Energy-Quality Scalable Analog-to-Digital Conversion and Machine Learning Engine in a 51.9 nJ/frame Voice Activity Detector
November 27, 2019
Publication
- Energy-Quality Scalable Analog-to-Digital Conversion and Machine Learning Engine in a 51.9 nJ/frame Voice Activity Detector
- November 27, 2019
In this work, voice activity detection (VAD) system
with dynamic energy-quality (EQ) scalability is presented. EQ
scalability is enabled through the insertion of multiple knobs at
different levels of the signal chain, starting from analog-digital
conversion and ending at the classification stage. Such knobs are
co-optimized at runtime to achieve a given quality target with
minimal energy. Such co-optimization is also shown to improve
the fit of the machine learning algorithm, allowing for more
graceful quality degradation. The proposed system, fabricated on
a 28nm test chip, classifies at 81.2% accuracy while consuming
51.9 nJ/frame in a 10dB noise context. Scaling up energy
consumption by 3.5x improves accuracy by 5.2%.
Researcher / Author: Jinq Horng Teo, Shuai Cheng and Massimo Alioto
IEEE (2019), 978-1-7281-0996-1/19; https://ieeexplore.ieee.org/document/8964767