Moonshine Micro, an open-source voice toolkit for microcontrollers, enables real time speech detection, transcription and synthesis on devices with as little as 470 KB of RAM, demonstrated on the Raspberry Pi RP2350.

The toolkit, released under the MIT License, includes components for voice activity detection, speech-to-text, custom word recognition and neural speech synthesis, all built on TensorFlow Lite Micro.

The demo runs on the RP2350 microcontroller, a low cost processor priced at 80 cents, and consumes approximately 3.6 MB of flash memory and 468 KB of provisioned SRAM. Voice activity detection uses about 89 KB flash and 36 KB SRAM with 0.8 million multiply-accumulate operations per frame. Speech-to-text requires 1.3 MB flash and 346 KB SRAM, operating at 36 million MACs per second. Neural speech synthesis consumes 1.8 MB of voice-pack data and 340 KB SRAM, delivering typical inference at 37 million MACs per reply (65 million MACs per second). The combined classification and speech pipeline completes in 0.7–1.0 seconds. These memory budgets fit within the RP2350’s 520 KB SRAM, with WiFi hardware using the remaining capacity.

The code repository provides a complete end-to-end example for establishing a WiFi connection using voice commands on the RP2350. Individual libraries for voice activity detection, speech-to-text, custom word recognition and neural speech synthesis can be used independently of one another.

Documentation sections cover voice activity detection, speech-to-text, custom word recognition, neural text-to-speech, and a WiFi setup example. The project is licensed under the MIT License, with third-party model licenses also under MIT.