TinyML Prototype of Infant Cry Classification System

Quick classification of infant cries is vital to determine the reason for a baby’s cry, especially in the first months of the baby’s life. This study aims to develop a lightweight TinyML model for rapid classification of infant cries that can be used for resource-constrained, low-power devices. Our contribution lies in achieving good accuracy with a minimal model size and suitable performance for tiny devices. We explored architectures including Convolutional Neural Networks (CNNs), Depthwise Separable CNNs (DS-CNNs), and a self-defined Residual Network (ResNet) tailored to be lightweight, unlike standard ResNet models, for resource efficiency. Mel Frequency Cepstral Coefficients (MFCCs) were extracted from cry signals and optimized by varying coefficients and frame lengths. We used two datasets, each with five categories, in our experiments. The datasets used were Baby Chillanto (DB1) and Donate a Cry (DB2). Each model was independently trained on each dataset in separate experiments. Then, each model was converted to TensorFlow Lite and quantized, with the unquantized self-defined ResNet achieving 96.26% and 93.7% accuracies on DB1 and DB2, respectively. After quantization, it maintains an accuracy of 94.71% on DB1 and 89% on DB2 with RAM usage under 30 KB and a model size of 93 KB. When deploying the quantized model trained on Baby Chillanto on a Raspberry Pi, it demonstrated an execution time of 1 s and an inference time of 3.44 ms, balancing accuracy, efficiency, and compactness for embedded systems.