Sonicert/Methodology
The honest engineering doc
Detection methodology
What the engine measures, what it covers, where it is still being tested, and where it can be wrong. Most detection sites hide this page; we think it's the most important one.
What the engine measures
The detector compares audio against reference fingerprints of known AI generators and human recordings, scoring four families of signals:
- Spectral statistics. Generators produce harmonics and frequency energy distributions that are statistically “cleaner” than any microphone-and-room chain produces.
- Micro-timing variance. Human performances drift; generated ones stay unnaturally consistent across the full track.
- Vocal synthesis artifacts. Phoneme-level traces — consonant onsets, breath placement, vibrato behavior — that follow model habits rather than human anatomy.
- Arrangement priors. Structural formulas and fill patterns the generators fall back on, measured against labeled corpora.
The signals are combined into a single AI-likelihood score with a confidence level. When the signals disagree or the audio gives them too little to work with, we report lower confidence — we never round a hedge into a verdict.
Coverage
| Generator | Coverage | Notes |
|---|---|---|
| Suno v3 – v5-era | Reliable | Full lineage through v5; strongest on full-length tracks |
| Udio | Reliable | Including mixed human/AI productions |
| Suno v6-era | In testing | Not yet in the engine's coverage list — see below |
| Other diffusion-style generators | Partial | Family-level detection where fingerprints overlap |
Suno v6 grey-box testing — in progress
Suno v6 is not yet in the detection engine's coverage list, and no tool on the market has published verified v6 results. We are running a grey-box evaluation: confirmed v6 samples across genres and track lengths, scored against the current engine, with per-sample results.
Status: collecting samples & running baseline
Sample-level results will be published on this page — including unflattering ones. Until then, treat any confident v6 claim, from us or anyone, as a guess.
Known limitations & false-positive directions
- Clips under ~15 seconds give the engine too little pattern — scores lean conservative and confidence drops.
- Heavily processed human music — hard-quantized EDM, aggressive autotune — can score as AI. Check the full track, weigh the confidence, and when it matters, use project files as the stronger evidence.
- Clean, short AI clips can score as uncertain. “Uncertain” means get more audio, not “it's human”.
- A score is a signal, not legal proof. Platforms and disputes want corroboration — stems, sessions, distribution records.
Data & privacy
Uploaded audio is analyzed and permanently deleted immediately after — including any transient copy in our transfer storage. No library, no training use, no third-party sharing, no ad trackers. The free checker runs without an account.
Methodology FAQ
Why don't you publish an accuracy percentage?
Because a single number hides the parts that matter — per-generator, per-length, per-genre behavior. Publishing an unbenchmarkable “99%” is exactly the move we built this site against. This page (and the v6 test results when they land) is our accuracy story.
What engine powers the detector?
A single production detection engine behind our own scoring and reporting layer. We keep the integration isolated so engines can be swapped or added as the field evolves — and so we can benchmark one engine against another and publish the difference.
Will you support batch screening or an API?
Yes, for professional use (labels, curators, platforms) — as paid volume on top of the free single-check tier, once the engine validation milestones are met.
Read the methodology? Now put it to work.
Check a track now