Investigation on Pleasure Related Acoustic Features of Affective Speech*

This paper presents our recent work on the investigation of affective speech along the pleasure-displeasure (P) dimension in the PAD emotional space. First, 76 conventional features are taken as candidates. And a novel feature, F0 Dominant Ratio, which is designed to reflect the dominant pitch, is also introduced. Correlative coefficients are calculated and a preliminary selection is done. Factor analysis and co-clustering are then employed. After another two rounds of selection, a set of 5 correlative features is selected. It includes Spectral Roll off, Spectral Low-high Ratio, Average F0, Min. F0, and the F0 Dominant Ratio, which can illustrate the acoustic differences between arousal-equaled emotion categories. Then a conversion experiment is carried out. Our model is built through stepwise linear regressions, and modifications are implemented carefully. Perceptual test shows that the conversion result of pleasure intensity is acceptable. Among the 5 features, the spectral ones show the highest favor. The dominant pitch appears relative, yet it requires further investigation. Meanwhile, intensity of arousal (A) in these sentences is agreed to be unchanged, i.e. the pleasure-oriented features are separated nicely from the arousal-oriented ones.