[dead]
Show HN: I trained a 125M model to autocomplete piano on-device
131–135 of 135 posts
Re: Show HN: I trained a 125M model to autocomplete piano on-device
#132WOW. I am curious to do the same for flute, mainly for Indian flute variations called Venu and Bansuri.
Re: Show HN: I trained a 125M model to autocomplete piano on-device
#133Hey @simedw this is an awesome project! A few quick ideas in case you hadn't tried them yet:
1. Have you considered representing different parts/voices independently? That would be harder to get good data for, but would likely help with voice-leading and phrasing and...
2. ... durational representation -> would it simplify your durational representation system if you modelled ties? E.g., then you wouldn't need a token to represent a duration worth 4 triplet eighth notes because you could model the tie of a quarter and a single triplet eighth. That would greatly reduce your durational vocab, right?
3. I've created a lot of academic tools for musical syntax analysis for humdrum, CRIM, VIS, and also made a couple of minor contributions to music21. I'd be happy to recommend a few automated analyses that might help the model with syntax or phrasing.
Re: Show HN: I trained a 125M model to autocomplete piano on-device
#134Omg! That's a great project. Will try doing it with some different instrument
Re: Show HN: I trained a 125M model to autocomplete piano on-device
#135Hearing the start of Für Elise, and then it being taken in an incredibly different direction, is surprisingly disconcerting.
I found it just... bad and out of whack.