Cough is a common symptom for various chronic respiratory diseases, and in its detection a measure of disease’s severity and progression can be made. As a result, researchers have dedicated several studies to investigate the cough detection and classification methods. One of the goals to improve these methods is the design and integration of cough detection algorithms into wearable devices. In this study, we address this goal by proposing a wearable TinyML cough detection system based on motion detected by an accelerometer sensor. An IoT platform, the Nordic Thingy:53, was used for accelerometry data collection from 5 subjects for cough and non-cough movements. The acquired signals were pre-processed, and 18 time-domain features were extracted. The extracted features were used to train a neural network composed of two hidden layers with 12, and 6 neurons respectively. The trained model achieved an accuracy of 94.38%, sensitivity of 93.92%, specificity of 94.84%, and F1 score of 94.35% for cough detection. This model was deployed under four different deployment options, using Edge Impulse platform, onto the nRF5340 SoC in Nordic Thingy:53 for on-device inference using the embedded machine learning model. The deployment options included floating-point, and quantized representation of the classification model, using either TFLite for microcontroller interpreter or EON compiler. Both quantized models provided an inference within 1 ms; and the EON model occupied 1.4 KB of RAM and 14.8 KB of Flash, while the TFLite model occupied 3.0 KB of RAM and 34.5 KB of Flash.