Estimation of vocal noise in running speech by means of bi-directional double linear prediction

The presentation concerns forward and backward double linear prediction of speech with a view to the characterization of vocal noise due to voice disorders. Bi-directional double linear prediction consists in a conventional short-term prediction followed by a distal inter-cycle prediction that enables removing inter-cycle correlations owing to voicing. The long-term prediction is performed forward and backward. The minimum of the forward and backward prediction error is a cue of vocal noise. The minimum backward and forward prediction error has been calculated for corpora involving connected speech and sustained vowels. Comparisons have been performed between the estimated vocal noise and the perceived hoarseness in steady vowel fragments, as well as between the estimated vocal noise in connected speech and sustained vowels produced by the same speakers.