Improving the Quality of a Noisy Speech Signal

The idea that a vocoder may be used to improve the quality of a noisy speech signal, has been around for about twenty years. To the best of our knowledge the first such proposal was made in 1960 by M. R. Schroeder.1 The basic idea of this proposal can be explained with the help of Fig. 1, as follows: Figure la shows a typical short-term magnitude spectrum of a voiced portion of a noisy speech signal. Let S(u) denote the envelope of this spectrum. (Recall that the "channel gains" of a vocoder are estimates of this envelope at the center frequencies of the channels. The fine structure of the spectrum is attributed to the harmonics of the fundamental voice frequency.) Figure lb shows a "formant equalized" version, S (cu), of the envelope. The peaks in S and 5 occur at the same frequencies but the peaks of S (unlike those of S) are all of the same amplitude. 1847 Fig. 1--Illustration of noise stripping by increasing the dynamic range between formant peaks and noise valleys, (a) Original spectral envelope and fine structure, (b) Formant-level equalized spectral envelope, (c) The product spectrum S2(cj)S(u) in which the ratios between formant peaks and valleys is larger than in the original spectrum.