Hugging Face's logo

Datasets:
Racoci
/
CORAA-v1.1

Tasks:
Automatic Speech Recognition
Text-to-Speech
Modalities:
Audio
Text
Formats:
parquet
Languages:
Portuguese
Size:
100K - 1M
ArXiv:
Tags:
Libraries:
Datasets
Dask
Croissant
License:
Dataset card Viewer Files Community
1
CORAA-v1.1 / README.md
Racoci's picture
Racoci
Upload CORAA dataset on Hugging Face format
439e5fc verified
preview code
|
raw history blame contribute delete
No virus
6.13 kB
metadata
license: cc-by-nc-nd-4.0
dataset_info:
  features:
    - name: file_path
      dtype: string
    - name: task
      dtype: string
    - name: variety
      dtype: string
    - name: dataset
      dtype: string
    - name: accent
      dtype: string
    - name: speech_genre
      dtype: string
    - name: speech_style
      dtype: string
    - name: up_votes
      dtype: int64
    - name: down_votes
      dtype: int64
    - name: votes_for_hesitation
      dtype: float64
    - name: votes_for_filled_pause
      dtype: float64
    - name: votes_for_noise_or_low_voice
      dtype: float64
    - name: votes_for_second_voice
      dtype: float64
    - name: votes_for_no_identified_problem
      dtype: float64
    - name: text
      dtype: string
    - name: audio
      dtype: audio
  splits:
    - name: train
      num_bytes: 63113404687.162
      num_examples: 382258
    - name: dev
      num_bytes: 1363924625
      num_examples: 7522
    - name: test
      num_bytes: 2594334946
      num_examples: 12676
  download_size: 66914186143
  dataset_size: 67071664258.162
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: dev
        path: data/dev-*
      - split: test
        path: data/test-*
task_categories:
  - automatic-speech-recognition
  - text-to-speech
language:
  - pt
pretty_name: coraa
size_categories:
  - 1K<n<10K

CORAA-v1.1

CORAA-v1.1 is a publicly available dataset for Automatic Speech Recognition (ASR) in the Brazilian Portuguese language containing 290.77 hours of audios and their respective transcriptions (400k+ segmented audios). The dataset is composed of audios of 5 original projects:

  • ALIP (Gonçalves, 2019)
  • C-ORAL Brazil (Raso and Mello, 2012)
  • NURC-Recife (Oliviera Jr., 2016)
  • SP-2010 (Mendes and Oushiro, 2012)
  • TEDx talks (talks in Portuguese)

The audios were either validated by annotators or transcripted for the first time aiming at the ASR task.

LICENSE

Attribution-NonCommercial-NoDerivatives 4.0 International

Metadata

  • file_path: the path to an audio file
  • task: transcription (annotators revised original transcriptions); annotation (annotators classified the audio-transcription pair according to votes_for_* metrics); annotation_and_transcription (both tasks were performed)
  • variety: European Portuguese (PT_PT) or Brazilian Portuguese (PT_BR)
  • dataset: one of five datasets (ALIP, C-oral Brasil, NURC-RE, SP2010, TEDx Portuguese)
  • accent: one of four accents (Minas Gerais, Recife, Sao Paulo cities, Sao Paulo capital) or the value "miscellaneous"
  • speech_genre: Interviews, Dialogues, Monologues, Conversations, Interviews, Conference, Class Talks, Stage Talks or Reading
  • speech_style: Spontaneous Speech or Prepared Speech or Read Speech
  • up_votes: for annotation, the number of votes to valid the audio (most audios were revewed by one annotor, but some of the audios were analyzed by more than one).
  • down_votes: for annotation, the number of votes do invalid the audio (always smaller than up_votes)
  • votes_for_hesitation: for annotation, votes categorizing the audio as having the hesitation phenomenon
  • votes_for_filled_pause: for annotation, votes categorizing the audio as having the filled pause phenomenon
  • votes_for_noise_or_low_voice: for annotation, votes categorizing the audio as either having noise or low voice, without impairing the audio compression.
  • votes_for_second_voice: for annotation, votes categorizing the audio as having a second voice, without impairing the audio compression
  • votes_for_no_identified_problem: without impairing the audio as having no identified phenomenon (of the four described above)
  • text: the transcription for the audio

Experiments:

Model trained in this corpus: Wav2Vec 2.0 XLSR-53 (multilingual pretraining)

Citation 

@misc{c2021coraa,
    title={CORAA: a large corpus of spontaneous and prepared speech manually validated for speech recognition in Brazilian Portuguese},
    author={Arnaldo Candido Junior and Edresson Casanova and Anderson Soares and Frederico Santos de Oliveira and Lucas Oliveira and Ricardo Corso Fernandes Junior and Daniel Peixoto Pinto da Silva and Fernando Gorgulho Fayet and Bruno Baldissera Carlotto and Lucas Rafael Stefanel Gris and Sandra Maria Aluísio},
    year={2021},
    eprint={2110.15731},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

Partners / Sponsors / Funding

References

  • Gonçalves SCL (2019) Projeto ALIP (amostra linguística do interior paulista) e banco de dados iboruna: 10 anos de contribuição com a descrição do Português Brasileiro. Revista Estudos Linguísticos 48(1):276–297.
  • Raso T, Mello H, Mittmann MM (2012) The C-ORAL-BRASIL I: Reference corpus for spoken Brazilian Portuguese. In: Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), European Language Resources Association (ELRA), Istanbul, Turkey, pp 106–113, URL http://www.lrec-conf.org/proceedings/lrec2012/pdf/624_Paper.pdf
  • Oliviera Jr M (2016) Nurc digital um protocolo para a digitalização, anotação, arquivamento e disseminação do material do projeto da norma urbana linguística culta (NURC). CHIMERA: Revista de Corpus de Lenguas Romances y Estudios Linguísticos 3(2):149–174, URL https://revistas.uam.es/chimera/article/view/6519
  • Mendes RB, Oushiro L (2012) Mapping Paulistano Portuguese: the SP2010 Project. In: Proceedings of the VIIth GSCP International Conference: Speech and Corpora, Fizenze University Press, Firenze, Italy, pp 459–463.