"Can we generate semi-interactive infinitely evolving ambient music?" Constraints: computationally cheap, no training data.

At the heart of what you are listening to is a custom software MIDI sequencer that generates semi-random clips. The sequencer is based on a Genetic Algorithm to evolve clips based on listener feedback. In the absence of listener feedback, the program calculates scores for each clip to prefer those that are more harmonically appealing.

The resulting MIDI stream can then be fed into any number of synthesizers. The above stream is the live output of a carefully tuned combination of software synthesizers. Along with note information, the generated MIDI stream also contains control change messages to subtly modulate sound.

This website is a simple interface to one instance of this setup. All listener feedback is combined to "democratically" influence the direction the melody is going. You may vote as often and frequently as you like. Clip evolution is relatively slow, so it takes about 30 seconds to a minute to hear changes. Repeated negative votes will eventually lead to "rolling the dice" and making changes more rapidly.