Classification of telephone signals with use of artificial neural networks

Automatic assessment of quality of voice and data transmission over telephone networks requires certain level of understanding of what types of signals are present on the line at each moment during which monitoring takes place. In particular, distinguishing whether the signal is a directly transmitted signal, or its echo, or just the line noise is a most desirable classification that must be performed prior to any subsequent analyses like, for example, those resulting from ITU recommendation G.168. In this paper we investigate suitability of neural networks for performing such categorisation. We search for low-cost network configurations that, after training, can reliably perform required classifications. A number of architectures of artificial neural networks are constructed, trained, tested and compared against each other. All presented experiments are performed with real telephone signals. The conclusions about the results are drawn and recommendations for further investigations are formulated.