Abstract

In recent years, there has been an increasing interest in statistical approaches and tools from Bayesian statistics and machine learning for the analysis of audio and music signals, driven partially by applications in music information retrieval, computer aided music education and interactive music performance systems. The application of statistical techniques is quite natural: acoustical time series can be conveniently modelled using hierarchical signal models by incorporating prior knowledge. Once a realistic hierarchical model is constructed, many audio processing tasks such as coding, restoration, transcription, separation, identification or resynthesis can be formulated consistently as Bayesian posterior inference problems.

In this talk, I will review our recent work in various signal models for audio and music signal analysis. In particular, factorial switching state space models, Gamma-Markov random fields and Nonnegative matrix factorisation models will be discussed. Some models admit exact inference, otherwise efficient algorithms based on Monte Carlo or variational approximation methods can be developed. I will illustrate applications in music transcription, restoration and audio source separation.