Speech Recognition and TTS in less than 500kb

Moonshine Voice is an open source AI toolkit for developers building real-time voice agents and applications. Moonshine Micro is a version designed specifically for embedded system processors like microcontrollers and DSPs, and uses the Raspberry Pi RP2350, which retails for just 80 cents, as its reference platform. It includes voice-activity detection, command recognition, and neural speech synthesis and can run in as little as 470 KB of RAM.
You can see a full walkthrough in the video below:
The memory and compute requirements are designed to fit resource-constrained systems. Figures below are for the RP2350 demo; the detailed memory budget breaks each one down:
| Component | Flash | SRAM (arena peak) | Compute |
|---|---|---|---|
| VAD (Voice Activity Detection) | ~89 KiB | ~36 KiB | ~0.8 MMAC/frame (~25 MMAC/s) |
| STT (SpellingCNN Speech-to-Text) | ~1.3 MiB | ~346 KiB | ~36 MMAC/s |
| TTS (neural diphone synth @ 16 kHz) | ~1.8 MiB voice pack | ~340 KiB | ~37 MMAC typical reply (~65 MMAC/s out) |
| TOTAL (Demo pipeline) | ~3.6 MiB | ~468 KiB provisioned* | classify + speak ~0.7–1.0 s |
Notes:
- Flash is
.text
+.rodata
measured witharm-none-eabi-size
on the defaultmoonshine_micro_echo
firmware (includes the embedded neural voice pack); SRAM is.bss
+ heap + stacks. - *VAD, STT, and neural TTS run sequentially and time-share one ~384 KiB TFLM arena, so SRAM is not additive — ~468 KiB is the total RAM provisioned on the 520 KiB RP2350 (
wifi_hardware
~491 KiB). - A MAC is one multiply-accumulate; MMAC/s = millions per second during the active (non-idle) stage.
The code is released under the permissive MIT License, suitable for commercial applications.
There's a complete end-to-end example showing how to set up a wifi connection on a microcontroller using voice on an RP2350 MCU.
The VAD, STT, and TTS libraries can be used independently of each other, relying on the included TensorFlow Lite Micro library for the neural computations.
- Voice Activity Detection
- Speech to Text
- Custom Word Recognition
- Neural Text to Speech
- Wifi Setup Example
This code, apart from the source in third-party/
, is licensed under the MIT
License — see LICENSE in this directory (also at the repository root).
The SpellingCNN and TinyVadCNN models in models/
are released under
the MIT License.
The code in third-party/
is licensed according to the terms of the open
source projects it originates from, with details in a LICENSE file in each
subfolder.

Facts Only

* Moonshine Voice is an open source AI toolkit for voice agents.
* Moonshine Micro is a version for embedded system processors, referencing the Raspberry Pi RP2350.
* It includes voice-activity detection, command recognition, and neural speech synthesis.
* The demonstration pipeline requires ~3.6 MiB of Flash memory and ~468 KiB of provisioned SRAM on the RP2350.
* VAD requires ~36 KiB of SRAM.
* STT uses ~346 KiB of SRAM.
* TTS requires ~340 KiB of SRAM.
* The total RAM provisioned for the demo pipeline is approximately 468 KiB on a 520 KiB RP2350.
* Flash memory usage for the demo pipeline is ~3.6 MiB.
* The code is licensed under the MIT License.

Executive Summary

Moonshine Voice is an open source AI toolkit designed for developing real-time voice agents and applications. Moonshine Micro is a specific version targeting embedded systems, utilizing the Raspberry Pi RP2350 as its reference platform. It incorporates voice-activity detection, command recognition, and neural speech synthesis capabilities. The design emphasizes low resource usage, allowing it to operate with as little as 470 KB of RAM. Performance metrics for the demo pipeline show that Voice Activity Detection requires approximately 0.8 MMAC/frame, while Speech-to-Text utilizes 36 MMAC/s, and the neural Text-to-Speech component operates at a typical reply rate of 37 MMAC, resulting in a total processing time of 0.7 to 1.0 seconds. The memory provisioned for the demo pipeline totals approximately 468 KiB of SRAM on the 520 KiB RP2350.

Full Take

The architecture of Moonshine Micro demonstrates a clear commitment to making advanced AI capabilities accessible on severely constrained hardware, leveraging concepts like TensorFlow Lite Micro to achieve this efficiency. The partitioning of memory across VAD, STT, and TTS functions highlights an engineering trade-off where functionality is achieved through careful time-sharing of a shared TFLM arena rather than simple additive memory allocation, which is a sophisticated approach to resource management in embedded systems. This pattern suggests a paradigm shift toward tightly integrated, highly optimized, low-overhead models for edge AI deployment. The existence of a complete end-to-end example and independent library components allows for flexibility, suggesting an intent to foster broad application rather than locking users into a single monolithic system. The implication is that real-time, complex voice processing can be democratized beyond high-end compute environments, pushing the boundary of what constitutes viable AI functionality on microcontrollers. What limits this realization are the real-world constraints on model complexity versus the physical limitations of the target processors; subsequent work must explore how to further compress these neural components without sacrificing necessary acoustic fidelity or latency.

Sentinel — Human

Confidence

The text reads like a detailed technical release or documentation, exhibiting high internal coherence and specificity consistent with human-written source material, likely from an open-source project context.

Signals Detected

Moderate sentence length variance; technical, detailed prose with some slightly awkward phrasing in the notes.

High internal consistency; flows logically from description to hardware specs to data tables. Lacks overt passion or hedging.

Clear, structured presentation of technical information and licensing details typical of open-source projects.

Claims about specific hardware components (RP2350 price, memory footprints) are highly specific and presented as factual claims related to an open-source tool release.

Human Indicators

Specific, deep knowledge of embedded systems metrics (MMAC/s, TFLM arena), which suggests domain expertise.

Detailed breakdown of memory allocation with specific paths (`.text`, `.rodata`) indicative of careful implementation detail.