18μW SoC for near-microphone Keyword Spotting and Speaker Verification

The first fully-integrated near-microphone Keyword Spotting (KWS) and Speaker Verification (SV) solution, directly interfacing with a passive or active analog microphone and not requiring any external memory. The 65nm SoC realizes speaker-specific keyword triggering while only consuming 10.6μW average or 18.3μW peak for real-time operation, or 10x below speaker-agnostic keyword spotting SotA. Low cost, power and good accuracy (99.5% SV TIMIT and 98.5% KWS TIDIGITS) are jointly achieved through a.) a fully integrated single-chip standalone solution; b.) optimized accelerators; and c.) HW-aware algorithmic tuning and task scheduling.