Communication sounds, such as human speech or animal vocalizations (calls), are typically produced with tremendous subject-to-subject and trial-to-trial variability. These sounds are also typically encountered in highly variable listening conditions—in the presence of noise, reverberations, and competing sounds. A central function of auditory processing is to extract the underlying meaningful signal being communicated so that appropriate behavioral responses can be produced. A key step in this process is a many-to-one mapping that bins communication sounds, perhaps carrying similar ‘meanings’ or associated with specific behavioral responses, into distinct categories. To accomplish this, the auditory system must generalize over the aforementioned variability in the production and transmission of communication sounds. We previously proposed, based on a model of visual categorization (Ullman et al., 2002), a theoretical model that identified distinctive acoustic features that were highly likely to be found across most exemplars of a category and were most contrastive with respect to other categories. Using these ‘most informative features (MIFs)’, the model accomplished auditory categorization with high accuracy (Liu et al., 2019). In that study, we also showed that features of intermediate lengths, bandwidths, and complexities were typically more informative for auditory categorization. In a second study, we further showed in a guinea pig (GP) animal model that neurons in the superficial layers of the primary auditory cortex (A1) demonstrated call-feature-selective responses and complex receptive fields that were consistent with model-predicted features, providing support for the model at the neurophysiological level (Montes-Lourido et al., 2021a). In this study, we investigated whether the feature-based model held true at a behavioral level, by determining whether the model, trained solely using natural GP calls, could predict GP behavioral performance in categorizing both natural calls as well as calls with altered spectral and temporal features.

Studies in a wide range of species have probed the impact of alterations to spectral and temporal cues on call recognition. For example, in humans, it has been shown that speech recognition relies primarily on temporal envelope cues based on experiments that measured recognition performance when subjects were presented with noise-vocoded speech at different spectral resolutions (Shannon et al., 1995; Smith et al., 2002). However, recognition is also remarkably resilient when the envelope is altered because of tempo changes—for example, word intelligibility is resilient to a large degree of time-compression of speech (Janse et al., 2003). Results from other mammalian species are broadly consistent with findings in humans. In gerbils, it has been shown that firing rate patterns of A1 neurons could be used to reliably classify calls that were composed of only four spectral bands (Ter-Mikaelian et al., 2013). In GPs, small neuronal populations have been shown to be resistant to such degradations as well (Aushana et al., 2018). Slow amplitude modulation cues have been proposed as a critical cue for the neuronal discriminability of calls (Souffi et al., 2020), but behaviorally, call identification can be resilient to large changes in these cues. For example, mice can discriminate between calls that have been doubled or halved in length (Neilans et al., 2014). This remarkable tolerance to cue variations might be related to the wide range of variations with which calls are produced in different behavioral contexts. For example, for luring female mice and during direct courtship, male mice modify many call parameters including sequence length and complexity (Chabout et al., 2015). In contrast, one study in GPs using naturalistic stimuli (a human footstep sound) demonstrated that while GPs discriminated time-compressed sounds from the natural sound on which they were trained, they did not distinguish between time-expanded sounds and natural sounds (Ojima and Horikawa, 2015). Along the spectral dimension, mouse call discrimination can be robust to changes in long-term spectra, including moderate frequency shifts and removal of frequency modulations (Neilans et al., 2014). Indeed, it has been suggested that the bandwidth of ultrasonic vocalizations is more important for communication than the precise frequency contours of these calls (Screven and Dent, 2016). Again, given that mice also modify the spectral features of their calls in a context-dependent manner (Chabout et al., 2015), it stands to reason that their perception of call identity is also robust to alterations of spectral features. In GPs using footstep stimuli, animals reliably discriminated sounds subjected to a band-reject filter from the natural sound (Ojima and Horikawa, 2015).

Overall, these studies suggest that for calls, in particular, information is encoded at multiple levels. Whereas the specific parameters of a given call utterance might carry rich information about the identity (Boinski and Mitchell, 1997; Miller et al., 2010; Gamba et al., 2012; Fukushima et al., 2015) and internal state of the caller as well as social context (Seyfarth and Cheney, 2003; Coye et al., 2016), call category identity encompasses all these variations. In some behavioral situations, listeners might need to be sensitive to these specific parameter variations—for example, for courtship, female mice have been shown to exhibit a high preference for temporal regularity of male calls (Perrodin et al., 2020). But in other situations, animals must and do generalize over this variability to extract call category identity, which is critical for providing an appropriate behavioral response. What mechanisms enable animals to generalize over this tremendous variability with which calls are heard and how they accomplish call categorization, however, is not well understood.

