In this study, an Edge AI solution has been proposed to recognize specific sounds created by human activities in smart office concepts. A convolutional neural network model has been developed on personal computers using collected data sets including 5 different office sounds made by humans including phone ringing, keyboard typing, door knocking, glass breaking and people talking. This trained CNN model is then deployed on an STM32 microcontroller to build standalone Edge AI recognition applications for smart offices. The evaluation results show the average recall of 95% obtained from the collected training datasets, and almost 90% obtained from the testing sounds simulated in a real office environment. The testing takes roughly 0.1 s per sample on the CNN model imported on STM32 microcontroller. This good recognition performance derived from the limited resources of memory and computational speed of the STM32F746NG MCU opens potential applications of Edge AI for human activity recognition while meeting the constraints of real time processing and other requirements.