Evaluating Noise-Robustness of Convolutional and Recurrent Neural Networks for Baby Cry Recognition

Reliable baby cry recognition plays a crucial role in infant care and monitoring, yet real-world environment poses challenges to system accuracy due to its background noises. This study proposes a novel CNN architecture for baby cry recognition under varying noise conditions, featuring three convolutional layers, a max pooling layer, and 0.5 dropout set, and compares its performance against standard RNN models. The models were trained for 100 epochs with a batch size of 64 and evaluated in both clean and noisy environments. To simulate real-world scenarios, recordings were transformed into audio signals and subjected to varying levels of background noise, particularly at different signal-to-noise ratios (SNRs). Results indicate that both models achieved high accuracy (>89%) in noise-free conditions. However, the proposed CNN maintained higher precision (93%) and overall accuracy (91%) than the RNN under 10dB noise, demonstrating its superior noise robustness for baby cry recognition. This improvement is attri buted to the CNN's capacity to capture spatial features in audio signals, making it susceptible to noise disruptions. These findings contribute to the development of more reliable and robust baby cry recognition systems.