Originally Posted By: shlind
When playing the converted midi one can roughly hear what song it is supposed to be.

Right. Even using the various transcription tools, it's easy to see why analysing music by AI is so difficult. Look at those spectral plots and you're seeing what the software sees, and it's complex!

There are not just simple overtones on each instrument, but very often all kinds of sympathetic resonances or 'dirty' synth sounds to contend with. The software has to decide which overtone, sympathetic and 'dirty' tones relate to which instrument and which note. Our ears and brains understand what a violin or a saxophone sounds like. Software just sees a series of frequency peaks that may or may not relate to any particular instrument.

Identifying the tempo is relatively easy. Perhaps the general rhythmic 'feel' may not bee too hard, though polyrhythms would likely be a big challenge. Generating MIDI is (IMHO) usually far from convincing. Analysing the song for a style? Pretty darned tough, I think.

Last edited by Gordon Scott; 12/24/21 11:36 AM.
Jazz relative beginner, starting at a much older age than was helpful.
AVL:MXE Linux; Windows 11
BIAB2025 Audiophile, a bunch of other software.
Kawai MP6, Ui24R, Focusrite Saffire Pro40 and Scarletts
.