In this study, based on our earlier modeling and neurophysiological results (Liu et al., 2019; Montes-Lourido et al., 2021a), we hypothesized that animals can generalize over this production variability and achieve call categorization by detecting features of intermediate complexity within these calls. To test this hypothesis, we trained feature-based models and GPs to classify multiple categories of natural, spectrotemporally rich GP calls. We then tested the categorization performance of both the model and GPs with manipulated versions of the calls. We found that the feature-based model of auditory categorization, trained solely using natural GP calls, could explain about 50% of the overall variance of GP behavioral responses to manipulated calls. By comparing different model versions, we could derive further insight into possible behavioral strategies used by GPs to solve these call categorization tasks. Examining the factors contributing to high model performance in different conditions also provided insight into why a feature-based encoding strategy is highly advantageous. Overall, these results provide support at a behavioral level for a feature-based auditory categorization model, further validating our model as a novel and powerful approach to deconstructing complex auditory behaviors.

The acquisition of learned categorization behaviors likely involves two underlying processes—the acquisition of the knowledge of auditory categories, and the expression of this knowledge by learning the association between categories and reward outcomes (Kuchibhotla et al., 2019; Moore and Kuchibhotla, 2022). Many models have been developed to characterize the latter process, that is, the association of a stimulus with reward. Examples include models of classical conditioning and reinforcement learning (Rescorla and Wagner, 1972), models that account for the motivational state of subjects during behavior (Berditchevskaia et al., 2016), models that account for exploratory drive in subjects (Pisupati et al., 2021), or models that account for arousal state (de Gee et al., 2020). However, the former process of how knowledge of auditory categories is acquired, especially when categories are ‘non-compact’ or evident only in multiparametric spaces, is far less understood. Thus, while we briefly describe the behavioral acquisition of a vocalization categorization task for completeness (Figure 1—figure supplement 1 ), in this study, our goal was to determine the computational principles underlying the former process, that is, what features animals may use to acquire knowledge about acoustic stimulus category, and in particular, vocalization categories.

GPs learn to report call category in a Go/No-go task

We trained GPs on call categorization tasks using a Go/No-go task structure. Animals initiated trials by moving to the ‘home base’ region of the behavioral arena (Figure 1A, B). Stimuli were presented from an overhead speaker. On hearing Go stimuli, GPs were trained to move to a reward region, where they received a food pellet reward. The correct response to No-go stimuli was to remain in the home base. We trained two cohorts of GPs to categorize two pairs of call categories—Cohort 1 was trained on chuts (Go) versus purrs (No-go), calls that had similar spectral content (long-term spectral power) but different temporal (overall envelope) structure (Figure 1C), and Cohort 2 was trained on wheeks (Go) versus whines (No-go), calls that had similar temporal structure but different spectral content (Figure 1D). GPs were trained on this task over multiple short sessions every day (~6 sessions of ~40 trials each, ~10 min per session; see Materials and methods). On each trial, we presented a randomly chosen exemplar from an initial training set of 8 exemplars per category. We estimated hit rates and false alarm (FA) rates from all trials in a given day and computed a sensitivity index (d′). GPs were considered trained when d′ reliably crossed a threshold of 1.5. On average, GPs acquired this task after ~2–3 weeks of training (~4,000 total trials,~250 trials per exemplar; Figure 1—figure supplement 1). On the last day of training, GPs displayed a mean d′ of 1.94±0.26 for the chuts versus purrs task, and a mean d′ of 1.90±0.57 for the wheeks versus whines task.

Figure 1 with 1 supplement see all

Download asset Open asset

To gain insight into possible behavioral strategies that GPs might adopt to solve the categorization task, we examined trends of behavioral performance over the training period. Initially, GPs exhibited low hit rates as well as low FA rates, suggesting that they did not associate the auditory stimulus with reward (Figure 1—figure supplement 1D). Note that this initial phase was not recorded for the first cohort (chuts vs. purrs task, Figure 1—figure supplement 1A). Within 2–3 days, GPs formed a stimulus-reward association and exhibited ‘Go’ responses for all stimuli but did not discriminate between Go and No-go stimulus categories. This resulted in high hit rates as well as FA rates, but low d′. For the remainder of the training period, hit rates remained stable whereas FA rates gradually declined, suggesting that the improvements to d′ resulted from GPs learning to suppress responses to No-go stimuli (Figure 1—figure supplement 1A, B, D, E).

