Deep learning resurgence ushered in the application of pattern recognition algorithms in high-impact research fields with impressive accuracy. In addition, deep neural networks (DNN) have recently been used to classify gestures for rehabilitation device control utilizing raw electromyography data. However, the computational resources required by a convolution neural network (CNN) are a constraint that often limits deployment to embedded devices for real-time inference. An optimized edge adaptive convolutional neural network using a short-time Fourier transform (STFT) spectrogram input was proposed in this study. The model’s classification accuracy was evaluated offline and on-device for inter-subject accuracy. Furthermore, an adaptive weight update approach was implemented to improve inference model accuracy due to degradation. The proposed model and optimization technique achieved offline accuracy of 92.19 % and 94.29 % for the raw and STFT input, respectively. However, the on-device accuracy for raw and STFT input to the model was 82.26 % and 85.19 %, respectively. On the other hand, the adaptive model update increased the respective accuracy by an average of 7% on-device. Finally, our study demonstrates the deployment of DNN on-device for real-time gesture classification inference.