We recruited 15 young adults (10 male; mean age: 25.6 ± 2.19) from the institutes' internal participant database. The sample size was chosen identical to a previous study (Baumgartner et al., 2017). One participant was excluded from the ALB and localization analyses due to a trigger box malfunction during the EEG recording and missing data. Externalization data were only collected from the last 11 participants. All participants had normal or corrected-to-normal vision as well as normal hearing, meaning that audiometric hearing thresholds did not differ more than 20 dB from their age-specific norm (Rodríguez Valiente et al., 2014) in the high-frequency frequency range of 1 to 16 kHz. All participants were monetarily compensated for their participation. The study applied the standard methodology of the Acoustics Research Institute, which the Ethics Board of the Austrian Academy of Sciences has approved.

The order of tasks was fixed across participants. Everyone began with directional localization followed by the looming bias task and, finally, the externalization paradigm. This order was chosen deliberately because naive participants experience high difficulties when performing a spatial task in anechoic conditions. Since the localization paradigm was the most interactive, it served as a good starting point for the testing session. Externalization was only evaluated after the looming bias paradigm as we aimed to avoid priming effects that would have come about by participants actively evaluating distance perception with novel spectral cues prior to the looming bias paradigm. All participants, but not the experimenter, were blinded regarding which experimental conditions were tested at what point within each paradigm.

We first recorded the acoustic transmission properties for each participant in terms of listener-specific HRTFs that take into account the complex listener-specific pinna, head and torso geometry [apparatus and procedure described in detail in Ignatiadis et al. (2024a)]. Thus, we captured the native spectrospatial properties of sounds arriving at the ear canal, which allowed us to subsequently manipulate and create sounds with distinct spectral shapes to our needs. The measured HRTF magnitude spectrum (spectral contrast factor C = 1; at 1.2 m distance from the listener) was spectrally smoothened (gammatone filter with bandwidth factor of 1) (Hassager et al., 2016) to obtain native spectral cues and then either mathematically inverted (C = −1) to obtain novel spectral cues or deprived of spectral spatial information (C = 0) to obtain flat spectra by applying a previously proposed manipulation method (Baumgartner et al., 2017):

where M₁(f) denotes the magnitude spectrum of the native HRTFs, N_f denotes the frequency bins within the manipulated frequency range of 1–16 kHz, and w′ denotes a frequency weighting function that approximates auditory frequency resolution by the across-frequency derivative of equivalent rectangular bandwidths (Glasberg and Moore, 1990). Representative HRTF spectra are shown in Figure 1a.

For the directional localization experiment, 500 ms Gaussian white noise bursts with 10 ms on- and offset ramps were manipulated using the spectral contrast procedure described above. This resulted in sets of stimuli that contain either the native spectral information, novel spectral information, or are devoid of any spectral information (i.e., flat).

For the externalization ratings, harmonic tone complexes (12 different fundamental frequencies evenly spaced between 100 and 140 Hz, bandwidth: 1–16 kHz, phase curvature: 0.5) (Schroeder, 1970) were corrected for constant loudness by iterative model-based adjustments (Moore et al., 2016) and subsequently filtered with either native, novel or flat HRTFs from four possible incidence angles, i.e. positions: Front-Up (FU) = 45°, Front-Down (FD) = 135°, Back-Up (BU) = 225° and Back-Down (BD) = 270° (Figure 2). We varied the fundamental frequency and the position of our stimuli, in order to avoid making the task too monotonous as well as enforcing listeners to base their judgements on spatial, rather than specific timbre, cues.

Looming bias was assessed using the same tonal stimuli as for the externalization ratings but combined in pairs. By transitioning between two stimuli with different spectral manipulations, we created flattened and shaped stimuli. The transition was initiated 600 ms after the onset of the first sound, with a jitter of ± 50 ms, and was implemented as a short 10 ms crossfade with raised cosine faders to ensure constant intensity for the highly correlated tonal stimuli. The transition was kept short to obtain a temporally clearly defined motion event for the ERP analysis (Baumgartner et al., 2017; Ignatiadis et al., 2024a).

All stimuli across all paradigms were presented binaurally using tubephones (ER-2, Etymotic Research) at an approximate sound pressure level of 70 dB at the ipsilateral ear.

