Automatic Music Composition with Simple Probabilistic Generative Grammars

Abstract

We propose a model to generate music following a linguistic approach. Musical melodies form the training corpus where each of them is considered a phrase of a language. Implementing an unsupervised technique we infer a grammar of this language. We do not use predefined rules. Music generation is based on music knowledge represented by probabilistic matrices, which we call evolutionary matrices because they are changing constantly, even while they are generating new compositions. We show that the information coded by these matrices can be represented at any time by a probabilistic grammar; however we keep the representation of matrices because they are easier to update, while it is possible to keep separated matrices for generation of different elements of expressivity such as velocity, changes of rhythm, or timbre, adding several elements of expressiveness to the automatically generated compositions. We present the melodies generated by our model to a group of subjects and they ranked our compositions among and sometimes above human composed melodies.