Use Youtube as speech set for speech recognition data

Question

DNNs are huge data hungry algorithms specially bidirectional-LSTM. For speech recognition models, there's a dearth of voice corpus for specific language/ accents. However, YouTube is a huge library for the same.

I would need to use some script/software which would automatically download YouTube videos(possibly only audio) and the related captions and thereafter slice the same into 20 seconds clip, so that the same can be used as training speech corpus.

Is there any way captions from Youtube can be downloaded? And given a audio and the scripts, how can I slice a audio into 20 seconds clips and have the same clipped in the transcript as well?

Answer 1

yt-dlp (or youtube-dl or another fork of youtube-dl)

1. Can download video and/or audio and/or subtitles and/or other data from YouTube and/or many other websites.
2. Easy to select the videos you want to target. Note options such as "--match-filter" to filter on almost any attribute and "--batch-file", which will download a list of URLs in the named file.
3. Easy to grab most information without downloading the video. See, for example, "--skip-download"
4. Many options explicitly for subtitles: see below. Especially note that you can distinguish between the automatic subtitles, which are machine generated, and files there were uploaded by a human, which doesn't mean they are accurate or generated by a human, just that they were not put in that location by YouTube.
5. You can discard the video and keep only the audio. See, for example, "--extract-audio"

Subtitle options:

--write-subs                     Write subtitle file
--no-write-subs                  Do not write subtitle file (default)
--write-auto-subs                Write automatically generated subtitle file
                                 (Alias: --write-automatic-subs)
--no-write-auto-subs             Do not write auto-generated subtitles
                                 (default) (Alias: --no-write-automatic-subs)
--list-subs                      List available subtitles of each video.
                                 Simulate unless --no-simulate is used
--sub-format FORMAT              Subtitle format, accepts formats
                                 preference, for example: "srt" or
                                 "ass/srt/best"
--sub-langs LANGS                Languages of the subtitles to download (can
                                 be regex) or "all" separated by commas.
                                 (Eg: --sub-langs "en.*,ja") You can prefix
                                 the language code with a "-" to exempt it
                                 from the requested languages. (Eg:
                                 --sub-langs all,-live_chat) Use --list-subs
                                 for a list of available language tags

FFmpeg

1. With a little cleverness, you should be able to create arbitrary length segments of audio and their associated subtitles. See, for example, this page describing how to remove the first five minutes of a video file and remove the first five minutes of the corresponding subtitle file.
2. Extremely good at converting media to your desired codec.
3. When configured properly (which isn't always obvious how to do), it is good at creating files with a precise time length.
4. Fast. Free. Open-source. Multiplatform. Extensible.

This question is three years old, so I imagine the OP already found many options. But, maybe this will help other people.