A Multimode 157µW 4-Channel 80dBA-SNDR Speech-Recognition Frontend with Self-DOA Correction Adaptive Beamformer

📋 论文概要

该论文提出了一种多模、超低功耗的4通道语音识别前端芯片，采用自DOA校正自适应波束成形技术，解决了传统固定波束成形在可变噪声环境下抑制效果差的问题。芯片在157µW功耗下实现了80dBA的SNDR，适用于耳塞、手机等设备的自动语音识别。

💡 主要创新点

• 提出自DOA校正自适应波束成形算法，无需外部校准即可动态调整方向，提升复杂噪声环境下的语音拾取质量。
• 实现多模工作模式，在低功耗与高性能之间灵活切换，适应不同应用场景。
• 集成4通道模拟前端与波束成形处理，在157µW超低功耗下达到80dBA SNDR，兼顾能效与精度。

🏷 关键词

📄 原文摘要

Mohammad R. Haghighat2, Michael P. Flynn1 *Equally-Credited Authors (ECAs) 1 University of Michigan, Ann Arbor, MI; 2Intel, Santa Clara, CA Beamforming with multiple microphones is essential for Automatic Speech Recognition (ASR) in earbuds, cell phones, and smart speakers. Although fixed delay-and-sum (DAS) beamforming is simple to implement, it only suppresses noise from a fixed direction of arrival (DoA) [1]; hence, it is ineffective in real varying noise conditions. Reference [2] implements ultra-low-power keyword spotting (KWS) with noise suppression, but the lack of an ADC and beamforming limit practical application. On the other hand, adaptive beamforming (ABF) actively adjusts nulls to suppress varying noise sources. Adaptive beamforming with a trained DNN is promising [3] but requires extensive training data and high power consumption and is not applicable for battery-operated systems.

👥 作者与机构

Taewook Kang*1, Seungjong Lee*1, Seungheun Song1,