Default speaker labels on and keep the cast list open

Bumps caption-local to d29fca6 (option ticked by default, clearer empty
panel hint, cast list survives polling), captionninja to 529779a (synced
capture bundle) and captions-coop-wiki to 44bb25e (guide update).

.env.production

@@ -1,3 +1,4 @@
 LETSENCRYPT_EMAIL=boop@autonomic.zone
 UI_HOST=app.captions.dev.autonomic.zone
 WIKI_HOST=wiki.captions.dev.autonomic.zone
+CAPTION_DIARIZATION=1

AGENTS.md

@@ -277,6 +277,12 @@ Syntax-check edited scripts with `node --check`.
 - **Cookies**: `caption_room`, `HttpOnly`, `SameSite=Lax`, `Path=/`, `Secure`
   by default on HTTPS. `CAPTION_COOKIE_SECURE` may force `1` or `0`; `auto` is
   the default. Do not move room sessions into localStorage.
+- **Speaker labels**: opt-in with `CAPTION_DIARIZATION=1` (plus
+  `CAPTION_SPEAKER_THRESHOLD`, `CAPTION_MAX_SPEAKERS`) in the Compose env file —
+  Compose `environment:` overrides `secrets/caption-local.env`. CPU-only
+  (`sherpa-onnx`, `caption-local/speakers.py`); no audio or voiceprints are
+  persisted. Cast lists persist per room; per-show names live in memory. The
+  relay whitelists the `spk` caption field, so deploy relay changes first.
 - **Offline LAN**: `.env.lan.example` binds the web UI to a LAN address. Plain
   `ws://`/`http://` relay URLs are allowed only with
   `CAPTION_ALLOW_INSECURE_LAN=1`, and only for private LAN hosts. Keep Whisper

CHANGELOG.md

@@ -9,6 +9,51 @@ Format loosely follows [Keep a Changelog](https://keepachangelog.com/).
 Commit references link to their repo: submodule changes link to the
 `caption-local`/`captionninja` forks, everything else to `captions-deploy`.
 
+## 2026-10-29
+
+### Fixed
+- **Identify speakers automatically** is now ticked by default when the service
+  offers it, and the empty speaker panel explains that it must be on when
+  captions start. The cast list no longer collapses every time the speaker panel
+  polls (`caption-local`
+  [`d29fca6`](https://git.coopcloud.tech/captions.coop/caption-local/commit/d29fca6),
+  `captionninja`
+  [`529779a`](https://git.coopcloud.tech/captions.coop/captionninja/commit/529779a)).
+
+### Docs
+- The speaker names guide covers the new default
+  ([`44bb25e`](https://git.coopcloud.tech/captions.coop/captions-coop-wiki/commit/44bb25e)).
+
+## 2026-10-28
+
+Speaker names for live captions.
+
+### Added
+- Optional automatic speaker labels (`CAPTION_DIARIZATION=1`, with
+  `CAPTION_SPEAKER_THRESHOLD` and `CAPTION_MAX_SPEAKERS`): CPU-only pyannote
+  segmentation plus TitaNet-small voiceprints via `sherpa-onnx`, numbered per room
+  in arrival order. `/transcribe?speakers=1` returns speaker segments; the capture
+  page offers **Identify speakers automatically** and publishes one caption per
+  speaker turn (`caption-local`
+  [`618f7ff`](https://git.coopcloud.tech/captions.coop/caption-local/commit/618f7ff)).
+- Persisted room cast lists and per-show speaker names through
+  `/api/rooms/<id>/speakers`, with a shared speaker panel on the capture page and
+  editor ([`618f7ff`](https://git.coopcloud.tech/captions.coop/caption-local/commit/618f7ff)).
+- Editor speaker field: fix one line, name a detected speaker for the rest of the
+  show (queued lines are relabelled), or press Alt+1–9 for a cast member. The
+  audience page shows speaker labels (`?nolabels` hides them), and the relay
+  accepts an optional `spk` field (`captionninja`
+  [`cb05906`](https://git.coopcloud.tech/captions.coop/captionninja/commit/cb05906)).
+
+### Changed
+- Compose passes `CAPTION_DIARIZATION`, `CAPTION_SPEAKER_THRESHOLD` and
+  `CAPTION_MAX_SPEAKERS` to `caption-local`; speaker labels stay off by default.
+
+### Docs
+- Speaker names guide, configuration, API and hardware notes in the wiki
+  ([`4386a23`](https://git.coopcloud.tech/captions.coop/captions-coop-wiki/commit/4386a23));
+  `docs/CONFIGURATION.md` and `AGENTS.md` cover the new settings.
+
 ## 2026-10-26
 
