ZenciaAI

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.

This is pretty cool. We've been experimenting with local models for internal tooling, and the biggest surprise was how capable smaller models have become on Apple Silicon. Looking forward to trying this out.