While these data were averaged over all sessions daily for further analyses, we noticed within-day trends in performance that might provide insight into the behavioral state of the GPs. We analyzed performance across intra-day sessions, averaged over 4 days after the animals acquired the task (Figure 1—figure supplement 1C, F). In early sessions, both hit rates and FA rates were high, suggesting that the GPs weighted the food reward highly, risking punishments (air puffs/time outs) in the process. In subsequent sessions, both the hit rate and FA rate declined, suggesting that the GPs shifted to a punishment-avoidance strategy. Despite these possible changes in decision criteria used by the GPs, they maintained consistent performance, as d′ remained consistent across sessions. Therefore, in all further analyses, we used d′ values averaged over all sessions as a performance metric. Note that these observations of possible motivational and other fluctuations are not intended to be captured by our model. Rather, our model solely focuses on what spectrotemporal features GPs can use to deduce stimulus categories.

A feature-based computational model can be trained to accomplish call categorization

In parallel, we extended a feature-based model that we previously developed for auditory categorization (Liu et al., 2019) to accomplish GP call categorization in a Go/No-go framework. Briefly, we implemented a three-layer model consisting of a spectrotemporal representation layer, a feature-detection (FD) layer, and a winner-take-all (WTA) layer. The spectrotemporal layer was a biophysically realistic model of the auditory periphery (Zilany et al., 2014). For the FD layer, we used greedy search optimization and information theoretic principles to derive a set of MIFs for each call type that was optimal for the categorization of that call type from all other call types (Figure 2A, B; Liu et al., 2019). We derived five distinct sets of MIFs for each call type that could accomplish categorization (see Materials and methods). We refer to models using these distinct MIF sets as different instantiations of the model.

Figure 2

Download asset Open asset

Call-specific MIF sets in the FD layer showed near-perfect performance (area under the curve, or AUC>0.97 for all 20 MIF sets [4 call categories×5 instantiations per category], mean=0.994) in categorizing target GP calls from other calls in the training data set. Similar to results from Liu et al., 2019, the number of MIFs for each instantiation of the model ranged from 8 to 20 (mean=16.5), with MIFs spanning ~3 octaves in bandwidth and ~110 ms in duration on average (Table 1). To assess the performance of the WTA layer based on these training data, we estimated d′ using equation 1 (Materials and methods). The WTA output also showed near-perfect performance for classifying the target from the distractor for both chuts versus purrs (mean d′=4.65) and wheeks versus whines (mean d′=3.69) tasks (Figure 2C).

Table 1

Call name	Instantiation	Number of MIFs	MIF duration (ms)(mean±std)	MIF bandwidth (octaves)(mean±std)
Chut	1, 2, 3, 4, 5	20, 20, 20, 20, 20	88±63, 106±53, 108±56, 109±64, 133±47	4.0±2.0, 4.4±1.2, 3.1±1.9, 3.7±1.9, 2.6±1.8
Purr	1, 2, 3, 4, 5	8, 9, 20, 20, 20	91±49, 83±43, 116±49, 116±56, 86±63	2.6±1.2, 2.8±1.2, 3.1±1.4, 3.2±1.5, 3.6±1.2
Wheek	1, 2, 3, 4, 5	8, 14, 13, 11, 12	144±47, 99±58, 104±68, 116±62, 114±65	2.3±1.6, 2.6±1.8, 2.9±2.2, 2.1±1.1, 2.5±1.7
Whine	1, 2, 3, 4, 5	20, 20, 15, 20, 20	109±55, 111±68, 133±37, 117±51, 108±70	3.5±1.8, 3.4±1.6, 2.6±1.4, 3.2±1.5, 3.9±1.6
Summary		16.5±4.7	109±57	3.2±1.7

Both GPs and the model generalize to new exemplars

To determine if GPs learned to report call category or if they simply remembered the specific call exemplars on which they were trained, we tested whether their performance generalized to a new set of Go and No-go stimuli (eight exemplars each) that the GPs had not encountered before. On each generalization day, we ran four sessions of ~40 trials each, with the first two sessions containing only training exemplars and the last two sessions containing only new exemplars. All GPs achieved a high-performance level (d′>1) to the new exemplars by generalization day 2 (Figure 3), that is, after being exposed to only a few repetitions of the new exemplars (~5 trials per new exemplar on generalization day 1). As an additional control to ensure that GPs did not rapidly learn reward associations for the new exemplars, for GPs performing the wheeks versus whines task (n=3), we also quantified generalization performance when the regular training exemplars and a second new set of exemplars were presented in an interleaved manner (400 trials with an 80/20 mix of training and new exemplars). GPs achieved d′>1 for new exemplars in this interleaved set as well, further supporting the notion that GPs were truly reporting call category.

