Spectral Extrapolation of Narrowband Speech

Abstract

It is known that human speech has components up to 7-8 kHz. However in telephony, speech is usually lowpass filtered at 3.4 kHz and sampled at 8 kHz. Although this filtering operation has little effect on intelligibility, it causes a perceptible degradation in speech quality. This paper deals with the reconstruction of the lost high frequency content of telephone speech using the low frequency spectral information. A reconstruction method based on a lookup table is described and preliminary results are presented.