Voice control toolkit comes to a Pico near you (Hackaday)

Julian Scheffers, Hackaday, 23 Jul 2026, on Moonshine AI‘s micro toolkit — a complete offline voice-control loop running on a Raspberry Pi Pico 2 W microcontroller (repo: moonshine-ai/moonshine/tree/main/micro). It is the far-edge endpoint of the spoke’s on-device thread: not a wrapper over a laptop-class engine like vocalinux, but the whole speech stack squeezed onto an MCU.

What it does

Three AI models chained into one loop: voice activity detection → STT (a “SpellingCNN”) → neural TTS. The runtime is C++: detect voice, listen for command tokens, process them, speak a reply, repeat. It handles up to 50 tokens and can be retrained for a specific vocabulary — i.e. a bounded keyword/command recognizer, not open-vocabulary dictation.

Why the footprint is the story

It uses 3.6 MiB of the 4 MiB flash and 468 KiB of SRAM, leaving almost nothing for anything else. Hackaday’s line — “fitting three different types of AI model needed to make this possible in such a space is quite impressive” — is the whole point: VAD, STT and TTS are each normally a service, and here all three run in single-digit-megabyte flash on a sub-$10 board with no network. That constraint is what makes it a distinct data point rather than another open-STT entry.

Where it sits

This extends the small/local pole the vocalinux page opened, one order of magnitude smaller. Vocalinux shows open engines replacing cloud dictation on a Linux desktop (vosk on a Pi was already the low end); Moonshine micro shows a purpose-built tiny stack on a microcontroller, where the design trades open-vocabulary accuracy for a 50-token retrainable command set that fits. It also makes moonshine (already the “edge, from 27M params” entry in open-source-stt-models) a paged subject, and is the spoke’s first instance of neural TTS on an MCU and of on-device VAD.

Caveats

Secondary hardware-hobbyist coverage (Hackaday), so T3 — no WER, no latency, no accuracy numbers, and the license is not stated in the article (the repo would resolve it). The impressive claim is the fit, which is concrete; recognition quality on a 50-token SpellingCNN at this size is unmeasured here and would be the thing to check before treating it as usable rather than a demo.

Related

moonshine · vocalinux · speech-to-text · text-to-speech · open-source-stt-models · speech-audio-ai