Fetch a text translation source and align it to the embedded track's timings #268
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: yolo/arr#268
Reference in New Issue
Block a user
Delete Branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
DESIGN.md §15. Found while testing a 2160p BluRay import (Batman Begins, movie 5): pt-PT can never materialise for that file, for two compounding reasons.
The deadlock
The release carries one English subtitle track,
hdmv_pgs_subtitle— bitmaps. §15 says an embedded track satisfies its language, so English reads as satisfied and arr never downloads an English SRT. §15 also says image tracks can never feed a translator. Sotranslation_source()(crates/arr-daemon/src/subtitles.rs:913) returnsNone— no sidecar with a path, no embedded text track — and the lane stops with "no text source to translate from".The file is stuck both ways: it cannot translate because it has no text, and it will not fetch text because it believes it already has English.
#206 describes the same dead end and proposes OCR, but its premise is "an English BluRay with no external subtitle available". That premise does not hold here — an English subtitle is readily available from OpenSubtitles, arr simply never asks for one. Closing the gap that way is far cheaper than OCR, and yields real text instead of OCR'd text that a translator would then turn into nonsense.
The timing objection, and why it does not require OCR
The embedded track's genuine advantage is that its timings are exact for this release, while a downloaded subtitle needs alass — a heuristic with no confidence value, which is why §15 already gates its output on plausibility.
That advantage is available without reading a single pixel: PGS timings are packet metadata.
ffprobe -select_streams s:0 -show_packetsreturns them directly, so a reference can be synthesised from the disc's own cue structure and used as alass's reference instead of the video's audio.Spike results
On the real file, over the whole film:
duration_timeisN/Afor PGS, but packets strictly alternate show/clear, so pairing them yields cues — every implied duration fell between 0.3s and 10s across all 1473.Alignment, measured on a synthetic 125-minute film (923 cues) whose subtitle was mistimed by a 23.976→25fps drift plus an offset, i.e. 316 seconds out by the end:
alass matches on interval structure, not words, so the two cue sets never have to agree — a sparse skeleton is still a strong signal. That matters because a retail disc's track and a downloaded SDH subtitle never carry the same cues.
The pairing is load-bearing, though. Treating every packet as a cue start instead of pairing them gives a consistent ~486 ms error — better than 152 s, but visibly wrong:
Work
WHERE path IS NOT NULL), so the fetched sidecar coexists with the embedded row — no schema change.Derivation costs a full demux, so it belongs at import — while the file is being read and hardlinked anyway — not on the translation path.
Not this issue
OCR (#206) stays filed for the case this does not reach: no provider has a text subtitle in any language, alongside #207. This issue should shrink #206's remaining scope to that case rather than replace it.