INTERSPEECH 2004 - ICSLP
8th International Conference on Spoken Language Processing

Jeju Island, Korea
October 4-8, 2004

Modified Realizable Frequency Warped ARMA Modeling and its Application in Synthesis Structures for Voiced Speech

Juan L. Navarro-Mesa, Pedro J. Quintana-Morales

Universidad de Las Palmas de Gran Canaria, Spain

Synthesis filters are of major concern in many speech applications. In this paper these filters are designed so as to track the natural variations, parameterize the perceptually relevant aspects of the pole and zero characteristics and their dynamics, and achieve reliable estimates of the filter coefficients. In this paper this is attained integrating three basic points of view in a general framework. First, frequency warping incorporates the perceptually relevant characteristics of human hearing. Second, direct implementations of efficient synthesis structures of these filters are realized with modified versions that overcome the problems associated to the incorporation of first-order all-pass sections for frequency warping. And third, reliable coefficients associated to the synthesis structures of several consecutive periods are directly estimated in the analysis. We show that the synthesis structures we apply and their coefficients offer good signal to reconstruction error ratios and error distribution in frequency.

Full Paper

Bibliographic reference.  Navarro-Mesa, Juan L. / Quintana-Morales, Pedro J. (2004): "Modified realizable frequency warped ARMA modeling and its application in synthesis structures for voiced speech", In INTERSPEECH-2004, 2461-2464.