Estimating band importance for environmental sound recognition using deep learning^a)

A. Participants

Data for this study were obtained from 46 listeners with typical hearing, defined as pure-tone air-conduction thresholds of 20 dB HL or better at octave frequencies from 250 to 8000 Hz in the better ear on the day of testing (ANSI, 2004, 2025). Participants were 18 to 26 years old (mean = 19) and received either course credit or a monetary incentive for participation. Forty-four of the participants were female, and two were male. Written informed consent was obtained from each participant prior to testing, and all procedures were approved by the Institutional Review Board of The Ohio State University.

B. Stimuli

The data analyzed in the present study were originally collected in three experiments reported by Johnson and Healy (2024a,b). In these experiments, ESR was measured in the presence of competing speech using a 25-alternative forced-choice paradigm. The set of environmental-sound signals consisted of 25 everyday sounds representing a wide range of categories typically used in environmental sound research (e.g., Gygi et al., 2007). Table I lists these environmental sounds. Several of these stimuli are best described as collections of repeated brief sounds rather than a single continuous acoustic event. For instance, the drum and cymbals produce successive percussive bursts, and the sound of fireworks typically includes multiple rapid explosions. Together, these environmental sounds capture a diverse set of auditory events spanning human, animal, mechanical, musical, and natural sources, consistent with the semantic and acoustic variety emphasized in ESR research.

Sentences from the standard recording of the Hearing in Noise Test (HINT; Nilsson et al., 1994) were used as maskers in these experiments. Speech represents one of the most common and ecologically valid competing signals in everyday environments, and its complex temporal and spectral structure makes it especially suitable for examining band-specific effects of audibility.

C. Procedure

On each trial in the experiments described by Johnson and Healy (2024a,b), an environmental sound was randomly selected from the set of 25 and mixed with a sentence from the HINT (Any of the environmental sounds could be paired with any of the HINT sentences used for testing.). The TMR is defined here as the level difference between the environmental sound (target) and the sentence (masker). Across the three experiments, TMRs spanned 21 values ranging from −70 to +36 dB. As described in the earlier publications, this large range of TMR values was needed to capture the full psychometric function relating TMR to ESR accuracy.

Prior to mixing, silence was trimmed from the beginning and end of each recording so that the onsets of the environmental sound and HINT sentence were aligned, eliminating any precursor containing only one signal. The environmental sound, which was always longer than the sentence, was then truncated to match the duration of the speech, aligning offsets and removing glimpses of the environmental sound at the end of the stimulus. Each environmental sound was scaled to achieve the desired TMR based on root mean square (RMS) values and mixed with the corresponding sentence. Each mixture was subsequently scaled to a common RMS value. Stimulus levels were set to 65 dBA (equivalent continuous sound pressure level of a representative looped stimulus) in each ear and verified using a Model 824 sound-level meter and AEC101 flat-plate coupler (Larson Davis, Depew, NH). Because the overall level was fixed, the relative levels of both the target and masker depended on the TMR. All stimuli were digitized at 44.1 kHz with 16-bit resolution.

Listeners were tested individually in a double-walled audiometric booth while seated in front of a P2225H computer monitor and MS116 mouse (Dell Technologies, Round Rock, TX). The experimenter was also present in the booth, seated approximately six feet away. Stimuli were played back from an OptiPlex 7090 Windows PC (Dell Technologies, Round Rock, TX), converted to analog form with a Fireface UCX audio interface (RME Audio, Haimhausen, Germany), amplified with a 1202-VLZ mixer (Mackie, Woodinville, WA), and presented diotically over HD 280 Pro headphones (Sennheiser, Wedemark, Germany). Prior to testing, listeners were familiarized with the environmental sounds by hearing them one at a time and identifying them, with feedback from the experimenter, until all sounds were correctly identified at least once.

On each trial, the listener heard the mixture and completed a 25-alternative forced-choice ESR task. Responses were made by clicking on one of 25 labeled images, displayed in a 5 × 5 grid on the computer screen. In Johnson and Healy (2024a) and in experiment 1 in Johnson and Healy (2024b), listeners performed ESR under divided-attention conditions (report both environmental sound and speech). In experiment 2 in Johnson and Healy (2024b), listeners performed the task under selective-attention conditions (report only environmental sound). In the present work, data from all three experiments were pooled to increase sample size—an important consideration for machine learning analyses—and because divided versus selective attention was found to exert only a minimal effect on ESR performance (Johnson and Healy, 2024b). The present dataset included responses from 14 listeners tested in Johnson and Healy (2024a), including two additional participants not reported previously.² From Johnson and Healy (2024b), data were included from 11 listeners in experiment 1 and 21 listeners in experiment 2, including one participant not reported in the original publication. Additional details may be found in Johnson and Healy (2024a,b). The pooled dataset for the present study consisted of 10 100 trial-by-trial responses.

