Diarization

Pass diarization=true (query param, form field, or SDK option) on a model whose provider page shows the diarization mark. Models that can’t diarize reject at connect time with unsupported_capability — never silently ignored.

What you get

Integer speaker labels on words, normalized across providers (vendors label "A"/"B", "Guest-1", or 0/1 — you always see 0, 1, 2 …):

{ "type": "transcript", "is_final": true,
  "text": "I disagree. Let me explain.",
  "words": [
    { "w": "I",       "start": 4.1, "end": 4.2, "speaker": 0 },
    { "w": "disagree","start": 4.2, "end": 4.6, "speaker": 0 },
    { "w": "Let",     "start": 5.0, "end": 5.1, "speaker": 1 }
  ] }

Granularity varies by model. Most diarizing models label every word. Segment-level models (e.g. openai/gpt-4o-transcribe-diarize) return one words-entry per speaker turn instead — same schema, coarser grain.

Failover semantics

Speaker indices are only meaningful within one provider’s run. After a provider_switched event, speaker_mapping_preserved: false tells you the labels restarted — speaker 0 after the switch is not necessarily speaker 0 from before. Treat each provider span as its own labeling epoch (or re-align in your application using overlap heuristics).

Pricing notes

Diarization is included in the model’s list price except where the vendor meters it separately — most notably Azure, where diarized streaming is its own product: use azure/conversation-transcription ($1.20/hr) — the $1.00 azure/speech-realtime refuses diarization=true rather than silently billing you a different meter.