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
Our self-hosted acoustic engine measures a spectral fingerprint from the first 300 seconds of audio and applies an open-weights logistic regression model, a band-correlation head and a frozen cepstral CNN. v2.3 adds a nonlinear classifier combining spectral, temporal and neural features from the opening eight seconds. It does not identify a generator or prove authorship. The score is uncalibrated: 99% on the dial is not a measured 99% chance of being right.
- · File bitrate no longer decides the result. Identical decoded audio receives the same assessment in a different container.
- · The v2.2 checks require a band-correlation score ≥0.9 with support from the earlier band head, or a median CNN score ≥0.9 across at least three complete, non-overlapping 10-second windows. This also checks earlier positive results.
- · v2.3 preserves those AI results. For other results, its eight-second classifier can add an AI signal at the frozen margin threshold 0.872304. Lower scores, short input and model unavailability retain the v2.2 result. This addition introduces no abstention rule.
- · A human result additionally requires no limited-spectrum flag and a strong negative band-head score (<0.15). Otherwise we withhold the verdict; damaged AI can fool both models.
- · Content Credentials are a separate provenance signal. Their presence or absence is not an acoustic verdict.
These safeguards currently apply to the single-song path. Batch screening still uses the original acoustic model; its migration remains pending. The new single-song evaluation is below.
Coverage
| Generator | Coverage | Notes |
|---|---|---|
| Earlier Suno versions | Model-declared | Historical raw scores are retained below; current file-quality safeguards can withhold those verdicts |
| Udio / MusicGen / Mustango | Challenge results published | The additional AIME challenge exposes missed AI, including false-human signals; broad coverage is not established |
| Suno v6-MINI | Development evaluation | 9 source files labeled v6-MINI; the archived v2.2 outcomes are included in the 76-file table |
| Other generators | Not established | YuE2 is not reliably detected by this engine; no broad coverage claim follows from the Suno samples |
Latest new-work evaluation · October 11, 2026
Production remains v2.3. A frozen round-three candidate was evaluated once on 192 newly acquired works: 108 AI originals and 84 human references from new author identities, including long human electronic music. Their six original/excerpt/transcoded forms make 1,152 correlated files. This is a local paired evaluation; these are not 1,152 independent works or production uploads.
| New originals | Production v2.3 | Unreleased candidate |
|---|---|---|
| AI detected / 108 | 36 | 50 |
| AI inconclusive | 7 | 7 |
| AI incorrectly called human | 65 | 51 |
| Human signal / 84 | 55 | 55 |
| Human inconclusive | 29 | 29 |
| Human incorrectly called AI | 0 | 0 |
Across all forms, AI detections rise from 176/648 to 240/648; 364 false-human and 44 inconclusive AI results remain. Human outcomes stay at 381 correct, 119 inconclusive and 4 false AI results across 504 files — the same four transcoded forms of one 2016 human chiptune that v2.3 already misjudges. No abstention or human error was added and AI detection strictly improved on originals, processed forms and the total, but the frozen gate counts every observed human error, so the candidate has not been deployed.
Read the complete round-three study, failure and source records
Earlier frozen evaluation · round-two candidate · October 11, 2026
A follow-up frozen candidate was evaluated once on its own 192 new works: 108 AI originals and 84 human references from new author identities. Their six original/excerpt/transcoded forms make 1,152 correlated files. This is a local paired evaluation; these are not 1,152 independent works or production uploads.
| New originals | Production v2.3 | Unreleased candidate |
|---|---|---|
| AI detected / 108 | 34 | 41 |
| AI inconclusive | 8 | 8 |
| AI incorrectly called human | 66 | 59 |
| Human signal / 84 | 72 | 72 |
| Human inconclusive | 11 | 11 |
| Human incorrectly called AI | 1 | 1 |
Across all forms, AI detections rose from 165/648 to 191/648, with no added abstention or human error, but the candidate preserved an existing v2.3 neural-recovery false positive and failed the predefined zero-observed-human-error gate. It was not deployed.