Figure 3 illustrates group-level lateral and polar localization response patterns for each cue-type in order to convey a first qualitative impression of the results. For the lateral angle, the scatter plots indicate high correspondence between target and response lateral angle (responses close to the main diagonal) for all three conditions. Larger differences across conditions are visible for the polar angle. In this dimension, high correspondence between target and response polar angle are only present for the native condition, whereas the flat and novel conditions led to a substantial degradation in relationship between target and response angles. For both flat and novel conditions, listeners exhibited a strong response bias toward the horizontal plane, primarily at their rear (around 180°).

For statistical evaluation, localization performance was quantified as lateral error, quadrant error rate, and local polar error (Figure 4). Spectral cue conditions significantly influenced the lateral error [F_{(2, 26)} = 12.89, p < 0.01, partial η² = 0.50]). The effect of order (first half presentation vs. second half presentation) failed to reach significance [F_{(1, 13)} = 1.10, p = 0.31], as did the interaction between order and cue type [F_{(2, 26)} = 0.17, p = 0.83]. Post-hoc comparisons revealed that the mean lateral error measure was significantly lower in the native condition compared to both the flat [d = −2.07°, 95% CI (−3.26°, −0.88°), p < 0.01] and novel [d = −1.84°, 95% CI (−2.87°, −0.81°), p < 0.001] conditions. No significant difference was observed between flat and novel [d = 0.23°, 95% CI (−1.08°, 1.53°), p = 0.89] spectral cues.

Considering the quadrant error rate, we again observe a large effect of cue manipulation [F_{(1.04, 13.56)} = 20.07, p < 0.001, partial η² = 0.61]. The effect of order again failed to reach significance [F_{(0.52, 6.78)} = 0.02, p = 0.90], as did the interaction between order and cue type [F_{(1.05, 13.56)} = 3.16, p = 0.08]. Post-hoc comparisons showed that performance in the native condition was significantly better than in both the flat condition [d = −14.25%, p < 0.01, 95% CI (−24.13%−4.38%)] and the novel condition [d = −18.99%, p < 0.001, 95% CI (−27.79%−10.19%)]. There was no significant difference in performance between flat and novel cue manipulations [d = −4.73°, 95% CI (−10.10°, 0.63°), p = 0.08].

Regarding the local polar error, cue manipulation again had a large effect [F_{(2, 26)} = 27.15, p < 0.001, partial η² = 0.68]. There was no significant effect of order [F_{(1, 13)} = 0.44, p = 0.52], nor the interaction between order and cue type [F_{(2, 26)} = 0.70, p = 0.50]. Post-hoc comparisons found that the local polar error was significantly lower for native, compared to novel [d = −10.05°, 95% CI (−12.87°, −7.23°), p < 0.001] and also lower for flat, compared to novel [d = −6.96°, 95% CI (−10.56°, −3.35°), p < 0.001]. There was no significant difference between native and flat [d = −3.09°, 95% CI (−7.55°, 1.37°), p = 0.20] spectral cues.

We also investigated whether, and to what extent, performance was better (or worse) than chance-level. To that end, we simulated a virtual localization experiment with fully randomized responses (Baumgartner et al., 2014) and used this as a reference for normalization. The division of real and simulated error measures thus shows how similar the performance is to chance level, with 1 indicating chance performance and 0 indicating lack of any errors. When simulating the virtual localization experiment, an equal response distribution across all possible response locations inside the target range was assumed, while considering the actual target locations of the experiment. For the lateral error, all cue-types performed better than chance-level (native: 0.29; flat: 0.33; novel: 0.33). Regarding the local polar error, performance with native and flat spectral cues was, albeit a bit worse overall, again better than chance (native: 0.76; flat: 0.82). However, performance with novel cues was close to chance level (novel: 0.95). The quadrant error rate exhibits a similar pattern: evident better than chance performance for native and flat spectral cues (native: 0.49; flat: 0.77) and close to chance level with novel cues (novel: 0.87). Overall, polar-angle localization performance was poor and around chance level for novel spectral cues.