 Self-service room links and a single-origin deployment.

docker-compose.prod.yml

@@ -11,6 +11,9 @@ services:
     environment:
       - PYTHONUNBUFFERED=1
       - CAPTION_ROOMS_DB=/data/rooms.db
+      - CAPTION_DIARIZATION=${CAPTION_DIARIZATION:-0}
+      - CAPTION_SPEAKER_THRESHOLD=${CAPTION_SPEAKER_THRESHOLD:-0.5}
+      - CAPTION_MAX_SPEAKERS=${CAPTION_MAX_SPEAKERS:-12}
       - CAPTION_COOKIE_SECURE=1
       - RELAY_ADDRESS=wss://${UI_HOST:?set UI_HOST in .env.production}/ws
       - RELAY_SITE=https://${UI_HOST:?set UI_HOST in .env.production}/

docker-compose.yml

@@ -14,6 +14,9 @@ services:
     environment:
       - PYTHONUNBUFFERED=1
       - CAPTION_ROOMS_DB=/data/rooms.db
+      - CAPTION_DIARIZATION=${CAPTION_DIARIZATION:-0}
+      - CAPTION_SPEAKER_THRESHOLD=${CAPTION_SPEAKER_THRESHOLD:-0.5}
+      - CAPTION_MAX_SPEAKERS=${CAPTION_MAX_SPEAKERS:-12}
       - CAPTION_MODEL=${CAPTION_MODEL:-small}
       - CAPTION_OFFLINE=${CAPTION_OFFLINE:-0}
       - CAPTION_ALLOW_INSECURE_LAN=${CAPTION_ALLOW_INSECURE_LAN:-0}

docs/CONFIGURATION.md

@@ -117,6 +117,25 @@ CAPTION_ALLOWED_ORIGINS=https://ui.example.org
 The env file is optional (`required: false`). Do not put `CAPTION_API_KEY` or
 relay tokens here; the shared-token API model has been replaced by room sessions.
 
+#### Speaker labels
+
+Automatic speaker labels are off by default. Compose sets these variables in
+`environment:`, which overrides `env_file`, so set them in the Compose env file
+(`.env.production`) or the shell rather than `secrets/caption-local.env`:
+
+| Variable | Meaning |
+|----------|---------|
+| `CAPTION_DIARIZATION` | `1` loads the CPU speaker models (~47 MB, cached in the model volume) and enables `/transcribe?speakers=1`; default `0` |
+| `CAPTION_SPEAKER_THRESHOLD` | Voice similarity to reuse a speaker, 0–1; default `0.5` (lower merges, higher splits) |
+| `CAPTION_MAX_SPEAKERS` | Automatic speakers per room; default `12` |
+
+Room cast lists persist in `rooms.db` (`cast_list` column, added automatically).
+Detected voices and per-show names stay in memory and reset with
+`POST /api/rooms/<id>/speakers/reset`, room closure or six idle hours. Speaker
+models add about 150 MB of memory and ~0.1 s of CPU per six-second window on two
+cores. The relay accepts an optional `spk` caption field (`S1`–`S99`); deploy the
+relay with or before pages that send it.
+
 ### `secrets/traefik.htpasswd`
 
 Only needed with `docker-compose.auth.yml`. It adds a Basic Auth perimeter to the
@@ -150,6 +169,8 @@ All routes are same-origin; room routes use the `caption_room` cookie.
 | `POST` | `/api/rooms/<id>/rotate` | room cookie; returns a new admin link once |
 | `DELETE` | `/api/rooms/<id>` | room cookie |
 | `GET` | `/api/rooms/<id>/config` | room cookie |
+| `GET`/`PATCH` | `/api/rooms/<id>/speakers` | room cookie; cast list and per-show speaker names |
+| `POST` | `/api/rooms/<id>/speakers/reset` | room cookie |
 | `GET` | `/api/public/rooms/<id>` | public, rate-limited; output read token only |
 | `GET` | `/api/admin/rooms` | admin key |
 | `POST` | `/api/admin/rooms/<id>/rotate` | admin key |