Read the complete round-two study, failure and source records
Earlier frozen evaluation · v2.3 · October 10, 2026
The model and decision rule were frozen before scoring 168 newly selected works: 72 generated tracks and 96 human-reference tracks from 96 new FMA artist identities. Each has an original file and four matched eight-second variants, totaling 840 correlated files. All were replayed through the integrated single-song detector; this is local validation, not 840 production uploads.
| New originals | Previous v2.2 | Current v2.3 |
|---|---|---|
| AI detected / 72 | 5 | 23 |
| AI inconclusive | 8 | 6 |
| AI incorrectly called human | 59 | 43 |
| Human signal / 96 | 80 | 80 |
| Human inconclusive | 16 | 16 |
| Human incorrectly called AI | 0 | 0 |
Across all five forms in that earlier test, AI detections rise from 7/360 to 81/360, with 21 inconclusive and 258 false-human results remaining. Human results remain 390/480 correct, 90 inconclusive and zero observed false positives. No file adds an abstention or loses a previously correct verdict. This remains limited coverage; zero observed false positives is not a population guarantee.
The earlier candidate's 1 human error and 57 misses are corrected in a consumed development regression, not a fresh test. Historical generator versions, unknown upstream training overlap and the FMA/AIME source confound limit generalization. A separate 12-file control from four OpenGameArt authors keeps its three human and nine inconclusive results.
Read the full v2.3 repair and limitations · Download all results and sources (JSON)
Archived development snapshot · v2.2 · October 10, 2026
76 files: 68 determinate, 8 inconclusive. We observed 0 false positives and 0 false negatives among the determinate results. Inconclusive files are not counted as correct detections. This is a local replay of the single-song pipeline, not a claim that every file was uploaded to production.
The set contains 59 source files and 17 transformations, not 76 independent works. It includes development data and is not a sealed blind test. Some exact source URLs and work identities remain unreconciled; the row notes expose those gaps. Labels come from the retained collection records. Audio is not redistributed.
Source files
52 determinate / 7 inconclusive / 59 files
Derived files
16 determinate / 1 inconclusive / 17 files
Historical 26 replayed with v2.2
22 determinate / 4 inconclusive / 26 files
| Same retained files | v2.1 | v2.2 archive |
|---|---|---|
| AI detected / 25 AI files | 20 / 25 | 25 / 25 |
| AI inconclusive | 5 | 0 |
| Human signal / 51 human files | 43 / 51 | 43 / 51 |
| Human inconclusive | 8 | 8 |
All 15 labeled AI source files received an AI signal in this replay. All five heavily transcoded AI controls received actual AI results. Their parents and every derivative were excluded from fitting the new head; these were known development targets, not unseen blind samples. The old weights remain frozen; the new correlation head and frozen CNN provide the additional evidence.The previous v2.1 snapshot remains available.
The historical 26-file pilot had 22 determinate and 4 inconclusive results in v2.2. An additional 104 existing codec variants yield 40 determinate and 64 inconclusive results, with no determinate errors observed. These variants overlap model development and reuse source works; they are a stress regression, not 104 independent works or a new blind test. Full rows and chain-level counts are in the JSON.
Download all unrounded scores, SHA-256 hashes and evaluation policy (JSON) · Read the failure, investigation and correction
All 76 files — including inconclusive results and source gaps
The original base score is archived separately from the routed score. Recovery may change the displayed score; the route and confidence determine the assessment. All scores and confidence are uncalibrated. Full routed scores are in the JSON.
