Speech, matching, and synthesis, all on-device

Say it however you like. It still works.

A voice agent that matches what you meant, not the exact words you used. Ask it to set up wifi any way you please and it runs the same multi-step conversation: speech recognition, phrase matching, and the voice answering you all running in this tab.

Start talking

Loading…

The conversation will appear here. Say “set up wifi”, or anything that means the same thing.

All of these start the same flow. None of them is the registered phrase except the first:

Or try your own wording

The flow, running live

This is the whole agent, in any of three languages. The highlighted line is the one waiting on you right now.

Why the wording does not matter

Registering a phrase does not register a string to match. Moonshine embeds your trigger phrases and whatever the user said into the same vector space and compares them by meaning, so “configure my wireless network” reaches a flow registered as “set up wifi”, sharing not one word with it, without you listing a single synonym. The same matching handles the answers inside the conversation, which is why “yeah, go ahead” counts as a yes. “Cancel” and “start over” work everywhere without being registered at all. It is a similarity score rather than magic, so a distant paraphrase can still fall short. How close is close enough is one number, triggerThreshold, which defaults to 0.7.