Transcripts for podcasts and audiobooks: the show's own, or speech-to-text on your server (whisper service)

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

docker-compose.yml

@@ -22,6 +22,7 @@ services:
       MS_CLIENT_ID: ${MS_CLIENT_ID:-}                       # Outlook.com mail (Microsoft app registration, see README -> Mail); leave empty to switch Mail off
       MS_CLIENT_SECRET: ${MS_CLIENT_SECRET:-}
       LYRICS_API_URL: http://127.0.0.1:9999                  # the bundled lyrics service (Lyrica, below, shares this container's network); remove this line and the service to skip it
+      TRANSCRIBE_API_URL: http://127.0.0.1:8000              # the bundled speech-to-text service (whisper, below): makes a transcript for a podcast or audiobook on request. Remove this line and the service to skip it
       SIGNAL_API_URL: http://127.0.0.1:8085                  # the bundled Signal bridge (below, shares this container's network); remove this line and the service to skip Signal
       OPENAUDIBLE_URL: ${OPENAUDIBLE_URL:-}                # audiobooks straight from OpenAudible (e.g. http://172.16.16.106:8000); the username and password in Settings -> Audiobooks sign-in are OpenAudible's. Leave empty to go through the audiobook player instead
       PODCAST_URL: ${PODCAST_URL:-https://podcast.necloud.us/}
@@ -48,6 +49,17 @@ services:
     volumes:
       - signal-data:/home/.local/share/signal-cli
 
+  # Speech to text (the "Transcribe" button on podcasts and audiobooks). faster-whisper behind an OpenAI-compatible API (the speaches project); the model
+  # is downloaded the first time it is used. CPU only: a one-hour episode takes roughly 5-15 minutes with the small model. Nothing leaves your server.
+  whisper:
+    image: ghcr.io/speaches-ai/speaches:latest-cpu
+    restart: unless-stopped
+    network_mode: "service:hub"       # reached by the hub at 127.0.0.1:8000 and by nothing else
+    depends_on:
+      - hub
+    volumes:
+      - whisper-models:/home/ubuntu/.cache/huggingface/hub
+
   # Lyrics and karaoke for the music player. Lyrica looks songs up on public lyrics sources (LRCLIB, Musixmatch via Lrcmux, NetEase, Megalobiz);
   # the song's artist and title are sent to them. Plex's own lyrics are used first when the song has them. See README -> Music.
   lyrica:
@@ -65,3 +77,5 @@ volumes:
     name: nehub-data   # fixed name so the data survives redeploys and renamed stacks
   signal-data:
     name: nehub-signal
+  whisper-models:
+    name: nehub-whisper