Figure 3

Download asset Open asset

Figure 3—source data 1

Hit rates, false alarm rates, and d′ values of GPs and the feature-based model for generalization to new vocalization exemplars.

https://cdn.elifesciences.org/articles/78278/elife-78278-fig3-data1-v2.xlsx

Download elife-78278-fig3-data1-v2.xlsx

Similar to GPs, to test model generalization, we quantified model performance for new call exemplars (Figure 3B, D). Models using different MIF sets, that is, all instantiations of the model for chut, purr, wheek, and whine classification achieved high categorization performance (d′>1) for the new exemplars. In summary, GPs as well as the feature-based model could rapidly generalize to novel exemplars.

Both GPs and the model exhibit similar categorization-in-noise thresholds

Real-world communication typically occurs in noisy listening environments. To test how well GPs could maintain categorization in background noise, we assessed their performance when call stimuli were masked by additive white noise at several signal-to-noise ratios (SNRs) for both Go and No-go stimuli. Experiments were conducted in a block design, using a fixed SNR level per session (~40 trials) and testing 5 or 6 SNR levels each day. At the most favorable SNR (>20 dB), GPs exhibited high hit rates and low FA rates, leading to high d′ (>2) for both call groups (Figure 4). With increasing noise level (i.e., decreasing SNR), we observed a decrease in hit rate and an increase in FA, as expected, with a concomitant decrease in d′. To determine the effect of SNR value on d′, we constructed a generalized linear model with a logit link function to predict trial-by-trial behavioral outcomes, with stimulus type (Go or No-go), SNR value, and an interaction term as predictors, and animal ID as a random effect (see Equation 3, Materials and methods). We compared this full model to a null model consisting only of stimulus type as a predictor and animal ID as a random effect (see Equation 5, Materials and methods). This comparison revealed a strong effect of SNR value of d′ (chuts vs. purrs: χ²=124.6, p=2.2×10^–16; wheeks vs. whines: χ²=182.5; p=2.2×10^–16). At the most adverse SNR (–18 dB) for both call groups, hit and FA rates were similar, suggesting that the animals were performing at chance level. To estimate the SNR corresponding to the performance threshold (d′=1) for call categorization in noise, we fit a psychometric function to the behavioral d′ data (see Materials and methods). We obtained performance thresholds (SNR at which d′=1) for both the chuts versus purrs (–6.8 dB SNR) and wheeks versus whines (–11 dB SNR) tasks that were qualitatively similar to human speech discrimination performance in white noise (Phatak and Allen, 2007).

Figure 4

Download asset Open asset

Figure 4—source data 1

Hit rates, false alarm rates, and d′ values of GPs and the feature-based model for vocalization categorization in noise (at different SNRs).

https://cdn.elifesciences.org/articles/78278/elife-78278-fig4-data1-v2.xlsx

Download elife-78278-fig4-data1-v2.xlsx

We also tested the performance of the feature-based model (trained only on clean stimuli) on the same set of noisy stimuli as the behavioral paradigm. Model performance trends mirrored behavior, with a higher threshold for the chuts vs. purrs task (–5.4 dB SNR) compared to the wheeks versus whines task (–15 dB SNR). Similar to behavioral data, we confirmed a significant effect of SNR value on model performance using a trialwise GLM analysis (chuts vs. purrs: χ²=402.3, p=2.2×10^–16; wheeks vs. whines: χ²=203.7; p=2.2×10^–16). To compare behavioral data with the model, we fit a line to the d′ values obtained from the model plotted against the d′ values obtained from behavior, and computed the R² value. We used the mean absolute error (MAE) to quantify the absolute deviation between model and behavior d′ values. Although the model over-performed for the wheeks vs. whines task, it could explain a high degree of variance (R²=0.94 for both tasks) of GP call-in-noise categorization behavior.

