PGS to SRT: ffmpeg can't, SeConv can
1,275 of 1,275
words right with Gemma 4, 150 subtitles
Every word: Gemma 4 in SeConv
SeConv is the free command line of Subtitle Edit. Point it at Gemma 4 in Ollama:
ollama pull gemma4:e4b
seconv movie.sup subrip --ocr-engine:ollama --ollama-model:gemma4:e4b
Keep the model flag: SeConv's default Ollama model is llama3.2-vision.
9.5 GB model download
63 to 78 s in 2 runs
The second run was slower than the first, so take times as rough.
No download: Apple Vision
Apple Vision is built into macOS. It read 1,273 words in 13.6 s:
seconv movie.sup subrip --ocr-engine:applevision
It misread 2 words: 1 am for "I am", and
TIl for "I'll".
8 ways to read one Blu-ray .sup
| Reader | Words | Lines | Time |
|---|---|---|---|
| Gemma 4 e4b | 1,275 | 150 | 63 to 78 s |
| Apple Vision | 1,273 | 148 | 13.6 s |
| pgsrip + RapidOCR | 1,272 | 148 | 20.3 s |
| Tesseract + OCR rules | 1,272 | 147 | 39.7 s |
| pgsrip + Tesseract | 1,269 | 146 | 7.4 s |
| nOCR + OCR rules | 1,262 | 141 | 2.9 s |
| Tesseract | 1,224 | 109 | 37.2 s |
| nOCR | 1,194 | 90 | 2.4 s |
Rows without pgsrip are SeConv. Every SeConv run kept all 150 start and end times within 42 ms, one frame.
ffmpeg has no OCR step
ffmpeg -i act1.sup -c:s srt out.srt
Subtitle encoding currently only possible from text to text or bitmap to bitmap
It exits 234. ffmpeg 9.0.2 decodes PGS, but PGS subtitles are pictures, and it cannot read text out of a picture.
Tesseract writes | for I
Plain Tesseract put 44 pipes where the text has a capital I. Three OCR rules fix that:
seconv movie.sup subrip --ocr-engine:tesseract --ocr-language:eng \
--fix-common-errors-rules:FixCommonOcrErrors,FixUppercaseIInsideWords,FixAloneLowercaseIToUppercaseI
That took it from 1,224 to 1,272 words.
Skip --fix-common-errors, which runs all rules. It joins
two-line subtitles, adds periods (cut for.) and
capitalizes (By the way). Apple Vision fell to 1,259 with
it, Tesseract to 1,257.
Only nOCR keeps italics
The 26 italic stage directions came out plain in every other engine. nOCR kept all 26, in 2.9 s:
seconv movie.sup subrip --ocr-engine:nocr --ocr-db:/path/to/Latin.nocr \
--fix-common-errors-rules:FixCommonOcrErrors,FixUppercaseIInsideWords,FixAloneLowercaseIToUppercaseI
Latin.nocr is in the Subtitle Edit repo, under Ocr. Without the OCR
rules, nOCR swaps Arial's I and l: tabIe,
AIgernon.
DVD VobSub: Apple Vision slips
Give SeConv movie.idx, with the .sub next to it. On these
smaller 720x480 subtitles, Tesseract with the OCR rules beat Apple
Vision.
| Reader | Words | Lines | Time |
|---|---|---|---|
| Gemma 4 e4b | 1,275 | 150 | 36.2 s |
| Tesseract + OCR rules | 1,270 | 146 | 35.5 s |
| Apple Vision | 1,259 | 135 | 12.2 s |
| nOCR + OCR rules | 1,250 | 131 | 3.7 s |
| Tesseract | 1,229 | 113 | 31.7 s |
pgsrip dropped "Oh!"
pgsrip is a Python tool that rips PGS subtitles. With Tesseract it dropped the line "Oh!". With RapidOCR it also dropped "Oh! ...".
pip install "pgsrip[rapidocr]"
pgsrip rip -l en --engine rapidocr movie.en.sup
It writes movie.en.srt next to the .sup. Keep the .en: with -l en, a plain movie.sup is skipped ("1 file filtered out") and no .srt appears.
How I tested it
Real Blu-ray subtitles are copyrighted, so I made my own: the first 150 subtitles of Act I of The Importance of Being Earnest (1895, public domain). SeConv drew them as a Blu-ray .sup and a DVD VobSub.
The method
- Truth: 1,275 words, 26 italic subtitles, 76 two-line subtitles.
- The .sup: 1920x1080, Arial 50, white with a black outline. The VobSub: 720x480, Arial 28. ffmpeg 9.0.2 decodes both.
-
SeConv 5.5.0-beta6 (MIT), Tesseract 5.5.4 from Homebrew, pgsrip
0.3.0 (MIT) with
--force -w 2, Ollama 0.40.16 with gemma4:e4b (Apache-2.0). -
Every SeConv run also had
--ocr-language:eng, and the Gemma runs had--ollama-url:http://127.0.0.1:11473/api/chat(Ollama on a spare port). One run at a time, lowest priority, on an Apple Silicon Mac. - Score: true words found in order, lines with the exact text, start and end within 42 ms. Curly and straight apostrophes count the same.
- Bias: SeConv made the files and read them back. pgsrip is the outside reader. In Arial, a capital I and a small l look the same.
- pgsrip with RapidOCR took 28.1 s on its first run, with the model download. The table shows the second run.
What I did not test
- Real film subtitles, other fonts, colors and languages.
- MKV input, the Subtitle Edit window, Windows and Linux.
Sources
Text in a scan, not a track?
Run it yourself. Our Apify Actor turns scanned PDFs and images into Markdown. It does not read subtitle tracks.
Or send us the scan. Tack, our AI, does the job for $29, up to 50 pages. First page free.
Questions people ask
Can ffmpeg convert PGS to SRT?
No. PGS subtitles are pictures, and ffmpeg has no OCR step. ffmpeg 9.0.2 exits 234: "Subtitle encoding currently only possible from text to text or bitmap to bitmap".
How do I convert SUP to SRT on a Mac?
Use SeConv, the Subtitle Edit command line, with Apple Vision: seconv movie.sup subrip --ocr-engine:applevision. On my test it read 1,273 of 1,275 words in 13.6 s.
How do I convert VobSub to SRT?
Give SeConv the .idx file, with the .sub next to it. Gemma 4 read all 1,275 words. Tesseract with 3 OCR rules read 1,270.
Does Subtitle Edit OCR keep italics?
Only the nOCR engine did: 26 of 26 italic subtitles. Apple Vision, Tesseract and Gemma 4 kept none.