Model: gemini-3.1-flash-live-preview (Live API, native audio)
Voice: Kore (same voice in all comparisons)
Language: Polish (pl-PL)
Summary
gemini-3.1-flash-live-preview speaks Polish with a strong English/American
accent. Earlier native-audio models pronounce the same language correctly,
with the same voice and the same system instruction.
Comparison (identical voice, identical prompt, Polish-only conversation)
- gemini-3.1-flash-live-preview → heavy English accent
- gemini-2.5-flash-native-audio-preview-12-2025 → correct Polish
- gemini-live-2.5-flash-preview-native-audio-09-2025 → correct Polish
(tested in AI Studio; 3.1 Live is not selectable there)
This is not a language-detection problem. The model speaks Polish and picks
correct words; the phonetics and prosody are English.
Why there is no workaround
Native audio models do not accept language_code, so the only available lever
is the system instruction. Adding an explicit instruction to use native Polish
pronunciation (“speak like a native Polish speaker, do not use English
phonetics”) produces only a small improvement and does not persist across
longer conversations. Switching prebuilt voices does not resolve it either.
Impact
This blocks production use for Polish-language voice agents. We had to choose
between 3.1 (better reasoning, longer conversations, more reliable tool
calling) and 2.5 (correct pronunciation, but noticeably slower and less
accurate). Both options are unsatisfactory for a product sold to Polish users.
Request
Please restore the Polish phonetic quality of the 2.5 native-audio models in
3.1, or expose language_code for native audio so the output language can be pinned explicitly.