Linear Frequency Residual Features for Infant Cry Classification

Abstract

Classification of normal vs. pathological infant cries is a socially relevant task as crying is the only known mode of infant communication. Due to quasi-periodic sampling of the vocal tract system, the spectrum formed by high pitch-source harmonics results in extremely poor spectral resolution for commonly used features. This paper investigates the effect of excitation source-based features captured using Linear Prediction Residual for classification of normal vs. pathological infant cries. The performance of Linear Frequency Residual Cepstral Coefficients (LFRCC) was compared for matched conditions (of train and test data) against state-of-the-art feature sets, namely, Mel Frequency Cepstral Coefficients (MFCC) and Linear Frequency Cepstral Coefficients (LFCC) using Gaussian Mixture Model (GMM) and Convolutional Neural Network (CNN) as classifiers. This study also investigated the effect of LFRCC on cross-database (i.e., mismatched conditions) and combined database evaluation scenarios. It was observed that LFRCC outperformed MFCC and LFCC by $24.9\%$ and $17.43\%$, respectively, for mismatched conditions and over 0.27%–1.11% for the combined database. The relatively better performance of LFRCC feature set maybe due to its capability in representing excitation source information, which is very prevalent in infant cry as formant structures are not well developed in the initial period of life.