Objective measures for the characterization of the basic functioning of noise suppression algorithms

Four objective metrics are presented for the characterization of noise suppression (NS) algorithms. These metrics are evolved of the work in (Paajanen et al., [1]) and in (Paajanen and Mattila, [2]), and of the discussion concerning the development of the ITU-T G.160 recommendation. The signal-to-noise ratio improvement (SNRI) metric assess the capability of a NS method to enhance the speech component of a noisy speech signal corrupted by additive noise. The noise power level reduction (NPLR) measures the reduction of noise level achieved due to NS during speech activity. The difference between SNRI and NPLR (DSN) is computed to monitor the consistency of a NS solution in attenuating background noise while causing a minimal effect on the speech level. These three metrics characterize the functioning of a NS method during speech activity. Finally, the total noise level reduction (TNLR) metric provides an indication of overall noise reduction level experienced during long speech pauses. The earlier metrics presented in (Paajanen et al., [1]) are included as characterization tools in recent 3GPP∗ [3] and TIA† [4] minimum performance requirement specifications for NS algorithms in mobile terminals. The metrics are proposed for a draft ITU-T G.160 recommendation [5]. This paper presents the revised objective metrics and discusses their robustness. In addition, a verification study to characterize four NS solutions is presented as an example, making use of real speech and background noise signals as well as of artificial signals.