A bottom up procedure for extracting repetitive structures in speech sounds has been developed on the basis of a temporally stable representation of periodic sounds (TANDEM) and adaptive spectral smoothing (STRAIGHT). The proposed method evaluates local periodic structures in the frequency domain to detect repetition in the time domain. A group of dedicated periodicity detectors are combined to construct the proposed procedure for a repetitive structure extractor called an excitation structure extractor (XSX). The proposed procedure is tested using a set of stylized test signals with artificial shimmer and jitter to investigate the applicability of such aperiodic signals. The test results indicated that the proposed procedure outperformed in descriptive power of those complex excitation modes over existing FO detectors. Finally, the proposed procedure is applied to analyze pathological voice examples to investigate the feasibility of voice quality restoration applications.

Index Terms. periodicity extraction, fundamental frequency, TANDEM-STRAIGHT, XSX. apei iodicity, pathological voice