FLAC vs WAV vs MP3: Which One Gives a Better MIDI Conversion?

None of them, measurably — and more literally so than we expected. We rendered one melody once, encoded that single master six ways, and converted each encode separately. All six returned exactly 60 notes: no misses, no extra notes, no octave errors. The note start times and durations matched to the millisecond across every encoding, and the pitches agreed to within 4 cents even at 32 kbps. If you are picking a format before converting, stop: bring the file you already have.

That is an anticlimactic answer, so here is the test behind it. Every number below comes from running the file through this site's converter, not from reading a specification.

One master, six encodings

We generated a 30-second melody at 120 BPM, 44.1 kHz, mono: 60 notes, one per quarter note, drawn from a 16-step pattern that spans MIDI 60 to 76 — C4 to E5, about 261 Hz to 659 Hz. It is deliberately monophonic. A single line isolates the codec, so any difference in the output is the encoder's doing rather than polyphony confusion.

The master was written once as 16-bit PCM, then every other file was encoded from that same file with ffmpeg, so no re-rendering or double conversion enters the chain. Then each of the six files was handed to the converter on its own, and the resulting note table was read straight out of the page and scored against the 60 notes we knew we had written.

Encoding	File size	Actual bitrate	vs. WAV size	Notes detected	Real notes found	Extra notes
WAV — PCM 16-bit	2,584 KB	705.6 kbps	100%	60	60 / 60	0
FLAC — lossless	598 KB	163.2 kbps	23.1%	60	60 / 60	0
MP3 — 320 kbps CBR	1,175 KB	320.7 kbps	45.5%	60	60 / 60	0
MP3 — 128 kbps CBR	470 KB	128.3 kbps	18.2%	60	60 / 60	0
MP3 — 64 kbps CBR	235 KB	64.2 kbps	9.1%	60	60 / 60	0
MP3 — 32 kbps CBR	118 KB	32.1 kbps	4.5%	60	60 / 60	0

One melody, one master, six encodes from that master, one conversion per encode. Desktop Chrome, headless, software WebGL renderer, nothing else running in the tab. Ground truth is known by construction — we wrote the notes, so we know exactly which 60 pitches were supposed to come back. These are measurements from one machine, one signal and one model. They are not a general accuracy claim.

Between the 2,584 KB WAV and the 118 KB MP3 there is a factor of 22 in file size. The note count does not move, in either direction. That is worth pausing on: every row is 60, not 60-plus-a-few-extra, because this converter writes a single melodic line rather than one note per frame of detected energy.

Nothing in the output moves either

Counting notes is a coarse test. The stricter one is to compare the note tables note by note — the same 60 entries, in the same order, and see how far apart the values are. If a codec were degrading the audio, this is where it would show: a pitch that drifts, an onset that slips, a duration that changes.

The tables are not identical as raw objects — the pitches carry several decimal places of float, and the encoder does perturb the signal slightly — so here is the largest disagreement in each column, measured against the WAV:

Encoding	Max pitch difference	Max onset difference	Max duration difference	Same pitch when rounded
FLAC	0.13 cents	0.000 ms	0.000 ms	yes
MP3 320 kbps	0.64 cents	0.000 ms	0.000 ms	yes
MP3 128 kbps	0.99 cents	0.000 ms	0.000 ms	yes
MP3 64 kbps	2.00 cents	0.000 ms	0.000 ms	yes
MP3 32 kbps	3.87 cents	0.000 ms	0.000 ms	yes

A cent is a hundredth of a semitone. The worst case in the entire ladder is 32 kbps, where the pitch moved by 3.87 cents — about four hundredths of the gap between two adjacent keys on a piano. Audibly that is nothing; in a DAW it will not change the note you see. The onset and duration columns are exactly zero at every bitrate, because the codec never moved a note to a different moment in time.

So the entire difference between the lossless WAV and a 32 kbps MP3, as far as this converter is concerned, is four cents of pitch on the most extreme encoding tested — and 2,466 KB of disk.

What low bitrates do not change

The thing that does not appear in these tables is as important as the things that do. With a pitch-per-frame model, a harmonic tone like ours generates a crowd of ghost notes: energy at 2×, 3× and 5× the fundamental gets reported as extra notes an octave or two above what was played. On the previous engine, this same melody produced 153 notes for 60 played — 93 of them ghosts — and the count wandered between 81 and 93 depending on the encoding.

This engine does not do that. It got exactly 60 notes, all six times, with zero extras. The harmonic stack is still there in the audio; the model simply reports the line it hears rather than every frequency component with energy in it. That removes an entire category of "did the codec make this worse" question from the table.

It also means the notes-detected column is now a fair test of the codec, where before it was mostly measuring the model's own overtone behaviour. And under that fair test, the codec does essentially nothing.

