• co-speech gestures
• Conditional Generative Adversarial Networks (cGAN)
• Generative Adversarial Networks (GAN)
• human-robot interaction
• humanoid robot
• machine learning

This thesis explores the integration of natural co-speech gestures in humanoid social robots to improve the realism and effectiveness of human-robot interactions, with the objective of developing a gesture generation system that produces synchronized and contextually appropriate gestures aligned with speech. This work further contributes towards the area of HRI by presenting a machine learning focused method addressing the limitations of predefined-scripts, data-driven and probabilistic approaches previously proposed. The primary challenges addressed include the creation of a comprehensive dataset from the Trinity Speech-Gesture dataset by aligning the body motion and the speech data, robust machine learning model training after the due preprocessing of said data, and extensive performance optimization. The focus is on the latter step; fine-tuning Generative Adversarial Networks(GAN) with a generator-discriminator architecture, and Conditional Generative Adversarial Network (cGAN) with an encoder-decoder architecture and a Sequence-to-Sequence (Seq2Seq) model to generate suitable gestures from given speech data. Results showed improvement in the gesture predictions with a high test R² score for the cGAN model despite the task complexity. The naturalness evaluation of the model's raw outputs was done by visualizing the generated gestures on the 3D modeling software Blender due to their difficulty in interpretation.