Speech enhancement for mobile telephony based on non-uniformly spaced frequency resolution

In this paper, we present a speech enhancement method based on spectral subtraction in non-uniformly spaced sub-bands. The main advantage of using different frequency resolutions for the various bands is the perception property of the human ear, which is able to separate low frequencies more precisely than high frequencies. The parameters of the noise reduction algorithm are chosen appropriately to the respective signal to noise ratio, which yields a performance superior to an uniform frequency resolution with the same number of sub-bands. A two-stage cascaded filter-bank is used for the decomposition of the signal.