AI Gateway transcription drops Gemini speaker diarization (word_info / spk_*)

Product: Vercel AI Gateway Feature: Speech-to-text / transcription Model: google/gemini-3.5-transcribe Endpoint: POST https://ai-gateway.vercel.sh/v4/ai/transcription-model Auth: AI Gateway API key (system credentials, not BYOK) Example generationId: gen_01M11K75WKZPWRXY37603K104G

What I need

Speaker diarization from Gemini 3.5 Transcribe: word-level speaker labels (spk_1, spk_2, …) so I can split a meeting transcript by speaker.

Google’s native Transcribe API supports this. Docs: https://ai.google.dev/gemini-api/docs/transcribe

Native request shape:

"mode": { "type": "verbatim", "diarization_mode": "speaker" }

What Gateway does
1. Native mode object is rejected
providerOptions.google.mode as an object (type + diarization_mode) returns HTTP 400: “invalid google provider options”.

Gateway does not accept the Google-native diarization config.

2. Flattened options are accepted, but speaker metadata is stripped
This is accepted (HTTP 200):

POST [https://ai-gateway.vercel.sh/v4/ai/transcription-model](https://ai-gateway.vercel.sh/v4/ai/transcription-model)
Authorization: Bearer $AI_GATEWAY_API_KEY
Content-Type: application/json
ai-model-id: google/gemini-3.5-transcribe
ai-transcription-model-specification-version: 4
ai-gateway-protocol-version: 0.0.1
ai-gateway-auth-method: api-key
{
  "audio": "",
  "mediaType": "audio/mpeg",
  "providerOptions": {
    "google": {
      "mode": "VERBATIM",
      "diarization": true,
      "languageCodes": ["ru-RU"]
    }
  }
}
Also tried (all 200, same empty speaker data):

diarizationMode: "speaker"
timestampGranularities: ["word"]
Response always looks like:

{
  "text": "",
  "segments": [],
  "warnings": [],
  "providerMetadata": {
    "google": { "usage": { "...": "token counts only" } }
  }
}
There is no word_info, no spk_*, no speaker field on segments. segments is always [].

The catalog entry for this model lists supported_parameters as chat-style fields (max_tokens, temperature, stop) and does not mention diarization.

Docs also don’t cover Google diarization on this endpoint: [https://vercel.com/docs/ai-gateway/modalities/speech-to-text](https://vercel.com/docs/ai-gateway/modalities/speech-to-text)

Ask
Is this a known limitation of the transcription abstraction, or a bug?
Can Gateway pass through Google’s word_info (or fill segments[] with speaker + timestamps) when diarization is requested?
If not yet, is there a supported providerOptions mapping for diarization_mode: "speaker", and a timeline to surface it?
This is a capability gap, not an integration bug on my side. I’m already billed on Gateway credits for these calls. Workaround today is a second Gateway chat request with the same audio, which doubles cost and is worse than native word_info.

"mode": { "type": "verbatim", "diarization_mode": "speaker" } What Gateway does 1. Native mode object is rejected providerOptions.google.mode as an object (type + diarization_mode) returns HTTP 400: “invalid google provider options”. Gateway does not accept the Google-native diarization config. 2. Flattened options are accepted, but speaker metadata is stripped This is accepted (HTTP 200): POST [https://ai-gateway.vercel.sh/v4/ai/transcription-model](https://ai-gateway.vercel.sh/v4/ai/transcription-model) Authorization: Bearer $AI_GATEWAY_API_KEY Content-Type: application/json ai-model-id: google/gemini-3.5-transcribe ai-transcription-model-specification-version: 4 ai-gateway-protocol-version: 0.0.1 ai-gateway-auth-method: api-key { "audio": "", "mediaType": "audio/mpeg", "providerOptions": { "google": { "mode": "VERBATIM", "diarization": true, "languageCodes": ["ru-RU"] } } } Also tried (all 200, same empty speaker data): diarizationMode: "speaker" timestampGranularities: ["word"] Response always looks like: { "text": "", "segments": [], "warnings": [], "providerMetadata": { "google": { "usage": { "...": "token counts only" } } } } There is no word_info, no spk_*, no speaker field on segments. segments is always []. The catalog entry for this model lists supported_parameters as chat-style fields (max_tokens, temperature, stop) and does not mention diarization. Docs also don’t cover Google diarization on this endpoint: [https://vercel.com/docs/ai-gateway/modalities/speech-to-text](https://vercel.com/docs/ai-gateway/modalities/speech-to-text) Ask Is this a known limitation of the transcription abstraction, or a bug? Can Gateway pass through Google’s word_info (or fill segments[] with speaker + timestamps) when diarization is requested? If not yet, is there a supported providerOptions mapping for diarization_mode: "speaker", and a timeline to surface it? This is a capability gap, not an integration bug on my side. I’m already billed on Gateway credits for these calls. Workaround today is a second Gateway chat request with the same audio, which doubles cost and is worse than native word_info.