InterSpeech 2021

Many features have been proposed for use in speech emotion recognition, from signal processing features to bag-of-audio-words (BoAW) models to abstract neural representations. Some of these feature types have not been directly compared across a large number of speech corpora to determine performance differences. We propose a full factorial design and to compare speech processing features, BoAW and neural representations on 17 emotional speech datasets. We measure the performance of features in a categorical emotion classification problem for each dataset, using speaker-independent cross-validation with diverse classifiers. Results show statistically significant differences between features and between classifiers, with large effect sizes between features. In particular, standard acoustic feature sets still perform competitively to neural representations, while neural representations have a larger range of performance, and BoAW features lie in the middle. The best and worst neural representations were wav2veq and VGGish, respectively, with wav2vec performing best out of all tested features. These results indicate that standard acoustic feature sets are still very useful baselines for emotional classification, but high quality neural speech representations can be better.