The monitoring of an escalating negative interaction has several benefits, particularly in security, (mental) health, and group management. The speech signal is particularly suited to this, as aspects of escalation, including emotional arousal, are proven to easily be captured by the audio signal. A challenge of applying trained systems in real-life applications is their strong dependence on the training material and limited generalization abilities. For this reason, in this contribution, we perform an extensive analysis of three corpora in the Dutch language. All three corpora are high in escalation behavior content and are annotated on alternative dimensions related to escalation. A process of label mapping resulted in two possible ground truth estimations for the three datasets as low, medium, and high escalation levels. To observe class behavior and inter-corpus differences more closely, we perform acoustic analysis of the audio samples, finding that derived labels perform similarly across each corpus, with escalation interaction increasing in pitch (F0) and intensity (dB). We explore the suitability of different speech features, data augmentation, merging corpora for training, and testing on actor and non-actor speech through our experiments. We find that the extent to which merging corpora is successful depends greatly on the similarities between label definitions before label mapping. Finally, we see that the escalation recognition task can be performed in a cross-corpus setup with hand-crafted speech features, obtaining up to 63.8% unweighted average recall (UAR) at best for a cross-corpus analysis, an increase from the inter-corpus results of 59.4% UAR.@en

The INTERSPEECH 2021 Computational Paralinguistics Challenge addresses four different problems for the first time in a research competition under well-defined conditions: In the COVID-19 Cough and COVID-19 Speech Sub-Challenges, a binary classification on COVID-19 infection has to be made based on coughing sounds and speech; in the Escalation Sub- Challenge, a three-way assessment of the level of escalation in a dialogue is featured; and in the Primates Sub-Challenge, four species vs background need to be classified. We describe the Sub-Challenges, baseline feature extraction, and classifiers based on the 'usual' COMPARE and BoAW features as well as deep unsupervised representation learning using the AUDEEP toolkit, and deep feature extraction from pre-trained CNNs using the DEEP SPECTRUM toolkit; in addition, we add deep end-to-end sequential modelling, and partially linguistic analysis.

@en