Nuance Labs has launched Emma, an AI you talk to face to face on live video. In its announcement on November 9, 2026, the Seattle startup calls her “an AI persona you’ll actually want to meet face-to-face” and says she is built on Nuplex 1.0, a full-duplex audiovisual foundation model. Full-duplex means the model listens and speaks at the same time, the way people do.
The launch post by co-founder and CEO Fangchang Ma passed 1.3 million views on X in about eight hours. His pitch: “Most AI personas stare blankly while you talk. Emma reacts while you’re still speaking.”
On this page · 8 sectionsOpen
Nuance Labs launched Emma on November 9, 2026: an AI persona built on Nuplex 1.0, a full-duplex audiovisual model that listens and speaks at once.
On 1,785 questions from real conversations, Nuance says Nuplex’s reply timing overlapped the person’s 89% of the time. The next closest system was at 35%.
Emma’s face and voice are licensed from a real actress. The beta API cannot clone a person, and you cannot upload a face or a voice.
Emma never claims to be human, and API video carries a visible watermark. AI girlfriends or boyfriends, healthcare and under-18 uses are not allowed.
It is English only and can misjudge when to speak. It is open in most of the US, the EU, the UK and Canada, with four US states and Quebec excluded.
§ 01What Emma is
Most voice AI takes turns. You speak, it waits until you stop, then it answers. Nuance says Emma does not wait: “Emma talks to you with her face, not just her voice.” Per the company, she reacts with small movements while you are still speaking, such as a long blink, a squint or a turn of the head, and she uses gaze as a signal for whose turn it is.
Her personality is designed too. Nuance describes her as curious and quick-witted, with opinions she will defend, and upfront that she is an AI.
| Question | Answer |
|---|---|
| What it is | An AI persona on live video, built on Nuplex 1.0 |
| What Nuplex stands for | Nuance Full-duplex: a system that listens and speaks at the same time |
| Where the face and voice come from | Licensed from a real actress who worked with the team |
| How to try it | emma.nuancelabs.ai, with no account needed |
| For developers | The Nuplex Beta API at platform.nuancelabs.ai, with a few personas |
| Language | English only, for now |
| Who can use it | Adults (18+) only |
| Where | The US except Colorado, Connecticut, Illinois and Texas; the EU; the UK; Canada except Quebec |
| Price | Not published in what we read |
§ 02The number Nuance leads with: 89% against 35%
Nuance argues that reply speed alone settles little. Fast open models answer quickly. Products such as GPT-Live and Gemini Live take longer, “most likely because they would rather not cut you off.” Its view: “Faster or slower is a product decision, not a score.”
So it measured something else, which is whether the model replies when a person would. The result, in Nuance’s words: “On 1,785 questions from our own recordings of real conversations, Nuplex’s reply timing overlaps the original speaker’s 89% of the time. The next closest system is at 35%.”
Three things to keep in mind about that number:
- It is Nuance’s own test, on its own recordings. No outside party has reproduced it in what we read.
- The text does not name the system at 35%. The post discusses PersonaPlex, Moshi, GPT-Live and Gemini Live. It notes that the last two were reached through their public APIs, so their timing includes network time.
- Nuance itself is cautious. It writes that “we believe benchmarks in the AV full-duplex space are not accurate to the user experience yet” and tells readers to try Emma and judge.
§ 03What Nuance will not let Emma do
Nuance says the first people who talked to Emma told it two things at once: “it was amazing, and it was a little unsettling.” Its answer is a list of limits on both Emma and the API. The first rule is the plainest: “You’ll always know you’re talking to an AI.”
| Rule | What Nuance says |
|---|---|
| Always an AI | Emma never claims to be human. Prompts that make a persona pose as a real person or company are refused |
| Watermark | Video served through the API carries a visible watermark |
| No cloning | You cannot upload a face or a voice. Custom faces are by arrangement, with the person’s explicit permission |
| No companions | No sexually explicit content, and no AI girlfriends or boyfriends |
| Closed areas | Healthcare and anything involving under-18s, including education, are not in the beta |
| During a call | Automated checks end the call for intent toward serious violence, slurs, sexual content, or weapons or gore on camera |
| The persona’s own words | The call ends if the persona asks for a passcode, card details, money or remote access to a device |
| Someone in crisis | Pointed to help, such as the Suicide & Crisis Lifeline, and not hung up on |
Nuance adds that these checks will miss things and will sometimes end calls that were fine.
§ 04What Emma cannot do yet
The announcement lists its own gaps:
- Timing still slips. “She can jump in while you’re still mid-thought, wait a beat too long, or miss a cue a person would catch.” Long pauses in the middle of a sentence are the hardest case.
- She can be wrong. Her reasoning comes from a language model, so she can make confident mistakes.
- English only, for now.
- Noise hurts. Background voices, a TV, a speakerphone or a weak microphone make turn-taking worse.
- Seeing is early. With your camera on she can talk about what she sees, and she can miss details or misread a scene.
- Visual glitches. Lip sync, gaze and head motion sometimes look off, and her expression can slip back to neutral.
§ 05Who is behind it
Nuance Labs is based in Seattle. Its three co-founders hold PhDs and previously worked at Apple: Fangchang Ma (CEO), Edward Zhang (CTO) and Karren Yang (chief scientist). The company’s press page lists a $50 million Series A announced on September 14, 2026, with Lightspeed, Accel and Nvidia named in one headline, and a $10 million round in September 2025.
Related on this blog: Tavus Griffin and Gemini 3.8 Live.
§ 06What to watch
- An outside test of the 89% timing figure, with the comparison systems named.
- API pricing, which Nuance has not published.
- More personas and languages, and whether the closed areas (healthcare, education, the four US states and Quebec) open.
§ 07Where CellCog sits
CellCog AI employees do not have a face. They join Google Meet, Zoom and Microsoft Teams calls with the camera off and speak in their own voice, and they can answer calls on their own phone number. The employee behind the voice is the same one that does the work, with its memory and its tools.
§ 08Sources
- Nuance Labs, Meet Emma, the First Nuplex Persona, November 9, 2026
- Fangchang Ma on X (17:59 UTC, November 9, 2026; 1.3 million views at 02:06 UTC on November 10)
- Nuance Labs, Nuplex Beta API and press page
- Kyutai, Moshi: a speech-text foundation model for real-time dialogue, one of the open full-duplex models Nuance compares against
Q1How do I try Emma?
At emma.nuancelabs.ai. Nuance says no account is needed. It is for adults only and is not available in every region.
Q2Where is Emma available?
The United States except Colorado, Connecticut, Illinois and Texas, the EU, the UK, and Canada except Quebec, per Nuance.
Q3Can developers build on it?
Yes. The Nuplex Beta API offers a few personas that you brief with your own prompt. Pricing is not published in what we read.
Q4Can I use my own face or voice?
Not by yourself. Nuance says the beta cannot clone people. Custom faces are by arrangement, with the person’s explicit permission.
Q5Do CellCog AI employees have a face?
No. They join calls with the camera off and speak in their own voice.