Detection of acoustic patterns by stochastic Matched Filtering

The detection of a pattern in an audio sequence is considered. An approach relying on the Stochastic Matched Filtering theory is proposed. It consists in first defining offline a basis from the statistics of the pattern and of the noise, then in isolating the pattern by means of a likelihood ratio test involving the online decomposition of the audio sequence on this basis. A simulated case study is proposed, which provides some guidelines to the tuning of the algorithm. Then, experimental results concerning the application of the method to voice activity detection are presented.