Machine-Learning-Based SAR ADC Featuring Smart Range Detection for Portable Voice-Activated IoT Devices

This article presents a novel machine-learning (ML)-based successive approximation register (SAR) analog-to-digital converter (ADC) designed for energy-constrained voice-activated Internet of Things (IoT) applications, such as portable voice assistants, wearable health monitors, and smart home devices. Voice interfaces in these devices require always-on operation while maintaining ultralow-power consumption to preserve battery life. The proposed ADC integrates a ML model into the SAR ADC search process to significantly reduce comparator activity and digital-to-analog converter (DAC) switching energy. Our approach leverages ML to predict the first-order difference between an unknown sample and its previously digitized sample, dynamically adjusting the SAR ADC search range and minimizing unnecessary conversions. To further optimize energy efficiency, the proposed method incorporates a dual-stream strategy: one optimized for signals with smaller ML prediction errors, and the other for signals with larger ML prediction errors. A time comparator-based error detection mechanism is introduced to adaptively select the most power-efficient search stream. These enhancements lead to significant power savings while maintaining high conversion accuracy. Measurement results from a fabricated 10-bit SAR ADC chip demonstrate a 70.05% reduction in comparator activity compared to the conventional monotonic approach. With a 5.6 kHz sinusoidal input, the ADC achieves an signal-to-noise and distortion ratio of 57.78 dB and an spurious-free dynamic range of 67.06 dB. The total power dissipation remains below $0.54~\mu $ W, with an additional estimation of 0.22 $\mu $ W for the digital part, making it an ideal solution to the always-on voice interfaces in battery-operated IoT devices.