Detection of audio events by boosted learning of local time-frequency patterns

It is often desired to detect some particular short sound events from an audio recording. For example, in music analysis and processing, one may be interested in detection of percussive events. In environmental audio analysis one may look for individual sound events related to some activity, for example, sounds of footsteps from a walking person. Generally these problems can be solved by matching some prototype time-frequency (TF) patterns to a TF representation of the input signals to obtain time-varying probability functions for the prototype events. The method introduced in this paper is based on a small number of locally collected event patterns that are used directly to define features for weighted cascade of weak classifiers that is trained using the AdaBoost algorithm. The results of a comparison to a traditional sound event classifier based on the mel-frequency cepstrum coefficients and a hidden Markov model are very encouraging. Copyright