In recent years, machine learning methods have been used increasingly for music analysis. A particularly interesting aspect is the modelling of the interaction of different musical parameters in the analysis process. When analysing the polyphonic structure (voice separation) of a piece, there are influences from  low-level perceptual processes, but there are also local and global expectations of melodic continuation that contribute to the formation of voices in the perception of the listener. Machine learning enables us to model the interaction of these processes in voice separation. The most used techniques for melody modelling are probabilistic models, mostly Markov and hidden Markov models. More recently, neural networks have shown very promising results. We will discuss the use and effectiveness of neural networks and probabilistic models in a joint melodic and polyphonic model.