1
"mode": { "type": "verbatim", "diarization_mode": "speaker" }
2

3
What Gateway does
4
1. Native mode object is rejected
5
providerOptions.google.mode as an object (type + diarization_mode) returns HTTP 400: “invalid google provider options”.
6

7
Gateway does not accept the Google-native diarization config.
8

9
2. Flattened options are accepted, but speaker metadata is stripped
10
This is accepted (HTTP 200):
11

12
POST [https://ai-gateway.vercel.sh/v4/ai/transcription-model](https://ai-gateway.vercel.sh/v4/ai/transcription-model)
13
Authorization: Bearer $AI_GATEWAY_API_KEY
14
Content-Type: application/json
15
ai-model-id: google/gemini-3.5-transcribe
16
ai-transcription-model-specification-version: 4
17
ai-gateway-protocol-version: 0.0.1
18
ai-gateway-auth-method: api-key
19
{
20
  "audio": "",
21
  "mediaType": "audio/mpeg",
22
  "providerOptions": {
23
    "google": {
24
      "mode": "VERBATIM",
25
      "diarization": true,
26
      "languageCodes": ["ru-RU"]
27
    }
28
  }
29
}
30
Also tried (all 200, same empty speaker data):
31

32
diarizationMode: "speaker"
33
timestampGranularities: ["word"]
34
Response always looks like:
35

36
{
37
  "text": "",
38
  "segments": [],
39
  "warnings": [],
40
  "providerMetadata": {
41
    "google": { "usage": { "...": "token counts only" } }
42
  }
43
}
44
There is no word_info, no spk_*, no speaker field on segments. segments is always [].
45

46
The catalog entry for this model lists supported_parameters as chat-style fields (max_tokens, temperature, stop) and does not mention diarization.
47

48
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)
49

50
Ask
51
Is this a known limitation of the transcription abstraction, or a bug?
52
Can Gateway pass through Google’s word_info (or fill segments[] with speaker + timestamps) when diarization is requested?
53
If not yet, is there a supported providerOptions mapping for diarization_mode: "speaker", and a timeline to surface it?
54
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.