Automated Question-and-Answer-Based Speech-In-Noise Hearing Test Using a Large Language Model

A conversational speech-in-noise test administered without human supervision is desirable to reduce the need to attend the clinic. This study aimed to create and evaluate a novel question-and-answer-based speech-in-noise test called 'CHAT', where stimuli were synthesised using text-to-speech (TTS), responses captured using automatic speech recognition (ASR), and responses evaluated using a large language model (LLM). Three tests were compared: the newly developed CHAT, the Adaptive Sentence List (ASL) and the Four Alternative Auditory Feature Test (FAAF). CHAT used synthetic speech while ASL and FAAF by default used human-recorded speech. Speech reception thresholds (SRT) of 20 participants with normal hearing and 20 with mild to moderate hearing loss were measured using these three tests. Validity was evaluated using receiver operating characteristic (ROC) analysis and correlations between SRTs and average pure-tone audiometry (PTA), across the three speech tests. Reliability was evaluated by comparing limits-of-agreement between SRTs achieved by repeated runs of CHAT and ASL tests using Bland-Altman analysis. Results revealed limits-of-agreement between repeated runs of CHAT and ASL were ±3.6 dB and ±3.1 dB, respectively. The area under the ROC curve was 0.97 for CHAT, 0.95 for ASL and 0.89 for FAAF. The correlation between SRT and PTA results was 0.86 for CHAT, 0.82 for ASL and 0.79 for FAAF. These findings indicate that CHAT demonstrates reliability and validity comparable to human-recorded ASL and FAAF, supporting the use of synthetic speech for speech-in-noise testing. Overall, the results suggest that AI tools can help create and administer conversational speech-in-noise tests.