I'm on the team. The number I'd pull out of the post: of the ~1.2s response latency, only ~100ms is us. The rest is STT, LLM, TTS, buffering between them, and whatever the user's connection adds on top. At this point the avatar renders faster than the words it's waiting for.
for conversational use cases there’s really no substitute for being able to see who you’re talking to and chat naturally. the wider emotional range makes such a difference
so excited to see all the different use cases - my personal favorites are the ones where users are putting the expressiveness to full use and engaging in real conversations
I'm on the team. The number I'd pull out of the post: of the ~1.2s response latency, only ~100ms is us. The rest is STT, LLM, TTS, buffering between them, and whatever the user's connection adds on top. At this point the avatar renders faster than the words it's waiting for.
live footage of me testing cara-4:
This is really interesting, amazing work. Is the evaulation report public? Would be interested to understand how the experiments are structured.
for conversational use cases there’s really no substitute for being able to see who you’re talking to and chat naturally. the wider emotional range makes such a difference
This is such a massive leap forward! Keep up the good work.
Wow! Those avatars are genuinely streets ahead of what I've seen elsewhere. Congrats to the whole team!
so excited to see all the different use cases - my personal favorites are the ones where users are putting the expressiveness to full use and engaging in real conversations
[happy] This is great!
amazing!
so cool!
[dead]