Best MP3 to MIDI Converter? A Two-Input Output Test
There is no single best MP3-to-MIDI converter for every source; choose by the notes your own input needs. In our two synthetic tests, this site's detector matched 5 of 5 target pitches on a solo melody but only 3 of 9 notes across three-note chords. An open-source AI model matched 5 of 5 and 6 of 9, respectively, but also produced 5 and 7 unmatched extra events. The chord test favored the model for pitch coverage, not for a clean, ready-to-use MIDI; neither result is a general ranking.
“Best” depends on whether your file is one clear melody or several simultaneous notes—and on how much correction you are willing to do. These controlled examples make that trade-off visible. They are not recorded songs, a survey of every online service, or a claim about current commercial converters.
The two-input result
Both methods received the same 192 kb/s MP3 in each row. We matched an output event to a target only when the MIDI pitch was identical and its onset was within 0.35 seconds; remaining target notes count as missed, and unmatched events count as extras. Mean onset error uses matched notes only.
| Input | Site detector | Open-source AI model | Practical read |
|---|---|---|---|
| 5-note solo melody | 5/5 matched; 0 missed; 0 extras; 160 ms mean onset error | 5/5 matched; 0 missed; 5 extras; 11 ms mean onset error | Both found the pitches. The model's closer starts came with one short extra event after each target note. |
| Three sequential 3-note chords | 3/9 matched; 6 missed; 0 extras; 151 ms mean onset error | 6/9 matched; 3 missed; 7 extras; 15 ms mean onset error | The model captured more chord tones here, but its MIDI still needs event cleanup. |
For the chord input, the site detector returned only the three low notes—MIDI pitches 48, 50, and 53. The model matched those plus pitches 67, 69, and 72; it missed the middle pitches 60, 62, and 65 and split several detected notes into extra events. A lower onset-error average does not cancel missing notes or fragments.
Download the inputs and MIDI outputs
- Solo melody: download the 192 kb/s MP3 input, the site-detector MIDI, or the model MIDI.
- Three-note chords: download the 192 kb/s MP3 input, the site-detector MIDI, or the model MIDI.
- The machine-readable measurements include the target note lists, detected events, hashes, settings, and matching rule.
How we ran the test
On October 1, 2026, this Windows machine generated two four-second mono WAVs as 16-bit PCM. The solo input contains five non-overlapping sine notes at MIDI pitches 60, 64, 67, 72, and 69. The chord input contains three successive triads, nine target notes total. Each tone has a 25 ms attack and 45 ms release. FFmpeg 20181217-f22fcd4 encoded each input as a 192 kb/s MP3 and decoded that same MP3 to 44.1 kHz mono float PCM before both analysis paths.
The site's unmodified detection and MIDI-writing functions ran in Balanced mode after their 11,025 Hz downsampling step. The second path used the official Basic Pitch open-source implementation, JavaScript package 1.0.1, with TensorFlow.js CPU inference, 22,050 Hz input, onset/frame thresholds of 0.25, and a five-frame minimum. Both used the same decoded MP3 for a given row; neither file was manually repaired before measurement.
To avoid hiding errors inside one score, we report target pitches matched, targets missed, unmatched output events, and the average absolute onset error among matched notes. Matching was greedy by nearest onset for the same pitch, with a 0.35-second maximum difference. This window is an analysis choice, not a universal musical tolerance.
Choose a converter with a short acceptance test
- Start with the audio you actually have. A clear, isolated melody and a chord-heavy passage are different tests. Do not infer full-song performance from a four-second synthetic signal.
- Keep a known reference. For a short phrase, write down the pitches and approximate starts before converting. Without a reference, extra notes can look plausible in a piano roll.
- Count omissions and additions separately. In these examples, the model captured more chord pitches but also created seven extra events; the site path produced fewer events but omitted six chord tones.
- Listen before editing. Check whether the source really contains each MIDI event, then decide whether correcting the output takes less effort than entering the notes manually.
- Choose by source type, not by one leaderboard label. Test a representative crop from your own recording and keep the input, output, tool version, and settings so the choice is repeatable.
What this comparison cannot tell you
- These are two deliberately generated sine fixtures, not vocals, piano recordings, guitar, drums, noisy audio, or commercial songs.
- Only two analysis paths were tested; this is not a ranking of all MP3-to-MIDI converters.
- The test calls the site's detector functions after FFmpeg decoding. It does not measure the browser file picker or browser audio decoder.
- Different preprocessing and note-extraction rules are part of each workflow; this does not isolate model architecture alone.
- We did not time manual correction, test a real user's editing workflow, or measure upload privacy, price, or service limits.
For a deeper look at one AI-versus-pitch-detection input, see the single-input comparison. If your file is a clear voice melody, read what voice-to-MIDI preserves; for general input preparation and note cleanup, use the MP3-to-MIDI guide.