Data Augmentation for Infant Cry Classification

Classification of normal vs. pathological infant cries is a challenging task as infant cry has higher fundamental (or pitch) frequency (F-0) and hence, distantly-spaced pitch source harmonics that sample vocal tract spectrum resulting in poor spectral resolution. This paper first presents infant cry classification system on DA-IICT infant cry corpus using two cepstral features, namely, Mel Frequency Cepstral Coefficients (MFCC), and Constant-Q Transform Cepstral Coefficients (CQCC) on statistical classifiers, namely, Gaussian Mixture Model (GMM). Results are then presented on cross-dataset scenarios on Baby Chilanto and DA-IICT datasets and found to degrade substantially. The classification results are also presented by merging these two corpora, where results were found to degrade for a certain parameter setting in CQCC. To mitigate this degradation in performance, we present three data augmentation methods, namely, signal perturbation using tempo, volume, and speed. The results obtained, indicated that the tempo perturbation enhances the performance by a small margin of 0.2 % and 0.98 % using MFCC and CQCC, respectively. However, results are found to degrade using speed perturbation by 3.49 % and 21.93 % using MFCC and CQCC, respectively. This indicates that F-0 and it's dynamics are very crucial acoustic cues for normal vs. pathological infant cries and which is why cry modes based on F-0 were used historically for infant cry classification in the literature. Additionally, the performance of the various feature sets are compared using J-statistics.