[Bug]: vertex_ai/ Gemini TTS fails on /v1/audio/speech — audio param missing from get_supported_openai_params() #21702

Check for existing issues

• I have searched the existing issues and checked that my issue is not a duplicate.

What happened?

The /v1/audio/speech endpoint does not work with vertex_ai/ Gemini TTS models, despite being documented in the Text-to-Speech docs. Requests fail with 400 INVALID_ARGUMENT from Vertex AI.

VertexGeminiConfig.get_supported_openai_params() in litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py (~line 275) does not include "audio" in its return list. The audio parameter is silently filtered out before reaching map_openai_params() (~line 659), which already has the correct audio → speechConfig mapping.

The speech-to-completion bridge correctly sends modalities: ["audio"] and audio: {"voice": "..."} to litellm.completion(), but by the time the request reaches Vertex AI, the audio param has been dropped. Vertex receives audio modality with no speech config and returns 400 INVALID_ARGUMENT.

Note: This is specific to the vertex_ai/ provider. The gemini/ (AI Studio) provider may not be affected.

Expected behavior: /v1/audio/speech with vertex_ai/gemini-2.5-flash-preview-tts should return audio, matching the documented curl example.

Related issues: #18178 (same provider, closed with docs-only workaround in #20255), #17735 (same provider, closed by #17851 which fixed a different layer). The underlying code bug — "audio" missing from VertexGeminiConfig.get_supported_openai_params() — was not addressed in either fix.

Steps to Reproduce

Config:

model_list:
  - model_name: vertex-tts
    litellm_params:
      model: vertex_ai/gemini-2.5-flash-preview-tts
      vertex_project: my-project
      vertex_location: us-central1
      vertex_credentials: /path/to/creds.json

Test 1 — /v1/audio/speech (fails):

curl http://0.0.0.0:4000/v1/audio/speech \
  -H "Authorization: Bearer sk-1234" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "vertex-tts",
    "input": "Hello world, this is a test.",
    "voice": "Kore"
  }' --output speech.wav

Result: 500 → logs show Vertex_aiException BadRequestError - {"error": {"code": 400, "message": "Request contains an invalid argument.", "status": "INVALID_ARGUMENT"}}.

Test 2 — /v1/chat/completions (also fails without workaround):

curl http://0.0.0.0:4000/v1/chat/completions \
  -H "Authorization: Bearer sk-1234" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "vertex-tts",
    "modalities": ["audio"],
    "audio": {"voice": "Kore", "format": "pcm16"},
    "messages": [{"role": "user", "content": "Say hello world"}]
  }'

Same 400 INVALID_ARGUMENT — the audio param is silently dropped.

Test 3 — with allowed_openai_params workaround (succeeds):

Adding allowed_openai_params: [audio] to the model config makes Test 2 return audio successfully. This confirms the mapping logic works — only the supported params list is missing "audio".

Note: the allowed_openai_params workaround does not fix /v1/audio/speech (Test 1), because the speech-to-completion bridge calls litellm.completion() internally and deployment-level config overrides aren't passed through.

Relevant log output

litellm.exceptions.BadRequestError: litellm.BadRequestError: Vertex_aiException BadRequestError - {
  "error": {
    "code": 400,
    "message": "Request contains an invalid argument.",
    "status": "INVALID_ARGUMENT"
  }
}
 LiteLLM Retried: 2 times

What part of LiteLLM is this about?

Proxy

What LiteLLM version are you on ?

v1.81.9-stable

Twitter / LinkedIn details

No response