What survives an MP3: compression and AI music detection
Nearly every file we get asked about has passed through a lossy encoder at least once. This page preserves a transformation study performed on a retired browser research build and separates those findings from the current third-party classifier.
Archived design · measurements published 24 August 2026 · historical build ensemble-1
Archived findings (historical build ensemble-1)
The study ran 168 detections across 12 reference tracks and the full encode ladder. These values show movement in that retired model’s numeric output, not accuracy and not current-provider behavior.
- Mean shift vs lossless: 2.4 pp — 1.8 pp at 320 kbps, 1.8 pp at 192 kbps, 2.3 pp at 128 kbps, 3.5 pp at 64 kbps.
- Worst single case: 26 pp, at 64 kbps. Averages hide individual files that move a long way.
- Mono downmix: 1.2 pp mean shift (worst 11 pp), even though stereo-field features are then unavailable.
- 10-second excerpt: 4.4 pp mean shift, up to 31 pp — a bigger effect than any bitrate in the ladder.
- Repeat determinism: 0.0 pp — the historical build showed no run-to-run variation in this sample, so every shift above is attributable to the transformation.
The full protocol, caveats, and machine-readable copy live on the benchmark page. One surprise: clip duration moved the historical numeric output more than bitrate did — we did not predict that, and the table below is left as originally written rather than adjusted after the fact.
Why transformation must be recorded
Popular explanations of AI music detection lean hard on the frequency spectrum: look for a hard ceiling, unnaturally smooth high-frequency texture, missing "air." Every one of those cues sits in exactly the band a lossy encoder attacks first.
An encoder’s psychoacoustic model spends its bit budget where human hearing is sharpest and discards whatever it decides you will not notice. The trouble is that “what you will not notice” and “where synthesis artefacts live” overlap almost completely. So a 128 kbps MP3 of a human orchestra and a 128 kbps MP3 of a generated track can end up looking more alike than either does compared to its own lossless original.
The archived study confirms that transformation can move a model output. It does not rank compression against other error sources for the current provider.
Archived expectations
The table below preserves expectations recorded for the historical study. Labels such as “reliable” describe that proposal and must not be read as current-provider guarantees.
Swipe the table sideways to see all columns.
| Source | Ceiling | Spectral features | Dynamics features | Expected verdict |
|---|---|---|---|---|
| Lossless (WAV / FLAC) | Source-limited | Intact | Intact | Full evidence available |
| 320 kbps | ~20 kHz | Largely intact | Intact | Reliable |
| 192 kbps | ~18 kHz | Degraded | Intact | Usable, reduced confidence |
| 128 kbps | ~16 kHz | Unreliable | Intact | Spectral evidence discounted |
| 96 kbps and below | ~15 kHz or lower | Destroyed | Mostly intact | Structure and dynamics only |
| Re-recorded (speaker to mic) | Room and device limited | Destroyed | Altered | Not analysable |
Study design
The completed historical arm used twelve procedural references rendered into a fixed transformation ladder and evaluated twice on the same retired build. The planned labelled accuracy arm described here has not been run.
- Fixed corpus. Identical recordings across every condition, so any change in the reading traces back to the encoder and nothing else.
- Output preservation. Historical component values were logged so the archived build could be studied without reducing every change to one headline.
- Confidence review. A future current-provider study must test whether confidence tracks observed outcomes instead of assuming that it does.
- Second-generation encoding. A separate arm re-encodes files that are already lossy, because that is exactly what happens when audio hops between platforms.
What this study cannot say about the current provider
The current provider does not disclose its internal feature weighting here. The points below are therefore interpretation cautions, not claims about its implementation:
- The spectral ceiling can be created by codecs, sample rates or deliberate filtering, so it cannot establish origin.
- Evidence of codec artefacts document transformation history; compression alone is not evidence of generation.
- Segment agreement may help describe provider-returned windows, but this application does not recompute or reweight them.
- Dynamics measurements such as crest factor mostly survive lossy encoding, but ordinary mastering can mimic them just as easily, so they can never carry a reading on their own.
What to do with a compressed file
Track down a better copy if one exists — a purchased download, a distributor file, the original export from whoever made it. Failing that, analyse what you have, but read the categorical outcome, component indicators and confidence separately. And do not bother converting a lossy file to WAV expecting an improvement; that changes the container, not what is inside it.
Keep the bigger picture in view, too. Provenance evidence — project files, upload history, someone who can walk through their own arrangement — outweighs every acoustic measurement on this page, regardless of bitrate.
Questions
Lossy encoding removes and reshapes information. In the archived browser-model study, transformed copies produced different numeric outputs from their lossless references. That does not reveal which features the current provider uses or guarantee a lossless result is correct.