NLP

A collection of resources for natural language processing. Mostly links to datasets for machine learning approaches.

Datasets

1. Mozilla Common Voice
   Link: https://commonvoice.mozilla.org/en/datasets
   Content:

   • 2185 hours validated english (of 2886 hours total) by 79398 voices
   • 1062 hours validated german (of 1133 hours total) by 16390 voices
   • 2000 hours validated kinyarwanda (of 2383 hours total) by 1055 voices
   • 826 hours validated french (of 902 hours total) by 16082 voices
   • 404 hours validated spanish (of 739 hours total) by 22741 voices
   • 162 hours validated russian (of 193 hours total) by 2452 voices
   • 310 hours validated italian (of 335 hours total) by 6576 voices
   • and many more languages

2. LJ Speech Dataset
   Link: https://keithito.com/LJ-Speech-Dataset/
   Content:

   • 24 hours english by 1 voice

3. CSTR VCTK Corpus
   Link: https://datashare.ed.ac.uk/handle/10283/3443
   Content:

   • ~400 sentences english each by 110 voices

4. Libri Vox
   Link: https://librivox.org
   Content:

   • unknown amounts of voices
   • 33270 books english
   • 2649 books german
   • 868 books french
   • 742 books spanish
   • 261 books italian
   • 430 books chinese

5. Libri Speech
   Link: https://www.openslr.org/12
   Content:

   • extracted from LibriVox (see 4.)
   • ~1000 hours english
   • ~585 hours in higher quality at https://openslr.org/60/

6. Vox Forge
   Link: http://www.repository.voxforge1.org/downloads/SpeechCorpus/Trunk/
   Content:

   • collection of speech files / transcripts by unknown amount of voices
   • 6319 files english
   • 2260 files french
   • 1419 files german
   • 1060 files italian
   • 630 files russian
   • 2248 files spanish
   • and some more

7. TED LIUM
   Link: https://www.openslr.org/51/
   Content:

   • 452 hours by unknown amount of voices

8. Torsten Müller
   Link: https://www.openslr.org/110/
   Content:

   • 300 phrases in 8 different emotions
   • ~3 hours german by 1 voice

9. Emotional Voices Database
   Link: https://www.openslr.org/115/
   Content:

   • collection of audio with 3-5 different emotions
   • ~7000 files english by 4 voices

10. Tatoeba
    Link: https://tatoeba.org/en/downloads
    Content:

    • sentences with audio files by unknown amounts of voices
    • 692348 sentences english
    • 113008 sentences spanish
    • 32951 sentences german
    • 8173 sentences french
    • 7575 sentences russian
    • 1747 sentences mandarin chinese
    • 1529 sentences japanese
    • 198 sentences italian
    • and many more

11. Spoken Wikipedia Corpora
    Link: https://nats.gitlab.io/swc/
    Content:

    • spoken wikipedia articles
    • 386 hours german by 339 voices
    • 395 hours english by 395 voices
    • 224 hours dutch by 145 voices

12. M-AILABS Speech Dataset
    Link: https://www.caito.de/2019/01/the-m-ailabs-speech-dataset/
    Content:

    • mostly extracted from LibriVox (see 4.)
    • 237 hours german
    • 45 hours british english
    • 102 hours american english
    • 108 hours spanish
    • 127 hours italian
    • 87 hours ukranian
    • 46 hours russian
    • 190 hours french
    • 53 hours polish
    • contains mixed data i.e. female and male speakers

13. VCTK Noisy Speech Database
    Link: https://datashare.ed.ac.uk/handle/10283/2791
    Content:

    • noisy and clean audio files by up to 56 voices
    • includes written transcripts
    • unknown amount of hours

14. American English Speech Corpus
    Link: https://www.magicdatatech.com/datasets/mdt-tts-e018-american-english-speech-corpus-for-tts-1631179203
    Content:

    • ~2 hours american english by 1 female voice

15. American Male Voice Dataset
    Link: https://www.magicdatatech.com/datasets/mdt-tts-e009-american-male-voice-tts-dataset
    Content:

    • 15 hours american english by 1 male voice

16. Facebook Vox Populi
    Link: https://github.com/facebookresearch/voxpopuli
    Content:

    • download instructions in README of repository
    • in 16 european languages including english, german, french and spanish
    • 1800 hours transcribed audio by unknown amount of voices

17. Multilingual Libri Speech
    Link: https://openslr.org/94/
    Content:

    • unclear if transcripts provided
    • extracted from LibriVox (see 4.)

18. Kensho SPGI Speech
    Link: https://datasets.kensho.com/datasets/spgispeech
    Content:

    • transcribed company earnings calls
    • ~5000 hours international business english by ~50000 voices

19. Free Spoken Digit Dataset
    Link: https://github.com/Jakobovski/free-spoken-digit-dataset
    Content:

    • 3000 recordings english by 6 voices
    • 50 recordings per digit per voice

20. Flickr Audio Captions Corpus
    Link: https://groups.csail.mit.edu/sls/downloads/flickraudio/index.cgi
    Content:

    • 40000 spoken image captions english of 8000 images
    • download original captions here https://www.kaggle.com/adityajn105/flickr8k