Detecting, classifying and predicting salient events using acoustic signals and markov models

The detection, classification and prediction of significant events by means of sounds has the potential to be a useful complement to video monitoring, motion sensing and other modes in applications including security surveillance, care of the sick, infirm and elderly. Here, we apply methods inspired by approaches used in the automatic recognition and understanding of continuous speech - Mel Cepstral analysis, use of principal components analysis (PCA) of spectral templates and Markov models - to detect, classify and model sequences of salient sound events, with considerable success, occurring during the relatively controlled situation of championship tennis matches. We propose generalisations of this to a wider range of applications, including those mentioned above, although we admit that less controlled environments are likely to present new challenges.