Bug: voiceName configuration is silently ignored in gemini-3.1-flash-live-preview (Multimodal Live API)

Gemini API
1

I am using the Multimodal Live API with ephemeral AuthTokens (authTokens.create → BidiGenerateContentConstrained WebSocket). My implementation strictly follows the API documentation for voice selection, but the gemini-3.1-flash-live-preview model completely ignores the configured voice and always defaults to a fixed voice.

Environment:

  • Model: models/gemini-3.1-flash-live-preview

  • Endpoint: generativelanguage.googleapis.com (Google AI Studio)

Payload Structure Used: I am wrapping the configuration correctly inside the token creation request as follows:

JSON

"bidiGenerateContentSetup": {
  "model": "models/gemini-3.1-flash-live-preview",
  "generationConfig": {
    "responseModalities": ["AUDIO"],
    "speechConfig": {
      "voiceConfig": {
        "prebuiltVoiceConfig": { "voiceName": "Kore" }
      }
    }
  }
}

What I’ve verified (Troubleshooting steps taken):

  1. The JSON schema is accepted by the API (no 400 Bad Request error).

  2. systemInstruction configured exactly the same way works perfectly.

  3. I have tested multiple valid Live API voices (Kore, Puck, Aoede). The result is always the same fixed default voice.

  4. I tested both baking speechConfig into the AuthToken and delegating it to the client WebSocket setup message (using lockAdditionalFields). Same issue.

Question: Is speechConfig intentionally hardcoded/disabled on the backend for the 3.1-flash-live-preview model, or is this a bug in the voice synthesis module? Any timelines on when custom voice selection will be enabled for 3.1?

Thank you!