Andy Halliday

All right.

I have a couple other little things here.

I mentioned earlier the release by NVIDIA of Personaplex, a full duplex conversational speech model that does the niceties, the natural turn-taking.

It doesn't interrupt impolitely.

And

And that still happens to me when I use a number of the voice interfaces.

I have to kind of start yelling at it saying, no, do not respond to me until I tell you this key word that means I'm done speaking.

Because otherwise it's just a fractured conversation.

Well, PersonaPlex is probably more natural in its turn taking.

So it'll handle interruptions, but it'll also not interrupt.

And how does it do this?

Well, it's positioned as an alternative to what currently is in operation, which is speech recognition on your device that then goes to transcription to the LLM.

And then text to speech, right?

The output of the model comes back as text to speech.

There's latency involved in that whole sequence.

Well, this is a pipeline that's quite different.

That's a long pipeline with multiple parties having to interact in order to make that happen.

But this system is a single model that listens and speaks concurrently.

And that aims to get to a more human conversational rhythm without sacrificing your ability to interrupt and control the conversation.

So yeah, I think that NVIDIA is always impressive by their ability to do not only the chip thing, but all the other things too.