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

1 Introduction

Automatic speech recognition (ASR) has become a routine component of virtual assistants, live captioning, and accessibility tools, and its accuracy under realistic acoustic conditions, rather than in the clean recordings on which most benchmarks are still built, increasingly determines whether these systems serve users equitably (Hinton et al., 2012; Ozarkar et al., 2020; Gong et al., 2023). Natural listening environments rarely offer clean speech: multiple talkers, environmental noise, and reverberation combine and fluctuate over time (Bregman, 1990; Dua and Dua, 2023), degrading machine transcription accuracy as measured by word error rate (WER; Kumalija and Nakamoto, 2022; Won, 2025). Recent models such as Whisper have narrowed the historical gap with human listeners for clean speech, but a gap persists in noisy or unpredictable conditions (Gong et al., 2023; Patman and Chodroff, 2024), raising the question of whether artificial systems process degraded speech by strategies comparable to those used by human listeners under the same constraints.

Direct human–ASR comparisons help specify where this gap remains. Kim et al. (2024) found that among four end-to-end systems tested against human listeners on second-language English in speech-shaped noise, only Whisper-large matched or exceeded human accuracy, and only at favorable SNRs; Patman and Chodroff (2024) reported that Whisper-large-v3 outperformed native listeners in most noise conditions but merely matched them in naturalistic pub noise; and Goswami (2026) showed that reverberation alone degrades Whisper’s WER by 2–16 percentage points depending on model size. Together with industry benchmarks placing Whisper among the more noise-resilient systems currently deployed (Achorn et al., 2025), these results suggest that the human advantage has narrowed in stationary noise but remains substantial in naturalistic and reverberant conditions. Most of this evidence, however, comes from English. Arabic is comparatively underresourced: its rich morphology, nonstandardized orthography, and dialectal variation degrade ASR accuracy even before noise is introduced (Besdouri et al., 2024; Salhab and Abusheikh, 2025), Modern Standard Arabic accounts for only a small share of training data for models such as Whisper, and regional dialects such as Najdi Arabic are represented more sparsely still (Talafha et al., 2025). Reported WERs for Arabic already exceed 40% in clean conditions for some transformer models (Gerazov et al., 2025) and decline further once realistic noise is added (Ezzine et al., 2025). Any human–ASR comparison conducted on Arabic stimuli should therefore be expected to expose larger machine deficits than an equivalent English comparison, a consideration that motivates the present design and constrains the interpretation of our findings.

Human listeners are not uniform in their robustness to noise, and this variability is itself informative. Hearing-impaired listeners (HILs) show well-documented difficulty separating speech from competing sound, reflecting reduced audibility together with deficits in temporal and spectral resolution that amplification alone does not resolve (Lorenzi et al., 2006; Healy et al., 2013; Nabelek et al., 2004); accuracy under noise, moreover, need not reflect equal effort, since the Framework for Understanding Effortful Listening (FUEL; Pichora-Fuller et al., 2016) predicts that two listeners can reach the same accuracy at different cognitive cost, a distinction ASR output, which returns only a transcript, cannot capture (Peelle, 2018). Blind listeners (BLs), by contrast, often show enhanced speech-in-noise performance (Bae et al., 2022; Pang et al., 2020; Hertrich et al., 2009), consistent with evidence that early blindness recruits occipital cortex for auditory processing (Lewis et al., 2010; Guerreiro et al., 2015; Sabourin et al., 2022). This asymmetry across listener groups suggests that placing ASR systems along a spectrum of human auditory variability, rather than against a single idealized listener, may be more informative than accuracy alone.

The present study asks whether contemporary ASR systems process Arabic speech in noise in ways that more closely resemble typical, impaired, or enhanced human listening. We compare normal-sighted, hearing-impaired, and blind listeners with two ASR systems (Whisper and TurboScribe) on Arabic sentences presented at four SNRs, holding linguistic content constant across groups so that differences in recognition accuracy can be attributed to how each listener, human or machine, handles acoustic degradation. Beyond its bearing on ASR evaluation, this comparison speaks to an applied concern: as ASR is increasingly embedded in captioning, hearing-aid processing, and telehealth transcription, its behavior at adverse SNRs has direct consequences for hearing-impaired and visually impaired users (Healy et al., 2013; Aguiar et al., 2025), and speech-in-noise performance is itself a more ecologically valid index of communication ability than pure-tone audiometry alone (Wilson, 2011; Jain et al., 2025).