D. Descriptions of models

To investigate how frequency-specific information supports ESR, two complementary deep learning model variants were developed and evaluated. These two model variants perform the same recognition task: they take acoustic waveform inputs (mixtures of environmental sounds and competing speech) and attempt to identify which of 25 possible sounds is present in each waveform. The model variants are also broadly similar in terms of the manner in which they process input waveforms. However, the two model variants were trained using different targets (supervisory signals). In practical terms, this means that any differences in model behavior or estimated BIFs can be attributed to differences in learning objectives rather than differences in what the models “hear.”

The first variant, referred to as the human-trained ESR model, was trained using human listener responses to the ESR task as labels. Because listeners did not always respond correctly, the training labels for this model did not consistently match the true identity of the environmental sound in the input waveform. The goal of this model is therefore not simply accurate sound identification but rather the reproduction of human ESR behavior under speech masking. As such, it serves as a computational proxy for human ESR performance, intended to capture systematic successes, confusions, and limitations that arise when listeners identify environmental sounds in competing speech. The rationale for this approach is that training the model to predict human listeners’ responses, rather than ground-truth sound identity, should encourage it to learn the acoustic features that shape human perception, including those associated with characteristic errors. Consequently, the resulting BIFs are more likely to reflect human-like auditory weighting strategies than those derived from a model optimized purely for sound-identification accuracy.

The second variant, the ground-truth-trained ESR model, was trained using correct sound identity labels. Its goal is to identify the sound in the stimulus as accurately as possible, independent of human response variability. This model represents a task-optimized recognizer that relies on the acoustic information that is most useful for correct sound identification. Previous research has shown that task-optimized neural networks can also replicate human auditory behavior (Kell et al., 2018).

Both variants were trained using five-fold cross-validation, yielding five fold-specific models per variant to assess robustness and reliability across data splits. Together, these two model variants support a central aim of this study: to estimate frequency BIFs for ESR and to determine how frequency regions that support optimal recognition compare with those emphasized in human-like recognition. The following sections provide detailed descriptions of these model variants, including their architecture, training procedure, and evaluation framework.

A. Human-trained ESR model variant performance

To evaluate generalization beyond the acoustic conditions represented in the human-subjects experiments, each fold-specific human-trained ESR model was tested on the set of 10 000 additional mixtures created for training the ground-truth model. These mixtures were excluded from all training and validation for the human-trained models and were created using TMR values that did not appear in the human-subjects dataset. This evaluation therefore assesses whether the learned recognition strategy transfers to previously unseen mixtures across a range of difficulty levels. Importantly, performance was evaluated using ground-truth sound labels rather than human responses. This was the metric used to evaluate human-subjects performance on the ESR task (Johnson and Healy, 2024a,b). As a result, all reported metrics reflect objective task accuracy, not the model’s ability to reproduce human response patterns.

Across folds, model behavior was largely consistent, with four of the five models achieving similar levels of 25-way environmental sound identification on the external set. Folds 2–5 achieved similar external-set accuracies (0.636–0.673) with comparable macro-averaged F1 scores (0.682–0.722), indicating that the human-trained variant produced a stable recognition strategy across cross-validation splits. Fold 1 was the notable outlier (accuracy = 0.488; macro-F1 = 0.596), suggesting reduced robustness under the external distribution shift for that fold. This fold-specific weakness was not limited to a single sound class but reflected a broader change in decision behavior (described below).

Figure 1 illustrates fold-to-fold consistency for the human-trained ESR model variant and compares model performance with human ESR accuracy. Evident is that for each fold-specific model, accuracy increased steeply as TMR improved, similar to human ESR performance. Also apparent is that the model for Fold 1 performed consistently more poorly than the other fold-specific models across a broad range of TMRs. In contrast, the models from Folds 2–5 showed similar performance to one another and approached human accuracy, especially at higher TMRs. In spite of the discrepancy between fold 1 and the others, pairwise correlations among the five fold-specific performance curves were uniformly high (r ≈ 0.94–0.99), indicating a high degree of consistency. Linear interpolation of the performance functions indicates that human listeners reached 50% accuracy at a TMR of −37.5 dB, whereas the best-performing fold-specific model (fold 2) reached this threshold at −34.2 dB. Thus, this fold-specific model’s 50%-accuracy threshold was only 3.3 dB higher (i.e., poorer) than that of the human listeners.

