Second-order Statistics Are Less Important for Audio Textures than Image Textures

This letter proposes to adapt the techniques of image texture analysis to the context of audio signals. Dynamics of audio frame-level features are particularly difficult to model, although they have been identified as crucial perceptive dimensions of timbre perception. Recent studies have given ample evidence that traditional means to model data dynamics, such as delta-coefficients, texture windows or Markov modelling, do not provide any improvement over the best static models. The situation is completely different for image texture analysis, where second-order statistical analysis, based on co-occurrence matrices have proven to be an efficient measure for texture similarity. This paper reports on experiments to adapt these image analysis techniques to audio signals. Results show that co-occurrence analysis of sound textures based on vector-quantized MFCC features do not provide any advantage over first-order histograms in the building of content-based similarity measures. This suggests that second-order statistics, in the form considered in this study, are not a factor as crucial for the perception of sound textures as it is for image textures. EDICS Category: AEA-AUEA,IMD-ANAL