Target-aware Neural Architecture Search and Deployment for Keyword Spotting

Keyword spotting (KWS) utilities have become increasingly popular on a wide range of mobile and home devices, representing prolific application field for Convolutional Neural Networks (CNNs), which are commonly exploited to perform keyword classification. Addressing the challenges targeting such resource-constrained platforms, requires careful definition CNN architecture overall system implementation. These reasons led growing need design optimization flows, able intrinsically take into account system’s performance when ported target platform. In this work, we present methodology based Architecture Search, combine exploration optimal network topology, audio pre-processing scheme, data quantization policy. The proposed flow includes target-awareness in loop, comparing different alternatives according model-based pre-evaluation metrics like execution latency, memory footprint, energy consumption, evaluated considering application’s processing We tested our obtain target-specific CNNs commercial platform, ST SensorTile. Considering two scenarios, enabling comparison with state-of-the-art efficient CNN-based models KWS, obtained up 1.8% accuracy improvement 40% footprint reduction most favorable case.