Issue No. 04 - July/August (2002 vol. 22)
Shlomo Dubnov , Ben-Gurion University
Ziv Bar-Joseph , Massachussets Institute of Technology
Ran El-Yaniv , Technion-Israel Institute of Technology
Dani Lischinski , Hebrew University of Jerusalem
Michael Werman , Hebrew University of Jerusalem
<p>In this paper we present a statistical learning algorithm for synthesizing new random instances of a sound texture given an example of such a texture as input. A large class of natural and artificial sounds such as rain, waterfall, traffic noises, people babble, machine noises, etc., can be regarded as sound textures--sound signals that are approximately stationary at some scale. Treating the input sound texture as a sample of a stochastic process, we construct a tree representing a hierarchical wavelet transform of the signal. From this tree, new random trees are generated by learning and sampling the conditional probabilities of the paths in the original tree. Transformation of these random trees back into signals results in new sound textures that closely resemble the sonic impression of the original sound source but without exactly repeating it. Applications of this method are abundant and include, for example, automatic generation of sound effects, creative musical and sonic manipulations, and virtual reality sonification. Examples are visually demonstrated in the paper and acoustically demonstrated in an accompanying web site.</p>
S. Dubnov, Z. Bar-Joseph, R. El-Yaniv, D. Lischinski and M. Werman, "Synthesizing Sound Textures through Wavelet Tree Learning," in IEEE Computer Graphics and Applications, vol. 22, no. , pp. 38-48, 2002.