Voice Quality: Functions, Analysis and Synthesis
August 27-29, 2003
The presentation concerns the segregation of vocal noise by means of linear predictive analysis of running speech. Forward double linear predictive analysis has been proposed earlier as a tool to isolate vocal noise from speech signals produced by dysphonic speakers. Double linear predictive analysis involves a short-term and a long-term prediction stage. Long-term linear predictive analysis concerns the prediction of the present speech cycle by means of past ones. Problems with this analysis method are the following. The short-term linear predictive analysis stage is less suited for clinical analyses of speech, because the short-term linear prediction error may not only increase with vocal noise but also with speech and speaker properties that are not related to vocal pathologies. Also, the long-term predictive analysis stage focuses exclusively on the short-term prediction error, discarding any vocal noise that remains in the short-term predicted signal (that is, the weighted sum of past speech samples). Finally, the forward long-term prediction error comprises inflated error transients near the beginning of the recording interval, as well as in the vicinity of the boundaries between voiced and unvoiced speech sounds. The purpose of the presentation is to show that these problems can be solved by a forward and backward long-term predictive analysis of speech, within which the short-term predictive analysis stage can be omitted. A vocal noise cue is calculated that enables comparing forward and bi-directional, as well as double and long-term predictive analyses on a subset of sentences of the "il est sorti avant le jour" corpus.
Full Paper Presentation (PDF) Presentation (Powerpoint)
Bibliographic reference. Bettens, Frédéric / Grenez, Francis / Schoentgen, Jean (2003): "Assessment of vocal noise via bi-directional long-term linear prediction of running speech", In VOQUAL'03, 69-72.