Feature Set Optimisation for Infant Cry Classification.

This work deals with the development of features for the automatic classification of infant cry, considering three categories: neutral, fussing and crying vocalisations. Mel-frequency cepstral coefficients, together with standard functional obtained from these, have long been the most widely used features for all kind of speech-related tasks, including infant cry classification. However, recent works have introduced alternative filter banks leading to performance improvements and increased robustness. In this work, the optimisation of a filter bank is proposed for feature extraction and two other spectrum-based feature sets are compared. The first set of features is obtained through the optimisation of filter banks, by means of an evolutionary algorithm, in order to find a more suitable speech representation for the infant cry classification. Moreover, the classification performance of the optimised representation combined with other spectral features based on the mean log-spectrum and auditory spectrum is evaluated. The results show that these feature sets are able to improve the performance for the cry classification task.