Introduction: Infant monitoring is important in early childcare. Most baby monitors only transmit audio or video but they cannot understand or respond to a baby's cry. In this study we propose a low power embedded sensor node that uses TinyML to classify infant cries, play a soothing lullaby, and send alerts to caregivers without cloud connectivity.
Methods: The system uses a MEMS microphone connected to a microcontroller device. The audio is processed in 1 second windows and Mel-Frequency Cepstral Coefficient (MFCC) features are extracted. The neural network was trained to classify sounds into two classes cry, and background noise. The model is deployed as a quantized INT8 classifier. Whenever cry is detected, the device plays a lullaby through a speaker and also send an alert to a caregiver's smartphone via Bluetooth Low Energy (BLE). The system also records temperature and humidity data for each event.
Results: The proposed TinyML model achieved 88.6% classification accuracy with an inference latency of 12 ms. The device operated continuously on a battery. Upon testing the system, the alert message delivered to caregiver in an average of around 1 seconds. This IoT based system logs cry and environmental data to help identify recurring discomfort patterns and further study.
Conclusion: The proposed system can detect and classify baby cry accurately without internet connectivity. Automatic music playback and caregiver alerts minimize the need for constant supervision. Future work would include increasing the number of cry categories and improving the classification accuracy with the larger and more diverse infant audio datasets. We will also work to make system identify crying patterns to determine if there are any health or behavioral issues and also integrate with other smart nursery devices such as lights, fans and white-noise machines to automatically respond to the baby’s needs.