Why MP3 to MIDI loses notes that WAV keeps

The short answer: an MP3 does not lose your high notes — it costs you the low ones. Pitch detection measures frequency in fixed slices, and at the analysis settings this converter uses one slice is 10.77 Hz wide. A semitone is only 5.95% of a frequency, so below 181 Hz (about F#3) two neighbouring semitones fall inside the same slice. The only thing separating them is an interpolation that reads the relative loudness of nearby slices. MP3 coding noise disturbs exactly that measurement, which is why an MP3 and a WAV of the same take can come back with different note names in the bass and identical ones up top.

What a pitch detector actually measures

Before talking about codecs, it helps to know what the conversion is doing. This converter resamples your audio to 22 050 Hz, then reads it with a 2048-point FFT stepped forward 256 samples at a time. Two numbers follow directly from that:

• Each FFT bin is 22 050 ÷ 2048 = 10.7666 Hz wide.
• Each frame covers 256 ÷ 22 050 = 11.610 ms, so there are 86.13 frames per second.

A note is not found by reading off the loudest bin. For a single line, the detector scores every candidate fundamental by summing its harmonics — up to 8 of them, weighted by 1 / k^0.7 — and then refines the winner with a parabolic interpolation across three neighbouring bins. That interpolation is what turns a 10.77 Hz grid into a usable pitch estimate, and it is the part that is sensitive to noise.

The 181 Hz dividing line

A semitone is a fixed ratio, 2^1/12, which means the gap in Hz grows with pitch: it is always 5.9463% of the frequency. Set that equal to one 10.7666 Hz bin and you get 181.06 Hz — MIDI note 53.63, roughly F#3. Below that pitch, two adjacent semitones are closer together than one bin.

Read the last column. At A6 a semitone spans nearly ten bins, so a small error in the peak position changes nothing. At A1 a semitone spans under a third of a bin, so the same absolute error moves you a whole note. This is not a flaw specific to one tool — it is arithmetic that any FFT-based detector on these settings has to live with.

Where the MP3 comes in

MP3 is a lossy codec. It spends bits on what a listener is likely to hear and discards the rest, and the residue shows up as coding noise spread across the spectrum. Three things follow:

1. Harmonic ratios get noisier. The harmonic-sum score depends on the relative magnitudes of a fundamental and its overtones. Noise perturbs those magnitudes, so the winning candidate can shift to a neighbouring one.
2. Parabolic interpolation drifts. It estimates the true peak from three bin heights. Perturb the heights and the estimated peak moves. Above 181 Hz that drift is smaller than a semitone; below it, it is not.
3. Stereo handling changes the signal. Joint and intensity stereo encode some of the image rather than two independent channels. Since the analysis step averages the channels to mono, anything the codec did to the channel relationship survives into the mono signal.

Note what is not on that list: high-frequency cutoff. Low-bitrate MP3 does low-pass the top end, which matters for listening. It barely matters here, because single-line detection only searches 55 Hz to 1800 Hz and polyphonic peak detection only goes up to 4000 Hz. A codec that trims above 15 kHz is cutting outside the region being examined. That is why "MP3 sounds fine to me" and "MP3 converts worse" are both true at once.

How to test it on your own file

You do not have to take the above on faith, and you should not trust any published accuracy figure you cannot reproduce. This takes about two minutes:

1. Start from a WAV you own. Export an MP3 from it at the bitrate you normally use.
2. Convert both files here with identical settings — same mode, same sensitivity, same shortest-note value.
3. Download the .csv for each. The columns are index,midi_note,note_name,start_sec,end_sec,duration_sec,velocity.
4. Diff the note_name columns row by row.
5. Sort the mismatches by pitch. If the pattern above holds, they cluster below MIDI 54.

That last step is the one worth doing. A raw mismatch count tells you little; a mismatch count that sits almost entirely in the bass tells you the cause is resolution, not the converter.

What actually helps

• Convert from a lossless file. WAV, FLAC or AIFF. If your source is already an MP3 you cannot undo the loss, but you can stop adding another generation of it.
• Do not re-encode on the way in. Exporting an MP3 from an MP3 to "make it smaller" stacks two codecs' worth of noise.
• Mono is cleaner than stereo for a single line. The analysis averages channels to mono anyway, so feeding it a mono file removes the channel-relationship step entirely.
• Fix the room before fixing the sensitivity. Reverb and noise hurt the same harmonic measurement that codec noise does. Raising sensitivity on a noisy file keeps more noise, not more notes.
• Check the piano roll, not the note count. Two conversions can return the same number of notes and disagree completely about what they are.

Convert a file now — runs in your browser, nothing is uploaded.