Effect of speech compression on the automatic recognition of emotions

This paper investigates the effects of standard speech compression techniques on the accuracy of automatic emotion recognition. Effects of Adaptive Multi-Rates (AMR), Adaptive Multi-Rate Wideband (AMR-WB) and Extended Adaptive Multi-Rate Wideband (AMR-WB+) speech codecs were compared against emotion recognition from uncompressed speech. The recognition methods included techniques based on three different types of acoustic speech parameters: Teage Energy Operator features (TEO), Mel Frequency Cepstral Coefficients (MFCCs), and Glottal Time and Frequency domain features (GP-T and GP-F). The results showed that in general, all three speech compression techniques resulted in the reduction of emotion recognition accuracy. However, the amount of degradation varied across compression methods and types of acoustic features. It was observed that the accuracy of emotion recognition using the AMR-WB technique was higher than the accuracy of the AMR-WB+ and the AMR codecs. Further, the TEO-PWP features showed much more robust performance under different compression rates than the MFCC, GP-T and GP-F features.