PersonaPlex marks a shift from dictation to conversation
Dictation turns speech into text. Conversation works in time. PersonaPlex marks the moment voice AI starts to operate in real-time dialogue.
I have written quite a bit about speech recognition, dictation, and voice over the past years. Looking back, I now see that I was often talking about different things under the same label.
PersonaPlex is a research model from NVIDIA that explores native, real-time speech-to-speech conversation, where listening and speaking happen at the same time.
It helps make a distinction visible that had been forming for a while already.
Not because earlier systems were wrong or outdated, but because what we mean by voice has started to change.
Dictation solved one problem
Models such as Whisper are excellent at dictation.
You speak. The system listens. You get clean, reliable text.
That alone was a major step. It removed friction between thinking and writing. It made meetings, interviews, and spoken notes usable at scale. For this purpose, dictation models are still extremely strong.
But dictation treats speech as something that already happened. Accuracy matters more than timing.
Rob Hoeijmakers
Conversation already feels solved, doesn’t it?
If you use ChatGPT with voice, or Gemini Live, it is easy to think: this already is full duplex.
You can interrupt them. They stop speaking immediately. The interaction feels fluid compared to older voice assistants.
From a user’s perspective, that intuition makes sense.
But under the hood, something else is going on.
How today’s voice systems actually work
Most production voice systems today rely on several fast components working together:
- One part listens for speech and detects interruptions.
- Another part reasons about what to say.
- Another part turns that into audio.
When you interrupt, a very fast detector notices this and simply cuts off the audio output. The system stops talking right away, even if another part is still finishing its thought elsewhere.
To you, it feels like the system was listening while speaking.
Technically, it mostly just stopped speaking very quickly.
This is not a flaw. It is a sensible engineering choice. It is cheaper, more stable, and easier to control.
But it is not yet what researchers mean by full duplex.
What “full duplex” really means
Full duplex simply means listening and speaking at the same time.
Not taking turns.
Not stopping first.
Not restarting after an interruption.
In a full-duplex speech system, incoming sound continues to shape what the system is doing while it is already talking. Interruptions are not just stop signals. They carry information: timing, tone, urgency.
Speech is no longer just an interface around reasoning. It becomes part of the reasoning itself.
That is the real shift.
Why this matters
Seen in that light, PersonaPlex is not just another voice demo.
It is a concrete example of what changes when speech is no longer treated as input that must first be stabilised, but as a medium in which interaction itself takes place.
Dictation systems listen, then act.
Most current voice assistants react quickly.
PersonaPlex listens while it speaks.
That distinction may sound subtle, but it changes what kinds of conversations become possible. Especially outside the assistant setting: in phone calls, service conversations, and other situations where flow, interruption, and timing matter.
PersonaPlex does not replace dictation models, nor does it invalidate today’s voice assistants. It shows that voice is becoming layered.
And it gives a first, working glimpse of what that next layer looks like.
NVIDIA ADLR
Rob Hoeijmakers