A 50μW 4-channel 83dBA-SNDR Speech Recognition Front-End with Adaptive Beamforming and Feature Extraction

Beamforming is essential for accurate Automatic Speech Recognition (ASR) in noisy environments. Commercial products, such as Google Home, Amazon Alexa, and Apple Airpods, beamform with multiple microphones. Delay-and-sum beamforming is amenable to low-power IC implementation (e.g. [1]) but can only suppress noise from a fixed direction making it ineffective for real situations with multiple constantly-changing noise sources. Adaptive beamforming (ABF) solves this limitation by automatically and adaptively suppressing noise from multiple, varying sources (Fig. 1). However, the high DSP power and large die area needed for conventional ABF DSP prevent practical implementation. Another challenge is that high angle accuracy is crucial for ABF to avoid severe distortion of the desired signal [2]. We combine the robust generalized sidelobe canceller (RGSC) algorithm [2] with bitstream processing for accurate and low-power ABF. We further reduce power and area through hardware sharing and optimized DSP clock rate. Our system combines multi-channel digitization, beamforming, automatic noise suppression, and feature extraction for a robust sub-mW single-chip speech-processing frontend.