README.md

metadata

configs:
  - config_name: all
    data_files: '*/*.tar'
    default: true
  - config_name: 3blue1brown
    data_files: 3blue1brown/*.tar
  - config_name: boxofficemoviesscenes
    data_files: boxofficemoviesscenes/*.tar
  - config_name: business
    data_files: business/*.tar
  - config_name: markrober
    data_files: markrober/*.tar
  - config_name: marvel
    data_files: marvel/*.tar
  - config_name: mitocw
    data_files: mitocw/*.tar
  - config_name: mkbhd
    data_files: mkbhd/*.tar
  - config_name: msftmechanics
    data_files: msftmechanics/*.tar
  - config_name: neoexplains
    data_files: neoexplains/*.tar
  - config_name: nvidia
    data_files: nvidia/*.tar
  - config_name: quantasciencechannel
    data_files: quantasciencechannel/*.tar
  - config_name: teded
    data_files: teded/*.tar
  - config_name: theinfographicsshow
    data_files: theinfographicsshow/*.tar
  - config_name: twominutepapers
    data_files: twominutepapers/*.tar
  - config_name: veritasium
    data_files: veritasium/*.tar
license: mit
task_categories:
  - automatic-speech-recognition
language:
  - en
size_categories:
  - 100K<n<1M

English Audio Dataset from YouTube

This dataset contains English audio segments extracted from various YouTube channels, along with corresponding transcription metadata. The data is intended for training automatic speech recognition (ASR) models.

Data Source and Processing

The data was obtained through the following process:

1. Download: Audio (.m4a) and available English subtitles (.srt for en, en.j3PyPqV-e1s) were downloaded from selected YouTube channels. This raw data, along with video metadata (metadata.csv), is stored initially in a data/{channel_id}/ directory structure.
2. Segmentation: The raw audio files were segmented based on the timing information in the .srt files.
   • Audio files are splitted by SRT segments and then combined to a maximum duration less than but close to 30 seconds per group for Whisper.
   • The corresponding audio portions for each group are extracted using ffmpeg and saved as .mp3 files at a 16000 Hz sample rate.
   • Metadata for each segment, including channel/video info and the text/timing of subtitles within the segment, is saved in a corresponding .json file.

Intermediate Dataset Structure (dataset directory)

Before being packaged into TAR archives for Hugging Face, the segmented data resides in the dataset directory with the following structure:

dataset/
└── {channel_id}/             # Directory named after the YouTube channel ID
    └── {video_id}/           # Directory named after the YouTube video ID
        ├── {video_id}_{group_name}.mp3  # Segmented audio file
        ├── {video_id}_{group_name}.json # Corresponding metadata file
        └── ...

• {channel_id}: The ID of the YouTube channel (e.g., greenbeanmediaofficial).
• {video_id}: The unique identifier for the YouTube video.
• {group_name}: Represents the subtitles included in the segment. It's either the index of the first subtitle (e.g., 1) if the group contains only one, or a range indicating the first and last subtitle indices (e.g., 1-5) if the group contains multiple subtitles.

Dataset Summary

The dataset comprises audio from the following channels:

Channel               |       Videos |      Duration | Percent
--------------------- | ------------ | ------------- | -------
3blue1brown           |   136 videos |   37.82 hours |   1.08%
boxofficemoviesscenes |  1626 videos |  153.06 hours |   4.38%
business              |   887 videos |  187.80 hours |   5.38%
markrober             |    97 videos |   21.77 hours |   0.62%
marvel                |   763 videos |   35.17 hours |   1.01%
mitocw                |  2844 videos | 1738.07 hours |  49.79%
mkbhd                 |   114 videos |   27.61 hours |   0.79%
msftmechanics         |   732 videos |  131.52 hours |   3.77%
neoexplains           |    35 videos |    8.06 hours |   0.23%
nvidia                |   134 videos |   19.42 hours |   0.56%
quantasciencechannel  |    93 videos |   13.60 hours |   0.39%
teded                 |  1768 videos |  145.53 hours |   4.17%
theinfographicsshow   |  3402 videos |  827.06 hours |  23.69%
twominutepapers       |   871 videos |   79.34 hours |   2.27%
veritasium            |   291 videos |   64.96 hours |   1.86%
--------------------- | ------------ | ------------- | -------
Total                 | 13793 videos | 3490.79 hours | 100.00%

Loading the Data

You can load the data using the Hugging Face datasets library:

import os

from datasets import load_dataset

ds = load_dataset(
    "OrcinusOrca/YouTube-English",
    "all",  # or channel_id as config
    split="train",
    streaming=False,  # or True
    num_proc=os.cpu_count(),
)