MP3 to MIDI Accuracy: Four Converters, Measured

On 3 November 2026 we ran the same six test clips through four MP3 to MIDI converters and counted three things per clip: how many reference notes were matched, how many notes appeared that no reference note explains, and how many reference notes never appeared in the output.

All four converters found 80 of 80 reference notes. The extra-note count is where they separate: 1 (Stayput), 43 (SoundToMIDI), 43 (Spotify's Basic Pitch demo), 1 (this site, after the change described below). None of them missed a reference note in this set.

Two converters show identical numbers in all six rows. That is not a coincidence, and the section on model identity below says why.

How to read the three numbers

• Matched — a reference note that has an output note of the same pitch starting within 0.15 seconds of it. Each output note can match at most one reference note, and each reference note can be matched at most once.
• Extra — an output note that matched no reference note. This is the number that separates the four converters above. An extra note is audible as a wrong note even when every real note was found.
• Missed — a reference note that no output note matched. All four scored zero here, so this column does not separate anything in this test.

How the six clips were built

The clips are synthetic, not recordings. Each was rendered from a written note list, so the reference notes are exact rather than estimated from a score. All six share the same eight pitches — MIDI 60, 62, 64, 65, 67, 69, 71 and 72 — one event every 0.5 seconds starting at 0.3 seconds, each note lasting 0.14 seconds. What changes between them is what sounds at those moments:

• Single notes — one pitch at a time, eight notes in sequence.
• Three-note chords — each of the eight pitches sounds as a major triad (root, +4 semitones, +7 semitones), three notes struck together, eight times, 24 notes total.
• Drum pattern — the same eight pitches, each rendered as a short pitched hit with a 0.09 second decaying noise burst layered on top.

Each of those three appears twice: dry, and with echoes — three delayed copies at 0.2, 0.32 and 0.41 seconds with gains of 0.53, 0.22 and 0.12. The files are 128 kbps mono MP3 at 44.1 kHz, five seconds long, 120 BPM.

Every rendered tone is a fundamental plus a second harmonic at 0.2 times the fundamental's amplitude. That figure is a setting in the generator, not something measured from the converters, and it matters here: a harmonic that a converter promotes into its own note lands in the extra column.

How the output was counted

Every downloaded .mid file was read by the same parser and compared against the same reference list in the same run, with no manual editing. The matching rule is pitch equality plus onset within 0.15 seconds, one-to-one. The 0.15 second tolerance is a choice we made; a tighter or looser window would move the numbers, and these figures are only comparable between sites measured under this same rule.

One caveat on the drum rows. A drum hit is a noise burst, not a pitched note, so the eight reference pitches for those clips are the pitches the generator was told to use, not pitches a listener would hear. A match there means the converter placed a note event at that pitch and time, not that it identified a pitch by ear. Read the drum rows as event placement, not as tuning accuracy.

Why two of the four columns are identical

SoundToMIDI and Spotify's Basic Pitch demo returned the same six cells, and their model files are byte-identical:

• Both serve a weights file of 742,392 bytes with SHA-256 b142a95737a52e1e412d5f92e73d8bb80dfe8d04941acc0702f11f4524fb377c.
• That same hash is what the published @spotify/basic-pitch npm package ships, and the same weights are embedded in this site's own page.
• Basic Pitch is published by Spotify under the Apache-2.0 licence.

So three of the four converters in this table run the same open-source model. What differs is the thresholds and the post-processing each site applies to that model's output, not the note detection itself. We did not identify Stayput's engine; that column is a black box in this test and we make no claim about what it runs.

What changed on this site, with the numbers

Before 3 November 2026 this site used an onset threshold of 0.25, a frame threshold of 0.25 and a minimum note length of 8 frames, with no merging step. On these six clips that produced 79 matched, 83 extra, 1 missed. The 83 was not spread evenly: 49 came from the three-note chords with echoes, and 16 each from the single notes and the drum pattern with echoes.

Two changes were tested separately on the same six clips:

• Thresholds only, moved to the Basic Pitch defaults of 0.5 / 0.3 / 5: 80 matched, 91 extra, 0 missed. Matching improved by one note and the missed note disappeared, but the extra count got worse, not better.
• Merging only, on top of those default thresholds: 80 matched, 1 extra, 0 missed.

The merging step is the one that mattered. The model often splits a single sounding note into several fragments of the same pitch separated by a few frames. The step joins any two fragments of the same pitch whose gap is 3 frames or less. At 22,050 Hz with a 256-sample hop, 3 frames is about 35 milliseconds. The site now runs the default thresholds plus this merge, which is the configuration in the right-hand column above.

The cost of merging, stated plainly

What this test does not cover

• Synthetic clips, not recordings. Six clips and 80 reference notes is a small, controlled set. It isolates one behaviour — fragmenting — and says little about a full mix.
• Note count only. Pitch bend, velocity and note duration accuracy were not scored. A converter can win here and still sound wrong.
• One snapshot. These are the results on 3 November 2026. Sites change their models and settings; the numbers describe that day, not a permanent ranking.
• Stayput's engine is unidentified, so the reason for its score is unknown to us. Only its output was measured.
• One run per clip. No repeat runs, so run-to-run variation is not characterised.

Checking this yourself

The method needs no special access: take audio whose notes you already know, convert it on any site, download the .mid, and match output notes to reference notes by pitch plus onset within a fixed window. Report all three counts — matched, extra and missed — because a converter that finds every note can still be the one adding the most wrong ones. We do not host the test audio or the resulting MIDI files on this page.

To convert your own file with the current settings, use the MP3 to MIDI converter, the WAV converter, or the piano converter. The mixed-song test shows what happens on a full recording rather than a controlled clip.