Lightweight test-score prediction methods for the Situated Phoneme (SiP) test based on the phoneme-level Speech Intelligibility Index (SIIP), the Phoneme Discriminability Level (PDL) and the Speech Intelligibility Index (SII).

OBJECTIVE

To introduce the Phoneme Discriminability Level (PDL) and compare its ability to predict scores on the Situated Phoneme (SiP) test with those of a phoneme-level version of the Speech Intelligibility Index (SIIP) and the original Speech Intelligibility Index (SII).

DESIGN

SII, SIIP, and PDL values were calculated for 24,444 SiP-test trials and used to train and evaluate logistic regression models. Models were trained on half of the study sample and statistically evaluated on the other half.

STUDY SAMPLE: Seventy-four adult native speakers of Swedish (mean age = 62 years), with hearing ranging from normal to severe loss. All had symmetric hearing without conductive components.

RESULTS

Models based on PDL and SIIP showed similar overall predictive power, and both outperformed the SII-based model. SIIP and SII were better at predicting average session scores, whereas PDL was better at predicting scores for individual test words. Adding random intercepts for test words improved the predictive accuracy of the SIIP and PDL models, but only at the level of individual words, not whole sessions.

CONCLUSIONS

All investigated models are useful for predicting SiP-test scores. SIIP and PDL appear suitable for optimising SiP-test settings and estimating critical differences.