Speech as an Emotional Load Biomarker in Clinical Applications

Resumo

Introduction: Healthcare professionals often contend with significant emotional burdens in their work, including the impact of negative emotions, such as stress and anxiety, which can have profound consequences on immediate and long-term healthcare delivery. In this paper a stress estimation algorithm is proposed based on the classification of negative valence emotions in speech recordings.

Methods: An end-to-end machine learning pipeline is proposed. Two distinct decision models are considered, VGG-16 and SqueezeNet, while sharing a common constant Q power spectrogram input for acoustic representation. The system is trained and evaluated using the RAVDESS and TESS emotional speech datasets.

Results: The system was evaluated for individual emotion
classification (multiclass problem) and also for negative and
neutral or positive emotion classification (binary problem). The results achieved are comparable to previously reported systems, with the SqueezeNet model offering a significantly smaller footprint, enabling versatile applications. Further exploration of the model's parameter space holds promise for enhanced performance.

Conclusion: The proposed system can constitute a feasible
approach for the estimation of a low-cost non-invasive biomarker for negative emotions. This allows to raise alerts and develop mitigating actions to the burden of negative emotions, being an additional management tool for healthcare services that allows to maintain quality and maximize availability.