Table of Contents
AI Subtitles
Open AI Subtitles in the left menu. The catalog scans on application startup; Scan libraries refreshes it after adding or changing subtitles. It uses Config → Jellyfin TV/movie directories, falling back to /jellyfin/tv and /jellyfin/movies (the existing Docker mounts). TV titles are grouped by their top-level folder and season folder or Sxx filename; movies by their top-level folder. External .srt and .vtt files can be translated when they contain UTF-8 dialogue. Existing Czech .ass, .ssa, .sub and .ttml sidecars are catalogued as well, but cannot be translated here. Embedded tracks and other encodings must be extracted/converted first.
Choose Anime TV Shows or Anime Movies. Titles appear in a compact list with search, an All / 0–9 / A–Z filter and pagination. Choose 20, 50, 100, 200, 500 or All entries per page; the default is 50. Alphabet filters use the first character of the title (accented letters match their base letter). Search and alphabet filters work together.
Click a title, then Queue file, Queue season, or Queue full series / Queue movie. English tracks are translated to Czech; Czech tracks (cs, ces, cze, cz, or detected Czech text) remain catalogued and are skipped. Language labels come from filename suffixes or a conservative English/Czech text estimate. Confirm or correct uncertain labels using the language dropdown before queueing. Other languages are catalogued but not translated by this English-to-Czech workflow. Existing Czech sidecars for the same track are also skipped, including across supported catalog formats.
In Config → AI Subtitles, choose Ollama or OpenAI API, then set the endpoint and model. Load available models fetches the current model names; you can also enter a custom name. Ollama defaults to http://192.168.88.15:11434 and translategemma:4b; install that model with ollama pull translategemma:4b if needed. KAIZOKU_OLLAMA_URL and KAIZOKU_SUBTITLE_MODEL supply initial defaults for existing deployments, while saved Config values take precedence. OpenAI defaults to https://api.openai.com/v1 and gpt-4.1-mini and requires an API token. The token is stored in .kaizoku config, omitted from config API responses, and can be replaced or cleared from Config. The OpenAI API uses separate billing from a ChatGPT subscription. Resume / retry retains the job's original provider, endpoint, model and saved cues; clearing or removing an unfinished job discards its cues so the next queue request uses current Config settings. OpenAI jobs read the current saved API token when they run.
Gemma 4 models use Ollama's chat endpoint so translation instructions are separate from subtitle dialogue; other Ollama models continue using the generate endpoint. Obvious instruction echoes and identical long responses to different cues are rejected before they become checkpoints. If an older saved cue contains an instruction echo, the job stops and identifies the cue; open Review on the failed job, correct that cue, then retry. Already completed Czech sidecars are not changed automatically.
The 17slever17/translate-gemma-4-sub-* Ollama fine-tune uses its documented single-segment chat prompt ([CURRENT_SOURCE] with optional previous source/translation context). Kaizoku saves each cue separately for this model, so a failed cue resumes without redoing earlier cues. Its model card says Czech was not among its specifically fine-tuned or extensively evaluated languages; review Czech output carefully. Existing saved cues are retained on direct retry; use Restart from beginning to discard bad partial output and apply the current Config model.
For a paused, stopped or failed file with saved cues, Restart from beginning on the AI Subtitles file row discards that file's saved Czech cues and immediately queues it again from cue 1 using the current Config provider/model. Resume / retry still keeps progress and the original provider/model. Clear saved progress on a stopped or failed file discards its checkpoints without starting a translation; the file then shows as Not queued and can be queued using current Config settings. These actions do not change the English source, completed Czech sidecars, or other files in a season/series job; the queue log records the reset. They are unavailable while a file is actively translating or after it has completed.
The Queue → AI Subtitles tab shows one job for each new single-file request or season/full-series request, with per-file status, combined cue progress and persistent logs for batch saves, retries, splits, failures and completion. Pause keeps a job ready to resume; Stop marks its unfinished files canceled, prevents further subtitle output, and makes Remove available. An in-flight model request may take up to three minutes to return, but its result is discarded after Stop. Remove hides the queue card and clears saved cues for its failed or canceled files; completed sidecars and translation history remain intact. Requeue a cleared file to start from cue 1 with the current Config provider/model. Stopped jobs can also be retried before removal, and Remove finished clears stopped cards. Existing jobs from earlier versions are preserved as individual queue entries because their original request grouping was not recorded.
After a rescan marks a subtitle missing, Remove missing hides its stale catalog entry. Pause any queued/running translation first. This action never deletes library files, checkpoints or queue history. A file that reappears is restored to the catalog on the next scan.
Use Review beside a translated file to open the English source and generated Czech cues side by side, with timing and length warnings. Select Edit beside a Czech cue, change its wording, then Save cue. A completed job updates its Czech sidecar and saved checkpoint; a paused, stopped or failed job updates its saved cue for the next resume. Active jobs and untranslated cues cannot be edited. Edits preserve cue numbers, timestamps and inline formatting, reject stale reviews or changed source files, and refresh length warnings. Unfinished jobs show their saved translations and blank cells for unfinished cues. The floating review window closes only with Close; outside clicks and Escape leave it open.
Translations run one file at a time in natural episode order, in batches of up to eight cues with preceding bilingual dialogue and recent examples from the same title. Add recurring names, preferred terms, honorifics, character relationships and speech styles under Terminology & character voices before queueing a title. Notes are captured when jobs are created, so a resumed translation uses the same notes. Translation memory, notes, catalog and per-batch checkpoints live in .kaizoku/ai_subtitles.sqlite3; keep this database with the rest of the persistent state.
The model receives dialogue with synthetic markers, never actual cue IDs or timestamps. Python preserves SRT numbering, timing, VTT identifiers/settings, headers, metadata, styling, BOM and newline format. Missing/duplicate/reordered cues, empty text, formatting changes, malformed responses and truncated output trigger up to three attempts per batch. Malformed batches are split into smaller groups, down to individual cues. If a single cue still loses inline formatting markers, Python keeps the original tags and whitespace and translates only the visible text spans, carrying preceding dialogue and already-translated spans as context. Empty or malformed span translations still fail validation; existing checkpoints remain available for retry. Resume / retry reuses saved batches; Pause preserves progress, although an in-flight model request can take up to three minutes to return. Pending/running jobs recover after restart; manually paused, stopped and failed jobs wait for you. Changed source files stop safely with their old checkpoints retained; restore the source to resume or use a new filename for a new job.
Cues exceeding two lines or 42 characters per line receive one shortening attempt. Remaining length warnings list cue positions for review. Translation quality and nuanced anime terminology still need human review, especially with a 4B model; structural validation cannot prove semantic accuracy. Completed translations are written alongside the source as Title.cs.srt or Title.cs.vtt (track qualifiers such as forced remain). Output is UTF-8, atomically published only after all cues validate, and never overwrites an existing file. Library mounts must permit creating files and hard links. Source subtitles, videos and existing Kaizoku state remain intact. Jellyfin discovers the new sidecars through its normal library scan; this page does not trigger a Jellyfin API refresh.
Rebuild/restart the Docker service after updating so the new page and worker are included. The existing Jellyfin and .kaizoku volume mappings are reused.