Conversational AI & Voice
The AI That Finally Listens
mlj.solutions
mlj.solutions Hlasová AI, která poslouchá a mluví zároveň. NVIDIA představila PersonaPlex - systém, který reaguje na přerušení, přitakává a drží kontext celé konverzace. Žádné čekání, než domluvíš. Žádné začínání od začátku. Full-duplex komunikace mění to, jak může AI fungovat v zákaznické podpoře, onboardingu nebo tréninku. 👉 Kde by tato technologie nejvíc pomohla vašemu byznysu?
View on Instagrammlj.solutions
mlj.solutions Zkusili jsme přerušit AI uprostřed věty. 🎙️ Klasický voicebot by začal znovu od začátku. Nebo zamrzl. Tenhle? Přirozeně pokračoval. Jako člověk. NVIDIA právě představila AI, která poslouchá a mluví zároveň. Reaguje na přerušení. Přitakává. Drží kontext. Konec éry hlasových robotů, kteří čekají, než domluví. 👉 Kde by tato technologie nejvíc pomohla vašemu byznysu? Celý příběh ve čtvrtek. 🎬
View on InstagramWhat the Study Is About
NVIDIA PersonaPlex is a full-duplex model that listens and speaks at the same time, handling interruptions and backchanneling just like natural human conversation. Role and voice can be set with a single text prompt, with no retraining of the model. It builds on the pretrained Moshi model, fine-tuned on fewer than 5,000 hours of targeted data.
100%
turn-taking success rate
On the FullDuplexBench benchmark — competitors ranged between 43% and 90%.
0.170 s
average latency
The lowest of all tested systems, including Gemini Live and Qwen 2.5 Omni.
< 5,000 h
of data for role-following training
PersonaPlex achieved task-following with minimal targeted data.
What This Means
- Voice and role can be set via prompt, with no retraining and no changes to the model.
- Full-duplex fundamentally changes interaction quality, the agent responds in real time and doesn't wait for the user to finish talking.
- Synthetic and real data complement each other — synthetic data covers scenarios, real recordings bring natural speech patterns.
- The model generalizes beyond its training, it handled a crisis scenario aboard a spaceship even though nothing similar appeared in training.
- NVIDIA PersonaPlex can be deployed to production right away, the code is available under an open-source license.
Our Take
The model is open-source and ready to deploy, but unfortunately it doesn't support Czech — a common gap in many of these solutions.
Michal Dobrovolný
Founder & CTO
Zdroj
Roy, R., Raiman, J., Lee, S., Ene, T., Kirby, R., Kim, S., Kim, J., & Catanzaro, B. (2026). PersonaPlex: Voice and Role Control for Full Duplex Conversational Speech Models. NVIDIA Research. https://research.nvidia.com/labs/adlr/personaplex/
Keep reading
Don't miss these posts
Want to implement AI in your company?
We'll go through your needs together and propose an approach that delivers measurable results.
Write to us
