BornilDB v1.0: Bangladeshi Sign Language Corpus

• BornilDB v1.0 is a large-scale, crowd-sourced BdSL video corpus with continuous signing videos and detailed, multi-level annotations for ASLR and SLT tasks.
• The dataset features a modular web-based framework for data collection, ensuring high-fidelity gloss and timeline annotations enriched by OpenPose and MediaPipe keypoint extraction.
• Benchmarking with transformer and two-stream fusion models demonstrates improved BLEU scores, underscoring the corpus’s potential for real-world sign language recognition and translation applications.

BornilDB v1.0 is the first large-scale, crowd-sourced continuous Bangladeshi Sign Language (BdSL) video corpus constructed end-to-end on the open-source Bornil platform. Developed to address the deficit of annotated sign language datasets, particularly for South Asian languages, BornilDB v1.0 enables modern deep-learning-based Automatic Sign Language Recognition (ASLR) and Sign Language Translation (SLT) through comprehensive, dialect-agnostic, and high-precision data acquisition and annotation workflows. With a focus on open-access infrastructure and methodological rigor, the resource is designed for scalability, robust benchmarking, and facilitating research in gloss-free recognition and translation.

1. Platform Infrastructure and Data Acquisition

Bornil provides a modular, web-based environment for large-scale sign language data collection, annotation, and validation (Dhruvo et al., 2023). The frontend is implemented in Next.js with server-side rendering and static HTML caching. Backend services are built in Go, leveraging goroutines for high-throughput request management. PostgreSQL functions as the primary database, while Amazon S3 stores video assets. The complete platform, including code, deploys via Docker, supporting reproducibility and scalability.

Data acquisition is orchestrated through a workflow in which administrators upload CSVs of scripted text or topic prompts. Contributors, after authentication, are randomly assigned either scripted texts or open-ended topic prompts, and proceed to record continuous signing via a browser-based camera interface. Recording occurs “in the wild”, with unrestricted settings regarding background, lighting, and device type. Contributors preview and trim their videos, supply metadata (camera angle, device, environmental lighting), and submit them for validation.

2. Annotation Pipeline and Quality Assurance

The Bornil framework implements a multi-tiered, crowdsource-driven annotation and validation regimen to ensure dataset precision (Dhruvo et al., 2023). Recordings that pass manual video validation—where validators assess semantic match and correct metadata—advance to annotation. Sentence-level annotation requires segmenting the video into timestamp-aligned utterances, while gloss-level annotation maps each individual sign (word) to start and end times within the clip.

Annotation is followed by a secondary round of timeline-based validation. Annotators and validators interactively review and adjust time-aligned subtitles, correcting segmentation or alignment errors. Videos and annotations are further enriched via extraction of body, face, and hand keypoints using OpenPose and MediaPipe pipelines, supporting downstream deep learning research. These combined controls yield high-fidelity, multi-level annotated corpora suitable for both sign recognition and translation.

3. Corpus Structure, Characteristics, and Accessibility

In a three-month initial data collection period, BornilDB v1.0 amassed 21 154 video samples recorded by three deaf native BdSL signers, including 21 104 scripted and 50 spontaneous topic-driven recordings, totaling 73 hours (Dhruvo et al., 2023, Arib et al., 14 Aug 2025). Metadata includes participant demographics (age, gender, locality), camera view, and device details. The dataset is structured as follows:

Attribute	Value	Notes
Language	Bangladeshi Sign Language	Focus of the initial release
Signers	3	Deaf, native BdSL users
Total video duration	73 hours (or 45 hours†)	†Conflicting reporting noted
Scripted video samples	21,104	Aggregated with spontaneous
Topic-driven samples	50
Total video–text pairs	21,154
Bengali words (total)	138,586	25,572 unique unimodal glosses
Avg. words per clip	7.7
Avg. clip duration	13.4 seconds
Data split	80:10:10 train:val:test	Standard protocol

Contextually, BornilDB v1.0 is the second largest continuous-sign dataset (after How2Sign) and the first of its kind for Bangladeshi Sign Language, representing a significant advance for South Asian SLT/ASLR research (Dhruvo et al., 2023, Arib et al., 14 Aug 2025). Licensing is CC BY-SA 4.0, with all raw videos, extracted keypoints, metadata, and annotations available at https://bornil.bengali.ai.

4. Benchmarking Protocols and Baseline Results

BornilDB v1.0 is benchmarked under modern ASLR and SLT frameworks. The baseline protocol in (Dhruvo et al., 2023) applies a transformer model (pre-trained on How2Sign) fine-tuned to the BdSL corpus, with BLEU-n (n=1 to 4) and sign/frame-level accuracy as metrics. Reported BLEU-4 scores on held-out test sets, stratified by sentence length, are:

• Length 12 ± 2: BLEU4 = 1.30
• Length 10 ± 2: BLEU4 = 0.79
• Length 8 ± 2: BLEU4 = 0.59

These results demonstrate the feasibility of high-precision recognition and translation from heterogeneous, in-the-wild crowd-sourced video, setting a consistent benchmark for South Asian SLT corpora.

Subsequent benchmarking in (Arib et al., 14 Aug 2025) introduces two-stream (RGB and keypoint) fusion models that integrate RGB–I3D–Transformer and Keypoint–STGCN–LSTM pathways. Notable design decisions include: three STGCN hops for keypoint graphs, direct summation fusion of feature tensors, and 6 encoder/3 decoder transformer layers. Evaluation uses standard BLEU-n with brevity penalty:

$$BLEU_n = BP \cdot \exp\left(\frac{1}{n}\sum_{i=1}^n \log p_i\right),$$

where $BP$ is the brevity penalty and $p_i$ is $i$-gram precision.

A reduced-BLEU (rBLEU) variant filters frequent function words before scoring. The model achieves:

Split	BLEU-4 (baseline)	BLEU-4 (proposed)	rBLEU (proposed)
Validation	0.27	0.72	7.62
Test	0.27	0.58	7.37

These gains are attributed to the two-stream fusion, label smoothing (ε = 0.1), and architectural ablations. Performance is robust despite BornilDB’s heterogeneous device, resolution, and environmental conditions (Arib et al., 14 Aug 2025).

5. Research Implications and Applications

BornilDB v1.0 establishes a reproducible benchmark supporting diverse ASLR and SLT tasks in a dialect-agnostic framework. The release of high-quality, gloss-rich, and gloss-free annotations enables: (i) training and validation of deep learning models for both recognition and translation, (ii) ablation studies on annotation granularity, (iii) stratification studies using demographic and device metadata, and (iv) transfer learning and domain adaptation with datasets such as How2Sign and RWTH-PHOENIX-2014T (Dhruvo et al., 2023, Arib et al., 14 Aug 2025).

The use of in-the-wild, uncurated environments makes BornilDB v1.0 particularly representative of real-world deployment scenarios, supporting robustness and generalizability evaluation.

6. Limitations and Future Directions

BornilDB v1.0’s initial release covers only the Bangladeshi Sign Language dialect, with recordings limited to three signers and single-camera, single-viewpoint data. While extensive demographic metadata is recorded, multiview and multisigner diversity is currently limited. The absence of audio and some pose metadata in benchmarks may constrain multimodal research (Arib et al., 14 Aug 2025). Future extensions may address increased signer diversity, dialectal expansion, and multiview capture.

A further plausible implication is that subsequent releases may support benchmarking on downstream tasks such as gloss-free translation, pose-based recognition, and cross-dialect adaptation, driven by BornilDB’s open-access design and methodological transparency (Dhruvo et al., 2023, Arib et al., 14 Aug 2025).