Automated digits-in-noise test using automatic speech recognition

In speech audiometry, the digits-in-noise (DIN) test is broadly used to assess speech perception in noise. In its most common setups: (1) onsite, clinicians run the test and score participants’ spoken responses, and (2) online, participants enter the digits they heard via a keypad. We propose a third setup, automatically scoring participants’ spoken responses via automatic speech recognition (ASR). The DIN test was conducted with native Dutch normal-hearing adults (Experiment 1; n = 30, 19-64 yr) and children (Experiment 2; n = 23, 5–17 years). Participants’ responses were decoded by the ASR Kaldi-NL, and automatically scored. Adults were tested via a computer. Children were tested via a humanoid NAO robot, We found an average of three triplets with decoding errors per participant in Experiment 1, and of eight triplets in Experiment 2. Experiment 1 results suggest our system provides reliable and comparable results to the current clinical setup. For children, it is not yet possible to evaluate the reliability of the system, as no intrasubject variability of the DIN test speech-reception-threshold (SRT) has been documented yet for normal-hearing children to use as a baseline.