2 Materials and methods

3 Results

Table 1 summarizes the mean speech recognition error rates for the three groups of human listeners (BLs, NSLs, and HILs) alongside the two ASR systems, Whisper and TurboScribe, across four SNR levels: +10 dB, +5 dB, −5 dB, and −10 dB.

The declines in the mean speech recognition accuracy across SNR levels are illustrated in Figure 2. Recognition accuracy decreased consistently as noise increased, yet the slope of decline differed markedly among the listener groups. BLs maintained near-ceiling performance at mild and moderate SNR levels and showed only a gradual drop under more adverse conditions. Even at −10 dB SNR, BLs achieved an average accuracy of 85%, suggesting exceptional resilience to background interference. NSLs exhibited strong performance at higher SNRs (95% at +10 dB) but showed a clear inflection point around +5 dB SNR, below which recognition degraded more rapidly. By −10 dB SNR, accuracy for NSLs declined to 60%, marking a 35-percentage-point drop across the range—nearly quadruple the drop observed in BLs. HILs performed the poorest at all noise levels. Their accuracy decreased sharply from 80% at +10 dB SNR to only 25% at −10 dB SNR, reflecting a severe deficit in speech-in-noise perception even under moderate masking conditions. Among artificial systems, Whisper outperformed TurboScribe across all SNRs but with similar overall trends. Both systems showed robust recognition in favorable conditions (above 90% at +10 dB SNR), comparable to NSLs. However, their decline between +5 dB SNR and −10 dB SNR was notably sharper than that of BLs or even some NSLs, indicating reduced noise generalization capability. Whisper maintained 55% accuracy at −10 dB SNR, whereas TurboScribe dropped to 50%.

A linear mixed-effects model was conducted to examine recognition accuracy across listener groups and noise levels. The analysis revealed a statistically significant main effect of noise level (p < 0.05), indicating that recognition accuracy varied systematically as the listening condition became more adverse. Pairwise comparisons further showed that BLs performed significantly better than HILs (p < 0.05). In addition, NSLs differed significantly from both BLs and HILs (p < 0.05), suggesting that the NSL group occupied an intermediate position between the two human listener groups. Because the three human groups differed in sex composition, sex was entered as an additional fixed-effect covariate in this model; it did not significantly predict recognition accuracy (p > 0.05), indicating that the observed group differences are not attributable to sex composition. No statistically significant difference was found between Whisper and TurboScribe (p = 0.071), nor did ASR systems differ significantly from NSLs (p > 0.05), indicating that the performance of Whisper and TurboScribe was statistically comparable to that of NSLs across the tested noise conditions. Figure 3 illustrates this differentiation between humans and ASR systems.

Table 2 disaggregates the recognition data by sentence type to facilitate direct comparison between the imperative and interrogative conditions for each listener group and ASR system across the four SNRs. The results for each sentence type reflect similar overall ordering of groups to that observed in Table 1: BLs achieved the lowest WER at every SNR, followed by NSLs and the two ASR systems clustered closely together, with HILs demonstrating the highest error rates throughout. The monotonic increase in WER as the SNR diminished from +10 to −10 dB was also replicated within each subset, confirming that the noise-level effect is not driven by one stimulus type. More substantively, interrogative sentences were recognized more accurately than imperatives by every human group, and this advantage increased as listening conditions deteriorated. For BLs, WER at −10 dB was 11% for imperatives but 4% for interrogatives; for NSLs, 32% (for imperatives) vs. 18% (for interrogatives); and for HILs, 49% (for imperatives) vs. 26% (for interrogatives). The HIL group thus demonstrated the largest stimulus-type effect in absolute terms, with interrogatives reflecting nearly half the error rate of imperatives at −10 dB.

The ASR systems demonstrated a more uneven profile. At favorable and intermediate SNRs (+10, +5, −5 dB), Whisper and TurboScribe transcribed imperatives marginally more accurately than interrogatives, suggesting that short, fixed-syntax imperative templates are well represented in their training distribution. At −10 dB, however, this advantage reversed sharply: Whisper fell to 25% WER for imperatives vs. 20% for interrogatives, and TurboScribe to 30% vs. 20%, respectively. This shift indicates that under severe masking, the ASR models also benefited from the additional acoustic information present in interrogatives—most likely, the more salient pitch contour and longer voiced segments—although the human-side gain was consistently greater than that for machines, particularly for HILs. The results in Table 2 indicate that sentence type interacts with both noise level and listener type: Humans extract more information from the interrogative structure than ASR systems do, and this differential widens precisely in the conditions where ASR robustness is otherwise weakest.

