I ended up with Glitchy BARD, which generates crappy poetry. It uses source text transformed into n-grams, while rhymes rely on the CMU Pronouncing Dictionary[0]. For the voice, I used the native SpeechSynthesis API instead of a fancy AI one.
Everything runs on the client, but you can share a song via URL, which includes the random number generator seed.
It's open source[1], though the code is a bit chaotic as I was experimenting a lot and never cleaned it up. A blog post about the CSS glitch effect[2] even gained some traction on HN.
I'd like to turn this into a proper talk, mostly because it was so much fun to make and touched a lot of different areas - language models, rhymes and poetry, pixel art (which is a pain do to on web), speech synthesis, CSS animations... I gave an internal presentation at my company and got very positive feedback.
So let me know what you think and if you have any questions about BARD, please shoot.
[0] http://www.speech.cs.cmu.edu/cgi-bin/cmudict
[1] https://github.com/Stanko/bard/
[2] https://news.ycombinator.com/item?id=44221907
edit: link formatting
nmoya•3h ago