Indian Sign Language Alphabet Recognition and Speech Synthesis Using a Hybrid Deep Learning Approach

Indian Sign Language (ISL) is vital for communication among India's hearing-impaired community. However, the lack of standardised datasets and reliable identification frameworks has hampered the use of ISL in modern assistive technology. This paper presents a deep learning-based solution to robust ISL alphabet identification, with an emphasis on both accuracy and practical use. A curated static ISL alphabet collection was created by combining authoritative visual references from the official Indian Sign Language website and the Ramakrishna Mission Vivekananda Educational and Research Institute (RKMVERI). Multiple deep learning models were trained and assessed, including CNN, ResNet-50, DenseNet-121, VGG16, MobileNetV2, and EfficientNet-B0, with a new hybrid CNN-ResNet architecture outperforming the others. 98% classification accuracy is achieved by the suggested approach, outperforming individual baseline models. Furthermore, the framework is expanded to support real-time applications, combining webcam-based capture with immediate conversion of recognized signs to textual and synthesized vocal output. Comprehensive performance evaluation, including confusion matrix analysis and ROC curves, demonstrates the solution's durability and practical applicability. This research enhances accessibility, promotes inclusive education, and prepares the path for scalable sign language translation systems in real-world human-machine interaction scenarios by enabling accurate and real-time ISL recognition with voice feedback.