We further analyzed the localizability of sounds under our different spectral cue manipulations and source positions in terms of the perceived degree of externalization. Figure 5a visualizes the distribution of the obtained externalization ratings across spectral cue types, indicating a gradual decay of externalization ratings with increasing deviation from the native reference. We analyzed those data by selecting the best mixed-effects model as outlined in Section 2.4.2. The model predicting externalization scores based on the two main effects of cue and position yielded the best fit (model #1 in Table 1). Compared to the intercept-only model, the BF approximation approach (Wagenmakers, 2007) revealed very strong evidence in favor of the cue-only model [BF = 27, 487.83; Pr(H₀|D)>0.99] and strong evidence in favor of the position-only model [BF = 69.21; Pr(H₀|D)>0.98], confirming a significant contribution of our main effects to explaining the data.

Regarding the main effect of spectral cue type, the estimated marginal means suggest a gradual decay of externalization perception from native, over flat to novel spectral cues. That is, native stimuli were rated as the most externalized [native = 0.67, 95-CI(0.61, 0.74)], followed by flat stimuli [flat = 0.48, 95-CI(0.41, 0.55)] and finally novel stimuli [novel = 0.28, 95-CI(0.22, 0.34)]. Post-hoc pairwise comparisons revealed that the differences between all three levels of cue types were significant (native—flat: z = 3.84, p < 0.01; native—novel: z = 8.26, p < 0.01; flat—novel: z = 4.10, p < 0.01). These differences in externalization perception between different spectral cue types are in-line with the model-predicted ratings (Figure 1c) and thus support the expected motion perception in the looming bias paradigm. This enables to test the Spatial Perception Hypothesis against the Acoustic Hypothesis.

Results regarding the main effect of position are shown in Figure 5b. The estimated marginal means indicated that stimuli presented from the frontal hemifield were rated as less externalized [FU = 0.32, 95-CI(0.25, 0.40); FD = 0.38, 95-CI(0.30, 0.46)] than stimuli presented from the rear hemifield [BU = 0.61, 95-CI(0.53, 0.69); BD = 0.61, 95-CI(0.53, 0.68)]. Post-hoc pairwise comparisons showed no significant differences between the positions within frontal and rear hemifields (FU–FD: z = –1.05, p = 0.72; BU–BD: z = 0.04, p>0.99), but significant differences between them (FU–BU: z = –5.13, p < 0.01; FU–BD: z = –5.09, p < 0.01; FD–BU: z = –4.02, p < 0.01; FD–BD: z = –3.98, p < 0.01).

Figure 6a shows the mean response times and the proportion of flattened-to-looming-match responses. In the native flattened condition, participants indicated “looming” in most cases (60%–80%) and in the native shaped condition, they primarily responded “receding” (50%–60%). In the novel flattened condition, however, participants responded mostly random (50% “looming” and 50% “receding”), whereas in the novel shaped condition they mostly responded with “looming” (60%–70%), indicating a perceptual switch between native and novel spectral cue conditions.

Before constructing the neurobehavioral model, results of the ALB paradigm were first investigated based on EEG data only. Cluster-based permutation tests showed no significant differences between flattened and shaped stimuli within cue-type (p = 0.072 for a negative cluster in the native condition ranging from 100 to 180 ms after correction for two-tailed testing). Comparing the differences between flattened and shaped stimuli between cue types, however, revealed a significant amplitude difference between the native and novel flattened-shaped contrasts in the N1-related range of 80–170 ms (clusterstat = –294.0344, p = 0.04) after motion onset. The native condition was characterized by a more negative N1-related-cluster amplitude for flattened, compared to shaped stimuli. In contrast, the novel condition was instead characterized by a more negative amplitude for shaped, compared to flattened stimuli (Figure 7a), in line with the externalization-based ALB elicitation hypothesis (Figure 1c) and participant's externalization ratings for the different spectral cue manipulations (Figure 5). Topographically, this difference across cue types emerged over left occipital electrodes and subsequently propagated toward central electrodes. At around 140 ms, the cluster became spatially bimodal and fronto-temporally lateralized (Figure 7b, right-most topographies). We further confirmed that both the native and novel evoked differences between flattened and shaped stimuli within the N1-related cluster of difference were significantly different from zero (native: t = −3.46, p < 0.01; novel: t = 3.16, p < 0.02, see Figure 7b, right panel). A two-way repeated measures ANOVA further revealed a significant main effect of spectral cue on the cluster amplitude of the evoked native-novel difference [F_{(1, 13)} = 16.4, p < 0.001, partial η² = 0.56]. There was no significant effect of position [F_{(3, 39)} = 1.52, p = 0.22], nor was there a significant interaction between spectral cue and position [F_{(3, 39)} = 0.46, p = 0.71].