| File / source | Recorded label | File kbps | Base score | Band-head score | Confidence | v2.1 → v2.2 / route | SHA-256 |
|---|---|---|---|---|---|---|---|
| human/Carefree.mp3 incompetech (CC-BY) source-link-recorded | human source-file | 256.1 | 7.7512e-9 | 0.00016928 | 100% | human → HUMAN signal human-unflagged | 8433b770a630… |
| human/Wallpaper.mp3 incompetech (CC-BY) source-link-recorded | human source-file | 320.1 | 1.7599e-7 | 0.00082213 | 100% | human → HUMAN signal human-unflagged | 9fa810e5193c… |
| human/Sneaky Snitch.mp3 incompetech (CC-BY) source-link-recorded | human source-file | 320.1 | 1.6769e-10 | 0.000019263 | 100% | human → HUMAN signal human-unflagged | db1710333d68… |
| human/Fluffing a Duck.mp3 incompetech (CC-BY) source-link-recorded | human source-file | 320.3 | 2.8191e-7 | 0.00030836 | 100% | human → HUMAN signal human-unflagged | cb23b1833c35… |
| human/Monkeys Spinning Monkeys.mp3 incompetech (CC-BY) source-link-recorded | human source-file | 320.1 | 0.00026436 | 0.094620 | 81% | human → HUMAN signal human-unflagged | a5bb345c2384… |
| human/Investigations.mp3 incompetech (CC-BY) source-link-recorded | human source-file | 320.4 | 1.1453e-11 | 7.4333e-7 | 5% | uncertain → Inconclusive weak-signal | 699f7a3cd227… |
| human/George Street Shuffle.mp3 incompetech (CC-BY) source-link-recorded | human source-file | 320.1 | 1.8983e-11 | 0.0000010057 | 100% | human → HUMAN signal human-unflagged | 15ee9995e01f… |
| human/Bumbly March.mp3 incompetech (CC-BY) source-link-recorded | human source-file | 332.7 | 4.6605e-10 | 0.00035741 | 100% | human → HUMAN signal human-unflagged | c43e98729a91… |
| human/The Complex.mp3 incompetech (CC-BY) source-link-recorded | human source-file | 320.1 | 0.000012194 | 0.031341 | 94% | human → HUMAN signal human-unflagged | c270fb44a698… |
| human/Achaidh Cheide.mp3 incompetech (CC-BY) source-link-recorded | human source-file | 326.7 | 4.6679e-15 | 1.6205e-7 | 100% | human → HUMAN signal human-unflagged | 414f949be039… |
| human/edison-1920-vocal.mp3 Archive.org (public domain) source-link-recorded | human source-file | 123.1 | 0.0011128 | 0.20355 | 5% | uncertain → Inconclusive weak-signal | e94e8f46bb85… |
| human/78rpm-vocal-1940s.mp3 Archive.org (public domain) source-link-recorded | human source-file | 183.8 | 0.014440 | 0.44667 | 5% | uncertain → Inconclusive weak-signal | 47da5bdcea6a… |
| compress/carefree-128k.mp3 ffmpeg transcode of a batch sample derived-from-retained-source | human derived-file | 128.1 | 1.0377e-7 | 0.00076286 | 100% | human → HUMAN signal human-unflagged | 1267253567c0… |
| compress/carefree-64k.mp3 ffmpeg transcode of a batch sample derived-from-retained-source | human derived-file | 64.0 | 0.0000037958 | 0.0072726 | 5% | uncertain → Inconclusive weak-signal | 22de7648cb91… |
| ai/ja1.m4a Generated by site author on Suno (2026-09) derived-from-retained-source | ai source-file | 140.0 | 1.0000 | 0.99991 | 100% | ai → AI signal ai-consensus | b0ab1c06adeb… |
| ai/ja2.m4a Generated by site author on Suno (2026-09) derived-from-retained-source | ai source-file | 133.9 | 0.99999 | 0.99581 | 99% | ai → AI signal ai-consensus | be845b0e160c… |
| ai/ja3.m4a Generated by site author on Suno (2026-09) derived-from-retained-source | ai source-file | 133.3 | 0.99999 | 0.99649 | 99% | ai → AI signal ai-consensus | 107f83900ccd… |
| ai/zh1.m4a Generated by site author on Suno (2026-09) derived-from-retained-source | ai source-file | 138.8 | 1.0000 | 0.99998 | 100% | ai → AI signal ai-consensus | c05768ba0906… |
| ai/zh2.m4a Generated by site author on Suno (2026-09) derived-from-retained-source | ai source-file | 137.6 | 0.99608 | 0.99754 | 99% | ai → AI signal ai-consensus | b022cff8ac8c… |