Why low bitrates survive this particular test

The melody sits between 261 Hz and 659 Hz. Even the sixth harmonic of the top note is about 3.9 kHz. A 32 kbps MP3 keeps usable content to roughly 11 kHz — three times higher than anything this melody puts out. Every frequency the model needs to identify the pitch is still in the file, and the encoder spent its limited budget on the low band where the music actually is.

That is the mechanism, and it also tells you when the answer changes. Bitrate is not uniformly cheap across all music. It is cheap for material whose information sits low and is expensive for material whose information sits high or is spread thin — cymbals, snare, breath, distortion, a dense mix. Our test deliberately chose the case where the codec has an easy job. A drum-heavy or heavily distorted recording at 32 kbps would not come through this clean, and we would not claim otherwise from this data.

What the test does establish is the baseline that surprises people: for a normal melodic line, the MP3 you already own is not a degraded input. The lossy step did not cost you a single note out of sixty.

What actually decides your conversion quality

If format does not decide it, something else does. The converter's own output tells you what the model saw, and the honest ranking from our testing is:

What changes the result	How much it matters
Whether the recording is a single line or a dense mix	Large — this is the dominant factor
Whether drums are present	Large, but not in the way you expect
Whether the instruments are clean or distorted	Moderate
The container format	None, as measured here
The bitrate, all the way down to 32 kbps	Almost none, as measured here

The drum row deserves a sentence, because it is the one place the old intuition misleads. On this converter drums do not turn into a low cluster of wrong notes; on a drum-only file you get an empty result, and on a melody with drums underneath the drum hits re-trigger the melody's own notes. Either way the fix is the same and it is upstream: strip the drums, or convert a section where the kit drops out.

The practical move is not to convert your file again in a better format. It is to give the converter a better take: a stem, an instrumental, a section where the arrangement thins out. You can see the result before you commit to it — the page shows the note count, the pitch range and the piano roll before you download anything, so a bad conversion is visible in seconds rather than after you have imported it into a DAW.

The container is not the problem. If your conversion sounds wrong, it is because of what is inside the recording, and the piano roll will show you which part.

Frequently asked questions

Which is better for MIDI conversion — FLAC, WAV or MP3?

In our own test they were indistinguishable, and more literally so than we expected. We rendered one 30-second melody once, encoded that single master six ways, and converted each encode separately. All six returned exactly 60 notes — no misses, no extra notes, no octave errors. The note positions and durations matched to the millisecond across every encoding, and the pitches agreed to within 4 cents even at 32 kbps. Give the converter whatever copy of the file you already have.

Does a lossy MP3 lose notes when you convert it to MIDI?

Not in our test, and not even at 32 kbps. Every encode from 320 kbps down to 32 kbps recovered all 60 melody notes. There were also no extra notes: this converter tracks a single melodic line, so the harmonic ghost notes a pitch-per-frame model produces simply do not appear. The worst difference at 32 kbps was a pitch shift of 3.87 cents — about a fortieth of a semitone — and the note timings did not move at all.

Is FLAC better than WAV for transcription?

They are the same audio, so they behave the same. FLAC is lossless, which means the decoded samples are bit-for-bit the samples in the WAV. Our two conversions produced the same 60 notes with identical start times and durations; the pitches differed by at most 0.13 cents. The practical difference is size: the FLAC was 598 KB against the WAV's 2,584 KB, 23.1% of the size for effectively identical output. If you are choosing, choose FLAC for smaller local files and storage and expect no change in the conversion.

Does bitrate matter at all?

For a single melodic line, much less than people assume. Our melody spans MIDI 60 to 76 — every fundamental between C4 and E5, roughly 261 Hz to 659 Hz. At 32 kbps an MP3's usable bandwidth is still around 11 kHz, comfortably above the harmonics the melody produces. The pitches the model needs are all still in the file. Bitrate starts to matter where the information is not — cymbals, breath, distortion, dense mixes — and there the loss shows up as jitter rather than as missing melody.

How much smaller can I make the file before converting it?

In our test, a great deal, with no cost to the notes. The ladder ran from 2,584 KB (WAV) to 118 KB (MP3 32 kbps) — a 22:1 reduction — and every step returned the identical 60 notes at the identical times. We would not use 32 kbps for anything except proving the point, but it does mean you are not obliged to hunt down a lossless copy of a track before converting it. The MP3 you already have is a valid input.

So which format should I convert?

The one you have. That is the honest answer from this test, and it is a useful one: you do not need to re-export from your DAW, re-download the track in a better codec, or avoid MP3. Format choice is not where your conversion quality is decided. What decides it is what is in the recording — a clean single line converts far better than a dense mix with drums, because that is a property of the music rather than of the container.