Moonshine

Moonshine is a family of edge / on-device STT models from Moonshine AI, built for speed and small footprint rather than leaderboard WER — noted in open-source-stt-models as an edge option from 27M parameters, the low end of the size axis where whisper Large is 1.55B. It anchors the small/local pole of this spoke’s STT field: “best” means runs on my hardware, offline, not lowest error rate.

The micro toolkit

Moonshine AI’s most extreme deployment is micro (moonshine-pico-voice-toolkit): a full offline voice-control loop — VAD → a SpellingCNN STT → neural TTS — running on a Raspberry Pi Pico 2 W in ~3.6 MiB flash / 468 KiB SRAM, with a retrainable 50-token command vocabulary. This is where the family’s design logic shows: at MCU scale you don’t ship open-vocabulary ASR, you ship a small retrainable recognizer that fits.

Placement

Where canary-qwen / whisper compete on Open ASR Leaderboard WER (LLM-fused, billions of params), Moonshine competes on fitting — the same axis vosk and whisper.cpp occupy on laptops, pushed down to microcontrollers. Provider maker: Moonshine AI (not separately paged — no distinct evidence beyond the model line and toolkit).

Related

moonshine-pico-voice-toolkit · speech-to-text · open-source-stt-models · vosk · whisper-cpp · vocalinux