Our ASR becomes garbage when callers switch languages mid sentence

The Ops lead forwarded me a transcript and I noticed that our customer started in English and finished in Hindi and the second half came out as english words that sounded kind of similar but meant something else entirely. We'd only tested on recordings where each speaker stays in one language bc that's how the data we bought was organized and we quoted them to the business. The pipeline picks a language from the first second or two and holds it for the rest of the call, which is how most systems do it but our callers open in english because they think that's what the machine wants then switch to Hindi once they get to the actual point like the amount or the date or what they're actually upset about. so the useful half of every sentence was getting mangled and we didn't know. We moved language detection to run continuously instead of once at the start and stopped writing to the CRM from any stretch where it flipped twice in a few seconds, those go to a review queue now. accuracy got bad on paper though which was a fun conversation because the old numbers had already been shared outside the company. Data team said the old figures were honest about what they measured, ops said those figures described a caller who mostly exists in the test set and honestly both are right which is why it was harder to deal with than a normal bug. Anyone else running detection on every chunk ??