We then investigated the behavioral results of the motion discrimination task by comparing different LBA models used to describe the speed of evidence accumulation in terms of drift rates. We started with identifying the best purely behavioral LBA model and then tested the benefit of adding the trial-by-trial cluster amplitudes (see Section 3.3) as an additional neural regressor:

Our comparison identified this full model where the drift rate was modulated by all five acoustic and behavioral factors (i.e., flattened-stimulus-to-looming-response match, cue type, motion direction, vertical and horizontal positions) and the additional neural marker as the best fit. The quality of fit was lower without the neural marker (dBPIC = 2) and the preceding selection of the purely behavioral reference model showed that it clearly outperformed simpler alternatives (intercept only: dBPIC = 2,999; match only: dBPIC = 2,701; match and direction: dBPIC = 2,318; match, direction and cue: dBPIC = 1,129) as well as an alternative regressand (response threshold instead of drift rate: dBPIC = 99). The benefit of adding the neural marker was also verified against surrogate data matching the mean and SD of the cluster scores (dBPIC = 12). In Figure 6a, we report a direct comparison between the real and simulated behavioral data obtained from the winning model.

We then used the model predictions to assess the strength of the ALB in terms of drift rates. The results demonstrate a dependence on perceived motion direction (Figures 6b, c). For correct responses, drift rates resulted higher for flattened compared to shaped stimuli [P_Δ>0 = 0.995, with Δ = 0.362s⁻¹, 95%-CI (0.093, 0.658)]. For incorrect responses, drift rates resulted smaller for flattened compared to shaped stimuli [P_{Δ < 0} = 0.970, with Δ = −0.339s⁻¹, 95%-CI (–0.648, 0.017)], indicating that looming responses consistently exhibited higher drift rates compared to receding responses. Furthermore, the analysis revealed that cue type modulates drift rates differently based on response match. For correct responses, the native condition yielded a greater drift rate compared to the novel condition [P_Δ>0 = 0.951, with Δ = 0.356s⁻¹, 95%-CI (–0.079, 0.793)]. Conversely, for incorrect responses, drift rates for the native condition reported smaller values than those from the novel condition [P_{Δ < 0} = 0.991, with Δ = −0.590⁻¹, 95%-CI (–1.088, –0.106)]. Finally, we inspected the magnitude of the fitted linear coefficient associated with the magnitude of the N1-related EEG cluster for the prediction of the model's drift rate which corresponded to –0.017 1/μVs with a 95%-CI (–0.039, 0.005), and with a 0.942 probability of it being a negative value. Figure 8 illustrates the linear relationship between the EEG predictor (N1-related EEG cluster amplitude) and the fitted drift rates, revealing that faster evidence accumulation is associated with a more negative EEG amplitude. This result, alongside the model comparison, further supports the notion that there is a coherent relationship between the neurophysiological marker and the behavioral outcome.

We demonstrated that localization performance with novel spectral cues was significantly impaired compared to performance with native cues across all error measures (lateral error; local polar error; quadrant error rate). This indicates that established spatial associations between stimulus characteristics and locations were successfully broken apart. Furthermore, we also observed diminished local polar localization performance for novel cues compared to flat cues, which is also reflected in the fact that performance with novel cues was close to chance performance, whereas performance with flat cues was still better than chance.

