WAV is uncompressed. Every sample you recorded is still there. MP3 throws away detail it thinks you will not hear, and at low bitrates that leaves smeared highs and small pre-echoes before sharp attacks. A transcription model can mistake those artifacts for notes.
At high MP3 bitrates the gap is small, so do not convert an MP3 to WAV hoping for a better result: the lost detail does not come back. But if you have a WAV from your own session, use it. It is the cleanest input you can give the converter.
Stems in, MIDI out: one part at a time
Transcription models handle one instrument far better than a whole band. When the bass, the keys and the vocal are each on their own track, the model never has to guess which overtone belongs to which part.
That is why the stem rack exists. Drop the stems you bounced from your DAW, give each lane a preset, and get one MIDI file per part. The files share the same tempo, so they line up when you drop them onto adjacent tracks.
If you only have the finished mix, use Full song mode on the MP3 to MIDI page. It separates the mix into stems first, then transcribes each one, at 2 credits per minute.
Velocity and pitch bends: what survives the trip
Each note in the MIDI carries a velocity from 1 to 127, estimated from how loud the note was in the recording. Soft ghost notes stay soft and accents stay accented, so the part keeps its dynamics when you play it back through a new instrument.
Fast mode can also write pitch-bend data. A bass slide, a vocal scoop or a bent guitar note keeps its shape instead of snapping to the nearest semitone. Turn bends off if you want clean notes for sheet music or for an instrument that does not respond well to bends.
What does not survive is timbre. MIDI has no idea what your recording sounded like; it only knows the notes. That is the point: you choose the new sound.
Sample rate, bit depth and headroom
The Fast model resamples everything to 22.05 kHz mono before listening, so a 96 kHz file gives it nothing extra to work with. Export at your project rate, usually 44.1 or 48 kHz, and keep the file size down.
Bit depth matters even less for transcription. A 16-bit file is fine. What matters is headroom: a clipped waveform adds harsh harmonics that turn into stray high notes. Leave a few decibels below zero and skip the limiter on the bounce.