What if you could run production-quality speech recognition and text-to-speech on a chip that costs less than a cup of coffee? Moonshine Voice Micro does exactly that, packing real-time voice activity detection, speech-to-text, and neural TTS into just 470KB of RAM on a $0.80 Raspberry Pi RP2350 microcontroller. The project has pulled in 8,700+ GitHub stars and 476 forks, and for good reason: this is the first production-ready speech AI stack that runs entirely on-device at microcontroller prices.
The implications are hard to overstate. Every voice interface you have used today almost certainly sends your audio to a cloud server for processing. Moonshine Micro changes that equation by running everything locally on a chip that costs less than a dollar, with no API keys, no cloud calls, and no latency from network round trips. For founders building IoT devices, wearables, smart home products, or privacy-first assistants, this opens a design space that simply did not exist before.
What Moonshine Voice Micro Actually Does
The Moonshine project is an open-source AI toolkit for building real-time voice agents and applications, created by Moonshine AI. The full library covers speech-to-text, text-to-speech, intent recognition, voice cloning, speaker identification, and conversational agents across desktop, mobile, and embedded platforms. But the Micro variant is the headline-maker here because it targets the most constrained environment imaginable: microcontrollers with kilobytes of RAM and megahertz-class CPUs.
The Micro package includes three core components. Voice Activity Detection (VAD) consumes roughly 89KB of flash and 36KB of SRAM, using about 25 million multiply-accumulate operations per second to detect when someone is speaking. Speech-to-Text (STT) uses a custom SpellingCNN architecture that fits in about 1.3MB of flash and 346KB of SRAM, running at 36 million MAC/s. And the neural Text-to-Speech component generates speech using a diphone synthesis model with a 1.8MB voice pack, consuming about 340KB of SRAM and 65 million MAC/s during active output.
The components run sequentially and share a single TensorFlow Lite Micro arena of about 384KB, meaning the total RAM provisioned comes to roughly 468KB on the 520KB RP2350. Total flash usage for the full demo pipeline is about 3.6MB. These numbers are extraordinary for what they deliver: real-time, on-device speech recognition and synthesis with accuracy that beats OpenAI's Whisper Large V3.
How It Compares to Whisper
The comparison to Whisper is worth unpacking because it reveals how different Moonshine's design philosophy is from the dominant approach. OpenAI's Whisper models operate on a fixed 30-second input window, which wastes computation on zero-padding when processing short phrases that are typical in voice interfaces. Whisper also does not cache anything, so every call processes the full audio from scratch, even when only a short new addition has arrived.
Moonshine addresses both issues directly. Its models support flexible input windows of any length with no zero-padding required, and the streaming models cache input encoding and decoder state so that incremental additions skip redundant computation. The result is dramatic: Moonshine's Medium Streaming model (245 million parameters) achieves a 6.65% word error rate with 107ms latency on a MacBook Pro, while Whisper Large V3 (1.5 billion parameters, 6x the size) achieves 7.44% WER at 11,286ms latency. On a Raspberry Pi 5, Moonshine Medium runs in 802ms while Whisper Large V3 is marked as N/A because it simply cannot run at all.
Even Moonshine's Tiny Streaming model at 34 million parameters achieves 12% WER at 34ms latency -- competitive with Whisper Tiny's 12.81% WER at 277ms on a laptop, while being smaller and an order of magnitude faster.
Who This Is For
Moonshine Voice Micro is primarily aimed at three audiences. First, hardware founders building voice-controlled IoT devices, smart home products, wearables, or industrial equipment where a cloud dependency is unacceptable due to latency, privacy, or connectivity constraints. Second, embedded systems developers who want to add voice interfaces to existing microcontroller-based products without redesigning the hardware. And third, AI researchers and hobbyists exploring the limits of on-device intelligence.
The license is MIT for the code and English-language models, which means commercial use is straightforward. Non-English models are released under the Moonshine Community License, a non-commercial license, so founders targeting international markets should check the terms carefully. The project includes complete end-to-end examples for the RP2350, with documentation covering VAD, STT, custom word recognition training, and neural TTS.
For solo founders, the most interesting takeaway is the product surface area this opens. An on-device voice agent that costs less than a dollar in BOM, requires no cloud compute, and delivers sub-200ms latency makes voice interfaces viable for products that would never have been economically feasible before. Smart light switches that understand natural language commands, wearable assistants that never need internet connectivity, accessibility devices for users who cannot use touchscreens -- the hardware constraint that has kept these products niche just got a lot smaller.