[nluken]

Neat idea for a project, and kudos for getting everything up and running!

Tried it once or twice, and the results are very hit or miss. "70s african funk without synthesizers" includes some stuff that matches the prompt, like Ofo the Black Company, but also includes more misses than hits: Herbie Hancock, War, Prince, the Ohio Players, and Sly & The Family Stone are all American. Seems like it latched onto "funk" and really went for it there. It's fun to give the bot a more vague prompt and see where it goes though.

Ultimately, my favorite musical "discoveries" have come from friend recommendations, or editorials like the Quietus or Bandcamp Daily. There's something about the social aspect of a personal recommendation that appeals to me, even if it's not as optimized as an algorithmic one. Many people don't have that sort of relationship with music though, so I could see something like this being pretty useful.

[dr_orpheus]

I also got the sense that when trying a prompt like this that it latched on to one or two parts of it. For my prompt of "upbeat rock with violins and bagpipes" most of the songs fell in to the "upbeat rock" or "rock with violins" (a lot of classic rock power ballads that came up here). Only one or two songs with bagpipes.

I think there may be something about the popularity of songs and artists in the data because most are more popular and that may skew the results a bit because maybe my celtic rock bands don't fall in to the more popular categories.

But still, kudos. The goal of it was to be able to find new music in a different way and it definitely got me there.

[clarge1120]

I suspect the training was overfit to certain kinds of music. A bot that's fine tuned with "funk", "70's funk", and other similar music will produce better results.