Best audio to MIDI converter

Jukeblocks and OpenMusic received the exact two original WAVs used for the Midilot checks below: the eight-note melody and the repeated C-major chord fixture. Each file is 4.64 seconds, mono PCM, at 44,100 samples per second. These are generated sine tones, not piano, singing or a full mix. We left the converters’ available settings at their defaults and used their file and download controls. The four newly saved MIDI files were inspected independently with mido.

Jukeblocks’ melody download is a 1,075-byte format-1 MIDI with two tracks and 220 ticks per quarter note. OpenMusic’s is a 135-byte format-1 MIDI with two tracks and 480 ticks per quarter note. Both contain the programmed eight-pitch sequence without extra note events. The timing rows above are computed from the saved event starts, not the playback animation or a vendor accuracy claim. Different file sizes and resolutions are properties of those exports, not quality scores.

The chord downloads contain 24 events each. Independent parsing found the unique pitches MIDI 60, 64 and 67, with a maximum overlap of three sounding notes. Jukeblocks’ file has 266 bytes; OpenMusic’s has 262. That is a materially different result from Midilot’s single C2 event on the same input. It supports trying either converter for this simple synthetic chord case. It does not demonstrate separation of instruments, preservation of a real piano performance or complete recovery of a finished song.

For each melody export we paired the eight same-pitch events in sequence and calculated sum(|exported start − programmed start|) / 8. Jukeblocks gives 0.039090909… / 8 ≈ 0.004886 seconds; OpenMusic gives 0.030416666… / 8 ≈ 0.003802 seconds. Every programmed start and exact exported start is retained in the event record below. These are mean absolute onset errors for one generated melody. The displayed means are rounded to six decimal places; they are not probabilities, full-song accuracy percentages or a general tool ranking.

Eight correct pitch labels can still need timing edits

Our reference melody contains the MIDI pitches 60, 62, 64, 67, 69, 67, 64 and 60, in that order. Each programmed tone lasts 0.42 seconds, then leaves 0.16 seconds of silence. The onset spacing is therefore 0.42 + 0.16 = 0.58 seconds. Eight such slots make a 4.64-second file. The original WAV generator and its event definition are linked below, so the intended starts do not depend on a listener guessing where each note began.

The saved output had 119 bytes and eight note events with that same pitch sequence. This passes a narrow pitch-sequence check for this particular input. It does not pass an exact timing check. The first exported note ends at approximately 0.417708 seconds, and the next starts at that same time. Its programmed start was 0.58 seconds. The second event is therefore early by 0.417708 − 0.580000 ≈ −0.162292 seconds.

The table uses the independently parsed event starts rather than the page’s two-decimal display. For each event we matched the same pitch in the same sequence, then calculated exported start minus programmed start. We rounded only the displayed columns to three decimals; the downloadable record keeps full precision. All events after the first begin early in this example. Counting eight matching pitch labels would conceal that editing work.

The mean absolute onset error is sum(|exported start − programmed start|) / 8. Using all eight exact pairs in the raw record gives 1.123333… / 8 = 0.1404166… seconds, or approximately 0.140417 seconds. The numerator is the sum of the eight absolute differences, not the total audio length. This is a calculation for these matched synthetic events, not a percentage accuracy score or a result for another converter.

A useful acceptance decision follows directly from this file: if you only need the pitch sequence as a starting draft, the result provides it; if you need the programmed rests and entrances preserved, this output still needs correction. The late part of the file does not become correct merely because the MIDI opens successfully. Compare the event list with the supplied source definition before quantizing or replacing the instrument.

A valid file can be the wrong output for a chord task

The other original WAV places MIDI pitches 60, 64 and 67 together during each sounding section. It has the same 4.64-second file length as the melody fixture. In this run Midilot produced a valid 56-byte MIDI file containing one note, MIDI 36, labeled C2, from 0 to approximately 4.458333 seconds. The three intended simultaneous pitches were absent from the output. This is a concrete failure for this chord input, even though the interface reported that the MIDI was ready.

That observation changes the shortlist for an overlapping-note task: our single-line converter is unsuitable for recovering this chord. It does not establish that any competing tool will recover it correctly. Use the downloadable chord WAV as an additional check when evaluating a tool that advertises polyphonic transcription. A candidate needs to return the simultaneous notes you require; one readable track or an attractive piano roll is not enough evidence.

Separate exported settings from detected performance

Both saved files use MIDI format 0, one track and 480 ticks per quarter note. Their tempo event is 500,000 microseconds per beat, so 60,000,000 / 500,000 = 120 BPM. Every note-on in these outputs has velocity 88. Those values match Midilot’s fixed exporter settings. They do not show that the original recording was performed at 120 BPM or with constant loudness. An instrument change can alter playback sound, but cannot add the missing chord events or restore the programmed silent gaps.

The first Jukeblocks attempt returned a busy message and no MIDI. A later follow-up on the same date produced native MIDI from both original WAVs. The current results are in the comparison table above. The original busy response remains part of the record; it is neither the final conversion outcome nor a detector-accuracy measurement.