Voice AI fits in 470KB on an 80-cent RP2350

Moonshine Micro is an open source version of the Moonshine Voice toolkit built for microcontrollers and DSPs, with the Raspberry Pi RP2350 as its reference platform. That chip retails for 80 cents, and the full demo pipeline — voice-activity detection, speech-to-text, and neural text-to-speech — can run in as little as 470 KB of RAM.

The project is aimed at developers building real-time voice agents and applications on tightly constrained hardware. In the RP2350 demo, the full stack uses about 3.6 MiB of flash and ~468 KiB of provisioned SRAM, with the components running sequentially rather than all at once.

Here’s the memory and compute breakdown for the demo:

• VAD (Voice Activity Detection): ~89 KiB flash, ~36 KiB SRAM, ~0.8 MMAC/frame (~25 MMAC/s)
• STT (SpellingCNN Speech-to-Text): ~1.3 MiB flash, ~346 KiB SRAM, ~36 MMAC/s
• TTS (neural diphone synth at 16 kHz): ~1.8 MiB voice pack, ~340 KiB SRAM, ~37 MMAC typical reply (~65 MMAC/s out)
• Total demo pipeline: ~3.6 MiB flash, ~468 KiB SRAM, ~0.7–1.0 s to classify and speak

According to the project documentation, VAD, STT, and neural TTS share one ~384 KiB TensorFlow Lite Micro arena, so SRAM use is not additive. The source also notes that ~468 KiB is the total RAM provisioned on the 520 KiB RP2350, or ~491 KiB with wifi_hardware.

The code is released under the MIT License for commercial use, except for software in third-party/, which keeps its original open source licenses. The repository also includes an end-to-end example for setting up a Wi-Fi connection by voice on an RP2350 MCU.

Recommended reading

8BitDo’s retro desk trio nails the NES look