A Speech Modification Method by Signal Reconstruction Using Short-Term Fourier Transform

The short-term Fourier transform analysis-synthesis technique is promising as a means to obtain high-quality synthetic speech, since the original speech can be reconstructed from the parameters of analysis. This paper proposes a speech modification method using this synthesis system, especially the pitch frequency modification. Among various methods of the short-term Fourier spectrum synthesis, the proposed method employs the minimization of squared error between the spectrum of the reconstructed speech and that of the target speech. The proposed method features the following two points: [1] In the pitch frequency modification, only the parameters exclusively related to the pitch frequency are modified; [2] the phase spectrum is not manipulated directly but is estimated by the speech reconstruction algorithm. As a result of evaluation of the proposed method by a hearing test, it was verified that a high-quality modified speech is obtained for both the uniform pitch modification over the whole word and the nonuniform pitch modification to change the accent position of the word.