Interactions API: speech_config.mode "conversational" (documented) now rejected with 400 for gemini-3.8-flash-tts

Gemini API
1

Hi,

Since 23 October 2026, multi-speaker TTS requests with conversational mode are rejected by the Interactions API.

The speech generation guide (https://ai.google.dev/gemini-api/docs/interactions/speech-generation, last updated 2026-10-12) says: “The REST API also supports “mode”: “conversational” inside speech_config for natural turn-taking cadence.” We used this daily until 22 October.

Request (POST https://generativelanguage.googleapis.com/v1beta/interactions):

{
  "model": "gemini-3.8-flash-tts",
  "input": [{ "type": "user_input", "content": [
    { "type": "text", "text": "Bom dia, Anna! Como está?", "annotations": [{ "type": "speech_metadata", "speaker": "Sofia" }] },
    { "type": "text", "text": "Bem, obrigada. E a professora?", "annotations": [{ "type": "speech_metadata", "speaker": "Anna" }] }
  ] }],
  "response_format": { "type": "audio" },
  "generation_config": { "speech_config": { "mode": "conversational", "speakers": [
    { "speaker": "Sofia", "voice": "pt-pt-assistant-7" },
    { "speaker": "Anna", "voice": "pt-pt-training-1" }
  ] } }
}

Response: 400 {“error”:{“message”:“Unknown parameter ‘mode’ at ‘generation_config.structured_speech_config’.”,“code”:“invalid_request”}}

What we tried: mode inside structured_speech_config, upper case CONVERSATIONAL, mode per speaker, Api-Revision headers 2026-05-20 / 2026-09-22 / 2026-10-12, and speechConfig.mode in generateContent. All return “unknown parameter”. The Interactions API reference no longer lists mode for speech either.

Without the mode the same request works, but dialogues sound noticeably less natural (turn-taking, cadence), which matters a lot for our language-learning dialogues.

Questions:

  1. Was conversational mode removed on purpose, or is this a regression?
  2. If it was moved or renamed, what is the new way to request it?

Thank you!

2

Hello @Valerie_pt ,

Thank you for reporting this. We have reproduced this issue and are looking into it. In the meantime, please omit “mode”: “conversational” from speech_config in your multi-speaker REST requests so your calls can complete normally.