Audio event classification

Audio sensors will have a significant part in the SMART system's ability to access the physical world. This will be done by processing the audio streams and identifying different events by the sounds which they generate.

We are currently creating tools for classification of continuous audio events (such as crowd noise and music). Those tools operate on the audio stream by first segmenting them into short, fixed length segments and then generating a representation of the audio within the segment using various features. The classification of the extracted features into the different event types is done using a multilayer perceptrons networks. The training of those networks employs Deep Belief Networks (DBN) techniques which have recently proven to be very useful for training speech classifiers in phoneme recognition tasks.