videocr

Extract hardcoded subtitles from videos using the Tesseract OCR engine with Python.

Input video with hardcoded subtitles:

import videocr

print(videocr.get_subtitles('video.avi', lang='HanS'))

Output:

0
00:00:00,000 --> 00:00:02,711
-谢谢 … 你 好 -谢谢
Thank you...Hi. Thanks.

1
00:00:02,794 --> 00:00:04,879
喝 点 什么 ?
What can I get you?

2
00:00:05,046  --> 00:00:12,554
休闲 时 光 …
For relaxing times, make it...

3
00:00:12,804 --> 00:00:14,723
三 得 利 时 光
Bartender, Bob Suntory time.

4
00:00:16,474 --> 00:00:19,144
Un, I'll have a vodka tonic.

5
00:00:19,394 --> 00:00:20,687
谢谢
Laughs Thanks.

API

videocr.get_subtitles(
        video_path: str, lang='eng', time_start='0:00', time_end='',
        conf_threshold=65, sim_threshold=90, use_fullframe=False)

Return the subtitles string in SRT format.

videocr.save_subtitles_to_file(
        video_path: str, file_path='subtitle.srt', lang='eng', time_start='0:00',
        time_end='', conf_threshold=65, sim_threshold=90, use_fullframe=False)

Write subtitles to file_path. If the file does not exist, it will be created automatically.

Adjustable Parameters

• lang

  Language of the subtitles in the video.

• time_start and time_end

  Extract subtitles from only a part of the video. The subtitle timestamps are still calculated according to the full video length.

• conf_threshold

  Confidence threshold for word predictions. Words with lower confidence than this threshold are discarded. The default value is fine for most cases.

  Make it closer to 0 if you get too few words from the predictions, or make it closer to 100 if you get too many excess words.

• sim_threshold

  Similarity threshold for subtitle lines. Neighbouring subtitles with larger Levenshtein ratios than this threshold will be merged together. The default value is fine for most cases.

  Make it closer to 0 if you get too few subtitle lines, or make it closer to 100 if you get too many duplicated subtitles.

• use_fullframe

  By default, only the bottom half of each frame is used for OCR in order to reduce errors. You can explicitly make the algorithm handle the full frame.