Real-Time Baby Crying Detection in the Noisy Everyday Environment

Baby crying detection is an important component in child monitoring, diagnostics, as well as emotion detection systems. This study proposed a real-time baby crying detection algorithm that monitors the noisy environment for baby crying on a second-by-second basis. The algorithm detected baby crying through five acoustic features – average frequency, pitch frequency, short-time energy (STE) acceleration, zero-crossing rate (ZCR), and Mel-Frequency cepstral coefficients (MFCCs). The thresholds for each feature in classifying an audio segment as "crying" were set by extracting and examining the distribution of the features of noise-free crying and non-crying samples collected from an audio database freely available on the Internet. Later, the algorithm was tested using noisy crying and non-crying samples downloaded from YouTube, where an accuracy of 89.20% was obtained for the offline testing. In order to test the robustness and performance of the designed algorithm, online testings were also conducted using three customly composed noisy samples containing both crying and non-crying segments. The online accuracy obtained was 80.77%, lower compared to the offline testing which was mainly caused by the extra noise introduced by the experimental settings. With more advanced equipment, it should be possible to increase the online testing to be closer to the offline testing accuracy, paving the way to use the designed algorithm for reliable real-time second-by-second baby crying detection.