Datasets:

erikbozik
/

SloPalSpeech

dataset_info:
  features:
    - name: id
      dtype: int64
    - name: start_time
      dtype: float32
    - name: end_time
      dtype: float32
    - name: session_num
      dtype: int64
    - name: snapshot
      dtype: date32
    - name: segment_number
      dtype: int64
    - name: text
      dtype: string
    - name: duration
      dtype: float32
    - name: audio
      dtype: audio
  splits:
    - name: train
      num_examples: 402966
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - automatic-speech-recognition
language:
  - sk
tags:
  - speech
  - asr
  - slovak
  - parliament
  - legal
  - politics
  - whisper
pretty_name: SloPalSpeech
size_categories:
  - 1M<n<10M
license: cc-by-4.0

SloPalSpeech

This dataset contains aligned and segmented Slovak speech–text pairs sourced from official plenary session recordings of the Slovak National Council (Národná rada Slovenskej republiky).
It was prepared by collecting raw audio from MediaPortál NR SR and matching it with official full-text transcripts from the Joint Czech and Slovak Digital Parliamentary Library.
A custom alignment and filtering pipeline segmented the recordings into short clips (≤30 seconds) with their corresponding transcriptions, producing a 2,806-hour ASR-ready dataset.

📌 Dataset Summary

Total duration: 2,806 hours
Number of segments: 402,966
Average segment length: ~25.1 seconds
Domain: formal Slovak parliamentary speech
Use cases: ASR model fine-tuning, speech research

📁 Data Structure

Each dataset entry has the following fields:

{
    "id": 101036,
    "start_time": 2949.985,
    "end_time": 2974.976,
    "session_num": 19,
    "snapshot": "2024-09-13",
    "segment_number": 108,
    "text": "všetkých členov výboru, ktorí neprišli na zasadnutie výboru, lebo sa rozhodli, že ho budú ignorovať, ako to urobili viackrát poslanci koalície, keď chceli znefunkčniť výbor? Potom asi tých by sme mali odvolať pri takomto postupe, lebo ja som riadne zasadnutie urobila, no ale koaliční poslanci sa rozhodli na niektorých nebyť, lebo sa im nechcelo. A potom tie ostatné dva body, no to je smiešne! Ďalší, tretí bod, že verejne spochybňujem a tým sa propagujem. No, prosím",
    "duration": 24.99,
    "audio": {
        "path": "19_2024-09-13_108.mp3",
        "array": [
            1.32518352e-07, -2.97130391e-08, -1.27864055e-07, ...,
            -1.25169004e-06, 1.16862184e-05, 4.28082467e-06
        ],
        "sampling_rate": 16000
    }
}

🔍 Field Descriptions

id: Unique integer identifier for the segment.
start_time: Start timestamp (in seconds) within the original session recording.
end_time: End timestamp (in seconds) within the original session recording.
session_num: Parliamentary session number in the election term.
snapshot: Date of the session.
segment_number: Sequential index of the segment in the session.
text: Slovak transcript (transcriber notes removed).
duration: Segment duration in seconds.
audio: Audio object containing local file path, waveform array, and sampling rate.

⚙️ Processing & Alignment

Collection
- Audio scraped to .mp3 format.
- Transcripts downloaded in .docx format.
Parsing
- Extracted speaker-segmented utterances.
- Removed transcriber notes.
Alignment
- Matched words from the ground-truth transcript with a pseudo-labeled transcript generated from the audio.
- Assigned timestamps to the matched words using the pseudo-labeled transcript, which had been force-aligned.
Segmentation
- Created segments ≤30 seconds based on the matched words.
Filtering
- Re-transcribed each segment.
- Discarded any with WER > 40%.
- Final dataset: 2,806 hours; 402,966 segments.

📊 Statistics

Metric	Value
Total hours	2,806
Total segments	402,966
Avg. duration	25.1 sec
Sample rate	16 kHz

🚀 Example usage

from datasets import load_dataset

ds = load_dataset("erikbozik/SloPalSpeech", split="train")

sample = ds[0]
print(sample["text"])
print(sample["duration"], "seconds")
sample["audio"]["array"]  # numpy waveform

📝 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}, 
}