Estimation of speech intelligibility using objective measures

In this paper, a method to estimate the subjective speech intelligibility scores of the Japanese DRT speech intelligibility test using some objective measures is proposed and evaluated. The objective measures tested were MOS calculated using the PESQ, SNRseg, fwSNRseg, and composite measures (Covl). These measures were mapped to its corresponding intelligibility scores using logistic functions. One function per phonetic feature was estimated, and these functions were used to map the objective values of another speaker of the same gender to predict its intelligibility. If the estimated intelligibility per phonetic feature is pooled, it is possible to estimate the intelligibility of the speech accurately. For speech mixed in white noise, the root mean square values (RMSE) between the subjective intelligibility and the estimated intelligibility were about 0.15, 0.14, 0.07, and 0.11 for MOS, SNRseg, fwSNRseg, and Covl, respectively. Other noise types showed similar values. The correlation between subjective and objective measures was over 0.91, 0.96, 0.98, and 0.96, respectively. The estimation accuracy was further investigated when the training and testing speaker gender or the noise types do not match. There was almost no decrease in accuracy with speaker gender, but a slight decrease with noise type. However, with fwSNRseg, the correlation between subjective and estimated intelligibility was mostly above 0.8, while other measures showed much lower correlation. This level of accuracy should justify the use of the proposed intelligibility estimation method, especially using fwSNRseg, to replace at least some of the expensive and time-consuming subjective intelligibility testing by screening out some test conditions.