metadata

dataset_info:
  features:
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: transcription
      dtype: string
    - name: duration
      dtype: float64
  splits:
    - name: train
      num_bytes: 4531568161.596
      num_examples: 5154
    - name: validation
      num_bytes: 1294272351.875
      num_examples: 1473
    - name: test
      num_bytes: 644375190
      num_examples: 737
  download_size: 6400544345
  dataset_size: 6470215703.471
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
      - split: test
        path: data/test-*
license: cc-by-3.0
language:
  - xh
task_categories:
  - automatic-speech-recognition
tags:
  - low resource languages

Xhosa Merged Audio

This dataset was cultivated from Beijuka/xhosa_parakeet_50hr. This dataset orginally came from NCHLT isiXhosa Speech Corpus (see below).

The original corpus contained audio and transcription in 3-5 word segments. This meant that the majority of the dataset was ~5 seconds long. Whisper can receive an input of 30 seconds. This meant that the dataset required substantial padding. To reduce the amount of padding, the audio segments were merged together sequentially until a limit of 30 seconds was reached.

Original Dataset

Name: NCHLT isiXhosa Speech Corpus
Size: Approximately 56 hours of transcribed speech
Speakers: 209 (106 female, 103 male)
Content: Prompted speech (3-5 word utterances read from a smartphone screen)
Source: Audio recordings smartphone-collected in non-studio environment
License: Creative Commons Attribution 3.0 Unported License (CC BY 3.0)

Citation

De Vries, N.J., Davel, M.H., Badenhorst, J., Basson, W.D., de Wet, F., Barnard, E. and de Waal, A. (2014). A smartphone-based ASR data collection tool for under-resourced languages. Speech Communication, 56, 119-131. https://hdl.handle.net/20.500.12185/279