Most AI video avatars stare politely while you talk and then answer. Emma, launched on November 9 by Seattle startup Nuance Labs, is designed to do what people do: nod, smile, glance away, and sometimes jump in before you’ve finished. “Most AI personas stare blankly while you talk. Emma reacts while you’re still speaking,” Nuance Labs’ Fangchang Ma wrote. Behind her is Nuplex 1.0, which the company calls a “full-duplex audiovisual foundation model.”
Quick Take
- What it is: Emma, a real-time AI video persona you can talk to in a browser with your mic (camera optional), no account needed.
- The tech: Nuplex 1.0 listens, sees, reasons, speaks and shows expressions at the same time instead of taking turns.
- The claim: On 1,785 questions from real conversations, Nuplex’s reply timing overlaps a human’s 89% of the time; the next closest system scores 35%: Nuance’s own test.
- For developers: A Nuplex Beta API (“Nuance Live”) with several personas, prompts, tool calls and LiveKit-based streaming.
- Guardrails: Always disclosed as AI, visible watermark on API video, no cloning, no explicit content, no “AI girlfriends or boyfriends,” no healthcare or minors in beta.
What happened
In a research post, Nuance Labs wrote: “After 18 months of research, we crafted Emma, an AI persona you’ll actually want to meet face-to-face.” Emma is available at emma.nuancelabs.ai, and developers can build on the “Nuplex Beta API at platform.nuancelabs.ai.” The company also posted a Show HN on Hacker News, admitting Emma “might be a little rough around the edges but we think it’s better than anything else out there (we’re biased, of course).”
The launch spread through tech X. Investor Justine Moore said Nuance “launched one today that’s actually fun to talk to,” and Robert Scoble posted a video interview with Emma: “First time I’ve interviewed an AI like this.”
Background: the race for real-time faces
Real-time AI avatars have improved quickly this year; i10x has covered the infrastructure behind live avatars and real-time video models like Vidu S2. Nuance itself names Tavus, LemonSlice and Keyframe Labs as “teams working on the same hard problem from different angles.” On voice, it compares Nuplex with open full-duplex models like PersonaPlex and Moshi and with GPT-Live and Gemini Live. Note: Nuance Labs is not Nuance Communications, the speech company Microsoft bought (a confusion Grok also had to clear up on X).
How it works
Nuance describes human conversation as “a simultaneous act. You perceive while you speak, react while you listen, and respond within a few hundred milliseconds.” Nuplex tries to model that in one system that “sees, hears, reasons, speaks, and expresses in the same moment.” Its homepage names “500ms” as “the hard floor for conversation latency.”
Rather than chase the fastest reply, Nuance says it asked “when does a person reply?” Open models like PersonaPlex and Moshi “are tuned to answer fast,” while GPT-Live and Gemini Live “take longer, most likely because they would rather not cut you off.” Nuplex “sits between them.” The company also stresses design over averages: Emma’s face and voice “come from a real actress,” with mannerisms like “a long blink for emphasis, a squint, eyes that wander and come back.” Her personality is built to “disagree with you, while being upfront that she’s an AI,” which Nuance contrasts with chatbots that are “agreeable to a fault.”
For developers, the Nuance Live docs describe sessions in which “your server holds the secret API key; your page or app only ever sees a short-lived session token.” Audio and video run over LiveKit rooms, and the persona can hand longer jobs to a backend that calls your tools “while it keeps talking with your user.” The API “is in preview: routes and events may change before 1.0,” and sessions run “on shared capacity.”
| What Nuance Labs says | |
|---|---|
| Model | Nuplex 1.0 (“Nuance Full-duplex”), audiovisual foundation model that listens and speaks simultaneously |
| First persona | Emma; face and voice licensed from a real actress |
| Research time | 18 months |
| Timing test | Reply timing overlaps the original human speaker 89% of the time on 1,785 questions; next closest system 35% |
| Access | Free in the browser at emma.nuancelabs.ai, no account; Nuplex Beta API at platform.nuancelabs.ai |
| Regions | US except Colorado, Connecticut, Illinois and Texas; EU; UK; Canada except Quebec |
| Limits | English only, adults 18+, no cloning, no sexual content, no healthcare or under-18 uses in beta |
Safety rules
Nuance calls its beta “intentionally limited (even annoyingly so).” Every developer prompt is screened, and calls end automatically if a user expresses intent toward serious violence, uses slurs, raises sexual content, or shows weapons or gore. The persona’s own words are checked too: “if it ever asks for a passcode, card details, money, or remote access to a device, the call ends.” Someone who seems to be in crisis “gets pointed to help, like the Suicide & Crisis Lifeline, not a hang-up.” Nuance says it tested Emma “against published benchmarks for sycophancy, manipulation and emotional reliance.”
