Car door closure sounds : characterization of perceptual properties through analysis-synthesis approach

The aim of this study is to identify perceptually pertinent parameters for the evaluation of car door closure sounds. For this purpose, perceptual properties of recorded sounds were first evaluated by sensory metrology. Then, an analysis-synthesis approach was chosen in order to identify perceptually pertinent signal parameters. The analysis part of this process first consisted in decomposing the sound in several independent impact sources using the Empirical Modal Decomposition method. Each impact is then modeled by a set of gains and damping factors in each critical band (ERB). These parameters were further used to synthesize sounds related to various aspects of the door closure sound with a real-time tool. This approach allowed for the generation of realistic, synthesized car door closure sounds that preserve perceptual properties with a reduced number of signal parameters. Listening tests finally allowed for the observation of the influence of the main signal parameters on the perceptual properties of such action-related impact sounds. INTRODUCTION Door closure sounds are of interest for impact sound research. From an acoustical point of view, such sounds are complex and composed by several impacts, such as the latch mechanism and the door/panel impact. From a perceptual point of view, it conveys complex perceptual properties: it tells whether the door is well closed and contributes to the overall impression of the car quality. The relations between sounds and perception have already been studied, but are still not well understood. A recent study [1] proposes timbre parameters as frequency balance and cleanness (only one temporal event is required) to predict preference ratings. This model is based on listening tests by paired comparison for similarity and quality ratings of 12 sounds and statistical links between signal parameters and perceptual dimensions. It appears, however, that this simple model is not sufficient to unambiguously predict perceptual properties. As shown for instrumental sounds [2], this issue could be addressed with synthetic sounds directly controlled with signal parameters, and especially designed to cover the perceptual space: on the one hand, we directly observe the effect of each signal parameter on perceptual properties in order to identify the true underlying parameters; on the other hand we mix the problem of finite number of data with a good repartition of stimuli in the perceptual space. Moreover we propose the sensory metrology instead of similarity rating to access perceptual properties. This methodology consists in extracting salient descriptors that characterize the sound space without any assumption on the continuity and orthogonally of dimensions. On the one hand, sensory descriptors that describe sounds per se are easier to link to signal parameters than impressions or events; on the other hand, they contain sufficient information to predict sound quality [3]. We thus propose to define an analysis synthesis model based on perceptual analysis of real sounds. The model has to reproduce the sound from a perceptual point of view and to be controlled by few parameters to allow an easy observation of the influence of signal parameters on perceptual properties.