On-device Customization of Tiny Deep Learning Models for Keyword Spotting with Few Examples

Designing a customized KeyWord Spotting (KWS) Deep Neural Network (DNN) for tiny sensors is time-consuming process, demanding training new model on remote server with dataset of collected keywords. This paper investigates the effectiveness DNN-based KWS classifier that can be initialized on-device simply by recording few examples target commands. At runtime, computes distance between DNN output and prototypes recorded By experimenting multiple TinyML models Google Speech Command dataset, we report an accuracy up to 80% using only ten utterances not seen during training. When deployed multi-core microcontroller power envelope 25 mW, most accurate ResNet15 takes 9.7 msec process 1 sec speech frame, demonstrating feasibility customization devices without requiring any backpropagation-based transfer learning.