Speaker-independent negative emotion recognition

This work aims to provide a method able to distinguish between negative and non-negative emotions in vocal interaction. A large pool of 1418 features is extracted for that purpose. Several of those features are tested in emotion recognition for the first time. Next, feature selection is applied separately to male and female utterances. In particular, a bidirectional Best First search with backtracking is applied. The first contribution is the demonstration that a significant number of features, first tested here, are retained after feature selection. The selected features are then fed as input to support vector machines with various kernel functions as well as to the K nearest neighbors classifier. The second contribution is in the speaker-independent experiments conducted in order to cope with the limited number of speakers present in the commonly used emotion speech corpora. Speaker-independent systems are known to be more robust and present a better generalization ability than the speaker-dependent ones. Experimental results are reported for the Berlin emotional speech database. The best performing classifier is found to be the support vector machine with the Gaussian radial basis function kernel. Correctly classified utterances are 86.73%±3.95% for male subjects and 91.73%±4.18% for female subjects. The last contribution is in the statistical analysis of the performance of the support vector machine classifier against the K nearest neighbors classifier as well as the statistical analysis of the various support vector machine kernels impact.