"Science fiction" is a bit harsh.

It's also not a binary [yes/no] question; it's more of a spectrum.

For instance, does it count if you can do real time voice to voice with 8xH100? That can be "local". You can download the model...It's just...Really expensive.

Similarly, what about quality? You might get a model running in real time, but it has occasional hallucinations or artifacts. It's possible you may not want to pick those up unintentionally.

I'd say we're probably 60-70% of the way to real-time accessible speech to speech models for casual conversation, and probably about 20-40% of the way to models of such quality and meta-cognition (with the ability to reflect on their own outputs for educational purposes, and be aware of their inflections, etc), that you would want to use them for language learning extensively.

It'll take a few more advancements, but we already know the way there, it's just we have to implement it.

Notably, as soon as someone trains a speculative decoding head for any of the existing speech models that's probably what we need to really make it mostly viable, but a Diffusion speech to speech model would probably be ideal.

I'd say we're maybe about a year out (at most) from real time speech to speech (with possibly some need to customize the pipeline to your needs and available hardware).

So, not quite 100% of the way there, but calling it science fiction isn't quite fair when all the tools are already there and just need to be put together in the right order.