Voice AI has reached a strange level of maturity. Speech recognition is often solid, language models are fluent, and synthetic voices can sound surprisingly natural, at least when you listen to demos. But the moment you interact with most voice systems, the illusion tends to crack, because the problem is no longer raw capability, it is conversational dynamics.
Most systems still rely on a familiar, rigid pipeline: ASR → LLM → TTS. It’s logical and modular, but it also forces a particular kind of interaction, where the system listens, then waits, then replies, and only then lets you continue. In other words, the technology might be powerful, but it still behaves like a walkie-talkie, and the result is that conversations feel stiff, slow, and subtly unnatural.
This is exactly the pain point NVIDIA is targeting with PersonaPlex-7B, their newly released open-source full-duplex conversational model. Full-duplex is the key phrase here, because it means the system can listen and speak at the same time, which sounds like a small detail until you realize it’s the thing that makes human conversation feel human. People don’t politely wait for a clean turn boundary; they interrupt, they overlap, they give backchannel signals like “mhm” and “right,” they self-correct mid-sentence, and they steer the conversation continuously rather than in discrete steps.
With full-duplex Voice AI, you finally get interactions that can breathe. You can have immediate feedback, smoother rhythm, and natural interruptions, meaning the system becomes less of a sequence of audio transactions and more of something that resembles dialogue. It’s not just a model and soon product improvement; it’s a shift in what voice interfaces can realistically feel like. Below an example.
Why this matters for interpreting
If you work in conversational interpreting (basically everything besides simultaneous), or in any industry where multilingual audio is mission-critical, this development is worth watching for one simple reason: conversational interpreting is not (only) turn-taking. Interpreting is the art, and the cognitive discipline, of operating inside conversational flow, often under time pressure, often with overlaps, repairs, and pragmatic messiness, and always with a sensitivity to timing that text-centric systems routinely underestimate. In fact, current conversational interpreting systems require the users to adapt (massively) their conversational patterns to the machine. Something which might be acceptable in some scenarios (a slow and short patient-doctor encounter), but that would fail on others, where natural conversational dynamics are key.
Today, most “AI interpreting” still depends on cascading systems: speech is recognized, converted into text, translated, and then spoken back. Frontier models even translate directly, without any intermediate text representation. This stack can work in controlled settings, but it carries structural weaknesses that are hard to escape, because delays accumulate, errors propagate, and the whole interaction is forced into an unnatural rhythm where speaking becomes sequential rather than interactive. The result is particularly painful in meetings, Q&A, negotiations, panel discussions, and any real conversation in which participants don’t behave like careful audiobook narrators, i.e. when they do not adapt strongly to the right turn-taking rules dictated by the system.
That’s why models like PersonaPlex signal something bigger than a single release, also for interpreting. It points toward a future where translation moves closer to direct and natural speech-to-speech interaction, meaning systems that are designed to support conversation in real time and dynamically rather than forcing, as introduced above, conversation to adapt to the limitations of a pipeline. And once that happens, some of the most stubborn problems in interactive multilingual communication, especially the turn-taking bottleneck, become much easier to tackle.
While none of this implies that interpreters suddenly become obsolete (if anything, it highlights what makes human interpreting valuable, because trust, accountability, nuance, and relational dynamics still matter enormously in real-world multilingual communication), it does suggest that Voice AI is finally moving in the direction it always needed to go: away from the “three modules passing the baton” approach and toward systems that behave more like participants in conversation. And while this technology will become common in the years to come, cascading models will not disappear any time soon, since there are many reasons while a less natural, but more controllable and auditable pipeline might be preferred (as I wrote in several articles such as here and here).
The most interesting aspect of this release, therefore, is not simply that PersonaPlex exists, that it is open source, or even that it is commercially usable, each of which, on its own, would already signal a future in which this kind of technology becomes easily accessible to any developer or company, and thus increasingly ubiquitous. The interesting part is that NVIDIA just removed the biggest awkwardness of Voice AI: the fact that it still didn’t know how to talk like someone who’s actually in the room.
Paper: https://research.nvidia.com/labs/adlr/files/personaplex/personaplex_preprint.pdf
The open weights can be downloaded here: https://huggingface.co/nvidia/personaplex-7b-v1