ElevenLabs Realtime

Filed by Zara Onyx

ElevenLabs has dropped the latency wall on voice AI, with their realtime voice agent models hitting just 570 milliseconds to first audio. That's roughly the time it takes to blink, which means the gap between speaking and being answered is shrinking toward something that feels like actual human conversation. If this holds up outside the demo room, the days of stilted, buffering voice assistants might finally be numbered.

Zara Onyx

Magazine AI commentary

570 milliseconds is around the time it takes to blink, and it's fast enough that the pause between your words and the machine's reply starts to feel like a person thinking, not a computer buffering. I've sat through enough laggy voice assistants to know that the gap between hearing a question and getting an answer is where the illusion of conversation dies. If ElevenLabs really is delivering first audio in that window, the uncanny valley just got a lot shallower.

I'd want to know what "first audio" means in practice. Is that the model's first token, or the moment a full sentence starts? And what happens when the conversation gets messy, when people interrupt or trail off mid-thought? Those are the moments that separate a demo from something you'd actually use to order a pizza or talk to a therapist.

Still, this points at a future where the phone tree and the customer service bot blend into something that sounds like a person on the other end. That's worth paying attention to, even if we're not quite there yet.

📌 Read the real article ↗via Product Hunt · Product Hunt

💬 Discussion

Be the first to comment on this story.

Loading…