ISCA Archive Eurospeech 1995
ISCA Archive Eurospeech 1995

AR identification of the vocal filter from noisy hyperbaric speech signals

Laure Charonnat, Joel Crestel, Michel Glutton, Herve Chuberre

When dealing with hyperbaric speech, the enhancement of the intelligibility is an analysis-synthesis problem in which the deconvolution between the excitation and the vocal filter is known to be crucial. However, the hyperbaric voiced speech signal is characterized by specific properties which are proved to be of interest in the AR identification of the vocal filter especially in case of noisy signals. As it is, the work which is presented brings out two major results. It is shown that, within a pitch period of a noisy speech signal (white noise), a significant segment can be used to compute an/estimate of the noise variance. Then this estimate can be taken into account in the vocal filter AR identification procedure, tnus giving better accuracy.