| Title |
A Multi-stationary Computing-in-memory Processor with Scalable-precision ADC for Keyword Spotting from Raw Audio Signal |
| Authors |
(Dongwook Kim) ; (Kyuho Jason Lee) |
| DOI |
https://doi.org/10.5573/JSTS.2026.26.5.407 |
| Keywords |
Keyword spotting (KWS); embedded DRAM (eDRAM); computing-in-memory (CIM); analog computing |
| Abstract |
Keyword spotting (KWS) is an essential technology for realizing always-on voice interfaces in edge devices, demanding extremely low-energy operation. Computing-in-Memory (CIM) architectures are promising for energy-efficient neural network processing by eliminating frequent external memory access. However, previous CIM-based KWS processors still suffer from low utilization in depthwise convolution (DWCONV) and significant energy consumption in analog-to-digital converters (ADCs). This paper presents an eDRAM-based CIM processor that achieves high integration density and ultra-low-energy end-to-end KWS operation through three key innovations. First, a frequency-selective stride convolution (FSSC) algorithm reduces the length of the raw audio input by 99.22%, thereby minimizing memory usage and computational complexity. Second, a multi-stationary CIM architecture maximizes hardware utilization in the DWCONV layer and reduces MAC energy by 95.26% by decreasing wordline driving activity. Third, a scalable-precision ADC optimizes capacitive digital-to-analog converter (CAP-DAC) switching energy according to the output range, achieving up to 75.71% lower conversion energy. As a result, the proposed processor achieves 95.1% accuracy on the 10-keyword Google Speech Commands dataset, with a macro-level energy efficiency of 34.8 TOPS/W and a system-level energy efficiency of 21.6 TOPS/W, consuming 0.62 μJ per decision |