FLAC vs WAV vs MP3: Which One Gives a Better MIDI Conversion?
Last updated 29 September 2026
None of them, measurably. We rendered one melody once, encoded that single master six ways, and converted each encode separately. All six returned the same 60 notes at the same 60 pitches. WAV and FLAC gave byte-identical note tables — same pitches, same starts, same durations, same velocities, same pitch-bend points. MP3 at 320, 128 and 64 kbps also recovered 60 of 60. Only 32 kbps differed at all, and only by producing fewer ghost notes. 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.
| Encoding | File size | Actual bitrate | vs. WAV size | Notes detected | Real notes found |
|---|---|---|---|---|---|
| WAV — PCM 16-bit | 2,584 KB | 705.6 kbps | 100% | 153 | 60 / 60 |
| FLAC — lossless | 598 KB | 163.2 kbps | 23.1% | 153 | 60 / 60 |
| MP3 — 320 kbps CBR | 1,175 KB | 320.7 kbps | 45.5% | 151 | 60 / 60 |
| MP3 — 128 kbps CBR | 470 KB | 128.3 kbps | 18.2% | 150 | 60 / 60 |
| MP3 — 64 kbps CBR | 235 KB | 64.2 kbps | 9.1% | 153 | 60 / 60 |
| MP3 — 32 kbps CBR | 118 KB | 32.1 kbps | 4.5% | 141 | 60 / 60 |
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 real-note count does not move.
The lossless pair is not merely close — it is identical
WAV and FLAC are the same audio. FLAC is lossless compression: decode it and you get back exactly the samples that were in the WAV, bit for bit. So the interesting question is not whether they score the same, but whether anything in the pipeline can tell them apart at all.
Nothing can. The two note tables are identical — not similar, identical:
| Property of the output | WAV | FLAC |
|---|---|---|
| Notes detected | 153 | 153 |
| Real notes recovered (of 60) | 60 | 60 |
| Pitch range written | 60 – 102 | 60 – 102 |
| Pitch-bend points | 3,875 | 3,875 |
| Complete note table | identical — same pitch, start, duration, amplitude and bend count for every note | |
We compared the two tables as whole objects, not field by field looking for near-misses. They match exactly.
So the entire difference between the two formats, as far as this converter is concerned, is 1,986 KB of disk. If you are choosing between them, you are choosing a file size and nothing else — and 23.1% of the size for identical output is not a hard call.
What low bitrates actually change
The one column that moves in the first table is the notes-detected count, and it moves in a direction that is easy to misread. Look at 32 kbps: 141 notes detected instead of 153. That looks like the codec destroying information. It is the codec removing something, but the something is not the melody.
The 60 real notes were found in every single encoding, including the smallest. What varies is the number of additional notes the model found — the harmonic ghosts. Our test tone is a harmonic stack, so the model picks up energy at 2×, 3× and 5× the fundamental and reports some of that as notes an octave or two above what was played. Those extra notes are not in the file the encoder damaged; they are a property of the model hearing overtones. Counted across the ladder:
| Encoding | Ghost notes detected | Real notes found | Real notes missed |
|---|---|---|---|
| WAV (reference) | 93 | 60 / 60 | 0 |
| FLAC | 93 | 60 / 60 | 0 |
| MP3 320 kbps | 91 | 60 / 60 | 0 |
| MP3 128 kbps | 90 | 60 / 60 | 0 |
| MP3 64 kbps | 93 | 60 / 60 | 0 |
| MP3 32 kbps | 81 | 60 / 60 | 0 |
The spread across all six is 81 to 93, and it is not monotonic — 64 kbps produced the same 93 ghosts as the lossless WAV, and 128 kbps produced the fewest of the MP3s. That is the signature of noise, not of a trend. If codec damage were eating the audio, the numbers would fall steadily as the bitrate dropped. They do not.
One honest caveat: ghost-note counts on a synthetic harmonic tone are sensitive to the model's frame-by-frame decisions, and a 12-note spread is within what changes between runs on the same input. Do not read the ordering of that column as a ranking of codecs.
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 — drums come back as clusters of low pitched notes |
| Whether the instruments are clean or distorted | Moderate |
| The container format | None, as measured here |
| The bitrate, above ~64 kbps | None, as measured here |
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. We rendered one 30-second melody once, encoded that single master six ways, and converted each encode separately. All six returned the same 60 notes at the same 60 pitches. WAV and FLAC produced note tables that were identical down to the value — same pitches, same start times, same durations, same amplitudes, same number of pitch-bend points. MP3 at 320, 128 and 64 kbps returned 60 notes out of 60 as well. Only 32 kbps differed, and only by dropping ghost notes, not real ones. 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, down to 64 kbps. Every encode from 320 kbps down to 64 kbps recovered all 60 melody notes with no misses. The difference showed up in the notes that were not really there: 93 harmonic ghost notes at 705.6 kbps WAV, 91 at 320 kbps, 90 at 128 kbps, 93 at 64 kbps and 81 at 32 kbps. The MP3 encodes did not lose signal — the artefacts they add are not the kind this model tracks. At 32 kbps the file was 4.5% of the original size and the real notes were still all present; what fell away was noise.
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 literally identical note tables — the same 153 detected notes in the same order with the same values. The practical difference is size: the FLAC was 598 KB against the WAV's 2,584 KB, 23.1% of the size for identical output. If you are choosing, choose FLAC for uploads and storage and expect no change in the conversion.
Does bitrate matter at all?
For the notes the model reports confidently, 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 fifth harmonic of the highest note. 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 and extra ghost notes 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 still returned all 60 notes. 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.