As is mentioned in a parallel comment this will change dramatically as LLM are introduced to voice assistants. LLM are much more robust in inferring semantic meaning from noisy language. As long as they don’t ask the LLM to solve partial differential equations and confine it’s originating prompt aggressively to only interpret voice text transcriptions to a finite set of actions, it’ll be amazing.
I don't doubt that voice recognition will continue to improve, but doesn't it seems kind of ridiculous for automotive manufacturers to pour millions of dollars into R&D (or spend millions licensing someone else's software) and add much more expensive processors to their cars at mass scale to solve an already solved problem? I can already adjust the climate control and other basic things without this sort of interface, and I don't have to look away from the road to do it.
If it sells more cars with more profit I fully expect them to, yes. I’d also note that wiring and layout of all those buttons as well as the manual work necessary to make a static user interface is fairly expensive. The slap a general purpose processor and use the existing AV in the car to control everything greatly simplifies both assembly and supply chain. Perhaps it’s counter intuitive, but buttons aren’t nearly as cheap as you might imagine, and computing resources are decreasing in cost rapidly. Finally LLM and voice recognition model computation costs are falling dramatically - you don’t need a GPT-4 to improve voice recognition outcomes, most of the smaller system scaled models would just as well. The more advanced models are only necessary for significantly more advanced semantic tasks.