Nuance Labs launches Emma, an avatar that nods and reacts while you are still speaking

Almost every video avatar you have talked to works like a walkie-talkie. You speak, the system waits for silence, transcribes your speech, sends the text to a language model, voices the reply, and only then does a separate model move the lips. A delay of 1.5–3 seconds is unavoidable in this setup. While the avatar is "thinking," it stares into space, and the feeling of a conversation disappears.

Nuance Labs, a US startup founded by former Apple researchers, is trying to get rid of the walkie-talkie itself. This week the company unveiled Nuplex 1.0, a foundational audiovisual model, and launched the Emma avatar.

What changed in the architecture

Instead of a chain of four models, Nuplex is a single end-to-end neural network. It perceives and generates audio and video together. The system continuously reads the tone of voice, pauses and hesitations of the person speaking, and streams speech and facial expressions in parallel.

The industry calls this full duplex: you can listen and speak at the same time, as in a real conversation. Until now this was mostly found in voice models; here a face has been added to it.

What it looks like in practice

Imagine telling Emma about a problem with your order. With a typical avatar, you talk into the void until you stop. Emma nods, changes her facial expression and gives active-listening signals such as "uh-huh" and "I see." All of this happens while you are speaking, not after. For support, training and onboarding services and for virtual conversation partners, this is the difference between "talking to a form" and "talking to a person."

In Nuance Labs' internal blind benchmark of 1,785 test questions, Nuplex matched the response speed of a live human in 89% of cases. The figure is interesting, but it comes from the company's own test. The methodology and the comparison with competitors are worth checking yourself.

Access and safety measures

Emma is open on the web with no sign-up or account, so you can assess the quality in a couple of minutes. Developers have been given early access to the Nuplex Live API. The company does not state prices, limits or exact latency in its announcement.

The launch comes with restrictions:

• Emma's appearance is licensed from a real actress;
• streams are labeled with mandatory watermarks;
• direct cloning of third parties' faces and voices is blocked.

For business this is a plus in terms of risk, but also a limitation: for now you cannot create a digital twin of an employee or a brand ambassador through this API.

What to do with this

If you are building a voice or video agent, the idea itself is what is useful in this news. The main problem of live dialogue is not picture quality but timing: when to come in, when to stay silent, when to nod. A cascade of STT, LLM and TTS structurally gets in the way, because each stage waits for the previous one.

A practical plan looks like this:

• open Emma and check how she behaves with interruptions, pauses and hesitant speech;
• compare the feel of the conversation with your current cascaded stack;
• request access to the Live API if you have a scenario where naturalness matters, not just answer accuracy.

Questions remain that the announcement does not answer. How smart is the avatar's "head" on complex tasks? How does the model hold context in a long conversation? How much does a minute of streaming cost? End-to-end models usually win on naturalness but lose to cascades on controllability: it is harder to connect tools, knowledge bases and rigid scripts. For now this is a strong demonstration of a direction, not a ready replacement for the whole stack. But the direction is clear: avatars are ceasing to be voiced chatbots and are becoming conversation partners.