Evaluation of the Oldenburg sentence test using speech recognition software

Introduction Language tests have a high clinical relevance for medical care. To realistically capture the speech understanding of hearing-impaired people, sentence tests are used, which are sometimes time-consuming. Automating the test using a speech recognition system (SRS) could result in considerable time savings and optimize resources.

Methods Included in the present prospective study were patients with the diagnosis of unilateral or bilateral profound inner ear hearing loss or surditas. As part of the study, the Oldenburg sentence test was carried out without noise and evaluated manually. At the same time, sounds were recorded, then evaluated with the SRS Dragon and the results compared with the manually evaluated language test.

Results The deviation of the evaluation by the SRS from the gold standard averaged 34 words or 22.7 % . This corresponds to a significant difference (p  < 0.0001). Subjects  <  60 years had a significantly lower error rate compared to the manual evaluation with a mean deviation of 24 words than subjects >/= 60 years with 41 words (p  < 0.05). No significant differences were found when differentiating between patients with unilateral or bilateral deafness and according to gender.

Conclusion A training phase with an SRS can significantly improve the quality and usability of the results. For follow-up studies, it would therefore be advantageous to activate the SRS during the test series, which are used as a training phase before the actually evaluated sentence list, and thus adapt the speech recognition system individually to the person to be tested.

Poster-PDF A-1412.pdf

Publication History

Article published online:
13 May 2021