Audio emotion recognition by perceptual features

A 9-D perceptual feature set has been used for audio emotion recognition. Performance tests have been performed on well known EMO-DB and VAM databases and the results are reported for different classifiers. Support Vector Machines, Gaussian Mixture Models and Learning Vector Quantization have been used in classification. Audio emotion recognition performance achieved by the perceptual visual features are compared to openEar and GerDa which are cited as state of the art audio emotion recognition systems. It is shown that the 9-D perceptual feature vectors are highly discriminative in continuous emotional space. It is concluded that the learning Vector Quantization increases the performance for natural records, while the Support Vector Machines provide the highest recognition rate for the acted records.