ADaptive Spectral Subtraction (ADSS) filter

Given digital adaptive filtering technology represents a modification of classic Wiener filter. It is based on the difference in temporal dynamics of variations of useful signal and noise.

Quality of speech enhancement

Noise type	Noisy speech (db,MOS)	Cleaned speech (db,MOS)
No	+inf , 4.60	30.2 , 4.42
Gaussian white noise	20.0 , 3.40	21.6 , 3.76
	10.0 , 1.73	15.1 , 3.00
	0.0 , 1.16	8.77 , 1.82
1KHz sinusoid	20.0 , 4.35	24.4 , 4.32
	10.0 , 4.07	19.8 , 4.24
	0.0 , 3.51	15.3 , 4.07

The speech quality was measured according to ITU-T P.861 recommendation.
All sample signals are digitized at 8 KHz sample rate and 16 bits per sample.
All noisy signals are performed by additive supplement of noise with different intensity and source signal.
Results here are for new version (2.0 beta) of filtering algorithm.

Hardware and software requirements

The following table contain computational requirements for filtering algorithm

Platform	CPU utilization (%)	RAM (Kb)	ROM (Kb)
Pentium III 800 MHz, Windows XP	2	48	13
MIPS VR4121 130 MHz, Windows CE 3.0	50	48	16

CPU utilization means the required part of computational resources of the system to invoke sound filtering in real time. The table contain values calculated with command line utility from ADSS SDK for 8 KHz source sample rate. These values have to be linear extrapolated for other sample rates.
RAM value is memory size required for variables and stack.
ROM value is memory size required for constants and program code.