You could make something basic using Markov chains. The principle is to first produce some unit of music (a single note, for example) and then, based on the last produced unit, randomly select the next unit.
First, pass through the input music. Each time you see a particular note/other unit of music, simply record in the table what came after it. When you have gone trough the entire input material, you will have a frequency table of which units follow which (After 'A', 'B' appeared 29 times, 'C' appeared 12 times and 'A' appeared twice; after 'B' ... etc).
Now select an initial note. Select the next one randomly according to the frequencies recorded in the table. Repeat until satisfied.
This will probably not yield good results if applied to individual notes, instead try short phrases. Also, the quality will improve if you have access to a large corpus of source music.