Non-intrusive intelligibility prediction for Mandarin speech in noise

Abstract

Most existing intelligibility indices require access to the input (clean) reference signal to predict speech intelligibility in noise. In some real-world applications, however, only the noise-masked speech is available, rendering existing indices of little use. The present study assessed the performance of an intelligibility measure that could be used to predict non-intrusively (i.e., with no access to the clean input signal) speech intelligibility in noise using only information extracted from the noise-masked speech envelopes. The proposed intelligibility measure (denoted as ModA) was computed by integrating the area of the modulation spectrum (within 0.5 Hz to 10 Hz) of the noise-masked envelopes extracted in four acoustic bands. The ModA measure was evaluated with intelligibility scores obtained by normal-hearing listeners presented with Mandarin sentences corrupted by three types of maskers. High correlation (r=0.90) was obtained between ModA values and listener’s intelligibility scores, suggesting that the modulation-spectrum area could be potentially used as a simple but efficient predictor of speech intelligibility in noisy conditions.