My experiments with Karpathy’s recurrent neural network (rnn) continue (see this, and this). Other folks have experimented with feeding music in ‘abc’ notation into the rnn, so I thought I’d give it a try too. I found a collection of Cape Breton fiddle tunes. I can fiddle, a bit; can my mac? Turns out, yes, yes it can. Here’s what it generated:
and you can listen to it here
The output from the rnn looked like this:
X:10 T:Mancis the Poeth R:strathspey C:traditional O:18th century Scottish B:Bremner's Collections, 1757 N:melodic code 1331 3555 Z: M:C| L:1/8 Q:333 K:D DFED EDB,D|CA,B,A, B,EA,|EGGB AGEG|EFGB AFDF| E2 EF DEFA|Bdef gfed|dBAF EB,E|D2:| |:c|defd edef|gedB GABA|G2 dB dbaf|g2 ef g2 bg|agfa g2 ef| gbga gdBd|bag
I changed that ‘333’ to ‘120’ because 333 beats per minute is just silly. This site converts ABC notation to MIDI file and makes the pdf of the score you see above (!); this site converts to mp3, which I then uploaded to soundcloud. All this after 2000 iterations; several thousand more to go. But already… pretty neat, eh?