Figure 4 visualizes the patterns reported in Table 2, displaying recognition accuracy (100 − WER) for the three human groups in the upper row and the two ASR systems in the lower row, with imperative and interrogative conditions plotted in adjacent columns. Two effects are immediately apparent. Among human listeners, every group performed better on interrogatives than on imperatives, and this gap widened systematically as the SNR worsened; this effect was most pronounced for HILs, where the difference in accuracy between imperative and interrogative sentences at −10 dB exceeded 20 percentage points. The two ASR systems, by contrast, showed much smaller and inconsistent stimulus-type effects, with imperative and interrogative trajectories overlapping closely at every SNR.

A linear mixed-effects model with sentence type (imperative vs. interrogative), listener type, and SNR as fixed effects and participant or system as a random effect was fitted separately to the human and ASR data to formally test these patterns. For human listeners, the main effect of sentence type was statistically significant (p < 0.05), confirming that interrogatives were recognized more accurately than imperatives across BLs, NSLs, and HILs. By contrast, the same contrast tested on the two ASR systems was not statistically significant (p = 0.30), indicating that Whisper and TurboScribe transcribed the two sentence types with comparable accuracy. The perceptual benefit conferred by interrogative structure was reliably present in human listeners but absent in the ASR systems tested.

In Figure 5, the WER values are decomposed into the three constituent error types—deletions, substitutions, and insertions—averaged across all SNR levels and stimulus types and are presented separately for each listener group and ASR system.

Two patterns are immediately apparent. First, the data reflect a similar relative ordering of groups by total error rate to rankings established in Tables 1, 2: BLs reported the lowest error counts in every category, followed by NSLs, with HILs producing markedly more errors of each type. Second, and more substantively, the qualitative profile of errors was opposite for human listeners and ASR systems. Across all three human groups, deletions were the most frequent error type (BLs = 11%, NSLs = 19%, HILs = 38%), substitutions were intermediate, and insertions were the least frequent. The two ASR systems demonstrated an inverse pattern: deletions were the least frequent error type (Whisper = 7%, TurboScribe = 8%), with substitutions and insertions occurring at roughly comparable higher rates (16–19%). A linear mixed-effects model with error type, listener type (human vs. ASR), and their interaction as fixed effects and participant or system as a random effect confirmed a significant main effect of error type, a significant main effect of listener type, and a significant error-type × listener-type interaction (all p < 0.05). Pairwise comparisons within the human data showed significant differences between deletions and substitutions and between deletions and insertions (both p < 0.05) but not between substitutions and insertions. The same pattern of pairwise contrasts was observed for the ASR systems: Deletions differed significantly from both substitutions and insertions, whereas substitutions and insertions did not differ. However, the direction of the deletion contrast was reversed: Deletions were the most frequent error type for humans but the least frequent for ASR systems. The interaction therefore captures a true qualitative dissociation in error profile rather than a simple difference in overall accuracy.

The WERs for all listener groups and ASR systems are presented, broken out by error type, in Table 3.

Figure 1 unpacks the deletion column of Table 3 by linguistic category, indicating which classes of function words were most often omitted by NSL listeners and by the ASR systems. Each English category label is sized in proportion to how often that category was deleted by the corresponding group, and the Arabic exemplars beneath each label illustrate the kinds of words involved (e.g., quantifiers and conjunctions). Two patterns stand out. NSL deletions were heavily concentrated in two categories—quantifiers and conjunctions—accounting collectively for approximately three-quarters of all NSL omissions; the remaining three categories contributed only marginally. ASR deletions were distributed almost evenly across all five categories, with no category dominating. The figure therefore makes visible a property the aggregated counts in Table 3 only imply: Human deletions are selective and follow linguistic structure, whereas ASR deletions are indiscriminate and appear to be driven primarily by acoustic rather than syntactic factors.

