Señatalk: Inclusive Web Application for the Recognition and Translation of Peruvian Sign Language (PSL) Using Recurrent Neural Networks (LSTM)

Peruvian Sign Language (PSL) recognition is approached as a dynamic sequence classification problem. SeñaTalk, an inclusive web application for real-time LSP recognition and translation, is presented using a model based on LSTM networks and spatial vectors extracted using MediaPipe. The system captures key points of the face, hands and body, generating sequences that are classified to predict signs associated with the alphabet and basic words. A dataset of 15 dynamic gestures was used and a comparative benchmarking was performed between CNN, Vision Transformer and LSTM models, evaluating accuracy, response time, robustness and generalization capacity. The results show that the LSTM model offers the best balance between 98.0% accuracy and temporal efficiency, being the most suitable for implementation in accessible, inclusive and real-time applications.