Mapping ASR systems onto the human auditory spectrum: speech-in-noise recognition of Arabic in normal-sighted, blind, and hearing-impaired listeners

Background/objectives This study examines how human listeners and automatic speech recognition (ASR) systems process speech under varying background noise levels. Methods Three groups of human participants—normal-sighted (NSLs), blind (BLs), and hearing-impaired listeners (HILs)—were compared with two state-of-the-art ASR systems, Whisper and TurboScribe. Forty Arabic sentences (20 imperative, 20 interrogative) were presented at four signal-to-noise ratios (SNRs: +10, +5, −5, −10 dB) using passenger boarding noise to simulate realistic acoustic environments. Results Accuracy declined with increasing noise for all groups, although rates varied substantially. BLs performed best, maintaining 85% accuracy even at −10 dB SNR, whereas HILs showed the steepest decline. The ASR systems achieved near-human performance in moderate noise but degraded sharply under severe masking, mirroring HIL trends. Statistical analysis confirmed significant group differences across SNRs ( p < 0.05). A linear mixed-effects analysis further revealed a significant advantage for interrogative over imperative sentences in human listeners ( p < 0.05), plausibly attributable to the rich prosodic cues of question intonation, but no comparable benefit was observed for ASR systems ( p = 0.30), which appear to weight segmental over prosodic information. Conclusion These findings highlight that, although contemporary ASR models approach human performance in favorable conditions, they lack the adaptive mechanisms that allow humans—especially BLs—to sustain intelligibility in complex acoustic scenes. By situating ASR performance within the spectrum of human auditory variability, this research underscores both the progress and the limitations of machine hearing and points toward neurobiologically inspired strategies for achieving noise-robust speech recognition. Clinically, this framework suggests that benchmarking ASR systems against the full spectrum of human auditory variability—including hearing-impaired and blind listeners—can inform the design of accessibility tools, hearing-aid signal processing, and supplementary diagnostic measures of speech-in-noise perception in audiology practice.