On the perception of affect in the singing voice: a study of acoustic cues

Abstract

This study addresses the perception of affect in vocal and glottal recordings of a singing voice. An experiment was made to rate the samples on four broad affect terms describing the two-dimensional model of emotion. A cross-tabulation between the singing expressions and affect scores revealed their relationship with affect dimensions. Prosodic as well as spectral acoustic cues were extracted and statistical analysis performed on 22 features revealed a set of cues whose means are statistically significant with respect to valence and arousal, namely SPR, F5, B1, B4, mean pitch, mean intensity, brightness, jitter, shimmer, mean autocorrelation, mean HNR, mean LTAS, RMS, SPL, LPH, and LTAS slope. Principal component analysis was made for vocal and glottal features: 2 components explained 78.1% and 73.5% of the original variance of prosodic cues, and 2 components explained 86.3% and 86.7% of the original variance of prosodic and spectral cues.