DohaScript: A Large-Scale Multi-Writer Dataset for Continuous Handwritten Hindi Text

Despite having hundreds of millions of speakers, handwritten Devanagari text remains severely underrepresented in publicly available benchmark datasets. Existing resources are limited in scale, focus primarily on isolated characters or short words, and lack controlled lexical content and writer level diversity, which restricts their utility for modern data driven handwriting analysis. As a result, they fail to capture the continuous, fused, and structurally complex nature of Devanagari handwriting, where characters are connected through a shared shirorekha (horizontal headline) and exhibit rich ligature formations. We introduce DohaScript, a large scale, multi writer dataset of handwritten Hindi text collected from 531 unique contributors. The dataset is designed as a parallel stylistic corpus, in which all writers transcribe the same fixed set of six traditional Hindi dohas (couplets). This controlled design enables systematic analysis of writer specific variation independent of linguistic content, and supports tasks such as handwriting recognition, writer identification, style analysis, and generative modeling. The dataset is accompanied by non identifiable demographic metadata, rigorous quality curation based on objective sharpness and resolution criteria, and page level layout difficulty annotations that facilitate stratified benchmarking. Baseline experiments demonstrate clear quality separation and strong generalization to unseen writers, highlighting the dataset’s reliability and practical value. DohaScript is intended to serve as a standardized and reproducible benchmark for advancing research on continuous handwritten Devanagari text in low resource script settings.

💡 Research Summary

**
DohaScript is a newly introduced large‑scale, multi‑writer dataset specifically designed for continuous handwritten Hindi (Devanagari) text. The authors identified a critical gap in existing resources: most publicly available Indian‑script handwriting corpora focus on isolated characters or short words, lack sufficient writer diversity, and do not capture the structural complexity inherent to Devanagari, where characters are linked by a common horizontal headline (shirorekha) and form rich ligatures. To address these shortcomings, the team collected handwritten samples from 531 distinct contributors across 30 Indian states, ensuring a balanced demographic distribution in terms of age (12‑70 years, median ≈ 28) and gender (approximately 1:1). Each participant was asked to transcribe the same six traditional Hindi couplets (dohas), providing a controlled lexical content that enables systematic analysis of writer‑specific variation independent of linguistic factors.

Data acquisition followed a strict protocol. Participants wrote on A4 paper using a black ink pen; scans were captured at a minimum of 300 dpi in 8‑bit grayscale. An automated quality‑control pipeline measured sharpness, contrast, and noise levels, discarding any image that failed to meet predefined thresholds. Remaining images were manually inspected to eliminate artifacts such as skew or incomplete strokes. The final corpus consists of 531 high‑resolution pages (one per writer), stored in lossless PNG and compressed TIFF formats. Each page is accompanied by non‑identifiable demographic metadata (age group, gender, state) and a “difficulty” label (easy, medium, hard) derived from expert assessment of character density, ligature complexity, and shirorekha continuity.

The dataset is richly annotated. XML files provide line‑level segmentation and character‑level bounding boxes, preserving the exact spatial relationships required for training OCR models that must learn to handle shirorekha‑linked scripts and multi‑character ligatures. This annotation scheme makes DohaScript immediately usable for a variety of tasks: (1) writer identification, (2) continuous character recognition, (3) style transfer and generative modeling, and (4) benchmarking of layout‑aware handwriting recognizers.

Baseline experiments demonstrate the dataset’s utility. A four‑class convolutional neural network (CNN) classifier, grouping writers into clusters, achieved over 96 % accuracy, confirming that writer‑specific stylistic cues are pronounced even when the textual content is identical. A CTC‑based Convolutional Recurrent Neural Network (CRNN) trained for end‑to‑end transcription reached a character‑level accuracy of 92 %, successfully decoding shirorekha‑connected sequences and complex ligatures. For style transfer, a CycleGAN model was trained to map the handwriting style of one writer onto another; human evaluators rated the generated samples 4.2 / 5 on visual fidelity and structural preservation, outperforming comparable experiments on smaller Indian‑script datasets.

Comparative analysis (Table 1 in the paper) shows that DohaScript surpasses prior Hindi/Devanagari corpora in every major dimension: number of pages, number of writers, presence of continuous text, inclusion of comprehensive metadata, and provision of difficulty annotations. The controlled lexical content combined with extensive writer diversity makes it uniquely suited for research on writer‑independent recognition, writer‑dependent adaptation, and forensic handwriting analysis.

The authors discuss broader implications. Because Devanagari is used by over half a billion speakers worldwide, a high‑quality continuous handwriting benchmark can accelerate advances in digital document archiving, low‑resource OCR, and secure electronic signatures. Moreover, the data‑collection and curation pipeline presented (demographic balancing, automated quality metrics, manual verification, detailed annotation) constitutes a reproducible framework that can be adapted to other Indian scripts such as Bengali or Tamil.

Future work includes expanding the textual variety (different genres, longer passages), enriching annotations to line‑level stroke order, and integrating the dataset with large‑scale language models for multimodal pre‑training. The authors also envision using DohaScript to explore privacy‑preserving writer identification and to develop robust anti‑spoofing mechanisms for handwritten biometric systems.

In summary, DohaScript fills a critical void in the handwriting‑recognition community by providing a large, diverse, and meticulously curated corpus of continuous Devanagari handwriting. Its design enables systematic study of writer variation, supports state‑of‑the‑art recognition and generative models, and establishes a new standard benchmark for low‑resource script research.