SPARSE TIME-FREQUENCY REPRESENTATIONS IN AUDIO PROCESSING,AS STUDIED THROUGH A SYMMETRIZED LOGNORMAL MODEL

摘要

Time-frequency representations are ubiquitous in speech andrnaudio signal processing, their use being motivated by both auditoryrnphysiology and the mathematics of Fourier analysis. Nonparametricrnstatistical models (or equivalently transform based signalrnprocessing methods) formulated in this space provide a principledrnway to decompose sounds into their constituent parts, as well asrnan effective means of exploiting the local correlation present in therntime-frequency structure of naturally generated acoustic signals.rnHere we describe how an appropriate generative statistical model,rneven under very simple assumptions, provides a means of exploringrnsparse time-frequency representations in audio. We introduce arnsymmetrized lognormal model for spectral coefficients, which showsrngood agreement across a broad range of speech samples taken fromrnthe TIMIT database, and demonstrate preliminary speech enhancementrnresults based on a maximum a posteriori shrinkage estimator.