Name speakers from 1:1 participants and unique voiceprints

[ComputelessComputer]

When a transcript has only you on the mic and one other participant, we should label their turns instead of leaving them as Speaker 1. In larger meetings, named voiceprints should identify people only when the match is unique and high-confidence.

The summary already inferred names (Marco (Speaker 1)), but the transcript did not:

• Enhance attribution required two named non-owner participants and two speaker clusters, so a John + Marco meeting never ran.
• The live participant list used for channel assignment also counted excluded people and calendar copies of the current user, so the “unique remote” path never fired.
• Voiceprints were collected only if “Remember speakers” was turned on, and extracted embeddings were never matched against known people.

Changes

• Assign the only other participant to the only unassigned remote/mixed speaker after enhance, without calling the LLM.
• Ignore excluded participants and owner/self-email duplicates when collecting human IDs for transcript rendering, live capture, enhance candidates, and the speaker-hint write guard. Self-copy matching uses coalesced human.email / participant.email, so a calendar copy whose address lives only on the participant row still drops from the remote count.
• Turn Remember speakers on by default for this update. An explicit off stays off — we do not overwrite a stored preference.
• After candidate extraction, match new embeddings against this meeting’s participants only (same model and capture domain). Assign a speaker only when they uniquely match one person above a score floor and margin. Skip your mic channel, already-named speakers, and 1:1 meetings (those still use the participant list). Thin margins or two speakers wanting the same person stay unlabeled.
• Score and write assignment against the extract-time words snapshot. If a hints-write conflict shows a newer diarization (words_json changed), abort instead of labeling the wrong clusters. Stored candidates are skipped when their speaker keys are missing from the current transcript.

Validation

• pnpm -F desktop typecheck
• pnpm -F desktop exec vitest run on speaker-attribution, content-queries, content-mutations, and stt/queries tests
• pnpm exec oxlint --quiet --format=github on the changed desktop files
• pnpm exec dprint check on the changed files
• cargo test -p tauri-plugin-transcription could not run here (gdk-3.0.pc missing)

  

[netlify]

✅ Deploy Preview for anarlog canceled.

Name	Link
🔨 Latest commit	41accbe
🔍 Latest deploy log	https://app.netlify.com/projects/anarlog/deploys/6a8e4e4686a3730008adb756

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 000e717. Configure here.}