AI MP3 to MIDI: A Same-Input Test Against Pitch Detection
AI can turn an MP3 into MIDI, but the label alone does not tell you whether its notes are cleaner. In one same-input test of a 192 kb/s MP3, an open-source AI model found all four intended pitches but added a second G4 event; this site's pitch detector returned four events but started the last three about 128–136 ms early. For this one synthetic melody, the AI model's four intended onsets averaged about 5 ms absolute error, while the duplicate note still needed removal. This is a single controlled example, not a ranking or a song-accuracy score.
If you are deciding whether to try AI-based transcription, the practical answer is: test a short phrase with known notes, then count both timing errors and extra events. A MIDI download is not proof that the notes match your recording.
What went into the comparison
The source was a newly generated 2.40-second, mono, 44.1 kHz four-note sine melody. Its MIDI pitches were C4 (60), E4 (64), G4 (67), and C5 (72), with intended starts at 0.000, 0.600, 1.200, and 1.800 seconds; each tone lasted 0.450 seconds. The source was encoded as a mono 192 kb/s MP3 with FFmpeg. Its SHA-256 was d8d525f87cc7dda8a062ba6b297e153e1f5886de7594f7f35c86026d7fc30c09.
Both methods received PCM decoded from that same MP3. The site's unchanged detector functions were run at Balanced sensitivity after their 11,025 Hz downsampling path. The AI model was Spotify's open-source JavaScript implementation, package version 1.0.1, using its CPU backend and documented inference functions after averaging the same decoded samples to 22,050 Hz. Its note extraction used thresholds 0.25 / 0.25 and a five-frame minimum. We did not upload the sample, manually edit either result, or tune thresholds to improve the comparison.
The project maintainers describe their model as machine-learning-based and discuss polyphony and pitch bends in their official technical introduction. This test used only straight, single-note sine tones, so it does not test those capabilities.
Measured note starts
To inspect the exact input and model output, download the 192 kb/s test MP3, the generated AI-model MIDI result, and the machine-readable measurements.
| Expected event | Target start | Pitch detector | AI model |
|---|---|---|---|
| C4 · MIDI 60 | 0.000 s | 0.000 s | 0.000 s |
| E4 · MIDI 64 | 0.600 s | 0.464 s (−0.136 s) | 0.604 s (+0.004 s) |
| G4 · MIDI 67 | 1.200 s | 1.068 s (−0.132 s) | 1.196 s (−0.004 s) |
| C5 · MIDI 72 | 1.800 s | 1.672 s (−0.128 s) | 1.811 s (+0.011 s) |
The AI output also contained a second G4 event at 1.045–1.196 seconds, before the intended G4 at 1.196 seconds. That made five events for four source tones. The pitch detector returned four events and no additional pitch in this run, but its last three starts were early. Averaging the absolute onset differences for the four intended events gives approximately 99 ms for the pitch detector and 5 ms for the AI output; the extra G4 is counted separately, not hidden inside that average.
So neither output was simply “correct” or “wrong.” One had fewer events but larger onset shifts; the other placed the intended starts closer to the targets but included an extra event. The useful comparison is the edit each result would need for the task—not the method label.
How to judge your own AI-to-MIDI result
- Use a short passage whose notes you can verify. Keep the original audio and a known pitch/onset list. A long finished song makes it harder to find the first error.
- Count intended, missing, and extra notes separately. A model may recover simultaneous notes yet add fragments; one note per source tone is not a sufficient score by itself.
- Compare each onset and release with the audio. In this fixture the AI output's starts were close, but the duplicate G4 would still need review. Check note lengths as well as pitches.
- Try the same short source through the tool you plan to use. Models, preprocessing, and settings differ. This comparison says nothing about every AI converter or a real vocal, piano, guitar, chord, or full mix.
- Keep the draft only if its edits are useful. Remove false notes, restore missing ones, correct timing, and compare the work with entering the phrase by hand.
Limits that matter
- This was one deliberately simple, synthetic monophonic melody—not a human performance, chord, or full song.
- The browser file picker and Web Audio decoder were not measured. The site's production detector functions ran locally after FFmpeg decoded the MP3.
- The two pipelines used different sample rates after decoding: 11,025 Hz for the site detector and 22,050 Hz for the AI implementation. Both started from the same MP3-derived PCM, but this is not a controlled study of model architecture alone.
- The result does not establish that AI is generally more accurate, faster, safer, or better for a particular instrument.
For the site's transparent non-model pitch-detection path, see the non-AI transcription explanation. For general source preparation and cleanup, use the MP3-to-MIDI guide; after conversion, the MIDI export check helps separate a file-format problem from a note-detection problem.