Neural Networks are promissing models for pattern recognition, applicationsin uncontrolled environments due to their generalization capability and noise insensitivity. Recently they have been applied to speech recognition,particularly to classify phonemes and isolated words. The work reported here is the description of a neural approach to a real application: speaker independent word recognition,offering a range ofndent word recognition,offering a range of performances which is very close to the one offered by the ndent isolated word recognition over the public telephone network.The feature needed for this application are speaker independence(naive speakers,accent variability) and robustness against noise and channel variations(different switching circuits). A large multi-speaker database of isolated words(10 digits) was collected through local and long distance telephone calls and processed using standard feature extraction techniques,to provide the neural network with a significative training and test set:the Multi-Layer Perceptron model was applied to these data, investigating several architectures. The results demonstrate obtained are discussed and compared with those achieved on the same data by continuous Hidden Markov Models. A second database, with the same characteristics but recorded in a simulated telecom environment, has been used to explore the gap which exists between alaboratory system and a real application. The result demonstrait that MLP plus a form of time-alignment preprocessins is a practical tool for speaker independent wordrecognition, offering a range of performances which is very close tothe one offered by the HMM technology with the advantage of a straight-forward parallelization on dedicated hardware.