Event-Driven Continuous-Time Feature Extraction for Ultra Low-Power Audio Keyword Spotting

Abstract

In the context of autonomous keyword spotting and sound detection, this paper proposes a low power feature extraction unit generating spectrograms that represent a unique signature allowing the classification of audio signals. This system is composed of a continuous-time digital signal processing feature extractor combined with a convolutional neural network engine. The study evaluates the hardware requirements to implement the feature extraction unit using an advanced CMOS process. Furthermore, a simulation of the complete system using Matlab ® reveals that the recognition accuracy remains higher than 90% while offering a power consumption 4000X lower than a conventional discrete time system.

Keyword Spotting System using Low-complexity Feature Extraction and Quantized LSTM

“…In [5], the authors report that the computational power is dominated by the FFT, which accounts for 72% of the total number of sums and multiplications of the FE block. To reduce the computationally expensive MFCC extraction, [7] presents a 32-channel analog filter bank employing a passive N-path filter topology consuming 800nW, while [8] introduces an event-driven approach, in which the system simulations show up to a 4000x lower consumption compared to a conventional discrete-time system.…”

“…Similar accuracy is obtained while the feature extraction method is much less complex than the MFCC computation. The presented method uses a 16-channel filter bank, with third-order filters and a quality factor of 1.3 that could be implemented with low-consumption techniques such as [7] or [8].…”