Audio texturedness indicator based on a direct and reverse short listening time analysis

Abstract

In this paper, we present an objective evaluation of audio texturedness level motivated by a subjective study which emphasizes the relevance of a direct and reverse short listening time analysis. This study was based on 77 undergraduate engineering students, where the concept of audio texturedness relied on audio and image analogies. Audio texturedness evaluation of a large audio database using a discrete [1 - 5] texturedness scale was performed. As a result, an objective audio texturedness indicator is proposed based on the cumulative average signal informational content change in the direct and reverse directions. This bidirectional cumulative entropy tracking was done in analogy with a classical multidirectional homogeneity method for image texture discrimination. As expected, the proposed indicator has ranked the class of noise on the high end of the [1 - 5] scale, whereas highly motivational speech signals ranked low on the proposed scale due to the large variations in their average informational content in the direct and reverse directions. This audio texturedness indicator offers a continuum of audio classes, in contrast with the classical noise, speech, and music sound categorization. The relevance of the proposed objective indicator auditory parameters was explored for a maximum objective-subjective cross-correlation and illustrated in a preliminary audio stream segmentation application.