A STATISTICAL PATTERN RECOGNITION FRAMEWORK FOR NOISE RECOGNITION IN AN INTELLIGENT NOISE MONITORING SYSTEM1

INTRODUCTION Actual noise monitoring systems have the shortcoming that although the intensity , duration, and time of occurrence of noises can be recorded, their source often cannot be identiied. Such information would be particularly useful when multiple noise sources are possible. This has led to research directed toward providing an \intelligent" noise monitoring system (Fig. 1) able to distinguish between the acoustic signature of diierent noise sources. Various techniques have been proposed for that purpose, including neural networks 1]]2], linear classiiers 3], and ad hoc methods 4]. In this cross-fertilization paper we intend to show how the theory of statistical pattern recognition 5]]6] provides a framework for building such a system. BASICS OF STATISTICAL PATTERN RECOGNITION In this section, we give a brief overview of pattern recognition theory. The concepts introduced here will be elaborated on in a companion paper 8]. Figure 2 presents a classical pattern recognition model for supervised learning. The signal prepreprocessor uses signal processing techniques to generate a set of features characteristic of the signal. For example, in our notional system the power spectral density of the acoustic signal is estimated and features are calculated to describe the relative amount of energy in diierent bands of the spectrum. Those features