Audio Feature Space Analysis for Emotion Recognition from Spoken Sentences

An analysis of low-level feature space for emotion recognition from the speech is presented. The main goal was to determine how statistical properties computed contours features influence signals. We have conducted several experiments reduce and tune our initial set configure classification stage. In process audio space, we employed univariate selection using chi-squared test. Then, in first stage classification, a default parameters selected every classifier. For classifier that obtained best results with settings, hyperparameter tuning cross-validation exploited. result, compared two different languages find out difference between emotional states expressed spoken sentences. show an containing 3198 attributes dimensionality reduction about 80% algorithm. most dominant at this based on mel bark frequency scales filterbanks its variability described mainly by variance, median absolute deviation standard average deviations. Finally, accuracy tuned SVM equal 72.5% 88.27% sentences Polish German languages, respectively.