A real-time AI avatar that answers you before the silence gets awkward is finally something you can poke at yourself. Nuance Labs — a US startup built largely by former Apple researchers — has released Nuplex 1.0, a full-duplex audiovisual model, and put Emma, a talking video avatar running on top of it, on the open web. No sign-up, no account, no waiting list.

What happened

Two things shipped at once. Nuplex 1.0 is the engine: one neural network that handles sight and sound in both directions. Emma is the face you actually talk to — an interactive video avatar that holds a natural back-and-forth in real time inside a browser tab.

The company is calling Nuplex 1.0 a base model and Emma a public demo of it, and it's opened preliminary access to a Nuplex Live API for developers. According to Unite.AI, Nuance Labs has also closed a $50M Series A with Nvidia taking part, which tells you serious money thinks duplex video conversation is a category rather than a party trick. Venture capital trackers list the company alongside that round.

The walkie-talkie problem every other avatar has

Most digital humans today are built like a relay race. Your voice gets transcribed by a speech-to-text model. The text goes to a language model. The answer gets read aloud by text-to-speech. Then a fourth model animates the lips to match. Four hops, four chances to stall, and roughly 1.5 to 3 seconds of combined lag.

That gap is why so many avatars sit there with a glassy stare while you finish your sentence. They aren't thinking. They're waiting for their turn.

Nuplex deletes the relay. Perception and generation of audio and video live in a single end-to-end network, so the model reads tone, pauses and the small hesitations in your voice continuously and streams matching speech and facial animation at the same time.

Emma nods while you're still talking

This is the part that feels genuinely different in use. Emma can listen and react simultaneously — nodding, smiling, dropping in the little "mm-hmm" and "got it" noises people make, and changing expression mid-sentence instead of after it.

Nuance Labs ran a blind benchmark across 1,785 test questions and says Nuplex matched the reaction speed of a live human in 89% of them. None of that proves Emma is a great conversationalist. But reaction speed is precisely what used to break the illusion.

What it means for you

Forget the architecture for a second. Here's where a face that answers in real time actually earns its keep.

At home

Language practice is the obvious one. Speaking drills fall apart with a chatbot that pauses for two seconds every turn, because you lose the rhythm of real conversation. A roommate-style avatar you can interrupt and be interrupted by is closer to the real thing. Same for a kid who wants a question explained five different ways without feeling like a nuisance.

At work

Rehearsal is quietly the biggest use case. Run a job interview out loud, pitch a deck to a face that reacts to your pauses, or practice a difficult conversation before the real one. Watching an avatar nod when you finally say the clear version of your point is brutal, useful feedback.

For business

A front-desk explainer on your site that can handle interruptions instead of dumping you into a scripted FAQ tree changes the feel of a landing page. Onboarding walkthroughs and product demos also get better when the presenter reacts to the viewer instead of reciting at them. Before you buy anything fancy, though, it's worth testing whether your idea even needs a face — the free AI tools at MyKreaTool will tell you fast and cheap what a plain chatbot can already do.

For study

There's solid evidence behind explaining things out loud to learn them. A patient avatar gives you something to explain to, on demand, at 11pm, without a tutor bill. The back-and-forth signals also keep you from zoning out the way you do watching a recorded lecture.

For creativity

Writers can hear dialogue spoken back instead of mouthing it silently. Animators and game teams can prototype a character's timing before committing to a rig. Table reads, storyboards, voice-first sketches — anything where the rhythm matters more than the polish gets faster.

For extra income

Talking-head explainers, localized product walkthroughs and interactive training clips are all easier to produce when the presenter doesn't need to be re-recorded for every tweak. Just stay on the right side of the rules — more on that below.

How to try it right now

It takes about a minute, and the free path comes first.

1. Open Emma at app.nuancelabs.ai/try in a desktop browser. Chrome or Safari on a laptop is the sane choice for a first run.

2. Skip any sign-up. There isn't one. You land straight in the demo.

3. Allow microphone access when the browser asks. Without it you can watch, but you can't talk.

4. Say a full sentence and keep going. Don't pause politely. The whole point is what happens while you're still speaking.

5. Interrupt it mid-answer. Talk over Emma, change your mind, ask something else. This is the honest test — the old cascade models visibly stumble here.

6. Try one question with no easy answer. You'll see where the seams are, which is more useful than another smooth demo.

7. If you build things, request provisional access to the Nuplex Live API through Nuance Labs.

Upsides and what changes

The shift isn't "avatars look more real now." It's that the turn-taking ritual disappears. Once a system can listen and speak at the same time, you stop managing it — no more waiting for the green light before you talk — and that's what moves these things from demo clip to daily tool.

Second, removing the four-model relay removes four failure points. Less to break, less to tune, cheaper to run at scale, easier for a small team to ship something that feels alive.

Third, the guardrails went in at launch rather than after a scandal. Emma's appearance is licensed from a real actress, outputs carry mandatory watermarks, and cloning a third party's face or voice is blocked outright. That's a sensible line, and it's refreshing to see it drawn on day one.

Limitations

Be clear-eyed about the fine print. Emma is a public demo of a base model, not a finished product — expect it to lose the thread on long, messy, multi-part questions and to be less charming than a 30-second clip suggests. The headline number is a company-run blind benchmark, not a peer-reviewed study, and 89% still means roughly one turn in nine lands off-pace; on a flaky connection that ratio will feel worse. Nuance Labs hasn't published how Emma handles heavy accents, crosstalk or a noisy room, and every one of these systems gets shakier when two people talk at once. Add the licensing limits: you can't drop in your own face or someone else's voice, so this isn't a shortcut to a synthetic spokesperson. Treat it as a genuinely impressive preview of where conversational interfaces are heading, and test it on your own use case before you build a business plan around the demo.

Conclusion

Nuplex 1.0 matters less as a model release and more as a proof point: the 1.5-to-3-second freeze that made every AI avatar feel like a walkie-talkie call has a real answer, and it's sitting in a browser tab with no login screen in front of it.

Your one action for today: open Emma, start a sentence, and deliberately cut in halfway through its reply. If it keeps its footing, you'll immediately understand what all the fuss is about — and if it doesn't, you've learned exactly which question to ask the next vendor who pitches you an avatar. Either way, you'll know in five minutes what took a research paper to explain last year.