Show HN: Wake-word spotter with a constant 5 KB state, running in the browser

Category: ai-ml

Tags: wake-word, keyword-spotting, streaming-inference, onnx, embedded, freertos, recurrent-neural-networks

Score: 6.5/10 (Innovation: 7, Technical: 7, Documentation: 8, Utility: 4)

O1-Sound is a multilingual wake-word spotter built on a 'liquid core' recurrent architecture with per-channel learnable time constants, achieving a constant 5,120-byte streaming state regardless of session length. The project stands out for its unusually honest documentation, including negative results, failed multilingual claims, and validated ONNX/C parity, but remains a research artifact rather than a production wake-word engine.

Target audience: ML researchers, embedded/edge AI engineers

Repository: https://awareliquid-o1-sound-demo.static.hf.space/index.html · Python · MIT

View on Hacker News