Unsupervised onset detection: A probabilistic approach using ICA and a hidden Markov classifier

We describe an onset detection system that takes a twostage approach, both of which are based on unsupervised learning in a probabilistic model. The first stage uses independent component analysis (ICA) to fit a short-term non-Gaussian model to frames of audio data. This model is used to generate a reduced signal to be interpreted as the ‘surprisingness’ of the original audio signal. Our hypothesis is that onsets and events generally are perceived as so because they are temporally localised surprises. The second stage uses a hidden Markov model (HMM) with Gaussian state-conditional densities to do unsupervised clustering of the ‘surprise’ signal as represented in a multidimensional embedding space. The clusters which emerge in this space can be associated the presence or absence of an onset, and so a trivial decision based on the current HMM state can be used to drive an onset detector.