Non-stationary acoustic objects as atoms of voiced speech

In spite of the undisputedly high degree of non-stationarity of speech signals, the present day determination of its acoustic features is based on the assumption that speech production can be described as a linear time invariant (LTI) system on the time scale of about 20 ms [1]. In automatic speech recognition, the wide sense stationarity of an LTI– system is used as prerequisite for the consistent estimation of Fourier spectra or of autoregressive models [1]. As an evolutionarily plausible supplement, speech perception is also assumed to be focussed on acoustic features which are obtained by using the LTI assumption [2]. Present day models of pitch perception are no exception [2, 3]. The empirical mode decomposition of Huang et al. [4] represents one of very few methods, which are suited to analyze signals without assuming their stationarity. In case of voiced speech, the formant frequencies have been shown to correlate well with frequencies of the empirical modes [5]. However, it is hard to imagine how the so called sifting process of Huang et al. [4] fits into the known function of the peripheral auditory pathway. The present study proposes a different method of empirical mode decomposition, which is based on well known features of the auditory pathway. As a further contrast to decomposition [4], the pitch perception oriented mode decomposition leads to mode reconstructions, which can be confirmed to have uncorrupted phases in comparison to the phases of underlying oscillatory subsystems [6, 7]. Furthermore, it is shown that voiced phones support a fast convergent iterative reconstruction. The empirical modes can be used advantageously to reconstruct a single fundamental oscillator, the frequency of which can be interpreted as the acoustic correlate of virtual pitch. Virtual pitch perception should be interpreted as an ingenious instrument of time scale separation, which separates the phonological time or frequency scales from the phonetically relevant ones. In contrast to numerous existing pitch tracking methods [2, 3], the virtual pitch oriented time scale separation does not rely on a frequency gap in the long term spectrum (being equivalent to temporary stationarity).