Stimulus information might be available to GPs in short-duration segments of calls

Several studies across species, including humans (Marslen-Wilson and Zwitserlood, 1989; Salasoo and Pisoni, 1985), birds (Knudsen and Gentner, 2010; Toarmino et al., 2011), sea-lions (Pitcher et al., 2012), and mice (Holfoth et al., 2014), have suggested that the initial parts of calls might be the most critical parts for recognition. We reasoned that if that were the case for GPs as well, and later call segments did not add much information for call categorization, we might observe a plateauing of behavioral performance after a certain length of a call was presented. To test this, we presented call segments of different lengths (50–800 ms) beginning at the call onsets (Figure 5A, D) to estimate the minimum call duration required for successful categorization by GPs. Trials were presented in a randomized manner in sessions of ~40 trials, that is, each trial could be a Go or No-go stimulus of any segment length. We did not observe systematic changes to d′ values when comparing the first and second halves of the entire set of trials used for testing, demonstrating that the GPs were not learning the specific manipulated exemplars that we presented. GPs showed d′ values >1 for as small as 75 ms segments for both tasks, and as expected, the performance stabilized for all longer segment lengths (Figure 5B, C, E and F). As with the SNR experiment, we used a comparison between a full GLM including stimulus length value and a null GLM to evaluate the statistical significance of the effect of stimulus length on behavioral performance. We confirmed a significant effect of segment length on behavioral performance (chuts vs. purrs: χ²=58.7, p=1.8×10^–13; wheeks vs. whines: χ²=120.2; p=2.2×10^–16). These data suggest that short-duration segments of calls carry sufficient information for call categorization, at least in the tested one-vs.-one scenarios. The fact that call category can be extracted from the earliest occurrences of such segments suggests two possibilities: (1) a large degree of redundancy is present in calls, or (2) the repeated segments can be used to derive information beyond call category (e.g., caller identity or emotional valence).

Figure 5

Download asset Open asset

Figure 5—source data 1

Hit rates, false alarm rates, and d′ values of GPs and the feature-based model for the segment-length manipulation.

https://cdn.elifesciences.org/articles/78278/elife-78278-fig5-data1-v2.xlsx

Download elife-78278-fig5-data1-v2.xlsx

Model performance, however, only exceeded a d′ value of 1 for ~150 ms call segments, and performance only plateaued after a 200 ms duration (Figure 5C, F). The effect of segment length on model performance was statistically significant (chuts vs. purrs: χ²=613.2, p=1.8× 0^–13; wheeks vs. whines: χ²=775.4; p=2.2×10^–16). This observation could reflect the fact that the MIFs identified for categorization were on average about 110 ms long. Despite these differences, model performance was in general agreement with behavioral performance for both the chuts versus purrs and wheeks versus whines tasks (R²=0.674 and 0.444, respectively).

Temporal manipulations had little effect on model performance and GP behavior

To investigate the importance of temporal cues for GP call categorization, we introduced several gross temporal manipulations to the calls. We first started by changing the tempo of the calls, that is, stretching/compressing the calls without introducing alterations to the long-term spectra of calls (Figure 6A, D). This resulted in calls that were ~0.45, 0.5, ~0.56, ~0.63, ~0.77, ~1.43, 2.5, and 5 times the original lengths of the calls. As earlier, we presented stimuli in randomized order and verified that d′ did not vary systematically between the first and second half of trials, suggesting that the GPs were not learning new associations for the manipulated exemplars. GP behavioral performance remained above a d′ of 1 for all these perturbations, showing high hit rates and low FA rates (Figure 6B, E) leading to similar d′ across probed conditions (Figure 6C, F). Similar trialwise GLM analysis as earlier, but including the sign of the manipulation (compression vs. expansion) as an additional predictor in the full model, revealed a weak effect (based on the value of the χ² statistic, which roughly corresponds to effect size) of tempo shift on performance (chuts vs. purrs: χ²=46; p=5.7×10^–10; wheeks vs. whines: χ²=13.6; p=0.003).

