Automatic classification of wideband acoustic signals

Until now, very few contributions were published in the field of wideband acoustic signal recognition, especially for handling impulsive noise signals such as glass breaking, detonations, or door slams, as encountered in security applications, where the signals are highly nonstationary and composed of higher frequency components. This paper shows how the audio alarm recognition problem can efficiently be tackled using either pattern recognition methods relying on Bayes classifiers, or on artificial neural networks (ANN). After extraction of filterbank coefficients in the acoustic analysis module, typical feature vectors are achieved from the concatenation of k consecutive signal frames, in order to exploit dynamic temporal information. The redundancy induced by this dynamic modeling requires a reduction of the feature space dimension, performed with a conventional principal component analysis. The performance of both systems is evaluated experimentally (7000 tests) for the classification of three types of noise events (glass breaking, door slams, and stationary noises). The score of correct classification reaches 99% for the statistical approach, and 98% for the ANN-based system. It is further noticed that errors do not occur on the same signal files for both methods. Thus robustness enhancement can be reasonably expected, when using hybrid methods or fusion strategies.