Spectral tilt as a perturbation-free measurement of noise levels in voice signals

Abstract

Acoustic analysis of voice quality proves useful in the objective assessment of voice disorders and for motivating new components for use in improving voice synthesis. A commonly used quantitative spectral index is the harmonicsto-noise ratio (HNR), which gives gross information regarding speech signal periodicity. However, as the measure is sensitive to all forms of waveform aperiodicities (not simply the additive random noise component of turbulent origin), it lacks specificity. Furthermore, the HNR of the radiated speech waveform has a fundamental frequency (f0) -dependence, increasing with fundamental frequency (for equal noise levels of the glottal source). Two spectral tilt measurements are applied to synthetically generated, aperiodic voice signals to investigate their sensitivity to the various forms of aperiodicity. The tilt measures are found to provide perturbation(jitter and shimmer) free measures of noise levels in speech signals. However, for radiated speech waveforms the tilt measurements are strongly f0-dependent.