4 Discussion

The results of this study clarify how current ASR systems approximate—but do not yet replicate—human auditory capabilities, particularly under adverse acoustic conditions. The comparative analysis of BLs, NSLs, HILs, and two ASR systems (Whisper and TurboScribe) across four SNRs provides empirical evidence of both human superiority in noise robustness and the distinct computational strategies employed by artificial versus biological auditory systems. Across all SNR levels, performance decreased consistently as noise increased, yet human listeners—especially blind participants—demonstrated significantly greater resilience than the ASR systems. These findings mirror observations from previous research showing that human listeners rely on contextual, linguistic, and cognitive redundancy to recover meaning from degraded acoustic signals (Mattys et al., 2012; Clonan et al., 2025). Whisper’s accuracy, which matched NSLs in mild and moderate noise conditions, underscores the strength of large-scale models trained via weak supervision (OpenAI, 2022; Radford et al., 2023). However, its rapid decline at −5 dB and −10 dB SNR suggests that noise robustness in machine speech recognition remains far from human adaptability. Even the most sophisticated ASR systems display considerable degradation in real-world noisy Arabic speech, confirming that generalization under adverse acoustic contexts remains a key challenge.

The superior performance of BLs under all noise conditions highlights the remarkable adaptability of the human brain. Consistent with neuroimaging and behavioral findings, early blindness leads to cross-modal reorganization in which visual cortical regions are recruited for auditory processing (Lewis et al., 2010; Guerreiro et al., 2015; Sabourin et al., 2022). We did not collect neuroimaging or electrophysiological data in the present study, so this mechanism is inferred from convergence with prior literature rather than demonstrated directly in our sample. Our BL sample is consistent with this developmental window: nine participants were congenitally blind and the remaining participant lost vision in early childhood, and none acquired blindness in adulthood; blindness history was based on participant and relative report, corroborated for the university-educated participants by disability-accommodation records that were not directly examined by the experimenters, and residual vision was not formally assessed, so a contribution of unassessed partial vision cannot be entirely excluded. This neural plasticity supports enhanced temporal precision and spectral resolution, allowing for more efficient segregation of speech from background noise (Hertrich et al., 2009; Bae et al., 2022). Recent electrophysiological evidence aligns with this account: Congenitally blind adults show shorter MMN and N2b latencies and larger P3a amplitudes during speech-sound discrimination, indicating faster auditory change detection and greater attentional engagement than sighted controls (Turgeon et al., 2025). Crucially, King (2015) emphasized that cross-modal plasticity following blindness reflects a balance of compensatory gains and developmental trade-offs rather than uniform enhancement, helping explain why BL robustness in this study was substantial yet not absolute. The sustained high performance of blind participants, even under extreme masking, is consistent with the possibility that cross-modal adaptation endows unique resilience, and may offer a useful source of inspiration for designing more adaptive acoustic models, though this behavioral parallel does not itself establish the underlying neural mechanism or demonstrate that it can be transferred to machine architectures.

By contrast, HILs exhibited pronounced difficulty across all SNR levels. Their reduced intelligibility, even at favorable listening conditions, is consistent with longstanding evidence that audibility restoration alone often does not fully compensate for compromised temporal fine structure or spectrotemporal resolution in sensorineural hearing loss (Lorenzi et al., 2006; Healy et al., 2013); because hearing status in our sample was self-reported rather than audiometrically confirmed, we cannot establish that this specific mechanism accounts for the HIL group’s performance here. A further, more parsimonious contributor is that HIL participants were tested without their hearing aids: some of the pronounced difficulty observed here may therefore reflect reduced audibility in the unaided condition rather than, or in addition to, a central deficit in processing degraded speech. Recent reviews of cross-modal plasticity in mild-to-moderate partial hearing loss have further indicated that adaptive neural changes do not consistently translate into perceptual gains in this population, in contrast to the clearer enhancements observed after total sensory loss (Aguiar et al., 2025). If our HIL group’s impairment falls within this range—which our self-reported classification cannot confirm—this distinction may help explain why these participants failed to recover in noise despite presumably preserved cognitive resources. Convergent evidence comes from a recent large-scale audiology study by Jain et al. (2025), who graded 100 listeners across five hearing-loss severities and showed that word-based recognition scores fell to floor at adverse SNRs precisely in the moderately severe and severe groups, whereas finer-grained phoneme-based scoring still discriminated severity at 98% accuracy under those same conditions. The collapse pattern they documented is broadly analogous to the floor we observed at −10 dB SNR in our HIL group, though without audiometric grading we cannot place our sample precisely within their severity bands; the parallel nonetheless suggests that the apparent saturation of HIL deficit at the most adverse SNR may partly reflect the granularity of WER as a measure rather than the absence of further auditory differentiation. The performance of HILs parallelled the steepest degradation observed in the ASR models, emphasizing both the technological proximity of ASR systems to impaired human processing and the extent of improvement still needed to achieve typical human robustness. This comparison reinforces the value of model–human benchmarking for identifying limitations in both machine-learning algorithms and human-assistive devices.

