metadata

dataset_info:
  features:
    - name: audio
      dtype: audio
    - name: text_ga
      dtype: string
    - name: text_en
      dtype: string
  splits:
    - name: train
      num_bytes: 4209155993
      num_examples: 15090
  download_size: 3452146050
  dataset_size: 4209155993
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Dataset Details

Synthetic audio dataset, created using Azure text-to-speech service. The bilingual text is a portion of the Wikimedia dataset, consisting of 7545 text segments. The dataset comprises two sets of audio data, one with a female voice (OrlaNeural) and the other with a male voice (ColmNeural). It includes more than 34 hours of audio data (34:23:12) for 15,090 utterances.

Dataset Structure

Dataset({
    features: ['audio', 'text_ga', 'text_en'],
    num_rows: 15090
})

Citation

@inproceedings{moslem2024leveraging,
  title={Leveraging Synthetic Audio Data for End-to-End Low-Resource Speech Translation},
  author={Moslem, Yasmin},
  booktitle={Proceedings of the 2024 International Conference on Spoken Language Translation (IWSLT 2024)},
  year={2024},
  month={April}
}