The United Nations has a project to acquire public domain translations, into as many languages as possible, of the Universal Declaration of Human Rights (UDHR). Librivox.org has a project to acquire readings of the UDHR in as many languages as possible.

This repository exists for the purpose of segmenting the librivox recordings, into chunks amenable for the training and testing of automatic speech recognizers and synthesizers, and then aligning them to the corresponding texts.

Segmented text and automatically generated phone transcriptions are distributed with the corpus (in the "text" and "phones" subdirectories, respectively). Audio is not distributed with the corpus. To get it, you'll need to have wget, unzip, ffmpeg, and python installed; then try the following steps:

pip install pycountry praatio librosa h5py
python udhrpy/prepare_data.py --audio

This will create a directory exp, with subdirectories:

• zip contains the zip files, downloaded from librivox
• mp3 contains the mp3, unzipped from the zip files
• wav contains the wav, converted from mp3 using ffmpeg.
• audio contains the segmented wav files.

You can then move audio out of exp, and delete the rest of exp.

import udhrpy
udhrpy.load_audio()
udhrpy.create_hdf5('UDHR.hdf5')
dataset=udhrpy.UDHR_Dataset('UDHR.hdf5')
print('Smallest melspectrogram has shape %s'%(str(dataset[0]['melspectrogram'].shape)))
print(' from uttid %s, language = %s'%(dataset[0]['uttid'][()],dataset[0]['languagename'][()]))
print(''.join(dataset.idx2phone[y] for y in dataset[0]['phones'][:])+'\n')
print('Largest melspectrogram has shape %s'%(str(dataset[-1]['melspectrogram'].shape)))
print(' from uttid %s, language = %s'%(dataset[-1]['uttid'][()],dataset[-1]['languagename'][()]))
print(''.join(dataset.idx2phone[y] for y in dataset[-1]['phones'][:])+'\n')

If you want to download the original source texts from the UDHR in Unicode project, do this:

python scripts/prepare_data.py --text

To convert the texts into international phonetic alphabet, using the languagenet G2Ps, install https://github.com/AdolfVonKleist/Phonetisaurus. Then do:

python scripts/prepare_data.py --phones

The source URLs are listed in the files in the conf subdirectory.

1. If the UDHR in Unicode project (https://www.unicode.org/udhr/) doesn't yet have text of the UDHR in your language, please donate a translation at https://www.unicode.org/udhr/contributing.html.

2. If librivox doesn't have an audio recording in your language, please consider donating one. Register for a librivox account at https://forum.librivox.org/ucp.php?mode=register, then mention your language ability and your interest on the "newbie forum" (https://forum.librivox.org/viewforum.php?f=17). You can also read the text and post your result in the "Short Works" forum (https://forum.librivox.org/viewforum.php?f=19), as soon as you've completed the "one minute sound check" described in the Newbie Guide to Recording (https://wiki.librivox.org/index.php?title=Newbie_Guide_to_Recording).

3. If you would like to contribute directly to the corpus, e.g., by segmenting data or by creating a train/test example, please go ahead and send us a pull request!