Complex Acoustic Environments: Review, Framework, and Subjective Model

Abstract

The concept of complex acoustic environments has appeared in several unrelated research areas within acoustics in different variations. Based on a review of the usage and evolution of this concept in the literature, a relevant framework was developed, which includes nine broad characteristics that are thought to drive the complexity of acoustic scenes. The framework was then used to study the most relevant characteristics for stimuli of realistic, everyday, acoustic scenes: multiple sources, source diversity, reverberation, and the listener's task. The effect of these characteristics on perceived scene complexity was then evaluated in an exploratory study that reproduced the same stimuli with a three-dimensional loudspeaker array inside an anechoic chamber. Sixty-five subjects listened to the scenes and for each one had to rate 29 attributes, including complexity, both with and without target speech in the scenes. The data were analyzed using three-way principal component analysis with a (2 3 2) Tucker3 model in the dimensions of scales (or ratings), scenes, and subjects, explaining 42% of variation in the data. "Comfort" and "variability" were the dominant scale components, which span the perceived complexity. Interaction effects were observed, including the additional task of attending to target speech that shifted the complexity rating closer to the comfort scale. Also, speech contained in the background scenes introduced a second subject component, which suggests that some subjects are more distracted than others by background speech when listening to target speech. The results are interpreted in light of the proposed framework.

Keywords: complex acoustic environments; complexity; hearing; perception; three-way principal component analysis.

Figure 1.

Scree plot of different component combinations of Tucker3 models. The triplets on the plot designate the number of components for scales scenes subjects. The amount of explained variation is measured as percentage of the sum of square differences between the original data set to the fitted one. The final model that was selected (2 3 2) is emphasized.

Figure 2.

Joint biplot of the scale and scene first and second components. The vectors represent the different scales, where names ending with (S) refer to target speech attributes in Part C of the questionnaire. The first two scene components are superimposed on the plot, with the large markers designating the scenes. If a marker is closer to a particular scale, then it is more closely related to that attribute. The biplot origin designates the average value of every scale over all scenes and subjects.

Figure 3.

The strongest correlation of any rating scale to the second subject component was with the listening effort rating during speech. Only the scenes with dominant background speech drove this differentiation. Hearing status does not appear to consistently predict the mean subjective ratings (four subjects with mild losses, >25 dB HL, were grouped separately from the slight losses for illustration purposes only).

Figure 4.

The scene complexity ratings averaged over all 65 subjects, ordered by increasing sound pressure level (dB SPL) from left to right. The error bars are the Tukey–Kramer 95% confidence intervals of pairwise comparisons of all ratings.