You clearly haven’t listened to enough ”state of the art” AI voices. Even in languages blessed with centuries’ worth of training data the butchered proper nouns and misemphasized syllables are grating.
As a misguided side effect of RLHF, the models which occasionally succeed in hiding their monotony will just indiscriminately vary the speech with a random patchwork of poorly timed pauses and melodic patterns. Good luck maintaining cohesion and semantic grouping of words.
And that’s for the five or so widely used easy languages.
God forbid your European car manufacturer, instead of using the completely adequate and predictably consistent Siri-style GOFAI solution, decided to use some cheap Elevenlabs-copycat API for localised driving directions.
Now you’re dealing with bandwidth and latency issues, pronunciation that constantly changes, and utterly unusable performance in your 10 million speaker ”niche” language. For many of these languages rule-based algorithms were good enough and our solar system will freeze before you accumulate enough training data for the current generation TTS to perform well.