Tolerance of FO adaptive time-frequency analysis for spectrographic representations

摘要

Spectrogram is a substrate for wide range of speech technologies. However, overlaid excitation information and response information make it difficult to extract and manipulate perceptually or physiologically relevant attributes directly. An FO adaptive procedure to remove locally periodic time-frequency structures from STFT (short term Fourier transform) based spectrographic representations enabled flexible manipulation of such attributes. The procedure is called TANDEM-STRAIGHT and is currently used in various applications. This article introduces tolerance of such representation to FO estimation errors and outlines an filternative design of TANDEM-STRAIGHT.