You can now run full voice AI on an 80-cent microcontroller: Moonshine Micro delivers speech recognition and neural text-to-speech in under 500KB

Voice AI on an 80-Cent Chip: Moonshine Micro Fits a Full Pipeline in Under 500KB

An 80-cent microcontroller can now listen, understand spoken commands, and talk back. Moonshine AI has released Moonshine Micro, an open source toolkit that runs voice activity detection, speech-to-text command recognition, and neural text-to-speech synthesis entirely on a Raspberry Pi RP2350 microcontroller, which retails for roughly 80 cents 1. The full pipeline provisions 468KB of SRAM on a chip whose total budget is 520KB, with neural computations executed through TensorFlow Lite Micro 1.

Nothing here touches the cloud. There is no API call, no round-trip latency to a data center, no subscription billed by the minute. Audio is processed on-chip and never leaves the device.

That 468KB figure works because the three neural stages run sequentially rather than in parallel. Voice activity detection, speech-to-text, and text-to-speech time-share a single ~384KB TensorFlow Lite Micro arena, so their memory footprints do not stack 1. The TinyVadCNN voice activity detector peaks at roughly 36KB. The SpellingCNN speech-to-text model reaches about 346KB. The neural diphone synthesizer approaches 340KB 1. Since the largest consumer stays within the shared arena, the total holds at 468KB instead of climbing past 700KB.

Compute demands stay modest. During active processing, the pipeline sustains roughly 36 million multiply-accumulate operations per second for speech recognition and peaks near 65 MMAC/s during synthesis 1. A full classify-and-speak cycle completes in 0.7 to 1.0 seconds 1. Total flash storage for the demo pipeline, including the embedded neural voice pack, comes to about 3.6 MiB 1.

Each component ships as an independent library, so a developer who needs only wake-word detection can use VAD without importing the speech-to-text or TTS modules 1. The repository includes a complete end-to-end demo showing how to configure a WiFi connection on a microcontroller using voice alone 1. The code carries the MIT License, which permits commercial deployment 1. Developer traction is visible in the repository's GitHub star count, which sits near 10,000 1.

Building a voice interface used to mean choosing among three expensive options. A dedicated audio processor added several dollars to the bill of materials and often required its own firmware stack. A more powerful system-on-chip could run cloud-connected middleware but drew more power and cost more per unit. Or the product could maintain a persistent connection to a cloud speech API, billed by usage, with reliability tied to network quality. Each path inflated costs, increased power draw, and created a dependency on infrastructure the device maker did not control. An 80-cent microcontroller running a local neural pipeline clears all three at once.

Embedded engineers can now add voice input and spoken output to products where the economics never worked before. The toolkit targets microcontrollers and DSPs by design, with commodity silicon as the reference platform 1. Thermostat controls, accessibility switches, medical monitoring devices, industrial sensors: the cost threshold for conversational interaction drops from a budget line item to a rounding error. Products that ship in the millions on razor-thin margins can absorb voice as a standard feature rather than defending it as a premium upgrade.

Privacy is structural, not a toggle. When recognition runs on the chip and audio never leaves the device, there is no stream to intercept, no server storing voice data, no transcript log to subpoena. The attack surface for voice surveillance contracts to the physical hardware itself.

Moonshine Micro exists because neural models can now be compressed small enough to run on commodity microcontrollers. That compression is the same force reshaping edge AI across the industry: model footprints are shrinking while silicon costs stay flat, and the gap between dumb chips and intelligent ones is closing from both directions. Companies selling voice recognition as a recurring cloud service face a question they could not have asked a year ago. What happens when their core capability runs on an 80-cent chip that a customer buys once and owns forever?

The answer is already on GitHub, MIT-licensed, and running on a chip that costs less than the cable you would use to program it.

NewsArticle JSON-LD: <script type="application/ld+json"> { "@context": "https://schema.org", "@type": "NewsArticle", "headline": "Voice AI on an 80-Cent Chip: Moonshine Micro Fits a Full Pipeline in Under 500KB", "description": "Moonshine AI's open source Moonshine Micro toolkit runs voice activity detection, speech-to-text command recognition, and neural text-to-speech on a Raspberry Pi RP2350 microcontroller that costs roughly 80 cents, using under 500KB of RAM with zero cloud dependency.", "author": { "@type": "Organization", "name": "ProvenBrief" }, "publisher": { "@type": "Organization", "name": "ProvenBrief" }, "about": { "@type": "SoftwareApplication", "name": "Moonshine Micro", "applicationCategory": "DeveloperApplication", "operatingSystem": "Embedded" }, "keywords": ["voice AI", "edge AI", "microcontroller", "RP2350", "Moonshine Micro", "embedded systems", "text-to-speech", "speech recognition", "open source", "TensorFlow Lite Micro"] } </script>