Extending the articulation index to account for non-linear distortions introduced by noise-suppression algorithms

Abstract

The conventional articulation index (AI) measure cannot be applied in situations where non-linear operations are involved and additive noise is present. This is because the definitions of the target and masker signals become vague following non-linear processing, as both the target and masker signals are affected. The aim of the present work is to modify the basic form of the AI measure to account for non-linear processing. This was done using a new definition of the output or effective SNR obtained following non-linear processing. The proposed output SNR definition for a specific band was designed to handle cases where the non-linear processing affects predominantly the target signal rather than the masker signal. The proposed measure also takes into consideration the fact that the input SNR in a specific band cannot be improved following any form of non-linear processing. Overall, the proposed measure quantifies the proportion of input band SNR preserved or transmitted in each band after non-linear processing. High correlation (r = 0.9) was obtained with the proposed measure when evaluated with intelligibility scores obtained by normal-hearing listeners in 72 noisy conditions involving noise-suppressed speech corrupted in four different real-world maskers.

Figure 1

Signal-processing framework used in the present study for analyzing non-linear operations in the presence of noise. The dashed block shows the additional stage used in most noise-reduction applications to compute parameters such as band SNR, modulation rate, etc. These parameters are in turn used to construct a noise-suppressive gain function. The function f(.) represents generally the gain function used in noise-reduction or the non-linear function (e.g., compression function) used in hearing-aid applications.

Figure 2

(Color online) Histogram of band SNRs for corresponding bands in which $\stackrel{\wedge }{S}$ > S after noise- suppression. Band SNRs were determined for each time-frequency (T-F) unit, and accumulated over the duration of a sentence.

Figure 3

(Color online) Panel (a) shows the wideband spectrogram of the IEEE sentence “The young kid jumped the rusty gate.” in quiet, and panel (b) shows the sentence processed via a spectral- subtractive algorithm. The input sentence was originally corrupted by babble at 0 dB SNR. Panel (c) shows the corresponding short-term fAI values computed every 50 ms. The resulting average fAI value was 0.032.

Figure 4

Scatter plot of speech intelligibility scores and predicted fAI values for 72 noisy conditions involving noise-suppressed speech in four different masker conditions (babble, car, train and street interferences) and two SNR levels.

Figure 5

Scatter plot of observed intelligibility scores (expressed in percentage) and predicted scores for the 72 noisy conditions tested.