Over a period of three years (2009 - 2011) the daily news and weather forecast airings of the German public tv-station PHOENIX featuring sign language interpretation have been recorded and the weather forecasts of a subset of 386 editions have been transcribed using gloss notation. Furthermore, we used automatic speech recognition with manual cleaning to transcribe the original German speech. As such, this corpus allows to train end-to-end sign language translation systems from sign language video input to spoken language. The signing is recorded by a stationary color camera placed in front of the sign language interpreters. Interpreters wear dark clothes in front of an artificial grey background with color transition. All recorded videos are at 25 frames per second and the size of the frames is 210 by 260 pixels. Each frame shows the interpreter box only.

WLASL is a large video dataset for Word-Level American Sign Language (ASL) recognition, which features 2,000 common different words in ASL. Source: https://github.com/dxli94/WLASL

CSL-Daily (Chinese Sign Language Corpus) is a large-scale continuous SLT dataset. It provides both spoken language translations and gloss-level annotations. The topic revolves around people's daily lives (e.g., travel, shopping, medical care), the most likely SLT application scenario. [1] Improving Sign Language Translation with Monolingual Data by Sign Back-Translation, CVPR, 2021.

The How2Sign is a multimodal and multiview continuous American Sign Language (ASL) dataset consisting of a parallel corpus of more than 80 hours of sign language videos and a set of corresponding modalities including speech, English transcripts, and depth. A three-hour subset was further recorded in the Panoptic studio enabling detailed 3D pose estimation.

Large-scale American Sign Language (ASL) - English dataset collected from online video sites (e.g., YouTube). OpenASL contains 288 hours of ASL videos in multiple domains from over 200 signers.

MSASL is a real-life large-scale sign language data set comprising over 25,000 annotated videos. Source: MS-ASL: A Large-Scale Data Set and Benchmark for Understanding American Sign Language

BosphorusSign22k is a benchmark dataset for vision-based user-independent isolated Sign Language Recognition (SLR). The dataset is based on the BosphorusSign (Camgoz et al., 2016c) corpus which was collected with the purpose of helping both linguistic and computer science communities. It contains isolated videos of Turkish Sign Language glosses from three different domains: Health, finance and commonly used everyday signs. Videos in this dataset were performed by six native signers, which makes this dataset valuable for user independent sign language studies. Source: BosphorusSign22k Sign Language Recognition Dataset

Content4All is a collection of six open research datasets aimed at automatic sign language translation research. Sign language interpretation footage was captured by the broadcast partners SWISSTXT and VRT. Raw footage was anonymized and processed to extract 2D and 3D human body pose information. From the roughly 190 hours of processed data, three base (RAW) datasets were released, namely 1) SWISSTXT-RAW-NEWS, 2) SWISSTXT-RAW-WEATHER and 3) VRT-RAW. Each dataset contains sign language interpretations, corresponding spoken language subtitles, and extracted 2D/3D human pose information. A subset from each base dataset was selected and manually annotated to align spoken language subtitles and sign language interpretations. The subset selection was done to resemble the benchmark Phoenix 2014T dataset. Our aim is for these three new annotated public datasets, namely 4) SWISSTXT-NEWS, 5) SWISSTXT-WEATHER and 6) VRT-NEWS to become benchmarks and underpin future research as the field moves closer to translation and production on larger domains of discourse.

Sign languages are the primary means of communication for a large number of people worldwide. Recently, the availability of Sign language translation datasets have facilitated the incorporation of Sign language research in the NLP community. Though a wide variety of research focuses on improving translation systems for sign language, the lack of ample annotated resources hinders research in the data driven natural language processing community. In this resource paper, we introduce ISLTranslate, a translation dataset for continuous Indian Sign Language (ISL), consisting of 30k ISL-English sentence pairs. To the best of our knowledge, it is the first and largest translation dataset for continuous Indian Sign Language with corresponding English transcripts. We provide a detailed analysis of the dataset and examine the distribution of words and phrases covered in the proposed dataset. To validate the performance of existing end-to-end Sign language to spoken language translation systems, we benchmark the created dataset with multiple existing state-of-the-art systems for sign languages.

An artificial corpus built using grammatical dependencies rules due to the lack of resources for Sign Language. Source: ASLG-PC12

