Confidence Measures for Nonintrusive Estimation of Speech Clarity Index

We present a confidence measure based on prediction intervals for a non-intrusive method to estimate the clarity index (C-50) of reverberant speech. The method employed to estimate C-50 is a data driven approach. Multiple features are extracted from a reverberant speech signal and these are then used to train a bidirectional long short-term memory model that maps from the feature space into the target C-50 value. The prediction intervals are derived from the standard deviation of the per frame C-50 estimates and the confidence measures are obtained by normalizing these prediction intervals. These confidence measures are shown to provide a high correlation coefficient with the absolute C-50 estimation errors, i.e., 0.95 when applying conditional averaging. The performance of the prediction intervals and confidence measure are shown to be consistent among different noisy reverberant environments.