This is a surprisingly tricky problem in multilingual voice systems. Once the language detector assigns a high enough confidence to one language, the conversation can get "stuck" there unless language preference is treated as an explicit part of the session state. Curious if that's what's happening here.
https://zencia.ai