The Azerbaijani Sign Language Dataset (AzSLD) is a comprehensive, large dataset designed to facilitate the development and evaluation of machine learning models for the recognition and translation of Azerbaijani Sign Language (AzSL). AzSLD is the first publicly available dataset focused on Azerbaijani Sign Language. It contributes to the global effort to improve accessibility for the deaf and hard-of-hearing community in Azerbaijan. The dataset aims to bridge the gap between technology and accessibility by providing high-quality data for researchers, developers, and practitioners working on sign language recognition or translation systems. The data collection costs are covered by the "Strengthening Data Analytics Research and Training Capacity through Establishment of dual Master of Science in Computer Science and Master of Science in Data Analytics (MSCS/DA) degree program at ADA University" project, funded by BP and the Ministry of Education of the Republic of Azerbaijan. Dataset Composition AzSLD is organized into three primary components: 1. AzSLD_Sentences This component contains video sequences of complete sentences in AzSL. It is designed to capture the fluidity and contextual nature of sign language, providing data for more complex language modeling tasks. It includes over 60 hours of high-definition video recordings, annotated with timestamped glosses for 500 distinct classes, enabling precise analysis and robust model training. Ground truth annotations of sentences for each class were added in a separate file. The videos were performed by 18 to 25 different signers, with a slight imbalance among them. 2. AzSLD_Words This component comprises a collection of short video samples representing frequently used words in Azerbaijani Sign Language. It is divided into two subsets: AzSLD_Words_100: Contains 100 commonly used words in AzSL. AzSLD_Words_200: Extends the first subset, including all 100 words from AzSLD_Words_100 along with an additional 100 words, for a total of 200 words. Folder names indicate the ground truth labels for the ease of word-level model evaluation. 3. AzSLD_Fingerspelling This component includes over 14,000 video and image samples of letters of the Azerbaijani alphabet. Each sign is captured from multiple angles to ensure comprehensive coverage of dactylology in AzSL. This component is ideal for tasks involving letter recognition and the integration of fingerspelling into broader sign language recognition systems. Key Features Double-View Recordings The dataset includes 10,104 synchronized video recordings from two camera angles to capture both frontal and side views of hand and body movements, ensuring that the subtle nuances of sign language are well-represented. Diverse Signers The dataset features recordings from a diverse group of native AzSL signers, encompassing variations in age, gender, and signing style. This diversity is crucial for training models that are robust to variations in signing. Detailed Annotations Each video is annotated with comprehensive metadata, including the sign’s label (dactyl, word, or sentence), signer ID, and timestamped glosses for sentence-level signs. High-Quality Data Format The dataset comprises RGB videos in high-definition (HD) resolution at 35 frames per second, accompanied by JSON files containing annotations and metadata. The data is systematically organized into folders by category for ease of navigation. Ethical Transparency All participants provided informed consent for collecting, publishing, and using the data, ensuring compliance with ethical research standards. Accessibility The AzSLD is available under Creative Commons Attribution 4.0 International with free access for academic research through Zenodo. Citation: When using AzSLD in your research, please cite the following paper: Alishzade, N., Hasanov, J. (2024). AzSLD: Azerbaijani Sign Language Dataset for Dactyl, Word, and Sentence Translation with Baseline Software. [Journal Name], [Volume(Issue)], [Pages]. DOI: [DOI link]. The preprint is available at: https://arxiv.org/abs/2411.12865 Contact: For questions, feedback, or contributions, please contact the project team at: slr.project.ada@gmail.com

A large-scale gloss-free sign language translation dataset with 1,985 hours of videos, approximately 86 times larger than the previous CSL-Daily dataset.

DailyMoth-70h is a fully self-contained ASL-to-English sign language dataset containing over 70h of video (48K clips) with aligned English captions of a single native ASL signer (white, male, and early middle-aged) from the ASL news channel TheDailyMoth. The primary purpose of the dataset is to be used as a benchmark and analysis dataset for (gloss-free) sign language translation. The dataset comes with four parts: - raw_videos: Contains the unsegmented DailyMoth videos (496 in total) with blurring applied to the burnt-in captions and advertisement breaks and news headline banners - blurred_clips: Contains the segmented video clips (48386 in total) with facial blurring applied. Each clip comes in its native frame rate (either 24, 29.97 or 30 fps) and as 224x224px region-of-interest (ROI) crops around the signer - unblurred_clips: Contains the unblurred segmented video clips (48386 in total). Each clip comes in its native frame rate (either 24, 29.97 or 30 fps) and as 224x224px region-of-interest (ROI) crops around the signer - manifests: Contains the manifest TSV files for training, validation, and testing. Also contains a combined manifest file and a tsv file with the start and end timestamps used to segment the raw video Detailed dataset statistics are listed in https://arxiv.org/abs/2402.09611. The dataset is available for download at https://github.com/facebookresearch/ssvp_slt?tab=readme-ov-file#dailymoth-70h.

LSA-T is the first continuous Argentinian Sign Language (LSA) dataset. It contains 14,880 sentence level videos of LSA extracted from the CN Sordos YouTube channel with labels and keypoints annotations for each signer. Videos are in 30 FPS full HD (1920x1080). * Download link (45GB compressed) * Visualization notebook * Presentation paper (preprint PDF)

Sign Language Datasets for French Belgian Sign Language This dataset is built upon the work of Belgian linguists from the University of Namur. During eight years, they've collected and annotated 50 hours of videos depicting sign language conversation. 100 signers were recorded, making it one of the most representative sign language corpus. The annotation has been sanitized and enriched with metadata to construct two, easy to use, datasets for sign language recognition. One for continuous sign language recognition and the other for isolated sign recognition. ## LSFB-CONT The dataset for continuous sign language recognition is made of over 25h of video clips. Each clip is associated with a time-aligned annotation file containing the start and the end of each sign along with a gloss (label) associated with all unique signs. Mediapipe pose and hands information were also computed for each video clip and these metadata are made available in the dataset. ## LSFB-ISOL The isolated version of the dataset contains only clips showing one isolated sign issued from the LSFB-CONT dataset. We chose to keep all the signs with at least 40 examples, leading to a dataset containing over 50 000 clips for 635 different glosses (labels). The Mediapipe metadata is also available for this dataset.

Mediapi-RGB is a bilingual corpus of French Sign Language (LSF) and written French in the form of subtitled videos, accompanied by complementary data (various representations, segmentation, vocabulary, etc.). It can be used in academic research for a wide range of tasks, such as training or evaluating sign language (SL) extraction, recognition or translation models. To build this corpus, we used videos from Média'Pi!, a bilingual online media with journalistic-type content in LSF with French subtitles. We collected 1230 videos dating from September 2017 to January 2022, representing a total of 86h. Based on the subtitles, we temporally segmented the videos into 50084 video segments (or extracts). We also automatically cropped the signer and harmonised the segments in terms of size (444x444) and frequency (25fps).

We provide separate training, development and test data. The training data is available right away. The development and test data will be released in several stages, starting with a release of the development sources only. The training data comprises two corpora, called FocusNews and SRF, see below for a more detailed description. The linguistic domain of both corpora is general news, and both contain parallel data between Swiss German Sign Language (DSGS) and German. The corpora are distributed through Zenodo.