Jasper Ooster1,2,Melanie Krueger2,3, Jörg-Hendrik Bach2,3,4, Kirsten C. Wagener2,3,4, Birger Kollmeier2,3,4,5, Bernd T. Meyer1,2,3
1 Communication Acoustics, Carl von Ossietzky Universität, Oldenburg, Germany; 2 Cluster of Excellence Hearing4all, Germany; 3 HörTech gGmbH, Oldenburg, Germany; 4 Hörzentrum GmbH, Oldenburg, Germany; 5 Medizinische Physik, Carl von Ossietzky Universität, Oldenburg, Germany
Background: Speech audiometry is an important tool to quantify impaired hearing of listeners. Established tests often require substantial resources since an experimenter is required to evaluate the subjects’ responses. In addition, calibrated audiometric hardware is required. Smart speakers provide a speech-based interface and could increase the accessibility of tests but could also introduce errors in the response logging. In our study (Ooster, et. al. 2020), the speech recognition threshold (SRT) is measured with a matrix sentence test implementation for Amazon’s Alexa, which is compared to a clinical setup.
Methods: 46 German subjects from 4 different groups are measured in 3 different acoustic conditions to evaluate the influence on the SRT accuracy with a smart speaker. The subject groups range from young, normal-hearing (NH) to elderly, moderately hearing-impaired (HI) subjects. The acoustic conditions are generated in a room with variable, electro-acoustically controlled, reverberation times, ranging from low to high reverberation.
Results: The SRT results were mostly constant over the acoustic conditions. The bias between the SRT measured with the smart speaker and the clinically measured SRT varied with the subject groups from +0.7 dB for elderly, moderately HI listeners to +2.2 dB for young NH listeners. The intra-subject standard deviation was close (0.2 dB) to the test-retest accuracy of the clinical test. An ROC analysis showed that the approach produces a high accuracy (AUC=0.95) when making a binary decision (normal or elevated SRT).
Conclusions: We developed a skill for performing a speech-based listening test, which is currently available on all Alexa devices in German-speaking countries (activation with “Alexa, starte Hörtest”). Although it is not as accurate as the clinical test, an AUC of 0.95 is achieved for deciding if the SRT is normal or not. This test could complement clinical testing.