Noise-Aware and Lightweight LSTM for Keyword Spotting Applications

Abstract

Keyword spotting (KWS) becomes increasingly prominent in voice-enabled IoT devices. To attain high KWS accuracy, neural network (NN) algorithm is commonly used in existing works which focus on developing NN-based ultra-low power KWS to support always-on feature but do not consider the performance under noisy environments. Thus, this work proposes a noise-robust, lightweight and accurate KWS. A long short term memory (LSTM) network is trained from scratch to adapt to different noisy scenarios. The proposed LSTM model is further pruned and quantized to reduce model size while achieving accuracy of over 75.4% for white noise with signal-tonoise ratio (SNR) from -5 dB to 20 dB. The final LSTM model is mapped to a hardware accelerator synthesized using a 40-nm CMOS technology, consuming only <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1.7\mu \mathrm{W}$</tex> at 0.6V.