The Science Behind Audio Forensics
Audio forensics is the discipline of extracting verifiable evidence from a sound recording — originally for courtrooms, now increasingly for catching AI-generated music. The core idea is simple: every process that touches audio, from a microphone preamp to a neural codec, leaves measurable statistical traces. This article explains what those traces are, how they are measured, and how a probabilistic verdict is built from them.
What a forensic analyzer actually measures
Forensic analysis never "listens" to a song the way you do. It converts the waveform into numerical features — compact descriptors of the signal's statistical behavior — and compares their distributions against reference populations of known human and known AI recordings. The workhorse features fall into four families:
- Timbral descriptors. MFCCs (mel-frequency cepstral coefficients, typically 13 coefficients plus their deltas) summarize the spectral envelope. Generative models produce MFCC trajectories that are statistically smoother than real ensembles of instruments recorded through real microphones.
- Spectral statistics. Spectral flatness (how noise-like vs. tonal the spectrum is), spectral contrast (peak-to-valley energy per band), and the harmonic-to-noise ratio. Synthesis tends to compress the natural variance of all three.
- Phase behavior. Phase coherence entropy measures how orderly the phase relationships between frequency components are. Real rooms, real reverb tails, and real analog chains scramble phase in characteristic ways that diffusion-based generators struggle to reproduce.
- Temporal micro-structure. Micro-timing deviations, tempo drift across a take, and the variance of note attacks. Human drummers drift; quantized generation drifts differently, or not at all.
The artifacts generative models leave behind
Beyond generic statistics, each generation architecture imprints specific defects — the forensic equivalent of toolmarks. Three are especially diagnostic:
| Artifact | Physical cause | What it indicates |
|---|---|---|
| Checkerboard patterns | Transposed-convolution upsampling in decoder networks | Periodic energy ripples typical of diffusion/GAN vocoders |
| Neural codec residue (5–8 kHz) | Quantization in codecs like EnCodec, used by MusicGen | Token-based generation pipelines |
| 16 kHz cutoff / resampling notches | Models trained or rendered at limited sample rates, then upsampled | Bandwidth extension masking a lower-fidelity source |
These marks matter because they are architectural, not stylistic. A Suno v5 track can sound convincingly human while still carrying decoder-level periodicities, which is why dedicated per-engine analysis exists — see the Suno detector and the MusicGen detector for how the same forensic pipeline is tuned to different architectures. If you prefer to start from what your own ears can catch, 5 Signs a Song Was Made by AI covers the audible counterparts of these measurements.
From 82 features to one verdict: Bayesian inference
No single feature is proof. A heavily mastered pop record can show reduced dynamic range; a lo-fi human production can have a low harmonic-to-noise ratio. The science is in the aggregation. ASC v8.3 extracts 82+ forensic signals and combines them through Bayesian inference: each feature updates a probability estimate according to how strongly its observed value discriminates between human and AI populations. The output is a calibrated probability, not a binary stamp.
Calibration is what separates forensics from guesswork. Against a holdout set of 50,000+ tracks, this approach reaches 99.1% accuracy with a 0.4% false-positive rate — and that false-positive number is arguably the more important figure. A tool that flags human musicians as AI does real damage; keeping that rate low means the model must be conservative when evidence is ambiguous.
One practical consequence: the analysis is codec-robust by design. Features are chosen so that MP3 compression or a YouTube re-encode shifts them far less than the human/AI gap itself, which is why analysis works on MP3, WAV, FLAC, OGG, and M4A files, or directly from a YouTube, Spotify, or SoundCloud URL.
Watermarks: forensics you can read directly
Statistical analysis infers provenance; watermarking declares it. Two standards are becoming relevant for music:
- C2PA attaches cryptographically signed provenance metadata to a file — who or what created it, and what edits followed.
- SynthID embeds an imperceptible signal into the audio itself, designed to survive re-encoding.
When present, a watermark is near-conclusive evidence, and ASC reads both. But watermarks are opt-in and strippable, and with the EU AI Act's labeling requirements taking effect in 2026, expect a messy transition period where some AI tracks carry credentials and most do not. Statistical forensics remains the backstop for everything unlabeled.
Why this is an arms race — and what stays stable
Each model generation erases some artifacts. Suno's progression from v3.5 to v5 sharply reduced audible tells; Udio v1.5 improved vocal naturalness over v1.0. So why does forensic detection keep working?
Because generators optimize for perceptual quality, not statistical realism. The loss functions that train these models reward output that sounds right to humans, which leaves sub-perceptual statistics — phase entropy, inter-frame similarity, cepstral peak prominence — unoptimized. Closing those gaps costs compute for no audible benefit, so model builders rarely do it. Detection engines, meanwhile, retrain on every new model release. Vocals remain the hardest frontier in both directions; How to Detect AI-Generated Vocals looks at that specific battleground in detail.
Where to go from here
If you review demos, curate playlists, or clear tracks for release, the practical takeaway is this: trust measurements over impressions, treat any single indicator as weak evidence, and treat converging indicators as strong evidence. That is the entire scientific method of audio forensics compressed into one sentence.
You can apply it in under a minute. Analysis runs on the extracted features only — audio features are computed in your browser via the Web Audio API, and no audio is stored without an account. You get 3 free analyses per day without signing up, or unlimited checks with a free account.