demorphus
Upload the final song when you need a result for your release, then check how much audio the service actually analyzed. A sample can help investigate a passage, but its result should stay attached to that excerpt. A low score on one chorus does not establish the result for the rest of the track.
A full upload is not always a full-length analysis
SH Labs documents a 300-second analysis limit. Its Spotify and ISRC lookups can use full audio or a roughly 30-second preview depending on access and availability; the audio_source field identifies that choice. Direct audio URLs use the supplied audio, subject to analysis limits. Check the response rather than assuming the song’s duration is the analyzed duration.
For a long track, record any uncovered portion explicitly. If you test additional passages, keep each report and timestamp separate. Do not present one excerpt’s result as a measurement of a whole album or extended mix.
When a sample is useful
Use a sample when the question is local: a replaced vocal phrase, an AI-generated instrumental section or a specific transition. Choose the passage before reading its result, and write down why you chose it. Keep the same start and end points when comparing original and processed versions.
For a Suno project, Studio offers Full Song, Selected Time Range and Multitrack exports. A selected range answers a different question from the complete mix; stems change the material again. Label these exports so a clip or isolated vocal cannot be mistaken for the full song.
Do not average excerpts into a new whole-song score
Separate excerpt predictions are not automatically components of a validated song-level calculation. A detector may combine evidence in a way your arithmetic average does not reproduce. SH Labs specifically notes that its optional spectral chunk values do not average to the overall AI probability.
Keep a list of passage-level findings instead: which file, which timestamps, which score and which verdict. A disagreement between sections is useful to investigate. Selecting only the most favorable section hides that information.
Use a repeatable sampling record
| Field | What to save |
|---|---|
| Parent file | Name and hash of the complete export |
| Passage | Start and end timestamps, or complete upload |
| Reason for selection | For example, a changed vocal or a dense chorus |
| Submitted audio | Format, duration and whether any effects were applied |
| Analysis coverage | Reported length or limitation; unknown if absent |
| Result | Service, date, version if provided, score and verdict |
Keep the original reference unchanged. For investigating an edit, export the matching passage from both versions using the same settings. This procedure is a way to organize evidence; it does not claim that a particular passage length guarantees accuracy.
What to check before sharing a result
Put the scope beside the claim: “this excerpt received this verdict,” or “this final export was submitted, with this analysis limit.” Keep the report with the exact file. If the displayed information does not state coverage, do not fill the gap with an assumption.
Use the Demorphus checker for your selected export and retain its report. When comparing another service, follow the detector disagreement checklist first. For the full experimental record, see our testing methodology.
Sources reviewed October 30, 2026. The comparison procedures are suggested workflows, not new benchmark results.
Prepared with AI assistance. Sources below.