feat: native Gemini (Google AI) TTS engine #33264

Before Submitting

• I searched open and closed issues and discussions for an existing request.
• I checked whether this already exists on the dev branch or latest source.
• I understand that maintainers want a well-written issue or discussion before any code pull request.
• This request is not a security vulnerability.

Problem

Gemini's TTS models (e.g. gemini-3.8-flash-tts) can currently only be used in Open WebUI through an intermediary (LiteLLM or OpenRouter) with the OpenAI TTS engine.

Pointing the OpenAI engine directly at Gemini's OpenAI-compatible API (https://generativelanguage.googleapis.com/v1beta/openai/) does not work: that layer offers no /audio/speech endpoint. Gemini TTS is only available through the native generateContent API with responseModalities: ["AUDIO"].

The intermediaries add their own failure points. For example, LiteLLM currently wraps the WAV that Gemini returns in a second WAV header (BerriAI/litellm#44541): the inner header is played as samples (a click at the start of every sentence), and the trailing metadata block is audible as noise at the end.

Desired Behavior

A native Gemini option in Admin → Settings → Audio → Text-to-Speech, similar to the existing ElevenLabs and Mistral engines:

• API key (Google AI Studio)
• Model (e.g. gemini-3.8-flash-tts, gemini-3.8-flash-lite-tts)
• Voice, ideally as a dropdown of the prebuilt voices (Kore, Puck, Charon, …)
• Optional style instructions

Internally: POST /v1beta/models/{model}:generateContent with responseModalities: ["AUDIO"] and speechConfig.voiceConfig.prebuiltVoiceConfig.voiceName. The existing transcoding to MP3 (added in #24145) could probably be reused for the returned audio.

Why This Matters

Gemini TTS sounds very natural and handles languages other than English well (we use it for German). Many users already have a Gemini API key for chat models; a native engine would let them use TTS without running or paying for a proxy, and without depending on how each proxy translates the audio format.

Examples or References

• Gemini speech generation docs: https://ai.google.dev/gemini-api/docs/speech-generation
• Gemini OpenAI compatibility (no /audio/speech): https://ai.google.dev/gemini-api/docs/openai
• issue: TTS treats PCM audio responses as MP3 #24143 / fix: transcode PCM audio responses to MP3 in speech endpoint #24145: PCM transcoding for Gemini TTS via OpenRouter/LiteLLM
• [Bug]: /v1/audio/speech returns a double WAV header for gemini-3.8 TTS models BerriAI/litellm#44541: double WAV header for Gemini TTS

Alternatives or Workarounds

• OpenAI engine → LiteLLM → Gemini: works, but currently produces clicks and noise (litellm#44541, fixes still open).
• OpenAI engine → OpenRouter (google/gemini-3.8-flash-tts): should work thanks to fix: transcode PCM audio responses to MP3 in speech endpoint #24145, but requires an extra account and routes all spoken text through a third party.