ufal/parczech4speech-unsegmented

ParCzech4Speech (Unsegmented Variant)

Dataset Summary

ParCzech4Speech (Unsegmented Variant) is a large-scale Czech speech dataset derived from parliamentary recordings and official transcripts. This variant captures continuous speech segments without enforcing sentence boundaries, making it well-suited for real-world streaming ASR scenarios and speech modeling tasks that benefit from natural discourse flow.

The dataset is created using a combination of WhisperX and Wav2Vec 2.0 models for robust automatic alignment and filtering. Segments are formed by aggregating consecutive well-aligned words until encountering a speaker change or misalignment.

The dataset is derived from the ParCzech 4.0 corpus (official transcripts of parliamentary sessions) and the corresponding AudioPSP 24.01 audios.

This dataset includes rich metadata and is released under the permissive CC-BY, allowing for commercial and academic use.

🔔 Note

📢 A sentence-segmented variant is also available ParCzech4Speech (Sentence-Segmented Variant), optimized for tasks requiring clean sentence boundaries and stricter control over segment quality.

Data Splits

Split	Segments	Hours	Speakers
Train	1,311,027	2631	527
Dev	20,352	43.43	30
Test	9,127	21.37	30

Dataset Structure

Each row represents a continuous speech segment with rich metadata:

Column	Description
true_text	Unnormalized official transcript.
rec_text	Whisper-based ASR output.
speaker	Speaker ID in the format Name.DateOfBirth.
dur	Duration of the segment in seconds.
vert	Name of the vertical file from ParCzech 4.0.
n_numbers	Number of numeric tokens in true_text.
n_true_words	Number of words in the segment.
seg_edit_dist	Levenshtein distance between true_text and rec_text.
align_edit_dist_max	Maximum edit distance among aligned word pairs.
true_char_avg_dur	Average per-character duration, computed at the word level.
start_token_id	Starting token ID from the vertical format.
end_token_id	Ending token ID.
wav2vec_rec	Wav2Vec 2.0 decoded transcript used for verification.
wav2vec_rec_edit_dist	Normalized edit distance between wav2vec_rec and rec_text.

Available Versions

• v1.0-full – Full dataset release with all intended data included. (Latest)

• v0.1-partial – Initial partial upload containing a subset of the training data. Kept for backward compatibility only. Users should prefer the full version.

By default, load_dataset() will download the latest version.
To explicitly load a specific version, you can specify the revision parameter:

from datasets import load_dataset

# Load a specific version of the dataset
dataset = load_dataset("ufal/parczech4speech-unsegmented", revision="v0.1-partial")  # or "v1.0-full"

Citation

Please cite the dataset as follows:

TODO