README.md

---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: text
    dtype: string
  - name: text_normalised
    dtype: string
  - name: words
    list:
    - name: char_e
      dtype: int64
    - name: char_s
      dtype: int64
    - name: time_e
      dtype: float64
    - name: time_s
      dtype: float64
  - name: audio_length
    dtype: float64
  - name: date
    dtype: string
  - name: speaker_name
    dtype: string
  - name: speaker_gender
    dtype: string
  - name: speaker_birth
    dtype: string
  - name: speaker_party
    dtype: string
  - name: party_orientation
    dtype: string
  - name: party_status
    dtype: string
  splits:
  - name: train
    num_bytes: 162874686121.866
    num_examples: 867581
  download_size: 179092718936
  dataset_size: 162874686121.866
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# The Croatian Parliamentary Spoken Dataset ParlaSpeech-HR 2.0

The master dataset can be found at http://hdl.handle.net/11356/1914.

The ParlaSpeech-HR dataset is built from the transcripts of parliamentary proceedings available in the Croatian part of the ParlaMint corpus (http://hdl.handle.net/11356/1859), and the parliamentary recordings available from the Croatian Parliament's YouTube channel (https://www.youtube.com/c/InternetTVHrvatskogasabora).

The dataset consists of audio segments that correspond to specific sentences in the transcripts. The transcript contains word-level alignments to the recordings, each instance consisting of character and millisecond start and end offsets, allowing for simple further segmentation of long sentences into shorter segments for ASR and other memory-sensitive applications. Sequences longer than 30 seconds have already been removed from this dataset, which should allow for a simple usage on most modern GPUs.

Each segment has an identifier reference to the ParlaMint 4.0 corpus (http://hdl.handle.net/11356/1859) via the utterance ID and character offsets.

While in the original dataset all the speaker information from the ParlaMint corpus is available via the `speaker_info` attribute, in the HuggingFace version only a subset of metadata is available, namely: the date, the name of the speaker, their gender, year of birth, party affiliation at that point in time, status of the party at that point in time (coalition or opposition), and party orientation (left, right, centre etc.).

Different to the original dataset, this version has also a `text_normalised` attribute, which contains the text with parliamentary comments (`[[Applause]]` and similar) removed.

If you use the dataset, please cite the following paper:

```
@inproceedings{ljubesic-etal-2022-parlaspeech,
    title = "{P}arla{S}peech-{HR} - a Freely Available {ASR} Dataset for {C}roatian Bootstrapped from the {P}arla{M}int Corpus",
    author = "Ljube{\v{s}}i{\'c}, Nikola  and
      Kor{\v{z}}inek, Danijel  and
      Rupnik, Peter  and
      Jazbec, Ivo-Pavao",
    editor = "Fi{\v{s}}er, Darja  and
      Eskevich, Maria  and
      Lenardi{\v{c}}, Jakob  and
      de Jong, Franciska",
    booktitle = "Proceedings of the Workshop ParlaCLARIN III within the 13th Language Resources and Evaluation Conference",
    month = jun,
    year = "2022",
    address = "Marseille, France",
    publisher = "European Language Resources Association",
    url = "https://aclanthology.org/2022.parlaclarin-1.16",
    pages = "111--116",
}
```