Figure 6

Download asset Open asset

Figure 6—source data 1

Hit rates, false alarm rates, and d′ values of GPs and the feature-based model for tempo-shifted vocalizations.

https://cdn.elifesciences.org/articles/78278/elife-78278-fig6-data1-v2.xlsx

Download elife-78278-fig6-data1-v2.xlsx

Similarly, model performance also remained above d′ of 1 for all tempo manipulations. We emphasize that the MIFs used in the model were trained only on natural calls, and were not temporally compressed or stretched to match the test stimuli. GLM analysis as above revealed a weak effect of tempo shift on model d′ for chuts versus purrs (χ²=43.3; p=2.14×10^–9), and no effect for wheeks versus whines (χ²=0.322; p=0.96). Note that while the model qualitatively captured GP behavioral trends, we obtained low R² values likely because of the weak effect of tempo shift values on both animal and model behaviors. The variability of the observed behavior was likely dominated by non-stimulus-related factors such as motivation, exploratory drive, or arousal (Berditchevskaia et al., 2016; de Gee et al., 2020; Pisupati et al., 2021), factors not explicitly modeled here. However, the relatively low MAE for the tempo manipulations (comparable with MAEs of the SNR manipulation which showed high R² values) suggests a broad correspondence between the model and behavior.

The tempo manipulations lengthened or shortened both syllables and inter-syllable intervals (ISIs). Because a recent study in mice (Perrodin et al., 2020) suggested that the regularity of ISI values might be crucial for the detection of male courtship songs by female mice, we next asked whether GPs used individual syllables or temporal patterns of calls for call categorization. First, as a low-level control, we replaced the ISIs of calls with silence instead of the low level of background noise present in recordings to ensure that GPs were not depending on any residual ISI information (silent ISI). Second, since many call categories show a distribution of ISI durations (Figure 7A, E), we replaced the ISI durations in a call with ISI values randomly sampled from the ISI distribution of the same call category (random ISI). The hit and FA rates for both silent and random ISI stimuli were comparable to the regular calls for both categorization tasks (Figure 7C, G), and thus, no significant difference in d′ values was observed across these conditions (Figure 7D, H; repeated measures ANOVA; p=0.536 for chuts vs. purrs and p=0.365 for wheeks vs. whines). Consistent with these behavioral trends, model performance was also largely unaffected by these ISI manipulations.

Figure 7

Download asset Open asset

Figure 7—source data 1

Hit rates, false alarm rates, and d′ values of GPs and the feature-based model for various ISI manipulations (silent ISI, random ISI, chimeric calls).

https://cdn.elifesciences.org/articles/78278/elife-78278-fig7-data1-v2.xlsx

Download elife-78278-fig7-data1-v2.xlsx

Finally, because the Go/No-go stimulus categories vary in their ISI distributions (Figure 7A, E), particularly for chuts versus purrs, we generated chimeric calls with syllables of one category and ISI values of the other category (e.g., chut syllables with purr ISIs). Since we combined properties of two call categories, we presented chimeric stimuli in a catch-trial design (see Materials and methods) and compared the Go response rates using syllable identity as the label for a category. While the response rates were marginally lower for the chimeric chuts (chut syllables with purr ISI values) compared to regular chuts (paired t-test; p=0.039), responses were unaltered for regular and chimeric purrs (paired t-test; p=0.415), chimeric wheeks (paired t-test; p=0.218), and chimeric whines (paired t-test; p=0.099) (Figure 7B, F). We did not test the model with chimeric stimuli because ‘Go’ and ‘Nogo’ category labels could not be assigned.

As a more drastic manipulation, we tested the effects of temporally reversing the calls (Figure 8A, D). Given that both chuts and purrs are calls with temporally symmetric spectrotemporal features, compared to natural calls, we observed no changes in the hit and FA rates (Figure 8B) or d′ values for reversed calls (Figure 8C; paired t-test; p=0.582). Wheeks and whines, however, show strongly asymmetric spectrotemporal features. Interestingly, reversal did not significantly affect the categorization performance for this task as well (Figure 8E, F; paired t-test; p=0.151). The model also maintained robust performance (d′>1) for call reversal conditions but with an ~18% decrease in d′ compared to behavior. Overall, these results suggest that GP behavioral performance is tolerant to temporal manipulations such as tempo changes, ISI manipulations, and call reversal, and this tolerance can be largely captured by the feature-based model.