Pros and cons
For users
- Pro: A free, account-less demo of conversation that feels closer to a video call than a chatbot.
- Con: Nuance itself warns she “can make confident mistakes,” can interrupt or wait “a beat too long,” speaks English only, and some users find the realism “a little unsettling.”
For developers
- Pro: A documented real-time API with prompts, tool hand-offs and SDKs for web and native apps.
- Con: Preview status, shared capacity that can turn sessions away, no custom faces or voices self-serve, and no healthcare or education use cases.
For companies
- Pro: Potential uses in customer service, training and games; custom personas are available “by arrangement.”
- Con: Strict content rules and regional gaps (several US states, Quebec) limit deployment; pricing was not published in the launch post.
What X is saying
The first AI persona you’ll actually want to meet face-to-face. Most AI personas stare blankly while you talk. Emma reacts while you’re still speaking. After 18 months of research at @nuance_ai, this is Emma, built on Nuplex, our full-duplex audiovisual foundation model.
— Fangchang Ma (@fangchangma) November 9, 2026 · View on X
Man we are so close to real-time AI avatars that are completely indistinguishable from humans. Insane progress in this space over the last few months around latency, body language, natural-sounding speech. And @nuance_ai launched one today that’s actually fun to talk to!
— Justine Moore (@venturetwins) November 10, 2026 · View on X
Damn, talking with the AI from @nuance_ai is pretty amazing. First time I’ve interviewed an AI like this.
— Robert Scoble (@Scobleizer) November 9, 2026 · View on X
Ma also explained the company’s stance on benchmarks: “We found that pushing our model higher on existing benchmarks made it feel worse to talk to. So we’re not asking you to believe a benchmark. You have one built in.” Scoble raised the business questions in a follow-up: “How expensive are these to run? How open are they to include in games, apps, websites?”
What Grok says
Asked how Emma handles hesitant users, Grok explained: “Nuplex detects tone, facial expression, and hesitation in real time as part of its full-duplex audiovisual model. For hesitant or uncertain users this lets Emma react while they speak—nodding, adjusting expression, or timing her reply—rather than staring blankly or cutting in.” It also clarified that Nuance Labs is “a 2025 Seattle startup” and not the company behind Dragon dictation, which “came from Nuance Communications, acquired by Microsoft in 2022.”
Critical analysis
Emma’s strongest claim is also its least verifiable. The 89% timing overlap comes from Nuance’s own recordings and its own metric, and the company concedes that “benchmarks in the AV full-duplex space are not accurate to the user experience yet.” Its answer (“simply meet Emma”) is honest but subjective. Independent evaluation of conversational naturalness barely exists, so early reviews from enthusiastic investors and influencers carry more weight than they should.
The safety design is notable for what it refuses: cloning, companion relationships, minors and healthcare. Those are exactly the markets where realistic AI faces could earn the most and do the most harm. Holding those lines will get harder as competitors offer fewer restrictions. The regional exclusions, including Colorado, Connecticut, Illinois, Texas and Quebec, also hint at how much state-level rules on AI and biometric likeness now shape product launches.
What’s still open
- What will the Nuplex API cost per minute once it leaves beta?
- Will independent researchers confirm the timing results?
- When will Emma support other languages and her own choice of language model?
- How will the safety checks perform at scale, and how often will they end harmless calls?
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated image. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based i10x analysis draws on Nuance Labs’ launch post, website and developer documentation, the company’s Hacker News post, and public posts on X through November 11, 2026. i10x did not test Emma directly for this article.
i10x Perspective
Voice assistants taught us to wait our turn. Full-duplex models like Nuplex aim to remove the turn-taking altogether, and adding a face raises both the appeal and the risk. The most interesting part of Emma’s launch is not the realism but the rulebook: watermarks, no cloning, no companions, hard stops on scams. If face-to-face AI becomes common in apps and customer service, those defaults should become industry norms, not a startup’s voluntary choice.