| ai/zh3.m4a Generated by site author on Suno (2026-09) derived-from-retained-source | ai source-file | 129.4 | 1.0000 | 0.99992 | 100% | ai → AI signal ai-consensus | 375791744964… |
| compress/zh1-192k.mp3 ffmpeg transcode of a batch sample derived-from-retained-source | ai derived-file | 192.1 | 1.0000 | 0.99997 | 100% | ai → AI signal ai-consensus | 0e13a15912f7… |
| compress/zh1-128k.mp3 ffmpeg transcode of a batch sample derived-from-retained-source | ai derived-file | 128.1 | 1.0000 | 0.99991 | 100% | ai → AI signal ai-consensus | 67cf304396f8… |
| compress/zh1-96k.mp3 ffmpeg transcode of a batch sample derived-from-retained-source | ai derived-file | 96.1 | 1.0000 | 0.99973 | 100% | ai → AI signal ai-consensus | 9c6c70abffab… |
| compress/zh1-64k.mp3 ffmpeg transcode of a batch sample derived-from-retained-source | ai derived-file | 64.1 | 0.99932 | 0.99656 | 99% | ai → AI signal ai-consensus | 75a09c2238a7… |
| compress/ja1-64k.mp3 ffmpeg transcode of a batch sample derived-from-retained-source | ai derived-file | 64.1 | 0.99829 | 0.99105 | 98% | ai → AI signal ai-consensus | 652b5a1a606b… |
| ai/en1-v6mini.mp3 Suno share page (V6-MINI) source-link-recorded | ai source-file | 128.1 | 0.99990 | 0.99819 | 100% | ai → AI signal ai-consensus | e412c00d3ec6… |
| human/Electrodoodle.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 320.1 | 7.7873e-7 | 0.0013044 | 100% | human → HUMAN signal human-unflagged | 75227ad15378… |
| human/Voltaic.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.1 | 0.000095561 | 0.035432 | 93% | human → HUMAN signal human-unflagged | 8c984f654a2a… |
| human/Digital Lemonade.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.1 | 4.1546e-9 | 0.000086848 | 100% | human → HUMAN signal human-unflagged | f9ebdf5e8d70… |
| human/Sardana.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.1 | 1.4801e-11 | 0.0000016121 | 100% | human → HUMAN signal human-unflagged | 2de923a476a6… |
| human/Crunk Knight.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.1 | 5.9477e-11 | 0.0000049486 | 100% | human → HUMAN signal human-unflagged | 702f3de96305… |
| human/Curse of the Scarab.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.1 | 0.000029434 | 0.035896 | 5% | uncertain → Inconclusive weak-signal | e1c825715a8e… |
| human/Ossuary 6 - Air.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.1 | 9.8705e-13 | 8.8839e-7 | 100% | human → HUMAN signal human-unflagged | bf57ba7ee56b… |
| human/Fireflies and Stardust.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.1 | 6.6181e-9 | 0.000041726 | 100% | human → HUMAN signal human-unflagged | 28d557e8d15e… |
| human/Tempting Secrets.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 320.1 | 1.9711e-16 | 1.5136e-9 | 100% | human → HUMAN signal human-unflagged | cd183a6f2584… |
| human/Pamgaea.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 320.1 | 5.9719e-11 | 0.0000064303 | 100% | human → HUMAN signal human-unflagged | 9f0ad72567c4… |
| human/Disco con Tutti.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 320.1 | 8.0023e-8 | 0.00011288 | 100% | human → HUMAN signal human-unflagged | ac268ea49ca8… |
| ai/suno-v6mini-ccl.mp3 Account capture labeled Suno v6-MINI in the retained manifest exact-source-link-not-recorded | ai source-file | 186.5 | 0.99999 | 0.99960 | 100% | ai → AI signal ai-consensus | 4ec51c7229a7… |
| ai/suno-v6mini-ica-b.mp3 Account capture labeled Suno v6-MINI in the retained manifest exact-source-link-not-recorded | ai source-file | 182.3 | 1.0000 | 0.99955 | 100% | ai → AI signal ai-consensus | 7ce04ddad4d5… |
| ai/suno-v6mini-ja-tonight.mp3 Account capture labeled Suno v6-MINI in the retained manifest exact-source-link-not-recorded | ai source-file | 184.3 | 1.0000 | 0.99989 | 100% | ai → AI signal ai-consensus | 37468d3ad37d… |
