Hacker News Reader

Comment by clvcooke | Hacker News Reader

I was pretty excited to try this out (there is an online demo linked), however after some playing around I found the experience lacking. Not that things sounded _bad_ persay but that I found it incompatible with actually playing anything that sounded nice.

I think part of my disatisfaction is due to pattern recognition. As humans we recognize patterns EVERYWHERE, even when they don't exist. So what happens (or what I believe might be happening) is that I think I recognize a pattern, try to play a note to continue it, then the NN decides otherwise. Although the resulting sound may be nice to an outsider it breaks my internal model, making it sound "bad" to me.

I do however like the idea of augmenting insturments to make them easier to play, I'm personally guitly of having many potential songs in my head but not having near enough the skill to execute them.

View on HN

I think the dissatisfaction you're talking about arises necessarily from projects that generate notes with a network trained on a bunch of MIDI data. Our ears expect music to have structure (chord progressions, etc) but models like this only really mimic small-scale features of the source. So any given phrase of 3-5 notes can sound musical, but over the course of a measure or two the illusion breaks and it sounds largely indistinguishable from a random walk of notes.

So very similar to a Markov text generator, then.

Makes sense, really, since music can convey much more than is evident merely by looking at the sequence of notes.

That's the analogy I usually make, yeah - it's like generating a story with a neural network trained on a bunch of stories. You'll get grammatical phrases, and sometimes sentences, but naturally none of it will make any sense.

>I do however like the idea of augmenting insturments to make them easier to play, I'm personally guitly of having many potential songs in my head but not having near enough the skill to execute them.

Yeah, we've been doing that for a while with music tech starting from basic sequencers, arpeggiators and looper pedals, down to programmable automatic accompaniment on keyboard workstations (you press a chord -- or even one or two keys -- and the virtual band is playing the backing track in the right key).

Anyway, for your problem of having songs in your head: hum them into a portable recorder (cell phone/whatever) whenever you have an idea. At home, try to pick up the melody on an instrument that you can play. With software, record the melody to a click track/metronome. If you can figure out the chords, layer them on top. There are plenty of drum loops lying around. Boom! You've transformed an idea into something others can listen.

IIRC that's how Michael Jackson composed. And if you can't play an instrument, you can hum to a friend who can play, and figure out things together. It really works!

The constraints need some tweaking. For example, it should be possible to repeat a note by repeating a keypress. The input could be interpreted more strictly as a Parsons Code [1].

[1] https://en.wikipedia.org/wiki/Parsons_code

Reply on news.ycombinator.com