SpectralAudit

Offline spectral analysis for music and final masters. A Python library by Eduardo J. Barrios.

SpectralAudit measures high-frequency energy, spectral cutoffs and regularly spaced spectral peaks in WAV/FLAC audio. It also compares a premaster and final master to reveal measurable differences between versions. Processing runs locally with NumPy and SciPy: no LLMs, AI models or external APIs.

Important: SpectralAudit is not a validated detector of AI-generated music. Frequencies above 17 kHz, low-pass cutoffs and periodic peaks do not prove AI origin. Results are measurements with stated limitations.

Install

Requires Python 3.11+.

pip install spectralaudit

Before the first PyPI release, install from the repository:

pip install git+https://github.com/edujbarrios/spectralaudit.git

Quick start

Analyze a master and export an HTML and JSON report:

spectralaudit analyze final_master.wav --html report.html --json report.json

Inspect frequency bands, compare two versions, or analyze a folder:

spectralaudit spectrum final_master.wav
spectralaudit compare premaster.wav final_master.wav --html comparison.html
spectralaudit batch ./music --json batch.json

Python API:

from spectralaudit import analyze, compare_masters

report = analyze("final_master.wav")
print(report.findings["high_frequency"]["candidate_cutoff"])
print(report.findings["spectral_comb"]["candidate_periodicity"])
print(report.interpretation)  # Origin inconclusive: ...
report.to_html("report.html")

comparison = compare_masters("premaster.wav", "final_master.wav")
print(comparison.spectral_changes)

What it measures

• Frequency bands: Welch power spectrum, relative band energy and candidate high-frequency cutoff, respecting the file's Nyquist limit.
• Periodic peaks: spacing of narrow spectral peaks; also possible in conventional harmonic music.
• Time: persistence of candidate patterns in successive audio windows.
• Alternative explanations: flags for sample-rate and processing-related ambiguities.
• Master comparison: estimated time offset, gain and band-level spectral differences.

The same inputs and settings produce deterministic analyses. Detection thresholds are exploratory heuristics, not calibrated AI-origin probabilities. Large files are read into memory; input audio is never modified. See Methods and Limitations.

Research, citation and license

If you use SpectralAudit in research, please cite Eduardo J. Barrios and the version used. See CITATION.cff and how to cite.

Selected scientific references:

• Afchar et al. (2025), A Fourier Explanation of AI-music Artifacts (paper).
• Cros Vila et al. (2025), The AI Music Arms Race (DOI).
• Welch (1967), The Use of Fast Fourier Transform for the Estimation of Power Spectra (DOI).

Full bibliography and implementation distinctions: REFERENCES.md.

Copyright © 2026 Eduardo J. Barrios. Open source under the Apache License 2.0; see NOTICE. Redistribution must preserve applicable notices. Academic citation is requested, not an additional legal condition of the license.