Random wavelet representation of unvoiced speech

The different types of speech noise are reviewed together with their first-order statistics. A theoretical framework is formulated for speech noise analysis in terms of time-frequency energy concentrations (spectral peaks, amplitude and phase modulation). This framework is based on the linear acoustic theory of speech production, and on the theory of random modulation. A synthesis scheme using random timed wavelets is proposed. It is based on the statistical properties of shot noise. Experimental results on synthesis of speech noise are given.<
>