Many natural sounds, such as those produced by rainstorms, fires, insects at night, or birds in a forest, result from large numbers of rapidly occurring acoustic events. Such “sound textures” are often temporally homogeneous, and in many cases do not depend much on the precise arrangement of the component events, suggesting that they might be represented statistically. To test this idea and explore the statistics that might characterize natural sound textures, we designed an algorithm to synthesize sound textures from statistics extracted from real sounds. The algorithm is inspired by those used to synthesize visual textures, in which a set of statistical measurements from a real sound are imposed on a sample of noise. This process is iterated, and converges over time to a sound that obeys the chosen constraints. If the statistics capture the perceptually important properties of the texture in question, the synthesized result ought to sound like the original sound. We tested whether rudimentary statistics computed from the responses of a bank of bandpass filters could produce compelling synthetic textures. Simply matching the marginal statistics (variance, kurtosis) of individual filter responses was generally insufficient to yield good results, but imposing various joint envelope statistics (correlations between bands, and autocorrelations within each band) greatly improved the results, frequently producing synthetic textures that sounded natural and that subjects could reliably recognize. In many cases, textures with audible features (raindrops, crackles, insect/bird calls) emerge from the imposition of much simpler statistical constraints. The results suggest that statistical representations could underlie sound texture perception, and that in many cases the auditory system may rely on fairly simple statistics to recognize real-world sound textures.

This talk represents joint work with Eero Simoncelli and Andrew Oxenham.