Estimating glottal aspiration noise via wavelet thresholding and best-basis thresholding

For the synthesized singing voice to sound breathy, aspiration noise is perceptually most important (see Klatt, D.H. and Klatt, L.C., 1990). We have recently proposed to use pitch-synchronous, amplitude-modulated Gaussian noise to model the aspiration component of the glottal excitation (see Lu, Hui-Ling and Smith, J.O., Proc. Int. Computer Music Conf., p.90-7, 2000). For proper parameterization of the noise residual model, we need to analyze the signal properties of the glottal aspiration noise estimated from real breathy singing recordings. Several wavelet thresholding and best-basis thresholding methods are studied and compared in order to obtain reliable estimates of the glottal aspiration noise. We conclude that the best-basis soft-thresholding method is the most effective way to extract the glottal aspiration noise.