### Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

### What happened?

When proxying a realtime (voice) session through LiteLLM with **server-VAD turn detection** and **input audio transcription enabled**, and with **no `realtime_input_transcription` guardrail configured**, the backend (Azure OpenAI GA realtime) returns an error on every user turn:

```json { "type": "error", "error": { "type": "invalid_request_error", "code": "conversation_already_has_active_response", "message": "Conversation already has an active response in progress: resp_XXX. Wait until the response is finished before creating a new one." } } ```

The same client connecting **directly** to Azure OpenAI realtime (no LiteLLM proxy) does **not** hit this - so the duplicate response is introduced by the proxy.

## Root cause

The proxy unconditionally injects a `response.create` to the backend when it sees `conversation.item.input_audio_transcription.completed`, **without** having disabled the backend's server-VAD auto-response. The backend's VAD already auto-creates a response for the user turn (`turn_detection.create_response…