Time-frequency representations of auditory signatures: Dynamic signal models

Abstract

Signal theory provides a mathematical framework within which to pose questions of representation. Such theory establishes the minimum conditions for which signals can be represented uniquely, for example. Time-frequency distributions are often used as representations of the temporally varying spectral characteristics of acoustic signals. When one desires a correspondence between the distribution of a signal and the auditory perception of that signal, however, it is not clear how to proceed. Several researchers have approached the auditory representation problem within the domain of time-frequency distributions [M. D. Riley, SpeechTime-FrequencyRepresentations (Kluwer Academic, Boston, 1989) and W. Heinbach, Acustica 67, 113–121 (1988)]. In the present paper, an alternative approach is provided based on dynamic signal models and psychoacoustic spectral processing theory. This approach is applied to characterizing acoustic transients such as automotive starter sounds. A dynamic signal model is introduced and experimental results are presented concerning the ability of this model to uniquely characterize the signals of interest. Application of Terhardt’s Fourier t transform is also investigated [E. Terhardt, Acustica 57, 242–256 (1985)] and results are compared with other time-frequency distributions. [Work supported by a Bell Communications Research graduate fellowship, the Ford Motor Company, and a National Science Foundation PYI award.]