Cochlear Features for Acoustic Segmentation
Abstract
This work explores an alternative set of features to the frequently used melfrequency coefficients (MFCCs). The cochlea features simulate the nerve fibre signal sent from the ear to the brain. In this study the usage of the cochlea features for acoustic segmentation is of main interest. Both the cochlea features and a variant of combining them with zero crossing with peak amplitude (ZCPA) have been used as input to an acoustic segmentation algorithm. Also experiments using the cochlea features as input to an artificial neural network (ANN) for classifying each vector as boundary/non-boundary have been performed. The results show that the features contain a great deal of information regarding the speech signal. Especially the combination of cochlea and ZCPA are giving good results.