An analytical method to convert between speech recognition thresholds and percentage-correct scores for speech-in-noise tests

Abstract

Speech-in-noise tests use fixed signal-to-noise ratio (SNR) procedures to measure the percentage of correctly recognized speech items at a fixed SNR or use adaptive procedures to measure the SNR corresponding to 50% correct (i.e., the speech recognition threshold, SRT). A direct comparison of these measures is not possible yet. The aim of the present study was to demonstrate that these measures can be converted when the speech-in-noise test meets specific criteria. Formulae to convert between SRT and percentage-correct were derived from basic concepts that underlie standard speech recognition models. Information about the audiogram is not being used in the proposed method. The method was validated by comparing the direct conversion by these formulae with the conversion using the more elaborate Speech Intelligibility Index model and a representative set of 60 audiograms (r¼0.993 and r¼0.994, respectively). Finally, the method was experimentally validated with the Afrikaans sentence-in-noise test (r¼0.866). The proposed formulae can be used when the speech-in-noise test uses steady-state masking noise that matches the spectrum of the speech. Because pure tone thresholds are not required for these calculations, the method is widely applicable.