fbpx


The MIDI Forum

  Wednesday, 07 September 2022
  4 Replies
  885 Visits
6
Votes
Undo
  Subscribe
I've used MIDI, well, since it began almost, but I use it sporadically and really know little about the technical details. I'm composing in MuseNet/GPT-2 artificial intelligence (it's free for anyone to use), and I save my files in MIDI since I insert tempo changes etc. using my own software in post-process, and also I have a nicer sounding piano VST. You can hear several of my MuseNet compositions on my YouTube channel "Timzart7"

All my piano files have come out fine, but today, wouldn't you know, I loaded my MIDI file to probably the best thing I've composed in two weeks, and it sounds muffled and bad, kind of low volume as well, and the piano VST makes it worse. Reading about Sekaiju here, I compared a MIDI file that is piano alone, just like the corrupted file, and they look identical to me in the parameters, but I don't know what I'm looking for. Any help or suggestions would be much appreciated. If anyone would like to look at the MIDI file itself, the text from the Sekaiju file or anything, I'd be happy to send it on. I've never had the problem of a corrupt MIDI before. Thanks in advance.
Tim set the type of the post as  Technical Question — 3 weeks ago
0
Votes
Undo
Can you attach a .zip with both files?
3 weeks ago
·
#15826
0
Votes
Undo
Thank you for your quick reply Clemens. I posted this problem up past my bedtime. When I encountered this problem, while I tried multiple things, I was kind of freaked out and was not rationally going through the obvious steps of looking at what I should have looked at. I woke up early this morning thinking about what I missed, and then I tried something that fixed 95% of the problem.

I looked at the velocity profile of the bad file in my MIDI software -- tall lines at the bottom of a piano roll view (I'm mentioning this for any beginners in MIDI) -- comparing it to the velocity profile of a normal file of a similar Chopin style piece I did on MuseNet. In the problem file, I was able to see the overall volume was much lower, and I brought the volume up on all notes, and that fixed the muted/muffled sound problem, that I was able to apply my good VST piano and it sounds good. Big relief.

There is a more subtle problem, in that the velocities of notes are identical height for long stretches, much more level at the tops than they are usually. There is some variance, but not nearly as much as there usually is on a file produced by MuseNet. Along with that, it has a mechanical evenness characteristic of MIDI rendered directly from music writing software, rather than the usual subtle variances in a human MIDI Keyboard to MIDI performance. Even though MuseNet is not using a human performance, it emulates it very well. I will mitigate it by putting in some tempo changes here and there.

In summary, I think this was failure on the MuseNet end, as it is a web interface, and I was doing something different a person probably shouldn't do, working in 10 tabs at once. That is, I had the MuseNet interface open in 10 tabs at once, working on 10 different pieces. Yet, when it came to saving the files, I was down to 2 tabs. But maybe that isn't even a good idea. My rationale for doing that is that I had a theory that as a composer, even using AI, it is better to discard compositions that are not going in a good direction, rather than trying to make the best of the direction they are going in.

Just to mention something about MuseNet for others wanting to try it, unlike some other artificial intelligence music composition tools, you can make measure-by-measure choices with MuseNet. I have generated as many as 50 measures, to choose only one measure of music among those 50, as I build a piece measure by measure. That generation time (it generates measures four at a time) can take up to two minutes, so that was my rationale of working in so many tabs also. Again, though, it was just an experiment. Working in three tabs at once means you won't have to ever wait while it is generating. I call them measures, but it generates in groups of notes, really, and number of notes is determined by the token size you choose. "Token" is a term in artificial intelligence for a piece of information, in this case, pieces of music. MuseNet works on a predictive model, predicting the next note that is most likely in a certain style you choose -- Mozart, Bach, jazz, etc. (the style choices are limited by what they offer).

Another thing about MuseNet that is almost unfathomable to any musician like myself, who in spite of my background in computers, starting programming in 1971, and I struggle to learn new music software and remember how I do various things when I don't use it for a while, MuseNet takes a few minutes to learn.
3 weeks ago
·
#15828
0
Votes
Undo
Note that if you want your recorded music to sound natural, as in retain the original natural sound from when it was recorded, you should try to avoid ANY messing with velocites after recording. It could be OK to 'fix' indiv notes. Maybe there was a global level that had got to be wrong, and not indiv notes, and you needed to correct the global level (which could have changed the indiv notes but kept the variations from before). rather than change by indiv notes?

Geoff
3 weeks ago
·
#15830
0
Votes
Undo
Finished it, and it came out well, will be posted on my YouTube channel within an hour as Diamond Etude on my channel Timzart7, for anyone interested in the pinnacle of what artificial intelligence can do in helping, along with me making literally hundreds of artistic choices as to what measure to use, in creation of an original work in the style of Chopin.
  • Page :
  • 1
There are no replies made for this post yet.