<![CDATA[The safety of toddlers at home is paramount for parents, but constant monitoring is difficult due to busy schedules. The limitations of camera-based monitoring solutions, namely privacy concerns and heavy processing, drive the need to develop monitoring systems that utilize sound recognition. This research aims to develop Smart Guardian, an Artificial Intelligence of Things (AIoT) system that can detect risky or emergency sound patterns from children and send real-time notifications to parents' mobile phones. The applied method includes the development of a YAMNet-based speech recognition AI model, installed on a Raspberry Pi as an edge computing device, with a microphone functioning to record environmental sounds. This system is designed to identify crucial environmental sounds such as breaking glass, explosions, screaming, water, fire alarms, smoke detectors, in addition to infant crying. The results of prototype trials under laboratory conditions indicate that the fire alarm and smoke detector classes have extremely high confidence levels (around 0.95 and 0.83). However, the glass class showed varying confidence levels (around 0.5), while cough, explosion, water, and screaming had lower confidence levels (median 0.15, 0.13, 0.25, and 0.4, respectively). The conclusion from these findings is that Smart Guardian has great potential as a privacy-focused toddler monitoring solution, although further optimization is needed to improve the speech recognition performance of events with low and varying confidence levels. ]]>
