DevaCAPTCHA-Audio: A Multimodal Devanagari CAPTCHA Dataset with Synthetic, Noisy, and Natural Speech

This Dataset presents DevaCAPTCHA, a comprehensive multimodal Devanagari CAPTCHA dataset developed to support research in secure authentication, pattern recognition, and multimodal machine learning. The dataset comprises 10 Devanagari numerals (0–9) and 23 selected characters, organized into three CAPTCHA classes: (i) character-only, (ii) digit-only, and (iii) alphanumeric combinations of characters and digits, with variable sequence lengths of 5, 6, and 7 symbols. Each CAPTCHA instance is represented through paired digitally rendered handwritten-style text images (JPEG) and corresponding audio files (MP3), enabling direct mapping between visual and acoustic modalities.
The dataset incorporates three complementary audio conditions: (i) synthetic clean speech generated using Google Text-to-Speech (gTTS), (ii) noise-augmented speech created by adding background noise to gTTS samples, and (iii) natural human speech recordings. The natural audio component is derived from an independently developed Devanagari speech dataset, which is publicly available on IEEE DataPort and Mendeley Data repositories, and further adapted to construct realistic audio CAPTCHA sequences. This integration introduces significant variability in pronunciation, speaker characteristics, and environmental conditions, enhancing real-world applicability.
The combination of digitally handwritten-style CAPTCHA images, multimodal audio variations, and variable-length sequences significantly increases the complexity of automated recognition, making the dataset a challenging benchmark for machine learning-based CAPTCHA-solving systems and bots. A structured metadata file in CSV format is provided for efficient dataset utilization, and a Python-based preprocessing pipeline ensures consistency and scalability.