SpectralAudit

Inspect unusual frequency patterns in music — offline, deterministically, without AI models.

SpectralAudit is an open-source Python library by Eduardo J. Barrios. It measures high-frequency energy, abrupt spectral cutoffs and narrow, periodically spaced peaks that persist across song excerpts. Research links some periodic spectral artifacts to music-generation architectures including ones evaluated on Suno and Udio [1]. SpectralAudit implements independent DSP heuristics, not a published detector or a verified Suno-specific signature.

Not proof of AI origin. A matching pattern does not establish that a song was generated by Suno, Udio or any other model. Instruments, synthesizers, EQ, compression and resampling can produce or obscure similar features. No AI-generation probability is returned.

Install

Python 3.11+:

pip install -U spectralaudit

Quick start

Analyze a WAV or FLAC file; export JSON, Markdown or plain text:

# Default output: readable text in the terminal
spectralaudit fingerprint song.wav
# Choose one output format:
spectralaudit analyze song.wav --json report.json
spectralaudit analyze song.wav --markdown report.md
spectralaudit analyze song.wav --text report.txt

Inspect only the persistent spectral fingerprint (5–16 kHz by default):

spectralaudit fingerprint song.wav --json fingerprint.json

Other available commands:

spectralaudit spectrum song.wav
spectralaudit batch ./songs --json batch.json

Python API:

from spectralaudit import analyze, analyze_fingerprint, load_audio

report = analyze("song.wav")
pattern = report.findings["synthetic_patterns"]
print(pattern["candidate_pattern"])
print(pattern["estimated_spacing_hz"])
print(pattern["persistence_fraction"])
print(pattern["interpretation"])
report.to_json("report.json")
report.to_markdown("report.md")
report.to_text("report.txt")

audio = load_audio("song.wav")
custom = analyze_fingerprint(audio.mono, audio.rate, low_hz=5000, high_hz=16000)

What does it measure?

• Band power and cutoff: Welch PSD, high-frequency ratios and possible rolloffs, respecting Nyquist.
• Persistent peak patterns: spectral-envelope residuals, peak frequencies and spacing, autocorrelation in frequency, and support from consecutive time windows.
• Possible alternative causes: limited bandwidth, filtering, clipping, harmonics.
• Auditable output: JSON, Markdown and plain text with the complete numerical evidence; no HTML reports. No network connection, LLM, classifier or pretrained model.

candidate_pattern=true means only that the recording satisfies experimental DSP thresholds. It does not certify AI authorship. No real-world Suno-versus-human sensitivity/specificity benchmark has yet been established for SpectralAudit. See methods, limitations, and validation.

Reproducible example

A real synthetic test WAV is committed to this repository. It is generated by conventional Python DSP, not by Suno or another AI system:

• Input: synthetic_spectral_comb.wav — 6 s, 44.1 kHz, mono PCM16.
• Measured report: JSON · Markdown · Plain text.
• Recreate the WAV with python examples/generate_demo.py; recreate the reports with python examples/synthetic_demo.py (run from the repository root).

spectralaudit fingerprint examples/assets/synthetic_spectral_comb.wav --json report.json

Actual numerical output from the included synthetic file:

candidate_pattern: true
peak_count: 21
estimated_spacing_hz: 441.43
persistence_fraction: 1.0
origin_attribution: not_supported

Its generator intentionally adds tones spaced 440 Hz apart over deterministic noise. Finding those peaks verifies the measurement algorithm, not real-world detection accuracy on Suno or human music.

Stereo improvement (v0.3): if channel averaging would nearly cancel opposite-phase audio, the analyzer uses the strongest channel and records this choice in analysis_signal.

Scientific references

These works motivated development. SpectralAudit does not claim to reproduce their detectors or reported accuracy.

1. Afchar, D., Meseguer-Brocal, G., Akesbi, K., & Hennequin, R. (2025). A Fourier Explanation of AI-music Artifacts. ISMIR 2025, 739–746. DOI: 10.5281/zenodo.17811475. Architectural periodicity and empirical examples using Suno and Udio.
2. Cros Vila, L., Sturm, B. L. T., Casini, L., & Dalmazzo, D. (2025). The AI Music Arms Race: On the Detection of AI-Generated Music. Transactions of the International Society for Music Information Retrieval, 8(1). DOI: 10.5334/tismir.254. Robustness and technical confounders.
3. Dugelay, E., Barand, T., Campeas, B., Laouar, A., Afchar, D., & Hennequin, R. (2026). Improved Robustness in AI-Generated Music Detection. arXiv:2607.27454. Spectral sensitivity to speed and pitch transformations.
4. Welch, P. D. (1967). The Use of Fast Fourier Transform for the Estimation of Power Spectra. IEEE Transactions on Audio and Electroacoustics, 15(2), 70–73. DOI: 10.1109/TAU.1967.1161901. PSD estimation.
5. Harris, F. J. (1978). On the use of windows for harmonic analysis with the discrete Fourier transform. Proceedings of the IEEE, 66(1), 51–83. DOI: 10.1109/PROC.1978.10837. Spectral windows.

Full bibliography, implementation independence and licensing caveats: REFERENCES.md.

Author, citation and license

Copyright © 2026 Eduardo J. Barrios. Licensed under the open-source Apache License 2.0; see NOTICE.

Research users: please cite Eduardo J. Barrios and the exact software version. GitHub provides citation metadata in CITATION.cff; see CITING.md. Academic citation is requested, not an additional Apache-2.0 restriction.