Bilingual subtitle test files (BookTranslator, 2026-11-04). Free to reuse.

sample-source.srt                 12-cue English dialogue written for this test
sample-es-bilingual.srt           English + Spanish output of the writer code described below
sample-zh-bilingual.srt           English + Simplified Chinese, same method
sample-ar-bilingual.srt           English + Arabic, same method (Arabic lines wrapped in U+2067/U+2069)
sample-es-translation-on-top.srt  the Spanish file after `flip`
sample-es-own-timing.srt          a Spanish track with its own timing (splits two cues, adds one)
sample-es-merged-by-time.srt      sample-source + own-timing after `merge`
sample-es-two-colors.vtt          the Spanish file after `to-vtt`
sample-vtt-settings.vtt           VTT with cue settings, a voice span and a NOTE
sample-vtt-settings-bilingual.vtt its output: settings and NOTE gone
frame-*.webp                      screenshots of Chromium's built-in <track> rendering, 1280x720,
                                  using a VTT copy of the SRT files
check_bilingual_subs.py           check, flip, merge, to-vtt (Python standard library)

How the bilingual files were made: BookTranslator's subtitle parse and write code
(apps/api/src/translation/plugins/subtitle, commit 94102c66d) was run locally. Logging,
file paths, the preview cutoff and character-encoding detection (UTF-8 only here) were
stubbed out, and translate() was skipped, so the translation pipeline did not run and no
model was called. The Spanish, Chinese, Arabic and French lines were written by hand and passed
to write() as the translations, so these files test the file structure, not translation
quality. Line counts in a real job depend on how the model breaks the translated cue.

Checker: `check` counts characters with tags removed, entities decoded and spaces included.
Default limits are 42 characters per line, 20 characters per second, and 3 lines on
screen. The 3-line default is our own choice. Set the target's limits with
--max-target-line and --max-target-cps.

Not tested: VLC, mpv, TVs, and hardcoding, because the ffmpeg build we used has no
subtitles filter.