A complementary pattern, equally instructive for the human–machine comparison, emerges from the sentence-type contrast. The mixed-effects analysis confirmed a significant advantage for interrogative over imperative sentences in human listeners (p < 0.05) but no comparable effect for the ASR systems (p = 0.30). This dissociation aligns with a well-established finding that the interrogative–declarative contrast is acoustically marked in Arabic, as in many languages, by a prominent rising pitch contour and lengthening over the final portion of the utterance (Peng et al., 2009; Shen, 2021) and that humans rely heavily on these prosodic cues to support intelligibility when the segmental signal is degraded by noise (Shen, 2021). The fact that BLs, NSLs, and HILs all benefited from interrogative prosody—and that HILs benefited most in absolute terms—suggests that prosodic top-down support is preserved even when peripheral auditory resolution is compromised and may serve as a partial compensatory route. End-to-end ASR systems, by contrast, are optimized primarily on segmental and lexical accuracy and weight prosodic structure only weakly during decoding (Tran et al., 2018; Wagner et al., 2015), which plausibly explains the absence of an interrogative advantage for Whisper and TurboScribe. The applied implication is that the human–ASR gap observed at low SNR will be larger for short, prosodically flat utterances such as imperatives than for utterances with rich intonational structure, an asymmetry worth factoring into the design of voice interfaces deployed in noisy real-world settings.

A third dissociation between humans and ASR systems is indicated by the decomposition of WER into its constituent error types (Figure 5). Human listeners, including BLs, produced predominantly deletions when the signal was degraded—omitting words they could not confidently identify rather than guessing. Substitutions and insertions were comparatively rare. The two ASR systems showed an inverse pattern: Deletions were the least frequent error type, with most errors taking the form of substitutions or insertions of phonetically or lexically plausible material that was not actually present in the signal. This contrast is consistent with documented properties of end-to-end ASR systems trained with weak supervision: Such models are driven to produce fluent, locally plausible output and exhibit a known tendency to hallucinate or paraphrase rather than to report uncertainty (Russell et al., 2024; Koenecke et al., 2024). Whisper in particular has been shown to insert content not present in the audio under low-SNR or silent conditions (Koenecke et al., 2024). The human strategy—omit rather than fabricate—reflects metacognitive awareness of perceptual uncertainty that current end-to-end ASR architectures lack. From an applied standpoint, the two error profiles have very different downstream costs: A deletion in an air-cabin announcement is recoverable through repetition or context, whereas a confident substitution may produce a fluent but incorrect transcription that the listener has no way to flag. Reducing ASR insertion and substitution errors—ideally by introducing calibrated uncertainty estimates that allow systems to abstain when the signal is too degraded—would therefore be a higher-priority target for safety-critical deployments than further reductions in overall WER alone.

The linguistic-category breakdown shown in Figure 1 sharpens this interpretation. NSL deletions were not distributed uniformly across function-word classes but were concentrated in two categories—quantifiers and conjunctions—which together accounted for approximately three-quarters of all NSL omissions, whereas articles, determiners, prepositions, and pronouns were retained at much higher rates. This selective omission profile is consistent with information-theoretic accounts of speech production and comprehension, which predict that listeners under processing load preferentially drop high-probability, low-information items that can be reconstructed from context (Levy, 2008; Mahowald et al., 2013). The pattern is also consistent with classic findings on shadowing and cued recall under noise, in which content words are preserved at the expense of grammatical morphemes (Marslen-Wilson and Welsh, 1978). The two ASR systems showed a markedly different profile: Deletions were spread evenly across all five categories at counts between 3 and 6, with no category dominating. This flat distribution suggests that ASR omissions are governed primarily by acoustic salience—weakly stressed or briefly articulated tokens are dropped regardless of their syntactic role—rather than by any learned sensitivity to linguistic predictability. The contrast in error structure therefore points to a deeper architectural divergence: Human listeners deploy linguistic prior knowledge to allocate effort selectively, whereas end-to-end ASR systems treat all tokens as equally informative until the acoustic signal forces a decision. Closing this gap may require ASR architectures that explicitly model word-level uncertainty and predictability, perhaps by integrating language-model surprisal signals into the decoding process so that confidence can vary with linguistic context rather than acoustic energy alone.

