Moonshine Micro runs voice recognition and TTS on-device in 470KB of RAM

Moonshine AI has released Moonshine Micro, an open-source voice toolkit trimmed down to run entirely on microcontrollers and DSPs rather than in the cloud. Its reference target is the Raspberry Pi RP2350, an 80-cent chip, and the full pipeline — voice-activity detection, speech-to-text, custom word recognition, and neural text-to-speech — fits in as little as 470KB of RAM. The trick to that footprint is sequencing: VAD, STT, and TTS never run at once, so they time-share a single ~384KiB TensorFlow Lite Micro arena instead of each claiming its own memory, keeping the total provisioned RAM around 468KiB on the RP2350’s 520KiB budget.

The three libraries are decoupled and can be used independently, each leaning on the bundled TFLM runtime for the neural math. Moonshine ships a complete end-to-end demo showing a device being configured for WiFi purely through voice, illustrating the intended use case: interactive voice interfaces on hardware too constrained for a network round-trip. Everything outside the third-party directory, including the SpellingCNN and TinyVadCNN models, is MIT-licensed and cleared for commercial use.

The significance is less about raw accuracy than about where the inference happens. Pushing a usable speech stack onto sub-dollar silicon means voice features can run locally on embedded devices without streaming audio to a server, which cuts latency and cloud dependency while keeping captured speech on the device rather than in a vendor’s data center — a meaningful shift for privacy-sensitive and offline products.