Wavelet-based automatic cry recognition system for detecting infants with hearing-loss from normal infants
Infant cry is a multimodal and dynamic behaviour that it contains a lot of information. Goal of this investigation is recognition of two groups of infants by new acoustic feature that has not used in infant cry classification. The cry of deaf infants and normal hearing infants is studied. ‘Mel filter-bank discrete wavelet coefficients (MFDWCs)’ have been extracted as feature vector. Infant cry classification is a pattern recognition problem such as ‘automatic speech recognition’, which in signal processing stage the authors performed some pre-processing included silence elimination, filtering, pre-emphasising and, segmentation. After applying the discrete wavelet transform on the Mel scaled log filter bank energies of a cry signal frames, MFDWCs feature vector was extracted. The feature vector, MFDWCs, of each cry sample has large length, so they used principle components analysis to reduce in feature space dimension, after training of neural network as classifier, they achieved to 93.2% correction rate in cry recognition of test data set. This result shows better efficiency in comparison with previous familiarised approaches.
