A Spectrum Is Not an Instrument
08 November 2026
Acoustic methods essay

Abstract
Praxymuse v0.6 places acoustics beside performance, instrument identity and production evidence. Its acoustic module lists modal relationships and signal descriptors, while explicitly limiting the claims that ideal models can support. This essay asks a narrower question than whether a generated song follows a plan: what can a measured sound feature establish about the source that produced it? The answer depends on the model, the recording and the ambiguity of the inverse problem. [1, original p. 3, M01–M02]

One frequency, several causes
An ideal tensioned string has frequencies related to length, tension and mass per unit length. The source also lists ideal open- and closed-pipe relations, together with a stiff-string correction. These equations describe different physical assumptions. A coincident fundamental does not make a string into a pipe; several mechanisms can produce similar frequency spacing. [1, p. 3, M01]

The distinction between forward and inverse use is crucial. Given a declared model and its parameters, one can predict a modal pattern. Given a recorded modal pattern, identifying the unique mechanism may be impossible without additional information. The recording system, room, playing technique and processing can all modify what reaches the microphone. A resemblance to a harmonic series is evidence about the spectrum, not a complete certificate of instrument identity.

What the window measures
Praxymuse includes a short-time Fourier transform and a spectral-centroid descriptor. A windowed transform trades temporal localization against frequency discrimination; the chosen window length and hop belong in the method record. A short attack and a sustained tone should not be treated as equivalent observations merely because an average descriptor agrees. The centroid compresses spectral magnitudes into one weighted frequency, discarding much of their arrangement. [1, p. 3]

Two sounds can therefore share a centroid while differing in partial spacing, noisiness, envelope and transient behavior. Conversely, equalization can shift the centroid without replacing the underlying instrument. This is why the source's timbral state includes multiple coordinates: partials, inharmonicity, flux, roughness, noise, envelope, formants, transients and spatial context among them. A richer descriptor remains a measurement representation, not the entire perceptual experience.

The physical model has a boundary
The modal oscillator equation includes damping and forcing. It is useful because those terms identify distinct influences on a response. Yet the source warns that nonlinear bow, reed or lip behavior and inharmonic plates are not proved by a harmonic-series formula. It separately identifies stable or passive discretization as an implementation requirement. [1, p. 3, M01]

This warning prevents an easy category error: listing a synthesis family does not demonstrate that a particular implementation accurately reproduces its physics. A numerical model must declare the equations actually solved, parameter ranges, time step and boundary treatment. If a neural or sample-based component supplies the sound instead, the result can still be useful, but its provenance should not be relabeled a physical simulation.

A proposed identification exercise
Use a rights-cleared set of recorded or synthesized tones with known production histories. Match fundamental frequency and approximate level, then compute descriptors under one declared analysis configuration. Include processed variants of the same source and different sources with similar spectral summaries. Ask whether the chosen features distinguish source families, recording conditions or merely processing choices. No dataset was assembled and no classification accuracy was measured for this essay.

Report confusion rather than hiding it. If an instrument label changes when equalization changes, the classifier may have learned a production cue. If descriptors separate only the attack portion, that finding should not be generalized to sustained sound. A held-out recording condition would provide a stronger test than randomly splitting near-identical variants of the same recording across training and evaluation.

Identity exceeds acoustics
The instrument ontology also records excitation, resonator, range, controls, techniques, culture, history and provenance. These fields answer questions that a waveform alone may not resolve. An acoustically convincing imitation does not establish the cultural or historical origin of the object it resembles. Nor does a name in a generator caption prove that an audible part came from the named instrument. [1, p. 3, M02]

Conclusion
Signal descriptors are powerful when their scope is clear. They can characterize a recording, support comparisons and expose changes. They become misleading when compressed measurements are promoted into unique explanations. The scientific opportunity lies in measuring which distinctions survive changes of recording and processing, while retaining uncertainty where multiple causes remain plausible.

Reference
[1] Praxymuse Theorem v0.6, Source-preserving mathematical integration, 6 November 2026, original PDF p. 3, modules M01–M02. Only that scientific page is supplied here, without archival attachments. The identification exercise is a proposed method, not an experimental finding.
