14.8 KASP: A 96.8% 10-Keyword Accuracy and 1.68μJ/Classification Keyword Spotting and Speaker Verification Processor Using Adaptive Beamforming and Progressive Wake-Up
Jianbiao Xiao, Xuhui Zhang, Shijian Zhu, Zhengwei Yang, Meng Du, Chunsheng Ji, Long Yu, Xiao Chen, Xiaoyu Miao, Liang Zhou, Liang Chang, Shanshan Liu, Jun Zhou
Abstract
Keyword spotting (KWS) processors have been proposed and used in voice-control applications such as smart homes, intelligent robots and smart wearables, as shown in Fig. 14.8.1. Existing KWS processors have the following issues: 1) they are sensitive to human-voice noise (e.g., nearby individuals talking, TV or radio), which affects their accuracy in real-life applications; 2) they do not sufficiently exploit domain-specific features for energy reduction and accuracy improvement; 3) they do not support multiuser speaker verification (SV) free of speaker-specific training. To address these issues, in this work, we have proposed a high accuracy and ultra-energy-efficient KWS & SV processor (named KASP) with the following features: 1) a dynamically reconfigurable KWS & SV processing architecture supporting KWS-driven adaptive direction of arrival (DoA) estimation and beamforming to improve the accuracy in the presence of human-voice noise; 2) an adaptive DoA frequency-channel selection technique and a lightweight frequency-domain beamforming (Lite-FDBF) technique to reduce the energy consumption and hardware overhead; 3) a four-stage progressive wake-up processing architecture with KWS-aware adaptive voice-activity detection (VAD) to reduce energy consumption and improve the accuracy under different SNR; 4) a lightweight X-Vector (Lite-X-Vector)-based SV for multi-user speaker verification with low energy consumption free of speaker training.