Beyond aggregate accuracy, we examined whether fold-specific models showed consistent patterns of successes and confusions across sound classes. A key observation was the emergence of a reliable sink class tendency for the ocean waves sound in four of the folds: when the model was uncertain, it tended to default to the ocean waves response. Importantly, this pattern mirrors the response bias observed in human listeners (see Johnson and Healy, 2024b, for details), where ocean waves also acted as a disproportionately frequent response option. In the present framework, this is a desirable property of the human-trained model variant because it indicates that the model is capturing not only which sounds are recognized but also the structure of human confusions and the bias to overuse particular categories under challenging conditions. This observation provides support for the likelihood that the human-trained model variant is relying on human-relevant cues for performing ESR.

However, fold 1 diverged sharply from this otherwise consistent structure. Instead of an ocean waves sink, fold 1 exhibited an extreme bias toward cymbals, with near-perfect recall for cymbals (0.998) but very low precision (0.075), consistent with over-predicting cymbals across many non-cymbals stimuli. (In the context of classification, recall refers to the proportion of instances belonging to a given class that are correctly classified. Precision represents the proportion of predictions assigned to a given class that are correct.) This atypical decision behavior provides a natural stress test for later analyses: if fold-averaged BIFs remain stable despite one fold’s idiosyncratic confusions, it strengthens the argument that the inferred frequency-importance patterns are not artifacts of a single split.

B. Ground-truth-trained ESR model performance

Performance of the ground-truth-trained model was evaluated on a held-out test set of 2883 mixtures (15% of the 19 200-mixture corpus) that was never used for training or model selection. Across folds, performance was both high and remarkably stable. Fold-specific test accuracies ranged from 0.715 to 0.740 [mean ± standard deviation (SD) = 0.729 ± 0.009], with macro-averaged F1 ranging from 0.724 to 0.750 (mean ± SD = 0.742 ± 0.010). Importantly, fold-to-fold reliability was very high across the full accuracy-by-TMR function. Pairwise correlations between fold-specific performance curves were consistently strong (r ≈ 0.98–0.99), indicating that the model’s susceptibility to masking level was highly reproducible across data splits.

Figure 2 shows accuracy for each fold-specific model of the ground-truth-trained variant, the mean model accuracy across folds, and human listener accuracy as a function of TMR. Here, again, accuracy increased steeply as TMR improved. For TMRs above −30 dB, the models approached ceiling performance (mean accuracy ≈0.990), reaching essentially perfect recognition at TMRs of −10 dB and higher in this test set.

A central observation was that the ground-truth-trained models were consistently more accurate than human listeners. Whereas the human listener psychometric function for ESR reaches the 50% threshold at TMR of −37.5 dB, the five fold-specific performance functions cross this threshold at TMR values ranging from −47.5 to −44 dB, an improvement of up to 10 dB compared to human-listener performance. Across the 10 100 behavioral trials, mean human recognition accuracy was 0.604. In contrast, all five ground-truth fold models exceeded 0.71 accuracy on the held-out test set, despite being evaluated on a corpus that included novel mixtures at TMRs outside the behavioral sampling.

The high accuracy of the ground-truth-trained model variant does not imply that it is more human-like than the human-trained variant; rather, it reflects the different learning objective and the fact that ground-truth supervision can leverage a larger pool of synthetically generated mixtures. Whereas the human-trained variant is explicitly encouraged to reproduce human confusion structure (including response biases), the ground-truth-trained variant is optimized for veridical identification and therefore serves as a task-optimized reference point. In this context, the expanded training set is best viewed as part of the ground-truth formulation: it allows the model variant to more fully realize the task-optimal solution under the same stimulus-generating process, whereas the human-trained model variant remains intentionally constrained by the information available from behavioral labeling.

