“Computer, Test My Hearing”: Accurate Speech Audiometry with Smart Speakers

Abstract

Speech audiometry based on matrix sentence tests is an important diagnostic tool for hearing impairment and fitting of hearing aids. This paper introduces a self-conducted measurement for estimating the speech reception threshold (SRT) of a subject, i.e., the signal-to-noise ratio corresponding to 50% intelligibility, based on a smart speaker. While the original measurement procedure is well-evaluated and provides a very high measurement accuracy (< 1 dB test-retest standard deviation), the measurement using a smart speaker differs in several aspects from the commercially available implementation, such as missing control over the absolute presentation level, mode of presentation (headphones vs. loudspeaker), potential errors from the automated response logging, and influence from room acoustics. The SRT measurement accuracy is evaluated with six normal-hearing subjects conducted with an Amazon Alexa application on an Echo Plus loudspeaker in a quiet office environment. We found a significant difference of 0.6 dB in SRT between the proposed and the commercially available testing procedure. However, this bias is smaller than the inter-subject standard deviation, and the measurement accuracy is similar to the original test for normal-hearing listeners, which indicates that smart speakers may become a helpful addition for the screening of hearing deficits.

Voice-Based Conversational Agents for the Prevention and Management of Chronic and Mental Health Conditions: Systematic Literature Review

“…The majority (n=10) of the selected papers developed and evaluated VCA [ 59 , 60 , 62 - 69 ], whereas 2 [ 61 , 70 ] aimed to report a criterion-based performance evaluation of existing commercial conversational agents (eg, Google Assistant and Apple Siri). Among the papers developing and evaluating a VCA, 6 [ 59 , 60 , 62 , 65 - 67 ] assessed the technology acceptance of the VCA, whereas 3 [ 63 , 64 , 68 ] assessed the system accuracy. Only one [ 69 ] assessed both performance and acceptance.…”

“…All studies (n=12) were nonexperimental [ 59 - 70 ], that is, they did not include any experimental manipulation. A total of 4 papers [ 61 , 66 , 68 , 70 ] did not explicitly specify the study design they used, whereas the other papers provided labels. One study stated conducting a feasibility evaluation [ 63 ], 1 a focus group study [ 65 ], 1 a qualitative assessment of effectiveness and satisfaction [ 62 ], and 1 a case study [ 69 ].…”

Speech Audiometry at Home: Automated Listening Tests via Smart Speakers With Normal-Hearing and Hearing-Impaired Listeners

“…1 Due to the increasing availability of smart speakers, an accurate screening procedure for hearing deficits could potentially lower the threshold for conducting tests for a large number of users and therefore have a positive effect on early provision of assistive hearing devices. In a previous pilot study, the smart speaker-based measurement was evaluated in a single office room with six young, NH listeners, where a similar performance to clinical laboratory results was found ( Ooster et al., 2019 ). However, the reliability for HI subjects was not part of the previous study although this aspect is crucial for speech audiometry.…”

“…An overview of the elements of the smart speaker application for the automated SRT measurement is shown in Figure 1 (adopted from Ooster et al., 2019 ). The setup differs from established clinical setups in several ways: (a) It uses synthesized speech instead of the original speech test recordings (which are protected by copyright), (b) when used at home, the sound is presented via the speaker in a reverberant environment, (c) the presented audio files are stored with lossy audio formats, and (d) the listener’s response is transcribed via ASR and not logged by an audiometrist.…”

“…The authors would like to thank Inga Holube and Theresa Nüsse for providing the synthesized matrix test stimulus sentences as well as Timmy Kröger and Dirk Eike Hoffner for their help with the conduction of the evaluation measurements. Data from a pilot study (covering responses from 15 listeners compared with 46 listeners presented here) have been presented at the ISAAR symposium 2019 in Nyborg ( Ooster et al., 2020 ).…”