2 ms·
One thing I was hoping to see from GPT-2 is phrasing. Most of the folk music they're using as input has a structure of internal repetition that lets you know wh
by jefftk 7y ago
One thing I was hoping to see from GPT-2 is phrasing. Most of the folk music they're using as input has a structure of internal repetition that lets you know where you are in the tune. Most commonly: AABB structure, 64 beats total, with every 2^n beats clearly grouped. Some of the examples seemed to have this at the 8-beat level, but I couldn't find any with it at the 16, 32, or 64 level.
Here's a random snippet of a real tune where you can clearly hear the AABB structure: https://www.jefftk.com/contras/tunes/cast64__starabovethegarter.mp3 https://www.jefftk.com/contras/tunes/cast64__starabovethegar...
- p1esk 7y agoYou don't really need any AI to do phrase repetition patterns. Typically you simply repeat it as is, or maybe change pitch. This can be done algorithmically after the phrase generating stage. It would be nice if the network learned these patterns (and transitions) on its own, but it's far more important to generate quality phrases. That is the hard part and as you can see from the samples we are not there yet.
- jefftk 7y agoGPT-2 for text is impressive because it does a surprisingly good job with larger scale patterns. If you give it a play it generates things with actor marks, it's sentences are a reasonable length etc. I had been hoping this would be similar to learning the structure of tunes, but it seems that requires something we haven't figured out yet. I do disagree, though, that phrase repetition is as simple as you say. A good 16-count phrase has patterns within it, and a good AABB tune has relationships between the A and B parts.