README.md

metadata

dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: transcription
      dtype: string
    - name: gender
      dtype: string
    - name: age
      dtype: int64
    - name: speaker_id
      dtype: int64
    - name: duration
      dtype: float64
    - name: md5sum
      dtype: string
    - name: pdp_score
      dtype: float64
  splits:
    - name: train
      num_bytes: 4926086264.08
      num_examples: 41871
    - name: test
      num_bytes: 452358128.68
      num_examples: 2802
  download_size: 6360680451
  dataset_size: 5378444392.76
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*
license: cc-by-3.0
task_categories:
  - automatic-speech-recognition
language:
  - zu
pretty_name: NCHLT isiZulu Auxiliary Speech Corpus

Dataset Details

This is the Zulu language split of the NCHLT speech corpus (nchlt-clean split). It containes 56 hours and 14 minutes hours of isiZulu speech from 210 (98 Male; 112 Female) native speakers of the language. Each recording consists of a speaker reading three words. These words were sourced from South African government websites. The full corpus can be downloaded from the SADILAR website.

For a full description of the corpus, see Barnard et al. 2014.

Citation

BibTeX:

@inproceedings{barnard2014nchlt,
  title={The NCHLT speech corpus of the South African languages},
  author={Barnard, Etienne and Davel, Marelie H and van Heerden, Charl and De Wet, Febe and Badenhorst, Jaco},
  year={2014},
  organization={Workshop Spoken Language Technologies for Under-resourced Languages (SLTU)}