| ai/suno-v6mini-ja-wangan.mp3 Account capture labeled Suno v6-MINI in the retained manifest exact-source-link-not-recorded | ai source-file | 186.0 | 1.0000 | 0.99997 | 100% | ai → AI signal ai-consensus | ab2374e7882b… |
| ai/suno-v6mini-nocturnal-a.mp3 Account capture labeled Suno v6-MINI in the retained manifest exact-source-link-not-recorded | ai source-file | 186.3 | 1.0000 | 0.99998 | 100% | ai → AI signal ai-consensus | 85b116e50a49… |
| ai/suno-v6mini-nocturnal-b.mp3 Account capture labeled Suno v6-MINI in the retained manifest exact-source-link-not-recorded | ai source-file | 189.4 | 1.0000 | 0.99974 | 100% | ai → AI signal ai-consensus | be7af17fd41f… |
| ai/suno-v6mini-zhroad-a.mp3 Account capture labeled Suno v6-MINI in the retained manifest exact-source-link-not-recorded | ai source-file | 178.8 | 1.0000 | 0.99990 | 100% | ai → AI signal ai-consensus | 562338094892… |
| ai/suno-v6mini-zhroad-b.mp3 Account capture labeled Suno v6-MINI in the retained manifest exact-source-link-not-recorded | ai source-file | 178.4 | 1.0000 | 0.99986 | 100% | ai → AI signal ai-consensus | 456593e15937… |
| human/Awesome Call.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 290.4 | 8.5265e-13 | 3.0211e-7 | 100% | human → HUMAN signal human-unflagged | a08b613cdbdd… |
| human/Black Vortex.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 320.1 | 0.0000039888 | 0.0074219 | 99% | human → HUMAN signal human-unflagged | a00af27dc5d6… |
| human/Dragon and Toast.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 323.6 | 4.6881e-9 | 0.000083722 | 100% | human → HUMAN signal human-unflagged | 4aa6c0135e80… |
| human/Guts and Bourbon.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.1 | 0.000049524 | 0.0021022 | 100% | human → HUMAN signal human-unflagged | 8cd93a36fb6e… |
| human/Hustle.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 320.1 | 5.4209e-9 | 0.000064384 | 100% | human → HUMAN signal human-unflagged | 8acbe21e7816… |
| human/I Knew a Guy.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 289.6 | 1.7098e-17 | 4.1295e-9 | 100% | human → HUMAN signal human-unflagged | 10e3480eefcb… |
| human/Kick Shock.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 332.0 | 6.3897e-16 | 1.0263e-8 | 100% | human → HUMAN signal human-unflagged | 762f9e87a78f… |
| human/Kool Kats.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 323.8 | 5.7116e-15 | 9.6569e-8 | 100% | human → HUMAN signal human-unflagged | 019ba5f2ebb6… |
| human/Militaire Electronic.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 325.9 | 2.7733e-10 | 0.00048097 | 100% | human → HUMAN signal human-unflagged | 8aaf3e0c8bd3… |
| human/Rynos Theme.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.1 | 0.000012638 | 0.0043491 | 99% | human → HUMAN signal human-unflagged | 4d01ed07cb2b… |
| human/Smooth Lovin.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.1 | 3.2585e-13 | 0.0000024776 | 100% | human → HUMAN signal human-unflagged | 97f914222378… |
| human/Spazzmatica Polka.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 320.2 | 1.2841e-13 | 8.5721e-7 | 100% | human → HUMAN signal human-unflagged | ac328eac4379… |
| human/Street Party.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.1 | 5.7997e-7 | 0.00099880 | 100% | human → HUMAN signal human-unflagged | b49adaacb7ed… |
| human/Take the Lead.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 320.1 | 2.2141e-8 | 0.00046246 | 100% | human → HUMAN signal human-unflagged | f2fb95ce220b… |
| human/The Lift.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.0 | 0.00011157 | 0.0053479 | 99% | human → HUMAN signal human-unflagged | ea72d7433ffc… |
| human/Volatile Reaction.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 256.1 | 0.00015591 | 0.035692 | 5% | uncertain → Inconclusive weak-signal | ced1148df201… |
