On the Use of EMD for Automatic Newborn Cry Segmentation
Cry segmentation is an essential preprocessing step in any infant crying diagnosis system. Besides crying sounds consisting of expiration phases followed by short periods of inspiration episodes, each recording of newborn cries also includes silence sections as well as other sounds such as speech of caregivers, noise and sound of medical equipments. This paper is devoted to a newly developed Empirical Mode Decomposition (EMD) application to cry segmentation. The goal of the segmentation is to detect cry episodes automatically from unimportant acoustic activities existed inside the recorded signals. EMD decomposes a multicomponent non stationary signal into a set of monocomponent signals called Intrinsic Mode Functions (IMFs). The cry signals are segmented using Hidden Markov Models (HMMs) applied to the features extracted by employing EMD combined with Mel-Frequency Cepstral coefficients to the recorded cry signals. The performance of the proposed approach is evaluated on a database of 200 cry signals recorded in a real clinical environment. The experimental results demonstrate the effectiveness and suitability of the proposed method for the automatic cry segmentation..
Keywords: Automatic Cry Segmentation, Empirical Mode Decomposition, Features Extraction, Mel Frequency Cepstral Coefficients, Classification, Hidden Markov Models.