BUG: Gemini 3.8 Live API overrides language_code / ignores system prompt accent formatting

Gemini API
1

Model Impacted: gemini-3.8-live (Regression from gemini-3.1-flash-live-preview)

Environment: Pipecat / WebSockets API integration

Description:
In previous preview versions (gemini-3.1), passing a non-US language code or system prompt instructions safely forced the Live API output voices to speak English with a regional accent (e.g., British or Spanish) while maintaining clean English text tokens.

With the stable release of gemini-3.8-live, this behavior is entirely broken/unstable:

  1. Passing a language_code to force an accent layer causes the audio generation weights to aggressively drift. The model will begin speaking the target language (e.g., Spanish) even when processing correct English text tokens.
  2. Traditional text-based system instructions to adopt an accent are ignored, defaulting the voice profiles to General American phonetics.

Impact on Use Case:
This breaks applications that require clean English spoken alongside other languages with an accented voice profile.

Steps to Reproduce:
Pass an English string to gemini-3.8-live with language_code set to a non-US variant. The model will either drift into speaking that language natively or fail to shift vocal phonetics away from the American base profile. Or pass clear instructions in the prompt and there is no accent change.