| human/Windswept.mp3 Kevin MacLeod / incompetech; collection tagged CC-BY in the retained manifest collection-link-only | human source-file | 320.1 | 1.3650e-11 | 0.0000022756 | 100% | human → HUMAN signal human-unflagged | 0d7445201046… |
| derived/postprod-h-Electrodoodle.mp3 Locally transformed control; not an independent work derived-from-retained-source | human derived-file | 165.3 | 3.0833e-26 | 6.1823e-12 | 100% | human → HUMAN signal human-unflagged | 50f65faa2a8d… |
| derived/postprod-h-Crunk Knight.mp3 Locally transformed control; not an independent work derived-from-retained-source | human derived-file | 183.4 | 4.7856e-22 | 1.5127e-10 | 100% | human → HUMAN signal human-unflagged | db7e3b8d0c32… |
| derived/postprod-h-Militaire Electronic.mp3 Locally transformed control; not an independent work derived-from-retained-source | human derived-file | 247.8 | 2.7687e-22 | 4.1731e-10 | 100% | human → HUMAN signal human-unflagged | c063586874f1… |
| derived/postprod-h-Kick Shock.mp3 Locally transformed control; not an independent work derived-from-retained-source | human derived-file | 213.4 | 7.8891e-22 | 3.4726e-10 | 100% | human → HUMAN signal human-unflagged | bafe1b5b1891… |
| derived/postprod-h-Black Vortex.mp3 Locally transformed control; not an independent work derived-from-retained-source | human derived-file | 172.1 | 2.6332e-22 | 1.2143e-9 | 100% | human → HUMAN signal human-unflagged | a9dab60f976e… |
| derived/postprod-ai-zh1.mp3 Locally transformed control; not an independent work derived-from-retained-source | ai derived-file | 95.9 | 0.10438 | 0.87760 | 100% | uncertain → AI signal cepstrum-ai | d5ec73bb4db1… |
| derived/postprod-ai-ja1.mp3 Locally transformed control; not an independent work derived-from-retained-source | ai derived-file | 97.0 | 0.0045210 | 0.25484 | 100% | uncertain → AI signal cepstrum-ai | 9e504e86f48e… |
| derived/postprod-ai-en1-v6mini.mp3 Locally transformed control; not an independent work derived-from-retained-source | ai derived-file | 51.7 | 0.39619 | 0.93501 | 87% | uncertain → AI signal correlation-ai | 1813fae96d41… |
| derived/postprod-ai-suno-v6mini-ccl.mp3 Locally transformed control; not an independent work derived-from-retained-source | ai derived-file | 103.3 | 0.22838 | 0.87794 | 98% | uncertain → AI signal cepstrum-ai | 13a36e7dbe4d… |
| derived/postprod-ai-suno-v6mini-nocturnal-a.mp3 Locally transformed control; not an independent work derived-from-retained-source | ai derived-file | 101.4 | 0.47501 | 0.78986 | 58% | uncertain → AI signal correlation-ai | cce6e2a4b2b5… |
| human/ia78-0.mp3 Historical recording labeled human in the retained manifest exact-source-link-not-recorded | human source-file | 290.6 | 0.000078390 | 0.14487 | 5% | uncertain → Inconclusive weak-signal | 683e857dccf8… |
| human/ia78-1.mp3 Historical recording labeled human in the retained manifest exact-source-link-not-recorded | human source-file | 288.9 | 1.2605e-7 | 0.00023394 | 100% | human → HUMAN signal human-unflagged | 45c5d30cf4e2… |
| human/ia78-2.mp3 Historical recording labeled human in the retained manifest exact-source-link-not-recorded | human source-file | 323.4 | 0.000013245 | 0.11292 | 77% | human → HUMAN signal human-unflagged | 2808c15a86fc… |
| human/ia78-3.mp3 Historical recording labeled human in the retained manifest exact-source-link-not-recorded | human source-file | 234.1 | 0.00013112 | 0.032474 | 5% | uncertain → Inconclusive weak-signal | cf7fc718a4a1… |
Historical raw-model pilot — 26 files, 26 matched labels
Published September 30, 2026 — using the raw base model before the current safeguards. These retained historical scores are not today's product outcomes. The archived v2.2 evaluation above includes every pilot file and its inconclusive result.
