A hands-on review of tunealart.com — testing real Suno (v3 → v5.5), Udio, four other AI generators, and a broad battery of real human music against the live detection API, plus a teardown of the architecture and the public-facing claims.
The core ML model is the strongest part of this product: it caught every Suno track from v3 through v5, and produced zero false positives on real human music across genres, voice, and heavy degradation. What undermines it for serious "gatekeeping" use isn't the model — it's fabricated UI detail, a false privacy claim, broken docs, an intermittent backend failure, and a narrow scope (weak on Udio, blind to other generators).
Each number below is from live calls to the production detection API. "Detection rate" = share of known AI tracks correctly flagged; "false-positive rate" = share of real human audio wrongly flagged as AI.
Detection holds across every Suno version tested, including v5 (and the v5.5 "fenix" model, beyond the stated "≤5" range). Udio is hit-or-miss. Every other AI generator tested passed straight through as "Real."
Filter the table. p(AI) is the raw probability from the API (0 = real, 1 = AI). The verdict pill is what the tool reported; the right column marks whether that matched ground truth.
| Sample | Category | Version / source | p(AI) | Reported | Correct? |
|---|
Ground-truth AI samples pulled from suno.com (CDN, version-verified via Suno's clip API) and HuggingFace datasets humair025/suno-audio, disco-eth/AIME. Real-music battery from sanchit-gandhi/gtzan, Cnam-LMSSC/vibravox, plus supplied tracks.
From the publicly-exposed /openapi.json, /docs, /health, response timing, and controlled signal probes.
Python FastAPI / uvicorn on AWS API Gateway. /health exposes model_available / model_cached — there's a genuine model artifact loaded by the worker, not a rules engine.
Deterministic: identical bytes → identical probability to 6 decimals, every time.
Content-sensitive: probabilities are high-precision floats (e.g. 0.00975462…) that track the audio — not round heuristic outputs.
Scales with duration: processing time runs ~22 ms (5 s) → ~1.6 s (245 s) — real per-frame feature extraction + inference.
The API self-describes as detecting "high-frequency neural-vocoder deconvolution artifacts." But low-passing a Suno track all the way to 4 kHz still scores p(AI)=1.0 — the signal lives in the low/mid band and survives aggressive HF removal. You can't evade it by filtering, but the stated mechanism is more marketing than literal.
Real recordings are recognized robustly: a track stayed "Real" after being degraded to mono / 8 kHz / 24 kbps, and even when mixed with 80% pure tone. Only fully synthetic signals (sine, silence, sweeps) hit the "AI" bucket — and those aren't real music.
Ordered by severity. These are about the product surface and reliability — not the model's core judgement, which tested well.
Black-box only — no access to source or model weights. AI ground truth comes from version-verified Suno CDN downloads and labeled public datasets; real-music ground truth from standard genre/voice corpora and supplied tracks. Sample sizes are small (a handful per category): the results are directional and reproducible, not a formal accuracy benchmark. The detection API used is the same one the public site calls (unauthenticated demo path). All "marker" / privacy / docs findings are verifiable from the live site and its exposed /openapi.json.
⚠ This review notes that the API description targets "Suno ≤5, Udio ≤1.5." Testing confirmed strong Suno coverage (incl. v5/v5.5) and materially weaker Udio coverage — consistent with that stated scope. The gap to flag for the founder is the distance between the homepage's broad "AI Music Detector / 94.7% accuracy" framing and the actual narrow, Suno-centric capability.