Subband Teager Energy Representations for Infant Cry Analysis and Classification

Due to quasi-periodic sampling of vocal tract system, spectrum formed by high pitch-source harmonics results in extremely poor spectral resolution for commonly used spectral features, such as Mel Frequency Cepstral Coefficients (MFCC). Therefore, classifying normal vs. pathological infant cry became technologically challenging signal processing problem. In that context, this study investigates the effectiveness of the spec-tral representations over cepstral representations for infant cry analysis. Furthermore, we show the spectrographic analysis for various spectral representations, such as Short-Time Fourier Transform (STFT), Mel Filterbank (MeIFB) coefficients, Linear Filterbank (LinFB) coefficients, and proposed subband Teager Energies (subband- TE). Experiments are performed on standard Baby Chilanto dataset with Gaussian Mixture Model (GMM) and Support Vector Machine (SVM) as classifiers. It is observed that spectral representations performs slightly better over the cepstral representations. Moreover, subband- TE representations with GMM classifier achieves 99.47% classification accuracy for normal vs. pathology cry and outperforms other architectures. This is due to the capability of TEO to accurately estimate the energies, especially in low frequency regions which consists of fundamental frequency
$(F_{0})$
and its harmonics.