Speech Emotion Recognition: A Multi-Dataset Experimental Study
Main Article Content
Abstract
Speech emotion recognition involves identifying emotions from speech recordings. When some emotions are more frequent than others in a dataset, overall accuracy can mask two kinds of errors: missing an emotion when it is present and predicting it when it is absent. In this study, we evaluated the Enhanced Speech Emotion Recognition Framework (ESERF) on eight emotion categories using recordings from RAVDESS, CREMA-D, and TESS. We assessed performance across three experimental scenarios using overall accuracy, confusion matrices, and class-wise precision, recall, and F1-score. Overall accuracy was 60.74%, 61.02%, and 49.22% in Scenarios 1, 2, and 3, respectively. Although the first two scenarios had similar overall accuracy, recall for Calm increased from 0.80 to 0.87, while precision fell from 0.64 to 0.44. Scenario 2 therefore identified a larger proportion of the Calm recordings, but fewer than half of the recordings it classified as Calm carried that label. The decline in recall from Scenario 2 to Scenario 3 also varied by emotion: recall fell by about 25 percentage points for both Fear and Happy, compared with only two percentage points for Sad.
Article Details
Issue
Section

This work is licensed under a Creative Commons Attribution 4.0 International License.