An accuracy-only reading of the HIL pattern, however, risks missing a layer of the phenomenon. Within the FUEL framework, the same percent-correct score can be achieved at very different cognitive costs, and HILs typically pay a higher cost even when their intelligibility scores match those of normal-hearing peers (Pichora-Fuller et al., 2016; Peelle, 2018). Recent work using EEG alpha-band oscillations and pupillometry has shown that elevated effort persists even at SNRs where accuracy appears intact and that effort increases sharply once the SNR drops below the individual speech-reception threshold (Wendt et al., 2018; Simeone et al., 2025). The pronounced HIL deficit observed here at −5 and −10 dB SNR is therefore plausibly an underestimation of the perceptual burden for these listeners, since accuracy had already collapsed before effort could be expended productively. In contrast, ASR systems have no analog of cognitive effort but also no graceful-degradation strategy when the signal becomes uninformative. The convergence between HIL and ASR error trajectories at low SNR may thus reflect a shared inability to recruit additional resources rather than a shared underlying mechanism, an interpretation consistent with the divergence between human top-down inference and the largely bottom-up statistical regularities exploited by current end-to-end models (Clonan et al., 2025; Hicks and McDermott, 2024).

A further dimension of these results concerns the language tested. The Whisper and TurboScribe declines observed at −5 and −10 dB SNR are steeper than several published English-stimulus benchmarks would predict (Patman and Chodroff, 2024; Kim et al., 2024), and this gap is consistent with the broader picture of Arabic ASR. Arabic is severely underrepresented in the training corpora of major end-to-end systems, and even on clean Arabic data, state-of-the-art models report markedly higher WERs than on comparable English material (Besdouri et al., 2024; Salhab and Abusheikh, 2025). Accuracy degrades further with ambient noise; for example, dialect-specific work in Moroccan Arabic under traffic noise has reported sharp declines tracking the SNR (Ezzine et al., 2025). Two implications follow. First, the apparent NSL–ASR equivalence at +10 and +5 dB in the present data should not be overgeneralized; it most likely reflects a ceiling on the relatively short, predictable imperative and interrogative sentences used here and would probably not extend to spontaneous or dialectal Arabic. Convergent cross-language evidence comes from Russell et al. (2024), who reported increases in Whisper’s WER from 2–3% on read audiobook English to approximately 19% on Switchboard telephone conversations and 27% on the RoomReader videoconferencing corpus, confirming that any benchmark obtained on short, well-formed read sentences sets an upper bound on what current ASR systems can achieve in conversational deployment. Second, the pronounced ASR collapse at low SNR cannot be attributed solely to acoustic masking; it is compounded by data-scarcity effects specific to Arabic. The implication for applied settings such as in-flight communication is that ASR accuracy benchmarks generated on English speech do not translate directly to Arabic-speaking passengers, and dedicated dialectal training data, rather than scaling alone, is likely to be the rate-limiting factor for closing the human–machine gap.

Overall, the pattern of results is consistent with current ASR models approaching human-level accuracy under moderate noise while declining more steeply than human listeners under more adverse conditions; we interpret this as suggestive of a difference in adaptive perceptual mechanisms and higher-order prediction between artificial and biological systems, though the present study measured recognition accuracy alone and did not test these mechanisms directly. The contrast between the performance of blind listeners and that of ASR systems is one motivation, among others, for exploring neuroplasticity-inspired approaches in future ASR development, such as integrating multimodal cues and dynamic reweighting strategies; whether such biologically inspired designs would in fact narrow the gap observed here remains to be tested. These findings reaffirm the practical limitations of current machine hearing under adverse conditions and the noise robustness of the human listeners tested here, and they suggest, rather than establish, directions for future noise-robust ASR development.

5 Concluding remarks, limitations, and future directions

This study compared normal-sighted, hearing-impaired, and blind listeners with two ASR systems (Whisper and TurboScribe) on Arabic speech-in-noise recognition across four signal-to-noise ratios. Recognition accuracy declined for all groups and both ASR systems as noise increased, but the rate of decline differed: blind listeners were the most resilient across all SNRs, ASR systems approximated normal-sighted listeners at favorable and moderate SNRs but declined more steeply at the two lowest SNRs, and hearing-impaired listeners, tested unaided, showed the most pronounced difficulty throughout.

