Elevated Pitch is a project started by John Reddin of Trinity College Dublin, working in the NCRA. Its homepage is here. Its aim is to generate short pieces of "elevator music" automatically, using Grammatical Evolution.
This page is about a series of experiments run using the Elevated Pitch software. The aim was to use EP to generate music automatically, and then get humans to listen to and evaluate the results. By comparing the results from different EP setups we can find out which setups give the best music.
We carried out four experiments. In each case, the experimental software played 10 pairs of pieces of music to the subject. The subject was presented with a GUI similar to that shown below, and used it to say which piece they preferred. The results were analysed to determine which music-generation technique was better, among four pairs of techniques as specified below. Between 10 and 25 users undertook each experiment.
Learning more(For a fuller explanation of methods and results please see the draft paper.)
It is possible to specify the structure of a piece of music using a formal generative grammar -- eg, a piece consists of a melody track and a chord track; a melody track consists of 16 bars; each bar consists of 4 beats; each beat is either a note or a rest, and so on. It is then also possible to use such a grammar to generate pieces at random. However, these pieces are unlikely to be very good. Instead, we can use an evolutionary process to generate many pieces, iteratively rejecting bad ones and recombining and mutating good ones, until we (ideally) arrive at a very good one. The measure of "bad" versus "good" pieces is far too complex to be expressed computationally, but as a first approximation we can simply penalise notes which are in the wrong scale.
Results so farIn the first experiment, then, the aim is to compare randomly-generated pieces against pieces which were evolved using this method. The results showed that a statistically significant number of people choose more "evolved" pieces than random ("unevolved") pieces, which is encouraging.
In the second and third experiments, the aim was to compare a grammar which imposed no structure on the music (ie notes were unrelated to their predecessors) against a grammar which did (some notes were produced by copying and transforming previous notes). The second experiment used very short pieces, and the advantage of the latter method was not apparent. In the third experiment, melodies were longer, and the meandering nature of the melodies where notes aren't related to their predecessors became apparent. In experiment 3, therefore, most people preferred the melodies using the ("transform") grammars over those produced using ("notransform") grammars.
In experiment 4 the aim was to study methods of calculating how good the melodies were. A simple method ("nomeasures"), which penalises notes in the wrong scale, as mentioned above) was used; and it was compared against a more complex method ("measures") which rewarded melodies whose overall shape and statistical properties were similar to those found in a corpus of good melodies. The hypothesis was that this more sophisticated method would produce more pleasing melodies, but the results showed the opposite! A lot of work remains to be done on automatically calculating how good melodies are.
In the paragraphs above, the six underlined terms represent keywords in the tests. So if you've done an experiment and you have the results file, you can open it in a text editor to find out which options you preferred in each case.