Speaker diarization API: confidence scores, and what Sume STT returns

A speaker diarization API labels which speaker said each part of a recording. AssemblyAI's changelog says that, when you set speaker_options.include_speaker_confidence to true, each word and each utterance gets a speaker_confidence value between 0 and 1. Sume STT returns word start and end times and optional sentence segments, with no speaker label and no confidence value.

AssemblyAI facts are from its changelog, read 2026-09-30. Sume facts are from the OpenAPI document for sume/stt-1.0.

What is a speaker confidence score for?

The changelog says it shows how confident the model is in a given speaker assignment. In practice you would use it to route low-score segments to a human reviewer, or to skip them when you build per-speaker clips. It only makes sense where a model assigns speakers in the first place.

What fields does a Sume STT result have?

Can I ask Sume STT to diarize?

No. The request body is closed, and the MCP tool description says diarize and tag_audio_events are fixed server-side, so sending them is rejected. A flag like AssemblyAI's would be rejected the same way.

What can I build from word timings alone?

Captions, paper edits and sentence-level cuts do not need speakers; see word timestamps. For a multi-person interview, record each person on their own file, transcribe each as a separate job and interleave by start, as in how to transcribe an interview. Speaker identity then comes from your recording setup, and there is no confidence score to read.