Sound Texture Synthesis with Hidden Markov Tree Models in the Wavelet Domain

In this paper we describe a new parametric model for synthesizing environmental sound textures, such as running water, rain, and fire. Sound texture analysis is cast in the framework of wavelet decomposition and multiresolution statistical models, that have previously found application in image texture analysis and synthesis. We stochastically sample from a model that exploits sparsity of wavelet coefficients and their dependencies across scales. By reconstructing a time-domain signal from the sampled wavelet trees, we can synthesize distinct but perceptually similar versions of a sound. In informal listening comparisons our models are shown to capture key features of certain classes of texture sounds, while offering the flexibility of a parametric framework for sound texture synthesis.