FIELD: digital communication systems using frames. SUBSTANCE: each frame contains given number N of voice samples, thus having N degrees of freedom. Each voice frame is approximated by parametric model, which decreases number of degrees of freedom to value below N and helps in calculation of spectral power density 
for each voice frame. Spectral subtraction function 
depends on estimation of 
and estimation of background noise spectral power density 
in non-voice frames. Background noise estimation is calculated by non-parametric method for estimation of power spectrum. EFFECT: improved noise reduction without degradation in sound quality. 10 cl, 7 dwg, 3 tbl
