Acoustic features analysis for recognition of normal and hypoacoustic infant cry based on neural networks

Abstract

This work presents the development of an automatic recognition system of infant cry, with the objective to classify two types of cry: normal and pathological cry from deaf babies. In this study, we used acoustic characteristics obtained by the Mel-Frequency Cepstrum and Lineal Prediction Coding techniques and as a classifier a feed-forward neural network that was trained with several learning methods, resulting better the Scaled Conjugate Gradient algorithm. Current results are shown, which, up to the moment, are very encouraging with an accuracy up to 97.43%. © Springer-Verlag Berlin Heidelberg 2003.