Project B-T2 – Learning the reflection characteristics of rooms

In most acoustic environments, walls and other objects reflect sound. This reverberation causes the signals captured by the ears of a listener to be very complex. Nevertheless, people are able to localize sound sources inside reverberant environments precisely. A possible explanation for this phenomenon is given by the precedence effect in the auditory system. The auditory system fuses direct sound with reflections arriving sooner than the so-called echo threshold and localizes the sound source using only the leading, direct sound. The echo threshold increases with prior exposure to thereflections [1]. The first aim of this project is to investigate if this precedence effect build-up also occurs with the multiple sound reflections of a realistic room, as a previous study suggests [2]. Changes in the reflection pattern may reset the build-up and reflections suddenly become audible, which may be an indication that the auditory system had adapted to a particular listening environment and that this adaptation breaks down when unnatural changes in the reflection pattern happen. A key aim of the project is to find out if the auditory system is able to extract, abstract and learn information about the room. In a later stage of the project, we will use numerical modeling to investigate possible mechanisms for extracting and learning room geometries from the information carried in the sound.

Objectives and description of the project

Previous studies show that room learning improves speech understanding [3]. We will further investigate the subject of room learning. In particular, we will address the question if only the pattern of reflections is learned or if room learning occurs on a higher level. This will lead to a better understanding of the representation of acoustic environments in the brain. For this project, we will use the highly accurate localization method of Seeber [4] and a sophisticated room simulation and auralization system, the Simulated Open Field Environment [5]. Furthermore, we will also model the observed effects. Djelani and Blauert [6] postulate that binaural information of sources and reflections is reduced by an onset triggered inhibition processs. We aim to develop a model that learns room properties from binaural signals. It has previously been shown that the room geometry can be inferred from a set of reflection patterns [7]. Our approach will exploit the assumption that structure is available in both the direct sound and the reflection pattern. The project will extensively draw on expertise in machine learning and BSS [8, 9], which will be translated to binaural auditory modeling.

[1]: Clifton and Freyman, Perception & psychophysics 1989.

[2]: Djelani and Blauert, Acta Acustica 2001.

[3]: Brandewie and Zahorik, J. Acoust. Soc. Am. 2010.

[4]: Seeber, Acta Acustica 2002.

[5]: Seeber et al., Hearing Research 2010.

[6]: Djelani and Blauert, Forum Acusticum 2002.

[7]: Dokmanic et al., PNAS 2013.

[8]: Hawe et al., IEEE Trans. on Image Proc. 2013