Unlike the human-trained variant, no fold-specific models of the ground-truth-trained variant exhibited an idiosyncratic collapse toward a single response category; instead, fold-to-fold variability was modest and primarily reflected small shifts in performance at the lower TMRs. Inspection of the fold-specific classification reports revealed a qualitatively different error structure from the human-trained variant. Several acoustically distinctive categories were recognized with consistently high precision across folds—most notably alarm, toilet, and windshield wipers (precision typically ≥0.94)—indicating that when the ground-truth model committed to these labels it was rarely incorrect. At the same time, some classes exhibited a more conservative pattern (high precision but moderate recall), such as helicopter and chainsaw, suggesting that the model required stronger evidence before selecting those labels.

Conversely, the few systematic confusions that remained tended to be modest and class-specific rather than global sink behaviors. For example, airplane was characterized by high recall but lower precision in several folds, consistent with overprediction of this class for some non-airplane stimuli. Importantly, ocean waves did not show the high-recall/low-precision signature that characterized the human-trained model’s ocean waves bias; instead, the precision and recall of ocean waves were both moderate-to-high, consistent with a model optimized for veridical identification rather than reproducing human response tendencies.

The contrast between variants is central for interpretation of BIFs. If the human-trained and ground-truth-trained models yield similar BIFs, it would suggest that the frequency regions that support human-like recognition closely track those supporting task-optimal recognition. Conversely, systematic divergences would indicate frequency regions where human recognition under competing speech departs from task-optimal use of acoustically available information. Section III C therefore quantifies fold-to-fold BIF reliability and compares average BIFs between the two model variants.

C. Estimates of band importance for ESR

Frequency BIFs were estimated separately for each fold-specific model by measuring the performance cost of selectively removing target information in one equivalent rectangualr bandwidth (ERB)-spaced band at a time (Moore and Glasberg, 1983). For each fold, baseline performance was first computed on a newly generated set of 2000 sound+speech mixtures spanning a wide TMR range (–69–35 dB). The same 2000 mixtures were then re-evaluated after applying a zero-phase, fourth-order Butterworth IIR bandstop filter to the target (environmental sound) in each of 35 ERB-spaced bands (125–12 000 Hz) prior to remixing with the unmodified masker. Band importance was quantified as the drop in accuracy relative to baseline (baseline − bandstop), such that larger drops indicate frequency regions whose removal most strongly impairs recognition.

Before any ablations, the human-trained ESR models achieved a baseline accuracy of 0.669 ± 0.011 across folds (macro-F1: 0.724 ± 0.005). The ground-truth-trained ESR models were consistently higher on the same set of 2000 mixtures, with baseline accuracy 0.754 ± 0.007 (macro-F1: 0.768 ± 0.011). Thus, even under identical evaluation conditions, the ground-truth formulation produced uniformly stronger recognition performance than the human-trained objective.

Figures 3 and 4 show fold-specific BIFs for the human-trained and ground-truth trained model variants, respectively. Across folds, BIFs were generally stable: the shape of the function (i.e., which frequency regions mattered most) was preserved even when the exact magnitude of the accuracy drops varied. Pairwise fold-to-fold correlations on the 35-element BIF vectors were positive and typically large for both model variants, supporting the interpretation that the observed BIF structure is not an artifact of a single train/validation split. Averaged across all 10 fold-pairs, Pearson correlations were similar for the two variants (both ~0.78), whereas rank-order agreement (Spearman) was somewhat higher for the ground-truth-trained variant (~0.69 versus 0.62). Importantly, fold-to-fold agreement for the ground-truth-trained BIFs was very high when considering folds 2–5 (mean Pearson ≈ 0.87; mean Spearman ≈ 0.73), indicating that the task-optimized BIF pattern was especially reproducible once a single lower-agreement fold was excluded. In contrast, the human-trained BIFs showed more modest rank-order stability across folds 2–5 (mean Pearson ≈ 0.77; mean Spearman ≈ 0.57), consistent with somewhat greater fold-to-fold variability in how importance was distributed across adjacent mid-frequency bands.

Overall, these reliability analyses support two conclusions: (i) both approaches yield BIFs that replicate across cross-validation folds, and (ii) the ground-truth-trained objective tends to produce a more internally consistent signature of band importance (particularly in rank ordering), whereas the human-trained objective shows slightly more variability in how importance is apportioned among neighboring informative bands.

Figure 5 displays BIFs, averaged across folds, for the human-trained and ground-truth-trained model variants. Apparent is that despite differences in training supervision and overall accuracy, the two model variants produced strikingly similar BIF shapes. In both variants, the fold-averaged BIF exhibited five prominent peaks at approximately the same center frequencies: 432, 774, 1460, 2615, and 9700 Hz.

