This paper summarises the work done to design, train and test neuronal networks for the process of identification of squeaks and rattles in vehicles. Looking for slim networks, it was necessary to find alternatives to reduce the number of inputs. Therefore the frequency spectrum of the noice signals was expressed in terms of their parametric representation. This alternative reduces the number of inputs from 5000 to 25, 50 or 100, depending on the type of parametric representation, and still conserves the unique representation of the noise signal. Analytical and experimental work was performed to verify that the RLS adaptive filtering process does not affect the type of parametric representation to be used. It was concluded that a parametric representation with 200 fundamental frequencies and a 50 Hz of bandwidth is the most suitable for this application. Since there is no standard procedure to design, train and evaluate neural networks, the design of a neural network to identify squeaks and rattles in vehicles was developed with due regard for the recommendation and suggestions of different authors. MLP neural networks with architecture 100-50-4, 50-25-4 and 25-10-4 were designed, trained and evaluated. A standard protocol to obtain the training pattern was established. It was found that it is necessary to remove from these training patterns those patterns whose variations with respect to the mean pattern are too large to be considered normal or expected. It was concluded that the training patterns to be rejected are those whose moments are outside the confidence interval with alpha = 95%. The neural networks were evaluated with patterns created arbitrarily and with patterns obtained from moving vehicles. Results showed that the three MLP neural networks are adequate for this application. Results showed that three MLP neural networks are adequate for this application. However these results can not conclude if this architecture is the optimal use. Work is underway to test other architectures. It was found that the network 25-10-4 presents a slight better performance. It was also found and that even though the neural networks were trained with much fewer training patterns than the recommended ones, they show an acceptable performance.