Datasets:

LangTech-MT
/

LSC-Parlament

language:
  - es
  - ca
multilinguality:
  - multilingual
pretty_name: LSC-Parlament
task_categories:
  - translation
license: cc-by-4.0
tags:
  - sign_language_translation
  - sign-language
size_categories:
  - 10K<n<100K

Dataset Card for LSC-Parlament

Dataset Description

Point of Contact: langtech@bsc.es, carlos.escolano@upc.edu

Dataset Summary

A multimodal dataset derived from videos of Catalan Parliament sessions. The dataset combines Catalan Sign Language (LSC) video recordings with synchronized audio and text transcriptions in Catalan or Spanish.

Supported Tasks and Leaderboards

This dataset enables research on sign language recognition, translation, and cross-modal alignment between visual signing, speech, and text.

Languages

The dataset contains two parallel data pairs: LSC to Catalan and LSC to Spanish.

Both pairs include video, audio and text examples.

Language pair	Codes	Split	Size (sentences)	Hours (Video)	Hours (Audio)
Catalan Sign Language - Catalan	lsc-ca	train	22717	60.67	59.35
Catalan Sign Language - Catalan	lsc-ca	validation	2000	5.36	5.22
Catalan Sign Language - Catalan	lsc-ca	test	2000	5.35	5.17
-------------------	-------	-------	-------------------	---------------	----------------
Catalan Sign Language - Spanish	lsc-es	train	3610	9.68	9.32
Catalan Sign Language - Spanish	lsc-es	validation	1203	3.31	3.21
Catalan Sign Language - Spanish	lsc-es	test	1203	3.34	3.25

Catalan Sign Language (LSC)

Catalan Sign Language (LSC) is the natural sign language used by the deaf community in Catalonia. It has its own grammar and structure, which is independent of spoken Catalan or Spanish. LSC is used by tens of thousands of people. Estimates of the number of signers range from around 9,000 to over 25,000 depending on how users are counted. LSC is an officially recognised language that plays a key role in accessibility and cultural identity in the region.

Dataset Structure

Data Instances

The dataset is provided in parquet format. Each instance contains a parallel sentence pair with the following structure:

{
  "signal": "80329797_01_4511_4624.mp4",
  "output": "Però hi han altres eines qué també estan a l'acord i que també té vostè: té la inspecció, té la sanció prevista per la llei si arriba fins a 900.000 euros, i si no hi ha prou amb la sanció, pot expropiar forçosament, expropiar-ne l'ús fins a set anys, fins i tot en aquests casos.",
  "signal_start": "89600.0",
  "singal_end": "103440.0",
  "audio_start": "84880.0",
  "audio_end": "101080.0"
}

Data Fields

signal: Name of the video file containing both sign language and audio
output: Textual translation of the segment.
signal_start: LSC segment start time in the video file (in milliseconds)
signal_end: LSC Segment end time in the video file (in milliseconds)
audio_start: Audio Segment start time in the video file (in milliseconds)
audio_end: Audio Segment end time in the video file (in milliseconds)

Data Splits

The dataset is split in two main subsets:

LSC - Catalan: Videos include LSC signing by a professional interpreter (hearing) and Catalan audio. Textual translations are in Catalan.
LSC - Spanish: Videos include LSC signing by a professional interpreter (hearing) and Spanish audio. Textual translations are in Spanish.

Train, Validation and Test splits are provided for both subsets.

Dataset Creation

Curation Rationale

Creating datasets for Sign Language is essential for advancing accessible technologies and supporting the deaf and hard of hearing communities. As a low- to mid-resource language, LSC lacks the large-scale, high-quality data needed to develop reliable systems for sign recognition, translation, and education. Building such resources not only enables technological innovation but also contributes to the preservation, standardization, and visibility of the language in digital environments.

This dataset has been automatically extracted and aligned from publicly available videos of the Catalan Parliament. Automatic annotation tools are useful for converting existing materials into training resources for our systems, making it feasible to scale up dataset creation without relying entirely on manual labeling or ad-hoc data creation.

Source Data

Initial Data Collection and Normalization

The corpus is a collection of sessions from the Catalan Parliament held between October 2021 and July 2024. All videos are publicly available on their website. To avoid altering the original meaning of the sentences or causing any misinterpretations, the videos contain the full interventions, and each segment is marked with timestamps within the intervention.

Data Filtering and Preprocessing:

The provided data underwent the following preprocessing steps:

Sign Window Filtering: An automatic script detects that a sing language interpreter window is located at the right side of the video. If the window is not detected, the video is discarded.
Sign Window Extraction: For all videos a 330x330px window is extracted. The window shows a sign interpreter wearing dark or black clothes over a white background.
Audio Segmentation: For all videos, audio is extracted using the ffmpeg library. Audio is then segmented and assigned timestamps using a Conformer based ASR model. Transcription is finally aligned with the offical session diary. All segments with a Character Error Rate (CER) higher than 0.3 are discarded.
Video Segmentation: All videos are segmented by sentences using a sign language segmenter.
Audio and Video Alignment: Video and audio segments are automatically aligned based on their timestamps with a maximum allowed delay of 6 seconds.

Who are the source language producers?

Parlament de Catalunya: this data belongs to Parlament de Catalunya and is released in accordance with their terms of use.

Annotations

Annotation process

The dataset does not contain any manual annotations beyond the parallel alignments, which were either preserved from source datasets or validated through automated alignment scoring.

Who are the annotators?

[N/A]

Personal and Sensitive Information

All data contained in this dataset comes from publicly available sources. Due to the nature of the task, the dataset includes both interpreter image and parliamentary speakers voice.

Considerations for Using the Data

Social Impact of Dataset

A Catalan Sign Language dataset can have a strong social impact by enabling the development of technologies that improve accessibility for deaf and hard of hearing communities. It supports tools such as automatic translation, captioning and sign recognition, helping to reduce communication barriers in areas such as education, the media and public services.

Discussion of Biases

No specific bias mitigation strategies were applied to this dataset. Inherent biases may exist within the datasets, reflecting the biases present in the source data.

Other Known Limitations

The dataset contains almost exclusively data from the administrative domain. Application of this dataset in a general domain or in other specialized domains would be of limited use.

Additional Information

Dataset Curators

Language Technologies Unit at the Barcelona Supercomputing Center (langtech@bsc.es).

Funding

This work has been promoted and financed by the Government of Catalonia through the Aina Project.

This work is funded by the Ministerio para la Transformación Digital y de la Función Pública - Funded by EU – NextGenerationEU within the framework of ILENIA Project with reference 2022/TL22/00215337.

This work has been funded by the Ministerio de Ciencia, Innovación y Universidades and the Agencia Estatal de Investigación through the project LLM4LS (PID2024-157855OA-C33).

Acknowledgements

We gratefully acknowledge the following entities for their valuable contribution of data to this corpus:

Parlament de Catalunya

Licensing Information

This work is licensed under a Creative Commons Attribution 4.0 International licence. While the dataset features public figures and interpreters in an official capacity, its use is restricted to research advancing assistive translation technologies; biometric surveillance or re-identification is explicitly prohibited.

Citation Information

[N/A]

Contributions

[N/A]