Speaker diarization / labels in transcript outputs

[frarredondo]

Summary

Interview episodes render as unlabeled dialogue, which makes quotes and “who said what” hard to follow.

Motivation / evidence

Shows like CoRecursive, Signals & Threads, and Beyond Coding are mostly multi-speaker. Current outputs have segments with timing/text only—no speaker field.

Proposal

Add optional diarization (flag/config), e.g. --diarize, producing speaker labels in text and JSON (SPEAKER_00 / SPEAKER_01, with optional name hints from title/description later).

podtx sync --diarize
podtx format episode.json --diarize   # only if re-derivable from existing data; otherwise sync-only

Out of scope

Perfect identity resolution, real-time streaming diarization, guaranteed name accuracy without metadata.

Acceptance criteria

• Multi-speaker transcripts can include stable speaker labels in outputs
• Feature is opt-in (default behavior unchanged)
• Performance/memory impact documented at a high level
• Docs / --help updated

[fasuizu-br]

BRAINIALL published a provider-neutral regression fixture that appears to match this issue's output-contract surface: 20.884 s of first-party synthetic PT-BR audio, 2 speakers, 4 turns, expected JSON/SRT/WebVTT/TXT, hashes, and a dependency-free local validator: https://github.com/fasuizu-br/brainiall-asr-diarization-fixture

If it would be useful, we can contribute one focused GPLv3-compatible fixture/test after an explicit opt-in from the maintainer. Limits up front: clean synthetic audio, no overlap/noise, and not a DER/WER or production-quality benchmark. We will not open a PR or follow up unless requested.

[frarredondo]

@fasuizu-br Thanks! This could be useful as a future smoke test asset once diarization lands. If we decide to use a GPLv3-adapted fixture, we'll reach out. No PR needed for now.