ESP32 voice assistant: 8x wake-word model compression with AIMET

carrycooldude · x · 2026-09-01

Ashwin shared a project building a small voice assistant for ESP32. The wake-word model was compressed 8x using Qualcomm's AIMET. While naive 4-bit rounding caused a 3.7% accuracy loss, AIMET's AdaRound successfully recovered the performance. The wake words were generated by GradiumAI TTS in 22 different voices.

Original post →

More from Embodied

Embodied channel →