My guess it's currently quite naive - automatically transliterate English to text (bad phonetic misspellings and all), then pass that to a language translation layer which garbles the output even more.
Incorporating a LLM layer would be able to help because LLMs operate rather like sequence prediction outputs that can go very far back in history. A LLM should do a better job of automatically smoothing over common mistakes, e.g., it should be able to un-twist the English transcribed "I'll have my meet medium wear" back to "I'll have my meat medium rare."
Adding in that extra step which is effectively a shortcut for NLU auto-correction should produce significantly better output. We've been finding that LLMs have been outperforming our prior NLU efforts.