These peaks were visible in the per-fold curves for each variant and remained the dominant maxima in the fold-averaged BIFs. Quantitatively, the magnitude of the fold-averaged accuracy drops at these bands was on the order of a few percentage points, with the largest peak near ~9.7 kHz (mean drop ≈ 0.055 for the human-trained models and 0.060 for the ground-truth-trained models). Notably, these five peak bands accounted for roughly two-thirds of the total summed importance across the 35 bands in both variants, indicating that importance was not broadly distributed across frequency; instead, it was concentrated in a small set of reproducible frequency regions.

Outside these peak regions, most bandstop filtering of the target produced minimal performance change. Median fold-averaged drops were <0.003 (i.e., <0.3 percentage points), and high-frequency bands above ~4 kHz contributed little except for the concentrated high-frequency peak near 9.7 kHz. In practical terms, this pattern suggests that for the specific set of 25 sounds studied here and under competing-speech masking, recognition is strongly constrained by access to a few frequency regions rather than by uniform contributions across the spectrum.

At the level of the fold-averaged BIF, the two variants were highly concordant: the correlation between the mean human-trained BIF and the mean ground-truth-trained BIF was very large (Pearson ≈ 0.97, Spearman ≈ 0.87). This is an important result conceptually. Although the supervisory signal differed (human-response labels versus veridical sound identity), both variants converged on nearly the same frequency regions as being most diagnostic for recognizing environmental sounds in competing speech. In other words, the frequency dependence of the task appears to be dominated by acoustic constraints of identifying the target sound from a speech masker, rather than being determined primarily by idiosyncrasies of the training objective.

Although the peak locations aligned, there were small differences in peak emphasis. Relative to the ground-truth-trained models, the human-trained models tended to place more importance in the mid-frequency region around ~1.3–2.6 kHz (including a stronger peak around ~2.6 kHz), whereas the ground-truth-trained models showed slightly greater importance in bands around ~1.1–1.9 kHz and in the high-frequency peak near ~9.7 kHz. One plausible interpretation of these small differences in peak emphasis is that the human-trained objective encourages reliance on frequency regions that human listeners weight more strongly in masked listening (mid-frequency regions where audibility and masking effects are perceptually salient), whereas the task-optimized objective more consistently exploits whatever frequency information most improves identification—including very high-frequency target cues that may be relatively less useful (or less accessible) to human listeners in everyday conditions.

In sum, both model variants produced BIFs that were reproducible across folds and characterized by the same five-band signature of importance, with the ground-truth-trained variant showing especially strong cross-fold stability. The close agreement between the two across-fold mean BIFs suggests that frequency importance is a generally robust property of ESR under competing speech, regardless of if a model is trained to predict human responses or simply perform task-optimal recognition.

The BIFs in Fig. 5 possess characteristics of classic functions for speech materials as well as substantial differences. With regard to similarities, the center of the spectrum at approximately 1500 Hz was found to be highly important for ESR. Also, substantial “microstructure” is observed in the current functions, in which a given band can display far more importance than the bands immediately below or above it in frequency. In the case of speech sentences, a band centered at 1370 Hz was found to possess more than twice the importance of the bands immediately below or above it (Healy et al., 2013). In the current case of environmental sounds, the band centered at 2615 Hz was found to possess approximately 14 times the importance of the immediately adjacent bands. This microstructure was found to be reliable for speech materials (Healy et al., 2013) and contrasts with the smooth functions found in classic literature (ANSI, 1997). With regard to differences between the functions for environmental sounds and those for speech, the current functions display particularly high importance for the very high frequencies (9.7 kHz), whereas speech functions typically do not. This difference between environmental-sound and speech importance can be understood in terms of differences between the sources of speech versus environmental sounds.

To situate the present band-importance estimates within the existing literature, it is useful to compare our masking-derived, narrowband ablation BIFs to earlier frequency-limiting studies of environmental sound identification in quiet. Gygi et al. (2004) measured identification of environmental sounds after low-pass, high-pass, and octave bandpass filtering and noted that listeners can often rely on multiple cue types, including temporal information, depending on the sound. When performance was averaged across sounds, the low-pass and high-pass functions intersected at an “isoperformance” cutoff near 1.3 kHz, consistent with a coarse boundary separating lower- from higher-frequency contributions. Their bandpass results further showed markedly poorer performance in the lowest octave bands (≈31% and 51% correct) and substantially higher performance in the highest bands (≈70–80% correct), indicating increasing usefulness of mid-to-high frequency information for many environmental sounds.

