I made a little python wrapper around it to serve an HTTP endpoint that’s OpenAI/elevenlabs compatible https://github.com/clayrosenthal/bootlegger

Given the tiny size of this, I wonder about possible future integration with esphome compatible hardware

So at that tiny 500kb size I imagine it could be compiled to web assembly, and run entirely in the browser right?

Couldn’t find a link, is that hard to do?

Quick link to the video where he demos it: https://www.youtube.com/watch?v=kMliOFYBiz4

Wow, it seems like this might beat out flite for very-low-memory TTS? I ended up abandoning a project of mine because I couldn't get high enough quality or low enough memory usage out of flite, so I'm very excited to try this out.

Flite for comparison: https://github.com/festvox/flite

I installed the command line version using uv

    uv init
    uv add moonshine-voice
    uv run moonshine-voice mic --language en

super nice to be able to run it to test it like this

good job on a clear readme.md tbh

For TTS I wonder how this compares to nanotts[1] with the en-GB voice, which is sort of unreasonably good.

[1] https://github.com/gmn/nanotts

This looks like an extreme point for AI-based TTS, as formant/tract modeling synths tend to be more accurate if you want TTS in a tiny amount of compute, but sound distinctly robotic.

TTS (neural diphone synth @ 16 kHz) ~1.8 MiB voice pack

This is in the realm of Microsoft Sam.

Do you have any accuracy benchmarks?

I’ve worked in this space. TTS in a small footprint isn’t the hard part —- it’s doing it accurately that’s hard.

Although for the use cases OP is targeting, lower accuracy may be good enough!

The voice activity detection alone here is compelling - very useful for doing things like highlighting a speaker who's transmitting in realtime. At that rate the impact on perf will be so minimal that you could easily run it in the browser across devices.

It looks great, thank you! I'll see if I can use it for my in browser AI assistant project's ( https://aidekin.com ) voice part. It's currently using Nemotron-3.5-ASR and supertonic-3 but overall it requires 1.2gb download.

this is good to see. i also trained a stt under 500kb for sub dollar chips. it had about 20 words that it could understand(like start, stop, left, right, go, up etc) and then the spell mode where you could say the word spell and then say the individual english alphabets and close with spell. it was super fun to work on. these tend to be extremely unstable though, like confusion between p and t (at least for my accent). will have to try this one now.

Very cool. I've done TTS on a 32K Arduino but it was pretty croaky. https://youtu.be/ErGDboTpwM0

Thank you for this. I love your work on Curb Your Enthusiasm.

ngl, it looks incredible

Flite for comparison: https://github.com/festvox/flite

For TTS I wonder how this compares to nanotts[1] with the en-GB voice, which is sort of unreasonably good.

[1] https://github.com/gmn/nanotts

I made a little python wrapper around it to serve an HTTP endpoint that’s OpenAI/elevenlabs compatible https://github.com/clayrosenthal/bootlegger

Given the tiny size of this, I wonder about possible future integration with esphome compatible hardware

https://esphome.io/

Quick link to the video where he demos it: https://www.youtube.com/watch?v=kMliOFYBiz4

Very cool. I've done TTS on a 32K Arduino but it was pretty croaky. https://youtu.be/ErGDboTpwM0

ngl, it looks incredible

So at that tiny 500kb size I imagine it could be compiled to web assembly, and run entirely in the browser right?

Couldn’t find a link, is that hard to do?

500k memory but not sure about disk.

Do you have any accuracy benchmarks?

I’ve worked in this space. TTS in a small footprint isn’t the hard part —- it’s doing it accurately that’s hard.

Although for the use cases OP is targeting, lower accuracy may be good enough!

> I’ve worked in this space. TTS in a small footprint isn’t the hard part —- it’s doing it accurately that’s hard.

This actually holds for everything in AI.

If you look at this chart here it seems the tiny model has a WER of ~12%… not sure about the micro model:

https://github.com/moonshine-ai/moonshine#when-should-you-ch...

This looks like an extreme point for AI-based TTS, as formant/tract modeling synths tend to be more accurate if you want TTS in a tiny amount of compute, but sound distinctly robotic.

TTS (neural diphone synth @ 16 kHz) ~1.8 MiB voice pack

This is in the realm of Microsoft Sam.

I installed the command line version using uv

    uv init
    uv add moonshine-voice
    uv run moonshine-voice mic --language en

super nice to be able to run it to test it like this

good job on a clear readme.md tbh

`uvx moonshine-voice mic --language en` That is even simpler.

Thank you for this. I love your work on Curb Your Enthusiasm.

I remember someone training smart kettle to use its speaker as microphone

IIRC the Alexa enabled voice remotes also used a similarly small model though perhaps not this small

500k memory but not sure about disk.

`uvx moonshine-voice mic --language en` That is even simpler.

IIRC the Alexa enabled voice remotes also used a similarly small model though perhaps not this small

I remember someone training smart kettle to use its speaker as microphone

If you look at this chart here it seems the tiny model has a WER of ~12%… not sure about the micro model:

https://github.com/moonshine-ai/moonshine#when-should-you-ch...

That's the error rate for STT, not TTS. TTS is generally easier than STT because you only need to produce one valid pronunciation and don't need to handle variation within and between individuals.

> I’ve worked in this space. TTS in a small footprint isn’t the hard part —- it’s doing it accurately that’s hard.

This actually holds for everything in AI.

Possibly Cousin Andy? https://curb-your-enthusiasm.fandom.com/wiki/Andy_David

Played by the great Richard Kind, who my wife swears she saw on the Highline in NYC.

Probably a joke that the author looks like someone on the show? I'm puzzled as well.

That's the error rate for STT, not TTS. TTS is generally easier than STT because you only need to produce one valid pronunciation and don't need to handle variation within and between individuals.

Possibly Cousin Andy? https://curb-your-enthusiasm.fandom.com/wiki/Andy_David

Played by the great Richard Kind, who my wife swears she saw on the Highline in NYC.

Probably a joke that the author looks like someone on the show? I'm puzzled as well.

Moonshine Micro — Voice Interfaces for Microcontrollers

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 with arm-none-eabi-size on the default moonshine_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.

Documentation

License

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.

Hacker Times

Hacker Times

Speech Recognition and TTS in less than 500kb

Discussion

Discussion

Moonshine Micro — Voice Interfaces for Microcontrollers

Documentation

License