Work-related stress is an uncompromising burden with compounding effects on individuals, communities, organizations, and the economy. Highly automated digital environments, such as smart factories and cyber-physical ecosystems, are significantly impacted by these negative effects due to the inherent constraints on human engagement and social interaction. Verbal communication between human operators is a critical resource in such environments, as it can be effectively utilized to address this challenge. A mix of statistical and artificial intelligence (AI) techniques have been reported in recent literature for the detection of stress-related indicators from audio recordings. However, none of these studies has focused on the challenges of detecting work-related stress in cyber-physical environments, such as smart factories, where verbal communication is not only constrained but also impaired by background noise and other disturbances common to factory settings. In this paper, we address these challenges by proposing a novel deep learning approach, based on the workings of the Convolutional Neural Network (CNN) and Growing Self- Organizing Map (GSOM) algorithms. In addition to this novel AI approach, sampling strategy and speaker diarization techniques are utilised for noise reduction. Pitch and speech augmentation techniques address the data imbalance issue common in most real-world datasets. The accuracy and effectiveness of the proposed approach is demonstrated using a benchmark dataset (DAIC-WOZ). We report F1 scores of 82% and 64% for normal and distressed classes respectively, which outperform the state-of-the-art models. We conclude with a discussion on the empirical evaluation of the proposed approach in cyber-physical environments and directions for future work.