Moonshine Micro: Full Voice AI on a 500KB Microcontroller
Speech Recognition and TTS in less than 500kb
I present Moonshine Micro, an open-source toolkit enabling real-time voice interfaces on ultra-low-cost microcontrollers like the Raspberry Pi RP2350. This system integrates voice-activity detection, speech-to-text, and neural text-to-speech, running entirely within 470 KB of RAM. By leveraging TensorFlow Lite Micro, developers can now build sophisticated voice agents on hardware costing just 80 cents.
"It includes voice-activity detection, command recognition, and neural speech synthesis and can run in as little as 470 KB of RAM."
HN discussion
- One practitioner notes that while TTS in a small footprint is feasible, achieving high accuracy is the true challenge, though lower accuracy may suffice for the targeted use cases.
- A developer who trained a sub-500KB STT model for sub-dollar chips reports that these systems are often extremely unstable, citing specific confusion between phonetically similar words like 'p' and 't'.
- Commenters highlight that Voice Activity Detection (VAD) alone is a compelling feature with minimal performance impact, making it viable for real-time browser-based speaker highlighting across devices.
- Users compare the project to legacy tools like Flite and Microsoft Sam, noting that while traditional formant/tract modeling synths are accurate in tiny compute, they sound distinctly robotic compared to this neural approach.