Voice Source Parameters Estimation by Fitting the Glottal Formant and the Inverse Filtering Open Phase

This paper presents two approaches to the problem of extracting the parameters of the LF source model directly from the speech waveform. The first approach relies on the glottal formant estimated from the anticausal contribution of speech. Indeed the ZZT technique has recently shown its ability to deconvolve speech into its causal and anticausal components. The second method is based on the glottal open phase obtained by inverse filtering. The notion of unanalyzable frames and the way to detect and correct them are also presented. Once source parameters are extracted, the coefficients of the ARX speech production model are estimated by spectral division. Decomposition on both synthetic and natural speech, as well as an analysis-synthesis test confirm the accuracy of methods exposed.