Deutsch W., Waubke H., Gygi B., Noll A.

Measures for describing noise data are usually selected from national and international standards or guidelines. These standards almost exclusively employ integrated levels with varying temporal and spectral weighting, primarily spectral A-weighting and temporal exponential smoothing with a 1s time constant. More specific descriptions use octave- and third-octave band spectra. Because of the averaging methods used, these parameters are not suitable to reproduce transients (isolated short time events) or to represent perceptual relevance sufficiently. A direct psychoacoustic evaluation using only annoyance estimates remains questionable, because of the large variability of annoyance estimates both between and among individuals. The temporal interpersonal variability is based (among other factors) on differential sleep patterns, attitudes towards the cause of the noise and short-time trends in the social environment. A solution that is independent from psychological variability, but characterizes noise events in more detail, uses multiple features that are derived from the waveform yet provide perceptual relevance. The international ISO standard MPEG7-4 was defined during the past years for semi-automatic description of audio in multimedia. The descriptors outlined in the MPEG7-4 standard are here tested in detail on a large number of recorded train segments. A method is proposed to reduce the large number of MPEG-7 descriptors to a smaller set that is relevant for noise events. Using MPEG-7 descriptors and some related acoustic measures, the similarity space for the train sounds is developed, and the features that best predict the structure of the space are isolated. This method singles out the features that best allow detection of acoustic events and also to identify general classes of trains with the highest risk of annoyance. Among these features are complex timbral and envelope descriptors. The project includes the installation of a monitoring system at a railway track to prove the relevance of the used feature set.