Automatic detection of stress in speech

We have developed software based on the Stevens landmark theory to extract features in utterances in and adjacent to voiced regions. We then apply two statistical methods, closest-match (CM) and principal components analysis (PCA), to these features to classify utterances according to their emotional content. Using a subset of samples from the Actual Stress portion of the SUSAS database as a reference set, we automatically classify the emotional state of other samples with 75% accuracy, using CM either alone or with PCA and CM together. The accuracy apparently does not depend strongly on measurement errors or other small details of the present data, giving confidence that the results will be applicable to other data.