Bandwidth Compression of Speech by Spectrum Sampling

Abstract

An experimental communications system has been built that filters speech into a number of narrow bands. The width of the bands, at the 3-db downpoints on the filter skirts, can be set at 25, 50, 100, 300, or 500 cps; the positions of the center frequencies of the filters are variable. A number of tests were conducted of the intelligibility of speech transmitted over a wide variety of combinations of filters having different bandwidths and different center frequencies. A few listening tests also were made of the effects on intelligibility of sequentially sampling in time the output of the various narrow band filters. The results of these tests show that intelligibility is a joint function of the number, width, and position on the frequency scale of the narrow passband filters. The most efficient combination of filters we have found to date, in terms of total effective bandwidth required for 80% PB word intelligibility, consists of five filters each 100 cps wide measured at the 3-db downpoints on the filter skirts. This represents a bandwidth saving over a low pass system (simulated telephone) by a factor of about 3 or 4 to 1; we previously reported [J. Acoust. Soc. Am. 32, 547–556 (1960)] a bandwidth reduction of 2 to 1 with three filters having bandwidths of 500 cps. (This work was performed under contract with the U. S. Army Signal Research and Development Agency.)