Determination of the Speech Spectrum through Measurements of Superposed Samples

Abstract

By recording on one magnetic tape the simultaneous utterances of several speakers saying unrelated things, we have obtained noise which is essentially time-invariant, has a continuous spectrum, and displays the average spectral properties of speech much more readily than methods employed heretofore. Spectral analysis of this speech noise has been performed. The advantages of the method are (1) the integration time needed for determining the long-time spectrum of speech can be reduced to a few seconds; (2) the combination of speech wave forms can be treated as a statistical noise constant in time; (3) this noise does not contain any line spectrum elements; (4) with normal speaking levels, better signal-to-noise ratios are obtained. The experiments suggest that different languages may be characterized by differences in the long-time average spectra. Spectra obtained by the present method were used to compute the average loudness level of a single speaker. The method is adaptable to measurement of the spectrum of a speech sound or of an individual mode of pronunciation.