MP3 to MIDI: When the Reverb Tail Outlasts the Note
A converter can only close one note and open the next where the audio leaves it a gap, and reverb is what fills that gap. One ratio decides whether note boundaries survive: how long the tail rings against how long the note lasts.
The ratio, in one box
- A tail is a decay rate, not a length. Its published measure is 60 dB of decay, so a 1.5-second tail falls 40 dB per second.
- At 120 BPM a sixteenth note lasts 125 ms. A 1.5-second tail spans twelve of them, and has fallen 5 dB when the next note starts.
- The file can place a boundary to one tick — 5.21 ms at 96 ppq and 120 BPM — 288 times finer than the blur it is trying to resolve.
- No edit restores a boundary the detector never found, because the tail was never in the MIDI file to begin with.
What the detector needs, and what reverb takes away
songscription describes the hard case as a model that has to "handle the partial overlaps when one note rings into the next". ai-midi warns that accuracy may be lower with "strong reverb". jukeblocks puts it on the free browser tool as "Polyphonic or very busy mixes may need note editing after import". All three name reverb; none puts a number on it.
The grid: how long a note lasts
A note value is a fixed fraction of the beat, and the beat is 60 seconds divided by the tempo.
- 60 BPMbeat 1,000 ms · eighth 500 ms · sixteenth 250 ms
- 90 BPMbeat 666.67 ms · eighth 333.33 ms · sixteenth 166.67 ms
- 120 BPMbeat 500 ms · eighth 250 ms · sixteenth 125 ms
- 140 BPMbeat 428.57 ms · eighth 214.29 ms · sixteenth 107.14 ms
- 180 BPMbeat 333.33 ms · eighth 166.67 ms · sixteenth 83.33 ms
The tail: a rate, not a length
The standard measure of a reverb tail is RT60: the time for the level to fall 60 dB after the source stops. Read as a rate, 60 divided by the tail gives decibels per second.
- 0.5-second tail120 dB per second · 15 dB in 125 ms
- 1.0-second tail60 dB per second · 7.5 dB in 125 ms
- 1.5-second tail40 dB per second · 5 dB in 125 ms
- 2.0-second tail30 dB per second · 3.75 dB in 125 ms
Set against the grid, a 1.5-second tail is twelve sixteenth notes long at 120 BPM. Only slower, longer values outlast it: at 60 BPM a half note runs 2,000 ms and clears it.
Why the file cannot be the limit
MIDI places note-off on a tick grid, and the division field that sets it is 15 bits wide, addressing up to 32,767 pulses per quarter note. Ableton exports at 96 ppq, where one tick at 120 BPM is 5.21 ms and a sixteenth note is 24 ticks. The tail above occupies 288 ticks. Quantising moves boundaries that exist; it cannot create one.
The lever that does work
It sits on the source. jukeblocks lists fine transient detail among what MP3 encoding removes, and the transient is the cue a detector uses to find an onset, so a lossy, reverberant file attacks the same cue twice. Given the multitrack, feed the driest stem; given only the mix, budget for moving boundaries by hand rather than for a better converter.
Questions
- Can I fix blurred note boundaries by quantising?
- No. Quantising snaps boundaries that already exist onto a grid, and at 96 ppq the grid step is 5.21 ms against a blur of 288 steps.
- Do longer notes survive a tail better?
- Yes. At 120 BPM a quarter note is 500 ms against a sixteenth's 125 ms, so one tail covers three of them instead of twelve.
The decay rule is the standard RT60 definition (60 dB, ISO 3382-1). Note lengths, decay rates and tick counts are arithmetic on that and on the published 96 ppq export resolution; tail lengths are stated as parameters, not measurements.