Thanks! There is a model that turns the voice into an embedding that is used to determine the voice. Unlike the STT and TTS, we won&#x27;t be releasing the weights of this voice cloning model, but we will provide it over an API so that we can do verification and prevent abuse.edit: Ah yes, and we do not store the voice sample on our server. The voice embedding is cached for 24 hours.

I thought this was fantastic! Surprised not more people are commenting on this. Is there a reason I am not aware of?To the author: what happens to my voice after I upload it?
What is your plan moving forward? I am too far left field to understand how to build a business and monetize an open source product like this, even though I found it fun to play around with.

Thank you! The TTS and ASR are our own unreleased models, but we&#x27;ll open-source them soon :)The latency is about 500ms once we detect that it&#x27;s the bot&#x27;s turn to speak (roughly 200ms for the LLM&#x27;s time-to-first token and 300ms for the TTS audio to start), plus a variable time for the semantic pause detection (VAD).If it&#x27;s clear that you&#x27;re done talking, like when you ask a question, the model will reply very fast. If you stop mid-sentence as if you have more to say, it will wait for longer to avoid interrupting you.

Way more entertaining than I would expect!
What TTS and ASR models do you use? What sort of latency do you get?

Incredible work. Short, sweet and simple. I hadn&#x27;t expected to enjoy this as much as I did. I can&#x27;t wait to see where it goes.

can&#x27;t wait for the open source release.

Upload a voice and write a personality prompt, or try the pre-made characters.Built by augmenting Gemma 3 12B with our new text-to-speech and speech-to-text models, both of which we will release as open-source soon. Stay tuned.

Show HN: Make your own voice AI in two clicks