This raises the question of why participants were still able to localize sounds in the polar dimension reliably at least to some degree with flat spectral cues. A possible explanation is that our HRTF manipulation only targeted frequencies above 1 kHz (see Section 2.3), thus low frequency information provided by torso reflections was still available to participants and possibly provided some usable elevation (Algazi et al., 2001), externalization (Zotter and Frank, 2018) and front-back discriminatory cues (Asano et al., 1990). The additional presence of contradictory, i.e., inverted, high frequency spectral cues in the novel cue condition could have potentially interfered with the information provided below 1 kHz, thus, leading to worse performance and a stronger degradation of externalization than in the flat condition, where low frequency cues were the only available ones. We also observed a response bias toward the horizontal plane with flat and novel spectral cues, primarily at participant's rear (around 180°). A response bias such as this is, also referred to as low elevation gain (slope of the stimulus-response relation) (Zwiers et al., 2003) under perceptual uncertainty, i.e. with impoverished spectral cues, is in line with previous work reporting similar patterns (Wanrooij and Opstal, 2005; Zonooz et al., 2018, 2019). A second possible explanation could arise from an implementation issue in the experimental design. Due to a miscommunication, participants were tested using a fixed order of spectral cue conditions—native, flat, and novel—which coincides with the observed performance ranking. However, since this fixed order was repeated three times, it also included native after novel blocks and thus is rather unlikely to account for the pronounced differences between conditions.

Moreover, lateral localization performance and quadrant error rate was also significantly reduced with flat spectral cues, compared to native spectral cues. Even though the impairment was rather small, reduced lateral localization abilities with flat spectral cues is in-line with a more recent, updated, view on the contribution of spectral cues for the lateral dimension. Traditionally, lateral sound localization was thought to almost exclusively rely on binaural cues, such as interaural time- and level differences, with monaural spectral cues thought to only contribute very little (Macpherson and Middlebrooks, 2002). However, recent work (Stevenson-Hoare et al., 2022) has challenged this idea by demonstrating an apparent front/back asymmetry in sound localization precision measures, which is especially strong for oblique azimuths. That is, the pinna introduces a high-frequency spectral cue that enhances lateral sound localization in the frontal hemifield only. Hence, contemporary computational models on human sound localization include both monaural and binaural cues as joint contributors to lateral and polar localization performance predictions (Barumerli et al., 2023). By flattening individualized HRTFs, these additional cues were no longer available to our participants, potentially explaining their diminished lateral localization performance with flat HRTFs compared to native HRTFs.

Furthermore, we also demonstrated that the spectral manipulation of individualized HRTFs degraded the perceived externalization of sound sources. An externalization breakdown was observed for both flat and novel HRTFs. Stimuli based on novel HRTFs were rated as even less externalized than stimuli based on flat ones, indicating a shift toward an internalized percept with novel, spectrally inverted HRTFs and a gradual decay of externalization. This decay is consistent with a template-based model of externalization under static listening conditions (Baumgartner and Majdak, 2021). Under this framework, perceptual externalization depends on comparing observed auditory cues and listener-specific templates in memory. When incoming sounds deviate from the established reference template, externalization perception is predicted to gradually decrease with increased deviation magnitude. In the present study this assumption would thus also predict a significant, but smaller, difference between native (e.g., template-matching) and flat (e.g., template-deviation) stimuli, compared to the difference between native and novel stimuli consistent with our observation.

Based on the present results we conclude that ALB elicitation was consistent with the listeners' perceived changes in externalization rather than triggered by inherent acoustical features independent of auditory space. In line with the template-based model predictions (Baumgartner and Majdak, 2021), participants reported less externalization for novel stimuli compared to flat and native stimuli. Furthermore, they tended to exhibit a low rate of shaped-stimulus-to-receding-response match, indicating that their perception of novel stimuli seems to have been reversed compared to native stimuli. Stimuli that exhibited a reduction of perceptual externalization during the short stimulus transition were perceived as looming (e.g. native flattened and novel shaped), and vice versa, stimuli that contained an increase of perceptual externalization were perceived as receding (e.g. native shaped and novel flattened). The speed of evidence accumulation, i.e. the drift rate, was also consistently higher for looming responses, regardless of stimulus-to-response-match. Taken together, our data support the hypothesis that ALB elicitation relies on spatial perceptual associations, rather than purely acoustic properties (see Figures 1b, c).

Although the current study demonstrates that ALB elicitation conformed with participant's spatial perception, due to its correlational nature we cannot clearly pinpoint the causal dependencies between bias elicitation, sound localizability and externalization perception in particular. A follow-up learning study investigating the generalizability from sound localization to ALB elicitation with novel spectral shape cues (Greif et al., 2024) will address this.

