Re:Show HN: Jukedeck – create unique, royalty-free music for...
by TheOtherHobbes on Wednesday December 09, 2015 @06:28PM
(#10707646)
The rules are different for different styles, and they're hard to get right, because musical structures are holistic, and multi-layered. The domain of interest is at least as long as a phrase, and can be as long as the entire piece - which means that potentially every note is related to every other, and the data structures you use have to understand how and why.
I'm finding the tunes the least successful element here. They sound like a fairly naive Markov sequence/chord spine mashup, with none of the character of really good tunes.
The sounds/production are much better. The drum parts are midway between the two.
>I find an interesting parallel between 1/2/4/8 repeating beats/parts in songs, and binary...
There's more to it than that.
Good algo-comp is PhD-level work. Kemal Ebcioglu, who made one of the best efforts to date, had to add a lot of hand-rolled heuristics to an expert system that was trying to produce classical counterpoint. He was last seen working for IBM research on compilers and architectures for parallelism - a very smart guy.
It's easy to underestimate how hard the problem is.
I'm finding the tunes the least successful element here. They sound like a fairly naive Markov sequence/chord spine mashup, with none of the character of really good tunes.
The sounds/production are much better. The drum parts are midway between the two.
>I find an interesting parallel between 1/2/4/8 repeating beats/parts in songs, and binary...
There's more to it than that.
Good algo-comp is PhD-level work. Kemal Ebcioglu, who made one of the best efforts to date, had to add a lot of hand-rolled heuristics to an expert system that was trying to produce classical counterpoint. He was last seen working for IBM research on compilers and architectures for parallelism - a very smart guy.
It's easy to underestimate how hard the problem is.