Hence my remark about ‘when it improves’. And god help those of us whose accent is not RP and therefore outside the training set that Apple et al trained their software on. I have seen video on youtube about for example lowland scots trying to get a voice recognition system to do the right thing.
I wonder if your Mercedes’ system is rubbish because of background noise. Siri doesn’t like it at all, everything has to be quiet, so no oriental tomcat yowling for smoked salmon (Somhairle) while Janet is making sandwiches and I’m trying to talk to Siri.
However as I said, Siri is good at speech recognition but just damn stupid, definitely still in the remedial class. It doesn’t have a proper parser for English and I suspect it relies on just a load of heuristics and hand-crafted rules, but that’s pure guesswork based solely on its failings when tested with non-trivial sentences. Give it another ten years and it might be something impressive.
If they had based it on something powerful like RRG (Role and Reference Grammar) then they might have a chance of really understanding sentences. However RRG only does syntax, specifically the syntax-semantics interface, and so needs a lexicon, phonology, morphology and pragmatics components from elsewhere and also more semantic analysis obtained from some other software shop. And a truly massive computer to run all that lot on, which is why Siri needs an internet connection, because she doesn’t have remotely enough CPU power available locally.
Forty years ago, I remember having a little chat with ELIZA on a DEC-10 mainframe. I said to her "If you don’t tell me what sex you are, I am long to kill myself" Answer: "We were talking about you, not me", Me: "I am going to my doom post haste", ELIZA: "Tell me more about your doom post haste."