The two ASR systems and the human listeners also differed in kind rather than only in degree: humans, including blind listeners, predominantly omitted words they could not identify, whereas the ASR systems more often substituted or inserted plausible but incorrect material. Taken together, these results indicate that current ASR systems can approximate typical human accuracy under moderate noise but do not yet match the resilience of our most robust human group under severe noise, and that they fail differently than humans do when they fail. We interpret the contrast with blind listeners as consistent with, though not a direct test of, neuroplasticity-based accounts of auditory reorganization after early blindness (see Future Directions for the research and clinical questions this raises).

A few limitations should be considered when interpreting these findings. The relatively small sample (n = 10–20 per group) limits statistical power and generalizability. Hearing loss in the HIL group was self-reported rather than audiometrically verified, leaving the precise degree and configuration of impairment uncharacterized. Although sex was entered as a fixed-effect covariate in the group-level model and did not significantly predict recognition accuracy, the BL group was exclusively male, limiting how confidently the present findings generalize to female blind listeners. The three groups also differed in mean age (NSL: 21 years; BL: 26 years; HIL: 31 years); age was not entered as a covariate in the present analyses, and although the NSL group’s younger age reflects its role as a normative baseline rather than a matched comparison group, age cannot be ruled out as a contributor to the observed group differences. Unlimited repetitions were permitted before response to give human listeners access to the acoustic signal comparable to the ASR systems’ access to the complete waveform (see Materials and Methods); however, repeated listening also affords memory, lexical prediction, and learning processes unavailable to the ASR systems, so the human–ASR comparison reported here should be interpreted as bounded by this asymmetry. Similarly, WER scoring relied on sequential review and adjudication across a research assistant, the experimenter, and an independent phonetician (see Materials and Methods) rather than on independent parallel scoring, so no formal inter-rater reliability statistic could be computed. The acoustic environment was restricted to a single noise type (aircraft passenger-boarding noise), only two ASR systems (Whisper and TurboScribe) were evaluated, and the stimuli were limited to imperative and interrogative sentences in MSA, which together bound the generalizability of the reported human–machine asymmetries. Finally, WER, used here as the sole accuracy metric, may underestimate residual speech-perception ability in the HIL group at the most adverse SNRs, where word recognition approaches floor.

Several directions follow from the limitations above. Future replications should recruit larger, gender-balanced samples with audiometrically confirmed hearing profiles rather than self-report, and should either match groups on age or enter it explicitly as a covariate. Broadening the range of background noise types and reverberation conditions, evaluating a wider set of ASR architectures, and extending the paradigm to additional languages, dialects, and speech genres would help establish how far the human–machine asymmetries reported here generalize. Complementing WER with finer-grained phoneme- or syllable-level scoring, which has been shown to discriminate hearing-loss severity even when word-level recognition is at floor (Jain et al., 2025), would also help characterize residual perceptual ability in listeners whose word-recognition scores saturate at the most adverse SNRs.

The BL–ASR contrast observed here raises a further, more speculative research direction: whether ASR architectures that incorporate adaptive, multimodal, or context-sensitive attention mechanisms—loosely inspired by the cross-modal reorganization documented in early blindness—can narrow the gap seen at low SNR. Testing this would require engineering work well beyond the present behavioral comparison and should not be read as already demonstrated by these data.

Finally, the human–ASR asymmetries documented here may be worth exploring for clinical and applied purposes, though we present these as hypotheses for future work rather than established applications. It remains to be tested whether ASR error profiles could serve as a low-cost adjunct to speech-in-noise testing in audiology, for example in screening or longitudinal monitoring; whether the deletion-versus-substitution asymmetry between humans and ASR systems has practical consequences for hearing-aid or telehealth transcription pipelines, where confidently incorrect ASR output could go unflagged at adverse SNRs; and whether benchmarking ASR systems against the full spectrum of human auditory variability, rather than against typical listeners alone, could usefully complement existing diagnostic tools (Wilson, 2011; Jain et al., 2025) for characterizing where a given ASR engine is likely to support, match, or fail particular patient populations.

Author contributions

YA: Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review & editing.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. I used AI for language editing and document organization in the Response to Reviewer file. All the content is mine. I used AI to edit the manuscript and check for any grammatical errors or language ambiguity. I used AI to enhance the quality of the figures and to introduce some small changes in the colors and patterns of the bars.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.