Chang et al. (2018) reported a similar pattern in normal-hearing listeners: environmental sound identification was perfect for high-pass cutoffs below 1 kHz and nearly perfect for low-pass cutoffs above 1 kHz, with a mean crossover frequency of 1.5 kHz for environmental sounds. They also reported that under high-pass filtering, the largest decrement in ESR performance in normal-hearing listeners occurred when the cutoff frequency was increased to the highest tested value (8 kHz).

Although these in-quiet, broad-filtering paradigms do not predict the fine microstructure visible in our bandstop-ablation BIFs under speech masking, they provide converging evidence that frequency regions above ~1–1.5 kHz often carry substantial information for environmental sound identification, supporting the plausibility of the broad spectral regions emphasized by our masking-condition BIFs.

To enable a coarse comparison with these frequency-limiting studies, we derived an analogous crossover frequency from our masking-condition BIFs by treating the cumulative accuracy drop across ablated bands as a measure of total importance and identifying the frequency at which half of the cumulative drop occurred below and half above. Using the fold-averaged BIFs, this crossover frequency was 1.32 kHz for the human-trained model variant and 1.53 kHz for the ground-truth-trained variant. Although this metric is derived from narrowband bandstop ablations under speech masking rather than broad cutoff filtering in quiet, the resulting values fall squarely within the range implied by Gygi et al. (2004) and Chang et al. (2018), i.e., ≈1.3–1.5 kHz, and reinforce the general conclusion that information above ~1–1.5 kHz often contributes substantially to environmental sound identification.

D. Class-specific band-importance analyses

To clarify which sound categories contribute most strongly to the peaks observed in the across-sound (25-sound) BIFs, we estimated class-specific BIFs separately for each of the 25 environmental sounds using a newly generated evaluation set of 12 500 mixtures. For a given sound class, baseline recognition accuracy was computed using only mixtures in which that sound served as the target (approximately 500 mixtures per class in the present evaluation set). Band importance was then estimated by applying the same bandstop filtering procedure used for the across-sound BIFs, and importance was again quantified as the accuracy drop for each ERB band for each individual sound.

Because fold-specific BIFs for each model variant were highly consistent (see above) and because a class-specific analysis across all folds would be computationally expensive, class-specific BIFs were computed using one fold-specific model per variant. For each variant, we selected the fold whose overall performance most closely resembled human performance. For the human-trained variant this corresponded to the best-performing fold (fold 2), whereas for the ground-truth-trained variant—whose performance exceeded that of human listeners—this corresponded to the poorest-performing fold (fold 1). The resulting class-specific BIFs are shown as clustered heatmaps in Figs. 6 (human-trained) and 7 (ground-truth-trained).

A general observation across both heatmaps is that band importance was fairly concentrated; for many sounds, a small subset of bands produced meaningful performance decrements, whereas the majority of bandstop manipulations produced near-zero changes in accuracy. This sparse structure provides a mechanistic explanation for the microstructure observed in the across-sound BIFs; large peaks can arise when several sound classes exhibit strong dependence on a narrow frequency region, even if other classes show little dependence at that same band.

E. Limitations

Several limitations should be considered. First, the listener sample was strongly imbalanced by sex (44 female, 2 male). This limits generalizability and prevents evaluation of sex-related differences in ESR behavior or in the response distributions used to train the human-trained model. Replication in larger, more demographically diverse samples (and in listeners with hearing loss) is needed.

Second, the reported BIFs are conditional on three factors: (i) the stimulus corpus and its distribution of sound types (i.e., the across-sound BIF summarizes this specific set of 25 sounds), (ii) the task (closed-set, 25-alternative sound identity recognition rather than detection, open-set labeling, or scene analysis), and (iii) the masker/listening condition (competing speech with the specific mixing and presentation choices used here). Different sound sets, tasks, masker types (e.g., stationary noise, multi-talker babble), or spatial configurations could shift the apparent importance of frequency regions.

Third, importance was defined via a specific ablation method (ERB-spaced target-only bandstop filtering followed by remixing with an unmodified masker). This operationalization is useful for probing functional reliance on target information under masking, but it is not a direct measure of physiological auditory weighting and may be influenced by bandstop-induced changes in effective band-specific TMR or time-domain structure. Finally, other model families or training regimes could yield partially different patterns.