Does MP3 Compression Affect AI Music Detection?
In this dated 15-pair test, converting 30-second clips from 192 kbps stereo MP3 to 128 kbps mono did not change any detector verdict. Probability scores still moved, so this result is evidence about one compression setting—not proof that every edit is harmless.
0 of 15
Verdicts changed after compression
1.3 points
Mean absolute probability change
6 points
Largest observed probability change
What we tested
We selected five Human, five Suno, and five Udio clips from the August 9, 2026 blind benchmark. Each original was a metadata-stripped, 30-second, 192 kbps stereo MP3. We created a paired 128 kbps mono version, submitted both versions to ACRCloud model h2zt2l57, and compared the returned verdict and AI probability.
Paired results
| Sample | Known source | Original AI probability | 128 kbps mono | Change | Verdict changed? |
|---|---|---|---|---|---|
| B001 | Human | 23% | 22% | -1 | No |
| B002 | Suno | 90% | 90% | 0 | No |
| B003 | Udio | 83% | 87% | +4 | No |
| B004 | Human | 24% | 23% | -1 | No |
| B005 | Suno | 90% | 89% | -1 | No |
| B006 | Udio | 88% | 86% | -2 | No |
| B007 | Human | 18% | 19% | +1 | No |
| B008 | Suno | 91% | 91% | 0 | No |
| B009 | Udio | 80% | 86% | +6 | No |
| B010 | Human | 18% | 18% | 0 | No |
| B011 | Suno | 91% | 91% | 0 | No |
| B012 | Udio | 85% | 86% | +1 | No |
| B013 | Human | 24% | 26% | +2 | No |
| B014 | Suno | 91% | 91% | 0 | No |
| B015 | Udio | 85% | 85% | 0 | No |
What the result means
All 15 classifications remained on the same side of the provider's decision boundary. The average score moved by 1.3 percentage points in absolute terms, and the largest move was 6 points. Compression therefore changed confidence more than classification in this small test.
What this test does not cover
- Other bitrates, codecs, clipping lengths, remasters, stems, noise, pitch changes, or time stretching.
- Every genre, language, generator version, or production workflow.
- Future versions of the provider model.
Read the complete 60-track benchmark, inspect the downloadable JSON data, or review our accuracy methodology before applying the finding broadly.