Quantifying Cognitive Load from Voice using Transformer-Based Models and a Cross-Dataset Evaluation

Cognitive load is frequently induced in laboratory setups to measure responses to stress, and its impact on voice has been studied in the field of computational paralinguistics. One dataset on this topic was provided in the Computational Paralinguistics Challenge (ComParE) 2014, and therefore offers great comparability. Recently, transformer-based deep learning architectures established a new state-of-the-art and are finding their way gradually into the audio domain. In this context, we investigate the performance of popular transformer architectures in the audio domain on the ComParE 2014 dataset, and the impact of different pre-training and fine-tuning setups on these models. Further, we recorded a small custom dataset, designed to be comparable with the ComParE 2014 one, to assess cross-corpus model generalisability. We find that the transformer models outperform the challenge baseline, the challenge winner, and more recent deep learning approaches. Models based on the 'large' architecture perform well on the task at hand, while models based on the 'base' architecture perform at chance level. Fine-tuning on related domains (such as ASR or emotion), before fine-tuning on the targets, yields no higher performance compared to models pre-trained only in a self-supervised manner. The generalisability of the models between datasets is more intricate than expected, as seen in an unexpected low performance on the small custom dataset, and we discuss potential 'hidden' underlying discrepancies between the datasets. In summary, transformer-based architectures outperform previous attempts to quantify cognitive load from voice. This is promising, in particular for healthcare-related problems in computational paralinguistics applications, since datasets are sparse in that realm.