Improved objective measures for characterization of noise suppression algorithms

Abstract

We present two improved objective measures and one new measure for the characterization of the effect of noise suppression (NS) algorithms. These metrics are an evolution of earlier work [5]. The Signal-to-Noise Ratio Improvement ( S N N measure describes the capability of an NS method to enhance the speech component of a noisy speech signal from an additive background noise. The Noise Power Level Reduction (NPLR) measure indicates the NS impact on the noise level in the vicinity of speech. In this paper, we enbance the formulation of SNRI and NPLR by replacing frame powers by their logarithms to improve the robusmess of the memcs and to increase the association with subjective perception. A new thud measure is the Difleerence between SNRl and NPLR (DSN) that monitors the consistency of an NS solution in attenuating the background noise while causing a minimal effect on the speech level. The earlier metrics presented in [5] are included as characterization tools in recent 3GPPl [I] and TIA2 [6] minimum performance requirement specifications for NS algorithms in mobile terminals. The enhanced and new measures are proposed for a draft ITU-T recommendation G.VED, "Voice Enhancement Devices" [Z]. In this paper, we present the revised formulation of the measures as well as results of a verification study, including a regression analysis between the presented measures and listening test results. The analysis suggests that the objective measures are relevant for the perception of overall acceptability of the effect of an NS algorithm.

A Noise Reduction Preprocessor for Mobile Voice Communication

“…The evaluation of noise reduction algorithms using instrumental ("objective") measures is an ongoing research topic [24,25]. Frequently, quality improvements are evaluated in terms of (segmental) SNR and the achieved noise attenuation.…”

An iterative noise cross-PSD estimation for two-microphone speech enhancement

“…We used two other objective measures to assess the performance of the noise estimation algorithms: Noise power level reduction (NPLR) [33] and signal to noise ratio improvement (SNRI) [33]. The NPLR is a measure for assessing the capability of the noise reduction methods to reduce the background noise level.…”

“…The NPLR is a measure for assessing the capability of the noise reduction methods to reduce the background noise level. NPLR is computed during short speech pauses and is defined as [33]:…”

“…Frame classification for computing SNRI and NPLR can be performed using the clean speech signal, by comparing the frames powers to the active speech level [34]. The details of SNRI and NPLR calculation can be found in [33,35]. The SNRI and NPLR results are shown in Fig.…”

Noise power spectrum estimation using constrained variance spectral smoothing and minima tracking

“…Two objective measures are used to assess the performance of the noise estimation algorithms: noise power level reduction (NPLR) (Paajanen and Mattila, 2002) and PESQ (perceptual evaluation of speech quality) (ITU-T P.862, 2001). These measures are used to assess the performance when the estimated short-time power spectrum of noise is used in combination with the MMSE log-spectral amplitude estimator (LSAE) noise suppression rule (Ephraim and Malah, 1985).…”