It seems like 3.8 live doesn’t understand words such as yes in hebrew (כן, pronouced as ken)… it asks multiple times for repetition, and it consistantly doesn’t understand basic words… Even when most of the system prompt and context are in hebrew. It even switches to other languages mid-call randomly.
It might be because there is no longer the ability to set a language in the API… you can only set a language in the prompt…
Hello @Zrizzy,
Gemini 3.8 Live is a native audio model that supports Hebrew (he) among its supported languages. Native audio models automatically detect and switch between languages during a conversation and do not support setting an explicit language code parameter; restricting the language the model speaks in is configured through your system instructions.
To help us reproduce and look into the Hebrew recognition and mid-call language switching behavior from our side, please share:
voice_name).