Building noise-immune speech interfaces for IoT
summary
Clear Command is an edge neural stack that trains a unified noise-robust model for vocabularies up to about 100 short phrases, fits code+model+RAM in tens–hundreds of kilobytes on Arm Cortex‑M, and reports phrase error rates more than 4x better than cloud under the same noisy conditions. Models range ~100k–200k weights; an M7 demo runs trigger at ~70 MHz and command nets at ~180 MHz, while tiny variants hit <90 KB total and deliver 80–90% accuracy at 0 dB versus ~60% for cloud.
AI-written summary · about
chapters (9)
- 0:00The challenges of speech
- 2:38Analyzing noisy speech
- 5:54Limitations of current tech
- 10:11Recasting the problem
- 13:37Neural network architecture
- 19:37Live demo and performance
- 23:30Hardware implementation
- 25:56Design considerations
- 28:10Lessons and takeaways
related talks
58mVoice Control Made Simple – Panel Discussion
36mVoice Enabling Mobile Apps and Custom Hardware Devices
31mKeynote: The next big thing in machine learning is tiny
47mAI and Edge Computing: Efficient, Lean, Dependency‑Free Language Processing
11mLightning Talk: Accelerating Experimentation...
21mMultimodal Open Source at Kyutai, From Online Demos to On-Device
21mDeveloping Machine Learning Based Speech Enhancement Models for Teams and Skype
17mServing Voice AI at Scale