Skip to main content

ybs Subtitles

Solving Sentence Break Issues​

Users who have used YouTube's automatically generated subtitles know that there has always been an issue with sentence breaks. These subtitles lack punctuation, different sentences stick together, and break off at some point (as shown in the main subtitle below). Since YouTube's subtitle translation is based on these broken sentences, it leads to poor translation quality.

13

We use machine learning to fix these subtitles, recombining the broken sentences, greatly improving the translation quality.

14

Generation Speed​

Since ybs subtitles are based on existing subtitles, this makes their generation extremely fast.

For example, for the 19-hour tutorial video of freecodecamp, we only need 30 seconds to generate all subtitles.

For the 48-minute video of Melanie Nakagawa, Chief Sustainability Officer at Microsoft, ybs takes 12 seconds, while whisper takes 3 minutes and 22 seconds (using RTX 4090)

Supported Languages​

Here are the languages supported by ybs subtitles:

AI Repair​

Since ybs subtitles are generated based on automatically generated subtitles, the quality of voice recognition technology used by YouTube directly affects the quality of ybs subtitles. Some languages have lower error rates, like English, while others frequently have errors, like Japanese. Currently, our solution is to use AI to repair these subtitles, which can fix some more obvious errors such as grammatical mistakes or missing vocabulary, but it may also change the sentence structure and wording.