youtube_tts_data_generator

Youtube Speech Data Generator

A python library to generate speech dataset. Youtube Speech Data Generator also takes care of almost all your speech data preprocessing needed to build a speech dataset along with their transcriptions making sure it follows a directory structure followed by most of the text-to-speech architectures.

Installation

Make sure ffmpeg is installed and is set to the system path.

$ pip install youtube-tts-data-generator

Minimal start for creating the dataset

from youtube_tts_data_generator import YTSpeechDataGenerator

# First create a YTSpeechDataGenerator instance:

generator = YTSpeechDataGenerator(dataset_name='elon')

# Now create a '.txt' file that contains a list of YouTube videos that contains speeches.
# NOTE - Make sure you choose videos with subtitles.

generator.prepare_dataset('links.txt')
# The above will take care about creating your dataset, creating a metadata file and trimming silence from the audios.

Usage

• Initializing the generator: generator = YTSpeechDataGenerator(dataset_name='your_dataset',lang='en')

  • Parameters:
    • dataset_name:
      • The name of the dataset you'd like to give.
      • A directory structure like this will be created:

        ├───your_dataset
        │   ├───txts
        │   └───wavs
        └───your_dataset_prep
            ├───concatenated
            ├───downloaded
            └───split
        

    • output_type:
      • The type of the metadata to be created after the dataset has been generated.
      • Supported types: csv/json
      • Default output type is set to csv
      • The csv file follows the format of LJ Speech Dataset
      • The json file follows this format:

        {
            "your_dataset1.wav": "This is an example text",
            "your_dataset2.wav": "This is an another example text",
        }
        

    • keep_audio_extension:
      • Whether to keep the audio file extension in the metadata file
      • Default value is set to False
    • lang:
      • The key for the target language in which the subtitles have to be downloaded.
      • Default value is set to en
      • Tip - check list of available languages and their keys using: generator.get_available_langs()
    • sr:
      • Sample Rate to keep of the audios.
      • Default value is set to 22050

• Methods:

  • download():
    • Downloads video files from YouTube along with their subtitles and saves them as wav files.
    • Parameters:
      • links_txt:
        • Path to the '.txt' file that contains the urls for the videos.
    • Usage of this method is optional. If you do not use this method, make sure to place all the audio and subtitle files in 'your_dataset_prep/downloaded' directory.
    • Then, create a file called 'files.txt' and again place it under 'your_dataset_prep/downloaded'. 'files.txt' should follow the following format:

      filename,subtitle,trim_min_begin,trim_min_end
      audio.wav,subtitle.srt,0,0
      audio2.wav,subtitle.vtt,5,6
      

    • Create a '.txt' file that contains a list of YouTube videos that contains speeches.
    • Example - generator.download('links.txt')
  • split_audios():
    • This method splits all the wav files into smaller chunks according to the duration of the text in the subtitles.
    • Saves transcriptions as '.txt' file for each of the chunks.
    • Example - generator.split_audios()
  • concat_audios():
    • Since the split audios are based on the duration of their subtitles, they might not be so long. This method joins the split files into recognizable ones.
    • Parameters:
      • max_limit:
        • The upper limit of length of the audios that should be concated. The rest will be kept as they are.
        • The default value is set to 7
      • concat_count:
        • The number of consecutive audios that should be concated together.
        • The default value is set to 2
    • Example - generator.concat_audios()
  • finalize_dataset():
    • Trims silence the joined audios since the data has been collected from YouTube and generates the final dataset after finishing all the preprocessing.
    • Parameters:
      • min_audio_length:
        • The minumum length of the speech that should be kept. The rest will be ignored.
        • The default value is set set to 5.
      • max_audio_length:
        • The maximum length of the speech that should be kept. The rest will be ignored.
        • The default value is set set to 14.
    • Example - generator.finalize_dataset(min_audio_length=6)
  • get_available_langs():
    • Get list of available languages in which the subtitles can be downloaded.
    • Example - generator.get_available_langs()
  • get_total_audio_length():
    • Returns the total amount of preprocessed speech data collected by the generator.
    • Example - generator.get_total_audio_length()
  • prepare_dataset():
    • A wrapper method for download(),split_audios(),concat_audios() and finalize_dataset().
    • If you do not wish to use the above methods, you can directly call prepare_dataset(). It will handle all your data generation.
    • Parameters:
      • links_txt:
        • Path to the '.txt' file that contains the urls for the videos.
      • sr:
        • Sample Rate to keep of the audios.
        • Default value is set to 22050
      • download_youtube_data:
        • Whether to download audios from YouTube.
        • Default value is True
      • max_concat_limit:
        • The upper limit of length of the audios that should be concated. The rest will be kept as they are.
        • The default value is set to 7
      • concat_count:
        • The number of consecutive audios that should be concated together.
        • The default value is set to 2
      • min_audio_length:
        • The minumum length of the speech that should be kept. The rest will be ignored.
        • The default value is set set to 5.
      • max_audio_length:
        • The maximum length of the speech that should be kept. The rest will be ignored.
        • The default value is set set to 14.
    • Example - generator.prepare_dataset(links_txt='links.txt', download_youtube_data=True, min_audio_length=6)

Final dataset structure

Once the dataset has been created, the structure under 'your_dataset' directory should look like:

your_dataset
├───txts
│   ├───your_dataset1.txt
│   └───your_dataset2.txt
├───wavs
│    ├───your_dataset1.wav
│    └───your_dataset2.wav
└───metadata.csv/alignment.json

NOTE - audio.py is highly based on Real Time Voice Cloning

References

SRT to JSON

Read more about the library here