# Speaker identity acceptance files

These files contain an original editorially constructed six-turn conversation with four named identities. They provide known text, speaker labels and timestamps for checking export preservation. They are not audio recordings, diarization results, transcripts generated by a service or measured product performance.


- **fixture.json** — unchanged canonical source.
- **constructed-turns.json** — all canonical turn/identity/role fields and explicit file provenance.
- **constructed-turns.txt** — turn ID, speaker ID, display name and millisecond anchors in a tab-separated heading, followed by verbatim text.
- **constructed-turns.srt** — numbered cues; each payload begins with a plain label such as [T01 | P01 | Maya], followed by the exact text.
- **constructed-turns.vtt** — turn/speaker IDs in the cue identifier, with the display name in a standard voice span.
- **fixture-verification.json** — independent parsing, exact field comparisons, negation/condition/correction checks, file hashes and encoding details.

Expected baseline: 6 turns, 4 distinct identities, 2 returning identities and 5 speaker transitions; 0–40,000 ms, with contiguous turn boundaries. The source wording preserves “Do not deploy build forty-two,” the budget approval condition, Sam/Maya ownership and the correction to build forty-three that still needs testing. Written numbers stay written numbers.


The SRT label and WebVTT voice annotation are explicit identity representations in these constructed files. Verification decodes them before comparing the canonical text. Product-result fields remain empty; preserving a fixture does not establish recognition accuracy or real-world speaker separation.