The non-destructive evaluation of composite materials is needed in order to check their mechanical viability. Several classical ultrasonic signal-processing methods may be used to perform this task, such as resonance spectroscopy and extraction of features from time echograms. Neural-network processing methods may allow the evaluation of structural parameters (ply thickness, interface quality and so on) from ultrasonic echograms. The aim of this work is to implement such techniques and to apply them to ultrasonic echograms obtained from theoretical models and industrial samples, so that meaningful comparisons may be performed in order to assess the validity of the neural-network method. The neural-network based method appears very flexible and less computationally intensive in the recall phase than are the other two methods. It gives a precision sufficient for most purposes and so constitutes a good candidate for evaluation of composite-material (and more generally multi-layer) structures.