A Dataset and TinyML Model for Coarse Age Classification Based on Voice Commands

This study explores the emerging sub-discipline of Tiny Machine Learning (TinyML) in the specific context of classifying audio data. New problems arise from the anticipation of voice prevalence in future interfaces with machines, especially embedded ones, and the need to have embedded intelligence into those machines, which can be captured by the idea of TinyML. In particular, there is a lack of studies on TinyML models for age classification based on voice commands, especially involving children as subjects. There is also a lack of datasets that include voice commands of both adults and children. This study aims to develop a TinyML model that is able to discriminate adults from children from their voice commands. The methodology follows the workflow of building deep learning models, including data collection and preprocessing, training, testing and deployment. Beyond the tools used for data collection and preprocessing, Edge Impulse was adopted as the development platform to train and test the model. The evaluation of the model showed a classification accuracy of more than 97% based on the custom dataset built to train and test the model. Finally, the tiny model was successfully deployed and evaluated on an Arduino Nano 33 BLE sense microcontroller.