Neural markers of ALB were identified between 80 ms and 170 ms after motion onset via cluster-based permutation testing and confirmed to be behaviorally relevant via LBA modeling. The latency and topography of that marker closely relates to the N1 ERP component and is consistent with previously reported results on ALB and motion perception (Ignatiadis et al., 2021, 2024a). The N1 component is an established marker of early perceptual processing, implicated to play a central role in attention allocation via sensory gain control (Getzmann, 2011; Getzmann and Lewald, 2010; Hillyard et al., 1998). That is, attended stimuli tend to exhibit an increased N1 amplitude compared to unattended stimuli. In the context of our study, sounds exhibiting a decay of externalization were predominantly perceived as looming, captured attention, and thus facilitated ALB elicitation. Interestingly, previous work has demonstrated this neural bias toward looming sounds also while attention was diverted, in a passive listening mode (Ignatiadis et al., 2024a), highlighting the high salience and attentional capture potential of approaching sound sources.

Incorporating the trial-by-trial N1-related EEG cluster amplitudes in the LBA model improved its explanatory power. Yet, the overall benefit was rather small (dBIC = 2), suggesting that this single-trial maker was quite noisy and supposedly incomplete. Alternative neurophysiological markers such as ERP latency, oscillatory power, or a joint metric, should thus be considered. Previous corticocortical connectivity analyses have identified bottom-up projections between temporal, parietal and the frontal regions as a possible neural network involved in the elicitation of the ALB (Ignatiadis et al., 2021). Specifically, connectivity among regions along the ventral stream, traditionally coined the “what” stream in visual and auditory processing (Saur et al., 2008), seems to play an essential role in ALB elicitation (Ignatiadis et al., 2024b). This makes it a suitable target for future work incorporating more sophisticated neural ALB markers (e.g. phase locking value). Implementing more flexible joint modeling methods with non-linear relationship assumptions (Palestro et al., 2018; Nunez et al., 2024) could also further improve the present model. Translating the present work into more immersive environments, e.g. via VR-based applications using more realistic sounds, could also lead to a stronger perceptual impression of spectral looms and recedes, potentially increasing the observed effects (Deng et al., 2019; Getzmann and Lewald, 2010) and strengthening the relationship between behavioral measurement and neural marker.

A key limitation of the present study is the relatively small sample size, which may reduce the statistical power to detect small or moderate effects. This limitation implies that non-significant findings, such as the lack of EEG differences between flattened and shaped stimuli within cue-type, should be interpreted with caution, as they may reflect insufficient power rather than a true absence of effect. Nevertheless, the study provides initial evidence that the emergence of auditory looming bias requires spatial perceptual associations to be established.

Another limitation concerns the low ecological validity of stimuli. As in previous studies (Baumgartner et al., 2017; Ignatiadis et al., 2024a), we here isolated spectral from intensity cues to hone in on the contribution of spatial perception to ALB, but it is worth noting that this isolation of cues does not happen outside of the laboratory and is expected to limit the extent of ALB (Bach et al., 2009) and attention control (Deng et al., 2019). Secondly, the spectral cue manipulations to elicit looming and receding sensations happened on a very rapid time scale. Such a rather instantaneous change in perceived distance has the advantage of providing a clearly defined event of interest for the analysis of EEG data, but it may limit the generalization of our results to the real world where changes in distance usually occur more gradually. Finally, our experimental setting did not allow participants to move during stimulus presentation because this would render the stimuli even more implausible. Listeners naturally engage in self-motion to resolve perceptual ambiguities (Wallach, 1940), consistent with principles of active inference (Friston et al., 2016). Future work is needed to develop more naturalistic experimental settings with high stimulus control for neuroscientific investigations.

The here presented results establish a strong link between spectrospatial associations and looming/receding perception, implying that any disruption of natural spectral cues would also disrupt motion perception in general, leading to potentially dangerous situations for anyone wearing hearing aid or hearing protection devices. Furthermore, the preservation of natural cues, or at least providing a sufficiently good approximation, seems integral for developing immersive virtual realities that provide realistic auditory experiences. Actively manipulating spectral cues provided to listeners could also provide unique and interesting opportunities for the development of gamified environments and/or audio design.