Investigation of test-retest reliability of a new close-to-everyday-life speech test using the Poisson binomial distribution

To enhance the ecological validity of speech intelligibility testing, this study introduces a novel speech test based on a conversation created with synthetic voices embedded in a 3D virtual cafeteria scene. 20 keywords from the conversation at constant signal-to-noise ratios (SNR) were used to assess intelligibility. The test was evaluated with 20 normal-hearing (NH) listeners. Test-retest accuracy was assessed and modeled based on simple and Poisson binomial statistics. Speech reception thresholds (SRT) were –9.7 dB ±2.5 dB SNR (average and standard deviation). After one familiarization run, no further significant training effect was observed. Due to variability in keyword intelligibility at fixed SNR, the “effective” keyword number was found to be n’=32, improving the test-retest reliability of the test compared to a test with 20 keywords of equal intelligibility. Confidence intervals were modelled and confirmed using Bootstrapping.