The “can I hum the melody in this clip?” test is the right first filter, and I agree that MP3 vs WAV is usually not the thing that decides it. A clean hummed hook or a short piano phrase will beat a pristine WAV of a dense master.

When the only source is already a finished mix, though, trimming to a “clearer” eight bars often still leaves drums, bass, and pads under the line. The converter then has to guess which of those should become notes. A more reliable extra step, before the short-section workflow you describe, is to isolate the intended part first (a vocal or lead stem from Demucs, UVR, or a similar separator) and run audio-to-MIDI on that file. You still treat the MIDI as a draft — fix durations, octaves, and extra notes in the DAW — but the input is closer to the “one clear musical idea” case the model actually handles.

If you were adding that to the table, would a pre-isolated vocal or lead stem sit with “usually good,” or do you still see enough bleed and reverb tails that it belongs with “needs cleanup”?