An 8-Element Frequency-Selective Acoustic Beamformer and Bitstream Feature Extractor with 60 Mel-Frequency Energy Features Enabling 95% Speech Recognition Accuracy

Abstract

A synergistic approach to beamforming and feature extraction, reduces processing complexity and die area, and delivers the high SNR required for reliable speech recognition. The 1.1mm2 IC combines frequency-selective bitstream beamforming, bitstream Mel frequency-band feature extraction, and an array of continuous-time sigma-delta modulators (SDMs) without area/power-intensive decimation. When coupled with a DNN, the prototype achieves 95.3% accuracy in recognizing spoken words from the Tensorflow dataset.