[nkeating-cmd]

What

Separate speakers for meetings held in a room, where everyone reaches the microphone. A new action on a meeting's Audio tab works out who said what from the microphone recording and labels each microphone line Speaker 1, Speaker 2 and so on. The labels reach the transcript view, copy, export, search, Ask and the notes model.

• Engine (src-tauri/src/diarise.rs), D-007:
  • Runs on Kōrero's own ort 2.0.0-rc.12, so there is one ONNX runtime, shared with transcribe-rs.
  • CPU only, after the meeting.
  • pyannote segmentation-3.0 is read for speaker changes, not just speech, and re-anchored every 10 s window.
  • WeSpeaker ResNet34-LM embeddings, then whole-meeting clustering.
• Models (D-008): downloaded on first use (about 32 MB) and kept only if their size and sha256 match the pinned values. Not bundled in the installer.
• Data:
  • TranscriptSeg.speaker is optional, with a serde default, so every saved meeting loads unchanged.
  • Meetings gain speakersFound, speakerNames (for naming in M3) and mergedFrom.
• Refused, with the reason shown:
  • imports;
  • meetings recorded before 1.26, whose lines have no timings;
  • combined meetings, whose lines come from two recordings.

Stacked on #1 (the brief, decision log and lessons log). Merge that first; this diff then shrinks to the feature.

Why

Bet B4 in docs/PRODUCT.md (in Now; product-check.py gate B4: PASS). Today a meeting's only "speakers" are the two capture channels, so everyone in a room is "You".

Evidence (main laptop, 16 Oct)

• cargo test --lib: 264 passed, 0 failed (13 new in diarise).
• bun test unit-tests: 92 passed (10 new in unit-tests/speakers.test.ts). These include checks that meetings without speakers give exactly the text they gave before.
• tsc: clean. eslint: 0 errors, and no warnings in the touched files.
• clippy: no findings in the new code (64 pre-existing elsewhere; the CI step is advisory).
• diar_ami_der (ignored; run locally, 177 MB of AMI audio):
  • Mean DER 21.5% with a 0.25 s collar on four AMI far-field test meetings, against a bar of 25%. For reference, pyannote's own pipeline scores 22.4% strict on all 16 test meetings; this scores 26.3% strict on 4 of them.
  • The check's own scorer agrees with pyannote.metrics to within 0.2 points.
  • 92 minutes of audio separated in about 65 s.
• cargo tree -i ort-sys: one ort.
• gitleaks on all 5 commits: no leaks. No personal paths or hostnames in the added lines.
• Independent review (read-only agent, briefed with every accepted cost):
  • No Critical or High findings, and no path that loses saved meetings.
  • Fixed in this PR: M1–M4, L1–L4, L6, L8, L10, L12 (see the second commit).

Not done yet

• Not clicked through in the running app. The UI is type-checked and unit-tested, but no installed build has run it. Next: a seeded hand-over build (dev-session section 3).
• Review follow-ups:
  • L5: overlapping same-speaker turns.
  • L7: the models are hashed on the async runtime on every run.
  • L9: the whole recording is held in memory, about 230 MB per hour.
  • L11: a line already open for editing when separation starts loses its edit.
  • L13: regenerate bindings.ts with tauri-specta. It was already stale; this PR adds one command by hand, matching the Rust exactly.
  • L14: CPU load if a recording starts during separation.
  • L15: verify the segmentation-3.0 (MIT) and kaldi-native-fbank (Apache-2.0) licences.
• Still in B4: naming, merge and split (M3); splitting lines where the speaker changes mid-sentence; mirroring the two model files to a Kōrero release (D-008).
• No version bump: 1.44 ships first.

Decision records: D-007, D-008 in docs/DECISIONS.md. Lessons: L-012 closed (check-added), L-013 added.

🤖 Generated with Claude Code

https://claude.ai/code/session_01M9niYsTjphhTSYCLhCwg5n