise, or a small EQ change. Level-match first and align timing before treating the leftover signal as evidence of nonlinear character. The residual also says nothing about authenticity. Two plugins can fail to null because they are different from each other while both remain inaccurate representations of the hardware they claim to model. Hardware measurements are needed if the question is analog modeling accuracy rather than simple plugin-to-plugin difference. Run the same tests across several input levels before concluding saturation. A plugin may appear almost linear around moderate operating levels yet produce much stronger harmonics when pushed. Another may add a fixed noise floor or static response difference that remains nearly unchanged with drive, which is a different phenomenon entirely. Finish with ordinary music only after the controlled tests tell you what to listen for. Level-match the renders, blind the instance identities if possible, and check whether the measured differences survive real program material. Small measurable variation can be genuine and still be musically irrelevant, while a subtle stereo or phase change may matter greatly on one source and disappear on another. [/QUOTE]

Name