Feature Request - Apple Vision API for OCR

[Brian]

It might be worth looking into the possibility of using Apple’s native Vision API for the Subler’s OCR tasks.

Tessaract is admittedly very fast on current Macs, but I have recently been testing tesseract vs Apple’s Vision API recently for various OCR tasks.

On an Intel Mac, the Vision API “fast” method seems to be the safe choice, but on an M1, the “accurate” mode neural based OCR has blown tesseract away for accuracy.

[Damiano Galassi]

I tried to integrate it in Subler already and accuracy was a bit worse than Tesseract for vobsubs. I guess I could make it a selectable option in the preferences.

[Brian]

I wonder if the reason I was having better luck was because I was OCRing scanned PDFs which is likely closer to Apple’s training data than most embedded subtitle images.

If you add the option, I’ll certainly give it a try on some varied content. It’ll be interesting to look at speeds on M1 and M1 Max for Tessaract vs Vision API too.