README.md

metadata

dataset_info:
  features:
    - name: session_num
      dtype: int64
    - name: snapshot
      dtype: date32
    - name: session_name
      dtype: string
    - name: segment_num
      dtype: int64
    - name: speaker
      dtype: string
    - name: first_name
      dtype: string
    - name: last_name
      dtype: string
    - name: role
      dtype: string
    - name: transcript
      dtype: string
    - name: transcriber_notes
      sequence: string
    - name: id
      dtype: int64
  splits:
    - name: train
      num_bytes: 516425482
      num_examples: 330183
  download_size: 287318835
  dataset_size: 516425482
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - text-generation
  - text-classification
language:
  - sk
tags:
  - legal
  - politic
  - transcripts
  - parliament
  - slovak
pretty_name: Slovak Plenary Transcript Corpus (2001–2024)
size_categories:
  - 100K<n<1M
license: cc-by-4.0

Slovak Plenary Transcript Corpus (2001–2024)

This dataset contains segmentated full-text transcripts of the Slovak National Council (Národná rada Slovenskej republiky) plenary sessions, covering a 23-year period from 2001 to 2024. The transcripts were collected by scraping official .docx documents from the Spoločná česko-slovenská digitálna parlamentná knižnica and subsequently parsing it.

📌 Dataset Summary

• Total of 66,412,216 words
• Total of 220,721,383 tokens (estimated using GPT-2 tokenizer)
• Structured into speech segments with speaker names, roles
• Includes contextual notes from transcribers (e.g., background noise, references)
• Suitable for speech modeling, political discourse analysis, and Slovak NLP

📁 Data Structure

Each dataset sample is a dictionary with the following fields:

{
    "id": 2183662,
    "session_num": 31,
    "snapshot": "2018-05-16",
    "session_name": "Autorizovaná rozprava",
    "segment_num": 56,
    "speaker": "Bugár, Béla, podpredseda NR SR",
    "first_name": "Béla",
    "last_name": "Bugár",
    "role": "podpredseda NR SR",
    "transcript": "Vážené panie poslankyne, páni poslanci, pristúpime k hlasovaniam. Prosím pána poslanca Laurenčíka, aby uviedol hlasovanie o návrhu uznesenia. (Ruch v sále.) Vážené panie poslankyne, páni poslanci, prosím o väčší kľud. Budeme hlasovať o návrhu na vyslanie príslušníkov ozbrojených síl Slovenskej republiky v rámci Posilnenej predsunutej prítomnosti NATO do Lotyšskej republiky, Litovskej republiky, Estónskej republiky a Poľskej republiky (tlač 977).",
    "transcriber_notes": ["(Ruch v sále.)", "(tlač 977)"]
}

🔍 Field Descriptions

• id: Unique identifier of the row
• session_num: Number of the session in the current election term
• snapshot: Date on which the session took place
• session_name: Name under which the session was listed
• segment_num: Index of the speech segment within the session
• speaker: Raw string capturing the full speaker line
• first_name: Speaker’s parsed first name. This may sometimes be an initial (e.g., B in the case of B. Bugár).
• last_name: Speaker’s parsed last name
• role: Speaker’s official role at the time of the session
• transcript: Full text of the speaker’s utterance
• transcriber_notes: A list of in-transcript contextual notes (e.g., background noise, document references)

💡 Note: The fields first_name, last_name, and role were derived from the original speaker field using regex or manual inspection. Similarly, the transcriber_notes field was extracted using regex matching.

A session is uniquely identified by the combination of session_num and session_name. In rare cases, these fields may be null. Such edge cases typically did not follow typical regular patterns and could not be reliably identified. Each segment_num represents a sequential speech segment within the session.

📝 Citation & Paper

For more details, please see our paper on arXiv. If you use this dataset in your work, please cite it as:

@misc{božík2025slopalspeech2800hourslovakspeech,
      title={SloPalSpeech: A 2,800-Hour Slovak Speech Corpus from Parliamentary Data}, 
      author={Erik Božík and Marek Šuppa},
      year={2025},
      eprint={2509.19270},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.19270}, 
}