| Batch | Samples | Result |
|---|---|---|
| Known-human productions | 14 | 0 false positives — including aggressive electronic productions, a 1920s Edison vocal recording, a 1940s 78rpm transfer, and two low-bitrate transcodes |
| Known-AI: Suno user-generated (2026) | 7 | 7/7 detected — 3 Japanese-vocal, 3 Mandarin-vocal, 1 English-vocal V6-MINI track |
| Known-AI: transcode ladder (64–192 kbps) | 5 | 5/5 detected — no verdict flips anywhere on the ladder |
Sample provenance: human tracks from incompetech (Kevin MacLeod, CC-BY) and Archive.org public domain transfers, according to the collection records. AI files are labeled Suno in those records; retained source links appear below. We have not independently audited the base model's training corpus. The V6-MINI file was captured from a share-page playback stream and converted through Opus and MP3; it is not an untouched download.
Read the scale honestly: 26 files form a small pilot with repeated works and transformations. These observations do not establish general accuracy or a sealed blind-test result. Current screening outcomes are published above.
Full per-track results — all 26 samples, with SHA-256 fingerprints
P(AI) is the engine's raw probability for each file. The SHA-256 prefix lets anyone holding the same file verify it is the exact audio scored. AI-track audio files are not hosted here — Suno's free-tier output terms are the site author's to honor, and publishing a ready-made evasion tuning set would be self-defeating. Human-track sources are linked.
| Sample | Provenance | Label | P(AI) | SHA-256 |
|---|---|---|---|---|
| Carefree.mp3 | incompetech (CC-BY) | human | 0.0000 | 8433b770a630… |
| Wallpaper.mp3 | incompetech (CC-BY) | human | 0.0000 | 9fa810e5193c… |
| Sneaky Snitch.mp3 | incompetech (CC-BY) | human | 0.0000 | db1710333d68… |
| Fluffing a Duck.mp3 | incompetech (CC-BY) | human | 0.0000 | cb23b1833c35… |
| Monkeys Spinning Monkeys.mp3 | incompetech (CC-BY) | human | 0.0003 | a5bb345c2384… |
| Investigations.mp3 | incompetech (CC-BY) | human | 0.0000 | 699f7a3cd227… |
| George Street Shuffle.mp3 | incompetech (CC-BY) | human | 0.0000 | 15ee9995e01f… |
| Bumbly March.mp3 | incompetech (CC-BY) | human | 0.0000 | c43e98729a91… |
| The Complex.mp3 | incompetech (CC-BY) | human | 0.0000 | c270fb44a698… |
| Achaidh Cheide.mp3 | incompetech (CC-BY) | human | 0.0000 | 414f949be039… |
| edison-1920-vocal.mp3 | Archive.org (public domain) | human | 0.0011 | e94e8f46bb85… |
| 78rpm-vocal-1940s.mp3 | Archive.org (public domain) | human | 0.0144 | 47da5bdcea6a… |
| carefree-128k.mp3 | ffmpeg transcode of a batch sample | human | 0.0000 | 1267253567c0… |
| carefree-64k.mp3 | ffmpeg transcode of a batch sample | human | 0.0000 | 22de7648cb91… |
| ja1.m4a | Generated by site author on Suno (2026-09) | ai | 1.0000 | b0ab1c06adeb… |
| ja2.m4a | Generated by site author on Suno (2026-09) | ai | 1.0000 | be845b0e160c… |
| ja3.m4a | Generated by site author on Suno (2026-09) | ai | 1.0000 | 107f83900ccd… |
| zh1.m4a | Generated by site author on Suno (2026-09) | ai | 1.0000 | c05768ba0906… |
| zh2.m4a | Generated by site author on Suno (2026-09) | ai | 0.9961 | b022cff8ac8c… |
| zh3.m4a | Generated by site author on Suno (2026-09) | ai | 1.0000 | 375791744964… |
| zh1-192k.mp3 | ffmpeg transcode of a batch sample | ai | 1.0000 | 0e13a15912f7… |
| zh1-128k.mp3 | ffmpeg transcode of a batch sample | ai | 1.0000 | 67cf304396f8… |
| zh1-96k.mp3 | ffmpeg transcode of a batch sample | ai | 1.0000 | 9c6c70abffab… |
| zh1-64k.mp3 | ffmpeg transcode of a batch sample | ai | 0.9993 | 75a09c2238a7… |
| ja1-64k.mp3 | ffmpeg transcode of a batch sample | ai | 0.9983 | 652b5a1a606b… |
| en1-v6mini.mp3 | Suno share page (V6-MINI) | ai | 0.9999 | e412c00d3ec6… |
