2003 UnsupervisedOnsetDetection

• (Adballah & Plumbley, 2003) ⇒ Samer Abdallah, Mark Plumbley. (2003). “Unsupervised Onset Detection: a probabilistic approach using ICA and and a hidden Markov classifier.” In: Proceedings of the Cambridge Music Processing Colloquium

Subject Headings: Onset Detection

Cited By

~14 http://scholar.google.com/scholar?cites=7661854296905884337

Quotes

Abstract

• We describe an onset detection system that takes a two-stage 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.

2 Generating a ‘surprise’ signal using ICa

• The aim of the first stage of processing is to build a short-term probability model for the signal so that we can generate a new signal which is the conditional negative log probability of a short segment.
• …
• The negative log-probability of a Gaussian density function is a quadratic form, and hence energy based methods of onset detection can be derived from Gaussian signal models. For example, a short term energy measure emerges if we model the signal as spherical (i.e. white) Gaussian noise, whereas a spectrally-weighted energy measure results if we use a more general non-spherical Gaussian model.
• Audio signals tend to be extremely non-Gaussian [1] and hence, we used ICA as an initial attempt to model that non-Gaussianity. ICA systems trained (see [3] for algorithm) separately on x and x1 allow computation of the the conditional density P(x2jx1). The results are shown in fig. 1, along with a comparison with other ways of generating a ‘surprise’ signal using other models.

References

---

,

	Author	volume	Date Value	title	type	journal	titleUrl	doi	note	year
2003 UnsupervisedOnsetDetection	Samer Abdallah Mark Plumbley			Unsupervised Onset Detection: a probabilistic approach using ICA and and a hidden Markov classifier			http://www.elec.qmul.ac.uk/people/markp/2003/AbdallahPlumbley03-cmpc.pdf