In this paper, we propose a new system which can reduce to a certain degree the noise component of voice signal. Experimental evaluation of its efficiency is also made. The principle of this system is based on the well known fact that the power spectrum and the autocorrelation form a Fourier transform pair with each other. An original noisy input signal is first processed with FFT, and then the square root of the resultant power spectra is extracted. Next the inverse Fourier transformation is made on it. Only a section of one pitch period of the output is then picked out as a useful part. Successive concatenation of these sections provides the desired output signal. It is easy to find that the final signal contains little noise component.
We have applied the present system to the detection of vocal sound elements such as /a/, /i/, /s/ etc., and confirmed successfully its validity through the listening test of normal conversation.
J-STAGEがリニューアルされました! https://www.jstage.jst.go.jp/browse/-char/ja/
