A Scalable Speech Recognizer with Deep-NeuralNetwork Acoustic Models and Voice-Activated Power Gating

📋 论文概要

该论文提出了一种可扩展的语音识别器，采用深度神经网络声学模型和语音激活电源门控技术，优先考虑语音活动检测（VAD）的准确性以减少误唤醒。通过集成三种VAD算法（能量、谐波、调制频率），在保证低功耗的同时提升了系统可靠性。

💡 主要创新点

• 提出了一种结合深度神经网络声学模型与语音激活电源门控的可扩展语音识别架构，有效降低系统平均功耗。
• 集成了三种不同原理的VAD算法（能量、谐波、调制频率），允许根据应用场景权衡准确性与功耗。
• 通过优先保证VAD准确性而非单纯追求低功耗，减少了误唤醒对下游系统的影响，提升了用户体验。

🏷 关键词

📄 原文摘要

Analog Devices, Cambridge, MA 1 Previous work such as [4] provided micropower VADs that can be used in quiet environments or in applications that tolerate false alarms. In our application, false alarms will unnecessarily wake up a larger downstream system, increasing timeaveraged power consumption and impacting the user experience. Hence, we prioritize VAD accuracy, even if it results in larger area and power for the VAD itself. Our test chip provides three VAD algorithms–energy-based (EB), harmonicity (HM), and modulation frequency (MF)–allowing us to evaluate the interaction of algorithm and circuit performance. 2 The applications of speech interfaces, commonly used for search and personal assistants, are diversifying to include wearables, appliances, and robots. Hardware-accelerated automatic speech recognition (ASR) is needed for scenarios that are constrained by power, system complexity, or latency. Furthermore, a

👥 作者与机构

Michael Price1,2, James Glass1, Anantha P. Chandrakasan1

Massachusetts Institute of Technology, Cambridge, MA