so it should be possible to do the musical equivalent of nonsense texts, where you start with some pattern and piece together fragments using n-grams to find common following phrases. has anyone done this yet?
David Cope experimented with Markov-chain-style generation of Bach pieces in the '80s, but he ended up switching to other models, like augmented transition networks (ATNs), because he wanted more global coherence, rather than locally coherent but wandering/aimless pieces.
He eventually ended up with a more complex system, EMI, that generates pieces in the style of around 100 composers, some of which have passed the "musical Turing test" in that scholars of the composer in question thought it might've been a genuine work.
There's a lot of pretty interesting experimentation with just about every possible generative grammar by other researchers, though, from Markov models to HMMs, context-free grammars, L-systems, cellular automata, etc. This 2009 book has a pretty good overview of what people have done, though it's textbook-priced: http://www.amazon.com/gp/product/3211999159/ref=as_li_ss_tl?...
Something similar: At the last Music Hack Day in Boston Joe Berkovitz of Noteflight, an online music score editor, implemented the Music Autocomplete idea using Peachnote's data, see http://www.youtube.com/watch?v=I4-eF02Wmdw.
By the way, all functionality in Peachnote is exposed vie an API, so it's easy to experiment with.
Generally, if anybody here is interested in music n-grams and in doing something cool with the largest symbolic music data set available, please get in touch!
text equivalent described here - http://www.decontextualize.com/teaching/dwwp/topics-n-grams-... (or http://www.theregister.co.uk/2008/10/02/sarah_palin_intervie...)
[searching around i found various interesting, related reports of work (http://peabody.sapp.org/class/dmp2/lab/markov1/; https://ccrma.stanford.edu/~jacobliu/254report/), but nothing using this corpus.]