Although you characterize it as only being able to "emulate old composers", I may have been referring to the program you are thinking of. Although it is true that it learns and extrapolates from musical input, so do human composers. Its algorithms can learn from anything, and by feeding it a mixture of styles, it can actually generate some fairly compelling and new sounding works. The programmer is also a composer and has trained the algorithm on some of his own works and the output sounds nothing like old composers.

See http://arts.ucsc.edu/faculty/cope/mp3page.htm for audio and http://arts.ucsc.edu/faculty/cope/experiments.htm for a description of the project.