Do word by word subtitle-audio sync

The usual subtitles files, such as .srt, do a line by line sync - meaning the subtitles appear when the person starts talking, says a few words, then the line disappears and a new one appears, etc.

1 00:02:17,440 –> 00:02:20,375
Senator, we're making
our final approach into Coruscant.

In this .srt example, at minute 2, second 17 those two lines of text appear and then they disappear at 2:20.

The task is to tag each individual word as is being spoken. This implies audio analysis. While in principle it doesn't seem terrible hard (since you just need to distinguish between individual words for which you at least have an ordered list) keep in mind that some times subtitles don't match audio 100%. For those words that do match, you need to provide a perfect audio-subtitle sync. For those words in the subtitle files that don't appear in the audio (this is a corner, yet possible, case) add some indicator. Finally for those words in the audio that don't appear in the subtitles, add a different indicator.

Focus on the challenging parts of the project, which is the sync itself. You can assume that the subtitle format is always .srt and don't deal with additional formats, since conversion tools exists. Similarly, you can assume that the audio is a .wav file and forget about dealing with video formats. FFmpeg can deliver a raw wav from almost any stream which is more than enough.

The solution needs to work in real time, meaning that it must be possible to pipe the subtitles and audio data into your program and get the word-by-word sync'ed version has it happens. So things like double pass are out of the question.

As a suggestion, take a look at this. You don't have to use it (you can if you want), but it's worth checking out for ideas and concepts.

  • public/gsoc/wordbywordsync.txt
  • Last modified: 2021/03/14 21:07
  • (external edit)