In this paper, an innovative approach utilizing artificial intelligence (AI) for the recognition of German Sign Language (GSL) gestures is presented, aimed at controlling an assistance robot. Sign language is an important method of communication for the hearing impaired and poses special challenges for automated recognition due to its complex and nuanced gestures. Leveraging advancements in deep learning techniques, particularly Long Short-Term Memory (LSTM) and The MediaPipe Holistic Landmarker for extracting hand, face, and pose landmarks, a robust GSL recognition system is proposed. The proposed model is trained to interpret a specified set of GSL gestures, focusing on common tasks or objects that an assistance robot can do or grab, respectively. Our findings demonstrate promising results, with the LSTM network achieving a validation accuracy of 96.55 % with minimal false positive classifications. This research contributes to the advancement of assistive technologies by harnessing the power of AI to pave the way for seamless integration of GSL into robotic control systems, empowering individuals with hearing impairments to interact intuitively with robotic platforms.

Speech recognition using artificial intelligence (AI) is widely used in everyday life, whether it's voice assistants for smartphones or smart speakers. Our voice command recognition systems, which are based on Deep Learning, are designed to recognize 10 specific German commands, which contain a maximum of two words, and convert them into a readable text. The two AI approaches used in this work are Bidirectional Long Short-Term Memory (BiLSTM) and convolutional neural network (ConvNet). These Networks have recently shown significant performance improvements in image and speech analysis. Our Models are realized with MATLAB® and finally optimized with suitable parameters. To enhance both recognition systems we used Mel Frequency Cepstral Coefficients (MFCC) Feature, instead of using raw wave signals. As a result, both approaches have reached a remarkable recognition accuracy. The best obtained validation accuracy was from the BiLSTM network, which has reached 98.22% and only one false positive classification by using it on a test data set. The purpose of this paper is to provide a comparative analysis and evaluation of two different deep learning approaches for a speech commands recognition system.