Launching Moonshine Micro

Long-time readers will know I’m convinced local voice interfaces and sub-$1 embedded chips will fundamentally change how we interact with everything in the physical world. That’s why I’m so excited to introduce Moonshine Micro, a version of the Moonshine Voice open source framework that can run a useful voice interface in just 520KB of RAM. It contains separate libraries for voice-activity detection, speech to text, and text to speech, all powered by tiny neural networks with an example bringing them all together on an 80 cent Raspberry Pi RP2350 chip.

I’m still working towards the end goal of the moonshot I started at Google Brain in 2017, a full ASR and TTS system on a 50 cent chip that can run on a coin battery for a year, but this is a big milestone on the journey. This release runs a 50-word command recognizer, that’s fully trainable for custom words, and a neural network-based text to speech engine, and can be used to set up a wifi connection. There’s still a lot of work to do to increase the scope of the recognition to full speech, rather than individual words, increase the text to speech quality, and to offer advanced intent recognition on this kind of system, but with the hardware improvements that are likely to come over the next few years, I think we’re getting a lot closer.

I’m looking forward to seeing applications I’d never thought of for this technology, so if you build something neat please tag me on Hackster, and for questions or issues let me know on GitHub.