Improving the prediction power of the speech transmission index to account for non-linear distortions introduced by noise-reduction algorithms

Although the speech transmission index (STI) has been shown to predict successfully the effects of linear distortions introduced by filtering and additive noise, it does not account for non-linear distortions present in noise-suppressed speech. In this study, the normalized covariance metric (NCM), a STIbased intelligibility measure, was modified to reduce the effects of non-linear distortions introduced by most noise-suppression algorithms for intelligibility prediction. This was done by designing a new definition of the output signal-to-noise ratio to compensate the biased estimation of the input SNR prior to the noise-suppression processing. The modified NCM measure was evaluated with intelligibility scores obtained by normal-hearing listeners in 72 noisy conditions involving noise-suppressed speech corrupted by four different maskers (babble, car, train and street interferences). Significantly higher correlation with intelligibility score was obtained from the modified NCM measure, in contrast to those from the original NCM measure.