Another real-time avatar. And it looks like it's good.
Nuance Labs publishes a full-duplex audio-visual model and avatar of Emma.
The American start-up Nuance Labs, founded by former Apple researchers, has developed the basic audio-visual model Nuflex 1.0 and released to the public the interactive video avatar Emma, capable of conducting a natural dialogue with a person in real time directly in the browser.
Most existing digital avatars are built on the walkie-talkie principle: the sound is transcribed using STT, transferred to LLM, played using TTS, and then a separate model activates lip movements.
This connection creates a delay of 1.5-3 seconds, due to which the avatar freezes with a glazed look, waiting for the end of the turn.
Noplex eliminates this layer by combining audio and video perception and generation into a single end-to-end neural network: the system continuously reads the tone, pauses and fluctuations in the interlocutor's voice and synchronously streams realistic facial expressions and speech.
The main difference between the Noplex-based Emma avatar is the ability to listen and respond at the same time. The model is able to nod, smile, give signals of active listening ("Ah-ha", "I see") and change facial expressions directly while the user is talking, without waiting for a pause.
In Nuance Labs' blind testing of a sample of 1,785 test questions, Noflex matched the response speed of a live human 89% of the time.
Surprisingly, Emma is available without registration or accounts directly online, and the company has opened initial access to the Noplex Live API to developers.
The launch is accompanied by marketing security filters: the performance is licensed by a real actress, streams are marked with mandatory watermarks, and direct cloning of faces and voices of third parties is strictly blocked.
app.nuancelabs.ai.