Generation of nasalized speech sounds based on branched tube models obtained from separate mouth and nose outputs

A new approach is discussed for the analysis and generation of nasalized speech sounds considering separate mouth and nose outputs. The mouth and nose signals are separated by an insulating panel encircling the head of the speaker. The separated signals are analyzed by a branched tube model each. The model parameters are estimated from the mouth and nose signals by iterative inverse filtering. The resulting areas are used for artificial nasalization of speech signals of nonnasalized vowels. The nonnasalized vowel is analyzed by an unbranched tube model which is extended by the results from the previous analysis of mouth and nose signals. For artificial nasalization, the enlarged model consists of a superposition of two branched tube models representing the mouth and nose signals which are excited by residual signals. Depending on the degree of mixture of mouth and nose signals, the effect of nasalization is very perceivable.