Implementation and evaluation of an accurate real-time voiceband signal classifier

Abstract

This paper describes the implementation and resulting accuracy of an economical voiceband signal classifier developed for use in the public switched telephone network. Companded digital signals are extracted from a 1.544 Mbps T1 digital trunk and then classified into either silence or twelve other active categories, including speech, four classes of data modem, three classes of fax, random binary data, fax signalling, ringback signal, and dual tone multifrequency (DTMF) digits. The classifier first derives from the baseband signal in each channel the central second-order moment and the first ten lags of the autocorrelation sequence, all normalized with respect to average power. Avoiding demodulation and Fourier transform steps in the classification process permits linear and quadratic discriminant functions to be computed in real time for all 24 T1 channels. Classification accuracies are reported for statistically optimal linear discriminant functions over classification intervals ranging from 31.5 to 256.5 ms. Also given are the greater accuracies achievable using statistically optimal quadratic discriminant functions and an automatically trained decision tree known as an adaptive logic network (ALN).