Utilization of Expert Knowledge in Automatic Classifiers of Noise Sources

1. INTRODUCTION Due to the progress in microprocessor and digital signal processing technologies , modern noise monitoring equipment offers increased performance. Current " sound-level meters " provide much more sophisticated measurements than old-fashioned sound-levels, like statistical and spectral analysis, or noise event recording. Moreover, there is actually a trend toward integrating noise monitoring systems (NMS) with personal computers to add further processing and memory capabilities. If this increased sophistication is generally welcomed, it has also some drawbacks. The noise control expert is provided with an ever growing wealth of information, and extracting the relevant features from the data becomes more and more difficult. Consequently, there are current research interests in the development of " intelligent " noise monitoring equipments, in order to simplify and automate as much as possible the data analysis task. One current research thread concerns the automatic recognition of environmental noise sources. In automatic recognition of environmental noise sources, the goal is to classify a noise event based on its acoustic signature [1]. That is, it is expected that the NMS will provide, in addition to the level and time of occurrence of a noise event, some information on the nature of its source (e.g., train passing by, airplane flying over,.. .). Such noise recognition capabilities can be obtained by adding a sound recognition subsystem to a classical NMS. Various classification approaches have been considered for the realization of this subsystem, including neural networks, statistical classi-fiers, and ad-hoc methods. Preliminary studies have shown the feasibility of automatic noise recognition [2][3][4]. In these studies, standard pattern classification methods were applied straightforwardly without reference to the specificities of the noise recognition problem or to the environmental acoustics framework in which it takes place. We believe that utilization of acoustical-domain knowledge in an environmental noise recognition would be beneficial. Therefore, in this paper, we try to answer the following question , " what are the desirable properties of an automatic noise recognition system? " from the point of view of the noise control practician. A method for implementing such properties in a practical system is then suggested.