We're rolling out to a new batch of beta testers! I think we were able to improve pretty much everything from what we learned.
In a nutshell:
Documentation and videos on https://doc.voxta.ai/
Virt-A-Mate scene was expanded: more actions, more details, better prompting, better animations.
Python is embedded, so we can add locally hosted packages like faster_whisper. It's important, both from a cost and privacy perspective, to allow using as much local services as we can. This was an important step to achieve that.
Prompt and prompt formatting have been greatly improved, you can expect much better results out of the box.
Action inference has been revamped and is now much more powerful, and works perfectly fine on small 13B models! We also support two new concepts, silent action inference (when the AI cannot speak, they can still act) and user function calling (not available in the beta, allows you to request commands to execute actions such as searching the internet).
Characters can now have voice fallback, so sharing a character has now more chances to work for everyone.
We added support for Deepgram, which is much cheaper than Azure and much better than Vosk. We also added silence detection to reduce the invoice of paid STT.
We support exporting Tavern V2 character cards, so expect characters sharing, with thumbnail!
There is a lot more, you'll find the release notes in the release channels on Discord. But we're not done. We want to stay focused on getting this core part right, but then we'll share more information about our tentative roadmap. There is so much we can do!
❤️ And big, big thanks to everyone who supports us, either through Patreon, by sharing your ideas or by providing feedback on the beta. We're pouring our hearts and souls into this, we're more than happy to see there's interest in what we're building!