MP3 TO MIDI · REPRODUCIBLE CHECK
MP3 to MIDI: input, output, limits, and cleanup.
Use a clear, isolated melody as the input, inspect the returned MIDI note events, treat short or silent material as a failure case, then clean timing and pitch in a MIDI editor. The four sections below use this site's recorded detector runs rather than a generic accuracy claim.
The evidence was run with the deployed detector in Node using deterministic mono PCM at 11025 Hz and sixteen-bit depth. Browser decoding and resampling were not measured, so this is a reproducible detector check, not a real-recording accuracy benchmark.
Input conditions
The WAV fixture used four separated notes at MIDI pitches 60, 64, 67, and 72. The tested run used a 11025 Hz mono signal, sixteen-bit depth, a minimum note duration of 100 ms, and a 2.8 second fixture. The recorded method and source hash are kept in the evidence record.
Output result
The detector returned four output notes for the four-note fixture: pitches 60, 64, 67, and 72. Their measured output windows were 0.0798–0.6585 s, 0.7184–1.2971 s, 1.3769–1.9556 s, and 2.0154–2.6141 s. The run status was “Conversion complete. Adjust the controls, preview, then download.”
Limits shown by the same run
The silence fixture returned no output notes and the status “No clear notes found. Try a louder, isolated melody or lower confidence.” The simultaneous-note fixture contained pitches 60 and 67 but returned one output note at pitch 48, so overlapping material can produce a result that needs inspection. Short notes also remain a boundary: the recorded run used a 100 ms minimum-note setting.
Post-processing
Open the MIDI in an editor, compare each note with the source, correct octave or timing errors, and remove false notes before assigning an instrument. The local MIDI file player can play the exported file in the browser; the MP3 to MIDI converter is the starting point for a new input.
Questions
Does a clean fixture prove real-song accuracy?
No. The evidence method explicitly says synthetic tests are not real-recording accuracy benchmarks.
What should I do when the output is wrong?
Keep the source beside the MIDI, inspect pitch and timing, and edit the returned note events before using them in a DAW.