We’ve open-sourced our 4B full-duplex speech model under Apache 2.0 !! Our model was trained on 40M hours and uses continuous tokens. Full-duplex AI for "Human like AI" is still largely unexplored. Let’s push the boundaries together! Check out our demos! 🚀 Website: Code: Model: Technical blog:

The swift implementation works inside JOSIE App, currently testing and optimising, I'll upload a clip with the finished model!!!!

This is huge fr 🙌

Continuous latents at 7.5 Hz instead of codec tokens is a bold call. Did that help most with the small "yeah"s and laughs, or with latency?

open weights on a full duplex model is great for phone agents, curious how it does with crosstalk on a noisy line

full-duplex is learned in the overlaps, and that's exactly where 13M hours separated from mono is the separator's guess, not what the speakers did

curious how you eval the full-duplex part, WER says nothing about whether it yields at the right moment when someone talks over it

So awesome, and thank you for making opensource! Does it support tool calling yet?

Apache 2.0 for a full-duplex model is a big deal for teams outside the US. The open question for me is how it does on languages that were barely in those 40M hours. That’s where open weights plus fine-tuning really pays off.

全二重音声だと、話し始めへの応答だけでなく、割り込みや言い直しで会話状態を崩さないかも試せるのが大きいですね。公開モデルなら、そうした挙動を用途に合わせて確かめられるのが良いです。

The separation model is doing fascinating work behind the conversational demo. If an overlap gets assigned to the wrong speaker, the model can learn the wrong social cue. Would love an ablation against the native two-channel fine-tuning, especially for laughter and backchannels.

Looking forward to seeing how this evolves

Can we port this to Mac and include it on Phosphene for all free and pro users, does the license apply?

releasing inference code with the weights makes independent testing much easier.