Voice Keyword Spotting on Edge Devices

One of the most challenging subjects in edge device voice recognition is the difficulty of detecting a given set of words in fluent speech. Convolutional Neural Networks (CNN) are proven to be extremely successful for cognitive tasks and are simple to deploy on edge devices. In this article, a Convolutional Neural Network (CNN) model for Keyword Spotting (KWS) is trained and deployed. Unwanted components are eliminated from audio recordings, and voice characteristics are retrieved using the Mel-Frequency Cepstral Coefficient (MFCC). The model is trained with 84 thousand parameters. The model uses 10.34 Kilobyte of RAM and Flash memory of 122.32 Kilobyte. The model is successfully deployed on STM32 Nucleo G474RE Board. The proposed work outperforms the existing works.