ADR-083: Speaker Identification and Voiceprints (exploration) - #1

Summary

Scopes bringing cross-video speaker identification and a per-person knowledge base into the catalog. The catalog today treats participants/transcript_text speaker labels as opaque, source-supplied strings with no identity that persists across videos or platforms (Fireflies and Zoom each diarize independently; nothing here does).

Uses two internal tools' already-working implementation of this pattern as a reference design — not code to port, since neither matches this stack:

• ListeningBuddy: pyannote.audio diarizes each recording and produces a per-speaker voice embedding; a clustering pass groups embeddings across the entire corpus, surfaces unnamed clusters with sample clips, and a human names each cluster once — the name applies retroactively to every member. New recordings get auto-suggested/confirmed against known voiceprints by cosine similarity.
• PeopleBuddy: once a voice is named, an LLM builds a running narrative dossier per person from everything they've said across every recording.

This ADR is Proposed (exploration) — it records the problem and the reference design without committing to an implementation. It explicitly leaves open, as decisions for maintainers before any Phase 1 code:

1. Where diarization compute runs (no GPU host exists in this project's infra today, unlike the reference tools)
2. Whether identity needs to unify across platforms (e.g. a Fireflies meeting and a raw Zoom recording of the same person)
3. Access/privacy — this catalog has multiple operator roles (Admin/Publisher/Viewer, ADR-010) vs. the reference tools' single-user context, so a persistent cross-video "who said what" index needs an explicit ownership decision, not an engineering default

Proposes a new event-sourced Speaker aggregate (following the ADR-008 pattern) with participants kept as a synced legacy field, the same precedent ADR-077 §1 set for destination_outcomes.

Test plan

N/A — documentation only (ADR), no code changes. docs/adr/README.md regenerated via scripts/gen-adr-index.sh and diff reviewed to confirm only the new entry was added.