metadata

license: cc-by-2.0
task_categories:
  - automatic-speech-recognition
  - text-to-speech
language:
  - cs
size_categories:
  - 100K<n<1M

ParCzech4Speech (Sentence-Segmented Variant)

Dataset Summary

ParCzech4Speech (Sentence-Segmented Variant) is a large-scale Czech speech dataset based on parliamentary recordings and official transcripts. This sentence-segmented variant is designed for speech recognition and synthesis tasks, offering clean audio-text alignment and reliable segment boundaries.

It is derived from the ParCzech 4.0 corpus and AudioPSP 24.01 audio collection. Using WhisperX and Wav2Vec 2.0 for automatic alignment, this dataset ensures high-quality segments and provides rich metadata for filtering and quality control.

The dataset is released under a permissive CC-BY, allowing unrestricted commercial and academic use.

🔔 Note

📢 A larger unsegmented variant of this dataset is now available! The unsegmented version provides longer, continuous speech segments that do not follow sentence boundaries, making it especially suitable for streaming ASR. You can find it under ParCzech4Speech (Unsegmented Variant) on Hugging Face.

Data Splits

Split	Segments	Hours	Speakers
Train	682,254	1131	525
Dev	5,094	10.14	29
Test	11,379	20.63	30

Dataset Structure

Each row corresponds to a sentence-level audio segment with accompanying metadata:

Column	Description
`true_text`	Official transcript from parliamentary stenographic records (unnormalized).
`rec_text`	Automatically recognized transcript using Whisper model.
`speaker`	Speaker identifier in the format `NameSurname.YearOfBirth`.
`dur`	Duration of the segment in seconds.
`vert`	Vertical file name from ParCzech 4.0 for backward compatibility.
`n_numbers`	Number of number tokens detected in `true_text`.
`n_true_words`	Number of true words in the segment.
`seg_edit_dist`	Levenshtein distance between `true_text` and `rec_text`.
`align_edit_dist_max`	Maximum word-level edit distance between aligned word pairs.
`true_char_avg_dur`	Average duration per character in `true_text` (ignoring whitespace).
`start_token_id`	Start token index (from vertical data) indicating the original source of the segment.
`end_token_id`	End token index (from vertical data).
`wav2vec_rec`	Transcript from Wav2Vec 2.0 model with greedy decoding strategy used as a secondary ASR reference.
`wav2vec_rec_edit_dist`	Normalized edit distance between `wav2vec_rec` and `rec_text`.
`speaker_text_cnt`	Frequency count of the given speaker-text pair, can be used for deduplication.

Citation

Please cite the dataset as follows:

TODO