State of the art

The first article I found is [paiement2007][1. Paiement, J. F., Grandvalet, Y., Bengio, S., & Eck, D. (2007). A generative model for rhythms. NIPS’2007 Music, Brain.]. In this paper, the authors only focus on rhythms. They try to predict the next subsequence given the previous one. The criteria used to evaluate there model is the distance between the predicted subsequence and the actual one, and is done on 2 type of music (jazz and hornpipe).

The model is done by dividing each bars into equal timeslot. for each  timeslot, there are 3 possible states:

1. note begins
2. note continue
3. rest

The considered approach

Let's consider a state machine. The states are notes. The note is characterized by:

• the frequency (noted by its  name, could be a rest)
• the duration

We can use lilypond notation (e.g. c'4)Let's count the number of changes from one state to another. let's also consider the following changes:

• changes from one note to another whatever duration
• changes from one duration to another whatever frequency

Misc

since I begun the redaction of this post, I discover one of the most advance porject on this theme[2. Emily Howell's work].

(Non-)Related articles

• lilypond: That gray area
• Markov Chains and Hidden Markov Models
• Artificial music: Computers that create melodies