A neural network estimating the psychoacoustical annoyance from physical data

An artificial neural network is able to work like an annoyance measuring device. The network is taught, in a supervised manner, with the field acoustical data and annoyance data. The latter correspond to a 10-valued function which is an objective synthesis of the behavior of 200 people listening to the noise data at the laboratory. After learning, the network delivers realistic values on the annoyance scale when new acoustical data are presented to it. The neural network simulated is a very simple one. It accepts as input a preprocessed signal: 8 statistical indices extracted from the original acoustical 10 minutes samples. More realistic situations require probably more sophisticated processing; but this preliminary result is very interesting : the psychoacousticians know it is difficult to define an explicit annoyance function. The neural network is an help for choosing the relevant parameters