Additional sources: successes and failures
The 76-file results above do not describe general accuracy. In the AIME challenge, 24 AI source files plus two transcodes each produce 72 files: 12 AI detections, 38 inconclusive results and 22 false-human results. These cover Udio, MusicGen Small, Suno v3 and Mustango. None entered fitting, but their results were visible during later policy iterations. Exact generator revisions and upstream training overlap are unknown.
Eight additional human works from four authors exposed failed candidates and became development controls. After the final policy froze, four works from four new human authors yielded 12 original/transcoded files: zero observed AI false positives, three human results and nine inconclusive results. These are four new works, not 12 independent songs; the poor human-transcode coverage remains a limitation.
Download failed candidates, sources, training groups and every additional result (JSON)
Suno v6 testing — historical signal, current limits
v6 is newer than the engine's official coverage list, so we tested it ourselves instead of claiming it. Result: a fresh V6-MINI track (English vocals, retained from a user capture) scored P(AI) = 0.9999 — and still 0.9999 after the double transcode described above. The same sample scored 0.99990 through the earlier production API. Today that retained 128 kbps file again receives an AI signal after removing the blanket bitrate rejection; the nine-file v6-MINI source group is listed in the archived v2.2 table.
Status: raw-model signal observed; current screening coverage remains limited.
v6 ships in several variants. Nine source files from a development collection do not establish broad coverage. Every new Suno generation and variant gets the same published treatment as it lands. Treat coverage claims without published tests, from anyone including us, as marketing.
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.
Comparison study: Suno's official C2PA checker vs acoustic analysis
Suno's official checker (August 2026) reads the C2PA content credentials embedded in new downloads — metadata only, as Suno itself notes. We put both tools through the same six-track gauntlet: fresh Suno v6 downloads, re-encoded copies, a stream-captured track, tampered variants, and human music. Tested October 2026 — every number below is from files anyone can rebuild with free tools.
| File | Suno official checker | Sonicert acoustic engine |
|---|---|---|
| Fresh Suno v6 downloads | Verified Suno — label, content ID, model version | 0.996–0.9999 AI likelihood |
| Same audio after one re-encode | No Suno provenance — the label is gone | 0.93 AI likelihood — the sound remains |
| Stream-captured track, as shared on social platforms | No Suno provenance | 0.9999 AI likelihood |
| Metadata markers copied into a clean file | Correctly rejected — not a valid label | Verdict unchanged |
| Genuine label on tampered audio | Fails validation | Verdict unchanged |
| Human-produced track | No Suno provenance | ≈0 — no false positive |
Reading the table: the official checker is authoritative while a file stays untouched — and blind the moment the label is stripped, which is what re-encodes, platform re-uploads and DAW exports do. The two methods answer different questions: the label states a file's declared origin; the engine measures the audio itself.
That is why Sonicert does both. A validated Suno credential surfaces as its own provenance signal on the Suno detector page — and underneath it, the acoustic score works on any track from any source, label or no label. Independent of every generator, free with no signup, audio deleted the moment analysis finishes: run your first track.
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