Some models can detect the language spoken. Some models are multilingual, with the possibility to stipulate a set of expected languages.
What happens the doc has no language, and the language hasn't been stipulated? We could auto detect, but we would then need to report the language back.
I'm tying this to multilingual models because we also need to somehow attach a language post-init, to results. This is typically done underneath as special output tokens.
A similar but separate concern is how we'll represent diarization -- it should have the same API shape probably.
Some models can detect the language spoken. Some models are multilingual, with the possibility to stipulate a set of expected languages.
What happens the doc has no language, and the language hasn't been stipulated? We could auto detect, but we would then need to report the language back.
I'm tying this to multilingual models because we also need to somehow attach a language post-init, to results. This is typically done underneath as special output tokens.
A similar but separate concern is how we'll represent diarization -- it should have the same API shape probably.