Keyword Sporting (KWS) is an algorithm used to detect specific keywords within audio stream. It is usually adopted to trigger further interactions between users and devices with specific keywords or directly act as an input to directly operating the device. However, the mainstream methods of KWS caused inconvenience or problems to users, such as latency caused by network problem or high cost for calculate device. It turns out to be cost-effective and fast-responded if KWS is able to run on a Microcontroller Unit (MCU), which is network-free and cheaper than commonly-used System on Chips (SoCs). In this paper, we proposed and lightened a convolutional neural network Based on Resnet, lightened the model until it can be successfully deployed on an MCU. Additionally, we adopted Rnnoise speech enhancement algorithm to denoise the input signal, purified the input of network under Mel Frequency Cepstral Coefficient (MFCC) domain, and significantly improved the performance of our model. With only 993.8KiB Flash memory required (needed 4.5MB before model quantization and structure improvement) on STM32H753ZIT6, our model can reach an average of 92.7% with 10 users to record 10 sound files each as input to extract features. The accuracy is even possible to increase further if the number of users is decreased.