metadata

dataset_info:
  features:
    - name: audio
      dtype:
        audio:
          sampling_rate: 16000
    - name: id
      dtype: string
  splits:
    - name: SR1
      num_bytes: 576956702.051
      num_examples: 3107
    - name: SR2
      num_bytes: 560538614.256
      num_examples: 2737
    - name: HR1
      num_bytes: 605630380.712
      num_examples: 2612
  download_size: 1718332618
  dataset_size: 1743125697.019
configs:
  - config_name: default
    data_files:
      - split: SR1
        path: data/SR1-*
      - split: SR2
        path: data/SR2-*
      - split: HR1
        path: data/HR1-*

Mak na Konac - A multi-reference Speech-to-text Benchmark for Croatian and Serbian

The data set consists of transcribed spontaneous speech samples from three sources (one Croatian and two Serbian) with a total duration of about 15 hours.

At this time, the multi-reference gold truth transcriptions are not publicly available. This submission only includes the audio and primary key.

Citation and more details


@inproceedings{Samardzic2024,
  author    = {Tanja Samardzic and Peter Rupnik and Mirjana Starović and Nikola Ljubešić},
  title     = {Mak na konac: A multi-reference Speech-to-text Benchmark for Croatian and Serbian},
  booktitle = {Proceedings of the Language Technologies and Digital Humanities 2024 conference (JT-DH 2024)},
  year      = {2024},
  address   = {Ljubljana, Slovenia},
  month     = sep,
  day       = {19--20},
}