Timbre description of the sound of air-treatment systems for predicting acoustic confort

This study aims at defining reliable acoustic cues for the measure, characterization and prediction of the acoustic comfort of air-treatment systems (ATS). To meet customers’ expectations, industrial products tend increasingly to follow a process of "sound design". In this process, the perceptual evaluation of sound quality is a necessary step to define acoustic specifications. In this context, this study aims at defining the main perceptual attributes of the sound of air-treatment systems in order to predict users’ preferences. The timbre space of a sound dataset extracted from a large recording database was thus identified through a similarity experiment where participants were asked to rate the resemblance between each pair of sounds. The results of this experiment were analyzed with a Multidimensional Scaling (MDS) method in order to extract the main perceptual attributes. Finally, these attributes were linked to relevant audio features through a regression method in order to define a reliable computable metric of acoustic comfort. This study was conducted through the Vaicteur Air2 project supported by OSEO. Similarity scaling experiment Goal: Whereas it is difficult for a listener to identify a sound’s most relevant acoustic features, it is much easier to rate how much 2 sounds of the same kind are different from one another Stimui : 16 loudness-equalized monophonic sounds = 16 different ATSs Apparatus : – GUI Labview 2010 – Interface audio RME Fireface 800 – Casque audio Sennheiser HD650 Procedure: Each possible pair of sounds (120 pairs) is presented to the listener who has to rate its similarity with a slider on a continuous scale MultiDimensional Scaling analysis [1] Context: – Similarity matrix: distances between the sounds = perceptual similarities – N element in a common geometrical space⇒ N − 1 dimensions required Goal: Model all distances with a space of much lower dimensionality, (2 or 3) Models: – MDSCAL single model ⇒ Rotationally invariant space – INDSCAL weighted model [2]: applies weightings to the dimensions ⇒ The space is no longer rotationally invariant ⇒ The dimensions are perceptually meaningfull MDSCAL model dij = √√√√ R ∑