Robust MCU Oriented KWS Model for Children Robotic Prosthetic Hand Control

There are few models of prosthetic hands in literature designed for children. The use of speech commands as control method was not found to be used in any of them. Control by voice based on Keyword Spotting (KWS) is a non-invasive method that offers many advantages over others. KWS based on Deep Learning models have proved to be the most accurate, but their implementation in microcontrollers (MCUs) is challenging due to MCUs low hardware resources. In this paper, a robust KWS model based on log-Mel spectrograms and CNNs is presented for deployment on MCUs. The model is trained to recognize 5 keywords using the Multilingual Spoken Words Corpus and UrbanSound8k datasets, including a large number of non-keywords and background noise in training to provide robustness. Some popular MCU platforms are evaluated to implement the model, and STM32 was chosen for its advantages. Inference time simulations were made on some model-compatible STM32 boards.