Objective quality evaluation of noise-suppressed speech: effects of temporal envelope and fine-structure cues

While temporal envelope and fine-structure cues are known to be good predictors for speech intelligibility, it is not clear how well they are correlated with subjective quality ratings, particularly those using noise-suppressed speech. The present work evaluated the performance of two objective measures (i.e., NCM and TFSS), which were originally developed with primarily envelope or fine-structure cue as speech intelligibility indices, when they were applied for predicting the subjective quality ratings of noise-suppressed speech along three dimensions of signal distortion, noise distortion and overall quality. We considered a wide range of distortion introduced by four types of real-world noises at two signal-to-noise-ratio levels and by four classes of noise-suppression algorithms. This work finds that the present envelopeand fine-structure-based measures poorly predict the subjective quality ratings of noisesuppressed speech. The PESQ measure is so far the best choice in terms of objectively evaluating both subjective quality ratings and intelligibility scores of noise-suppressed speech.