Subgen Setup Guide - ajgillis04/GillisDockerDepot GitHub Wiki
Subgen creates subtitles with Whisper (speech to text) for videos that have none. It can work on its own, watching your folders or reacting to Plex/Jellyfin events, or as an AI provider for Bazarr Setup Guide.
Is it right for you? Whisper writes out everything said, so it makes full subtitles for a whole movie. It doesn't make "forced" subtitles that only cover foreign-language scenes in an otherwise English film. It's most useful for foreign-language films and shows with no subtitles available anywhere.
compose/server7/subgen.yaml:
services:
subgen:
container_name: subgen.${HOST_NAME}
image: mccloud/subgen:latest
environment:
PUID: ${PUID}
PGID: ${PGID}
TZ: ${TZ}
WHISPER_MODEL: small.en
WHISPER_THREADS: 4
TRANSCRIBE_DEVICE: cpu # or cuda with a GPU
CONCURRENT_TRANSCRIPTIONS: 1
WEBHOOKPORT: 9000
PROCADDEDMEDIA: "True" # new media from Plex/Jellyfin webhooks
PROCMEDIAONPLAY: "True" # media when someone presses play
ONLY_SKIP_IF_SUBGEN_SUBTITLE: "True"
MONITOR: "True"
MODEL_PATH: /subgen/models
#TRANSCRIBE_FOLDERS: "/tv|/movies" # batch the backlog once, then turn off
ports:
- "${SUBGEN_PORT}:9000"
volumes:
- ${DOCKERDIR}/subgen/models:/subgen/models
- ${MEDIASHARE}/Videos/TV:/tv
- ${MEDIASHARE}/Videos/Movies:/movies
restart: alwaysThe media paths inside the container must match what Plex or Bazarr report, or Subgen can't find the files.
In Bazarr, Settings → Providers → Whisper, endpoint http://subgen.<host>:9000. Bazarr then asks Subgen for a subtitle when no provider has one.
- On a CPU, transcribing a full movie takes a long time, even with
small.en. KeepCONCURRENT_TRANSCRIPTIONSat 1 and let it work through the queue. -
.enmodels only handle English audio. For foreign-language audio use a multilingual model (small,medium) with translation to English. - Batch mode (
TRANSCRIBE_FOLDERS) processes everything in those folders; turn it off once the backlog is done.