Evaluation of Designs for Loudness-Matching Experiments

In a study of the assessment of loudness of music and speech, several different experimental designs were compared. The effect of using a fixed-reference design, compared to using a balanced pair-matching design with various degrees of redundancy and balance, was investigated with a technique inspired by resampling statistics. In a loudness matching experiment using a fixed-reference design, the subjects match all stimuli against a single sound segment selected in advance. Alternatively, in a balanced pair-matching design, both stimuli of a pair are drawn from the same segment collection. In balanced pair-matching, each of the segments do not need to be matched with every other segment; the specification of the set of segment pairs to include in the experiment corresponds to the degree of redundancy and balance in the design. The results indicated that the choice of specific fixed reference segments did influence the accuracy of the assessments – in particular a 1 kHz tone was inferior. When using characteristic samples of speech material, rock/pop, and classical music, as fixed reference, no significant difference in accuracy between the genres was indicated. Applying the balanced pair-matching method with only a small degree of redundancy yielded considerably more accurate results than the designs without repetitions evaluated. 1. INTRODUCTION The perceived loudness of music and speech can be measured by means of a controlled listening experiment. The loudness of homogeneous sound segments, with a duration of several seconds, may be compared because the overall loudness of each segment is perceived to be fairly constant. The property of the sound which is assessed is its long-term loudness. The results from a loudness assessment experiment could be studied to gain knowledge of the perception of loudness (psychoacoustics). The results could also be utilized in the development of objective measures or algorithmic models of loudness. Measuring the perceived loudness in a listening experiment implicates several choices regarding the experimental design and procedure. In particular, a loudness matching experiment could either be based on a fixed-reference method, or on what we shall call the balanced pair-matching method. When using the fixed-reference method, a single sound segment is selected in advance as the reference stimulus. The subjects then match all (other) stimuli against this reference. In contrast, when using the balanced pair-matching method, no sound segment is distinguished as the reference; both sound segments in a pair of stimuli are drawn from the same collection. The composition of the set of pairs to be matched is said to be balanced, basically because the frequency of occurrence of the different segments is the same. In [1] we described an investigation of loudness assessment of music and speech. A pilot experiment was carried out using half the number of subjects and a scaled-down