Abstract

The term stress can designate a number of situations and affective reactions. This work focuses on the immediate stress reaction caused by, for example, threat, danger, fear, or great concern. Could measuring stress from speech be a viable fast and non-invasive method? The article describes the development of a system predicting stress from voice – from the creation of the database, and preparation of the training data to the design and tests of the regressor. StressDat, an acted database of speech under stress in Slovak, was designed. After publishing the methodology during its development in 1, this work describes the final form, annotation, and basic acoustic analyses of the data. The utterances presenting various stress-inducing scenarios were acted at three intended stress levels. The annotators used a ``stress thermometer'' to rate the perceived stress in the utterance on a scale from 0 to 100. Thus, data with a resolution suitable for training the regressor was obtained. Several regressors were trained, tested and compared. On the test-set, the stress estimation works well (R square = 0.72, Concordance Correlation Coefficient = 0.83) but practical application will require much larger volumes of specific training data. StressDat was made publicly available.

Keywords

Cite this article

@article{cai-6895,
  author  = {Marián Trnka and Sakhia Darjaa and Róbert Sabo and Milan Rusko and Meilin Schaper and Tim Stelkens-Kobsch},
  title   = {Prediction of Stress Level from Speech -- from Database to Regressor},
  journal = {Computing and Informatics},
  volume  = {42},
  number  = {5},
  pages   = {1233--1254},
  year    = {2023},
  issn    = {1335-9150},
  doi     = {10.31577/cai_2023_5_1233}
}
Download .bib