Digits-In-Noise Hearing Test Using Text-to-Speech and Automatic Speech Recognition: Proof-of-Concept Study

This proof-of-concept study evaluated the implementation of a digits-in-noise test we call the 'AI-powered test' that used text-to-speech (TTS) and automatic speech recognition (ASR). Two other digits-in-noise tests formed the baselines for comparison: the 'keyboard-based test' which used the same configurations as the AI-powered test, and the 'independent test', a third-party-sourced test not modified by us. The validity of the AI-powered test was evaluated by measuring its difference from the independent test and comparing it with the baseline, which was the difference between the Keyboard-based test and the Independent test. The reliability of the AI-powered test was measured by comparing the similarity of two runs of this test and the Independent test. The study involved 31 participants: 10 with hearing loss and 21 with normal-hearing. Achieved mean bias and limits-of-agreement showed that the agreement between the AI-powered test and the independent test (-1.3 ± 4.9 dB) was similar to the agreement between the keyboard-based test and the Independent test (-0.2 ± 4.4 dB), indicating that the addition of TTS and ASR did not have a negative impact. The AI-powered test had a reliability of -1.0 ± 5.7 dB, which was poorer than the baseline reliability (-0.4 ± 3.8 dB), but this was improved to -0.9 ± 3.8 dB when outliers were removed, showing that low-error ASR (as shown with the Whisper model) makes the test as reliable as independent tests. These findings suggest that a digits-in-noise test using synthetic stimuli and automatic speech recognition is a viable alternative to traditional tests and could have real-world applications.