Figure 8

Download asset Open asset

Figure 8—source data 1

Hit rates, false alarm rates, and d′ values of GPs and the feature-based model for natural and reversed vocalizations.

https://cdn.elifesciences.org/articles/78278/elife-78278-fig8-data1-v2.xlsx

Download elife-78278-fig8-data1-v2.xlsx

Spectral manipulations cause similar degradation in model performance and GP behavior

Because temporal manipulations did not significantly affect GP behavioral or model classification performance, we reasoned that categorization was primarily driven by within-syllable spectral cues. To ascertain the impact of spectral manipulations on call categorization, we varied the fundamental frequency (F0) of the calls from one octave lower (–50%) to one octave higher (+100%) than the regular calls without altering call lengths (Figure 9A and D). As earlier, we verified that d′ did not vary systematically between the first and second half of trials, suggesting that the GPs were not learning new associations for the manipulated exemplars. Both increases and decreases to the F0 of the calls significantly affected behavioral performance, revealed by a GLM analysis as described earlier (chuts vs. purrs: χ²=106.7; p=2.2×10^–16; wheeks vs. whines: χ²=158.2; p=2.2×10^–16). Particularly, we saw a rise in FA rates (Figure 9B) as the F0 deviated farther from the natural values, leading to a significant drop in d′ values for several conditions (Figure 9C). Model performance was also significantly affected by F0 shift magnitude (chuts vs. purrs: χ²=93.9; p=2.2×10^–16; wheeks vs. whines: χ²=456; p=2.2×10^–16). As earlier, MIFs used in the model were trained only on natural calls, and were not frequency-shifted to match the test stimuli. Model performance mirrored behavioral trends as evidenced by high R² (0.723 and 0.468 for chuts vs. purrs and wheeks vs. whines, respectively) and low MAE values.

Figure 9

Download asset Open asset

Figure 9—source data 1

Hit rates, false alarm rates, and d′ values of GPs and the feature-based model for F0-shifted vocalizations.

https://cdn.elifesciences.org/articles/78278/elife-78278-fig9-data1-v2.xlsx

Download elife-78278-fig9-data1-v2.xlsx

Finally, because wheeks and whines differ in their spectral content at high frequencies (Figure 1D), we asked whether GPs exclusively used the higher harmonics of wheeks to accomplish the categorization task. To answer this question, we low-pass filtered both wheeks and whines at 3 kHz (Figure 10A), removing the higher harmonics of the wheeks while leaving the fundamental relatively unaffected. Although GP performance showed a decreasing trend for the filtered calls (Figure 10B, C), it was not significantly different from regular calls (paired t-test; p=0.169), indicating that the higher harmonics might be an important but not the sole cue used by GPs for the task. Similar to behavior, the model performed slightly poorly but above a d′ of 1 in the low-pass filtered condition.

Figure 10

Download asset Open asset

Figure 10—source data 1

Hit rates, false alarm rates, and d′ values of GPs and the feature-based model for natural and low-pass filtered vocalizations.

https://cdn.elifesciences.org/articles/78278/elife-78278-fig10-data1-v2.xlsx

Download elife-78278-fig10-data1-v2.xlsx

Feature-based model explains a high degree of variance in GP behavior

The feature-based model was developed purely based on theoretical principles, made minimal assumptions, was trained only on natural GP calls, and had no access to GP behavioral data. For training the model, we used exemplars that clearly provided net evidence for the presence of one category or the other (Figure 3C; green and red tick marks in Figure 11A, D). We tested the model (and GPs), however, with manipulated stimuli that spanned a large range of net evidence values (histograms in Figure 11A, D), with many stimuli close to the decision boundary (blue ticks correspond to an SNR value of –18 dB). Despite the difficulty imposed by this wide range of manipulations, the model explained a high degree of variance in GP behavior as evidenced by high R² and low MAE across individual paradigms (call manipulations) as well as overall (Figure 11B–F; R²=0.60 for chuts vs. purrs and 0.37 for wheeks vs. whines; using model and behavior d′ values pooled across all tasks).

Figure 11 with 2 supplements see all

Download asset Open asset

To determine whether simpler models based on the overall spectral content of vocalizations could explain some of these results, we built a support vector machine (SVM) classifier that attempted to classify vocalizations based on the long-term spectra of calls (see Materials and methods). Similar to the feature-based model, a WTA stage was implemented by comparing the outputs of the target-call SVM model and the distractor-call SVM model for each input call (response=Go if target-SVM output>distractor SVM output). The spectrum-based SVM model achieved high performance on natural vocalizations (d′=3.48 for novel chuts vs. purrs and 2.65 for novel wheeks vs. whines), but failed to capture many aspects of GP responses to manipulated calls, such as the robust performance at low SNRs and modulation of performance by F0-shifted calls. Overall, this resulted in a much lower fraction of the variance explained (Figure 11—figure supplement 1). These analyses illustrate that although a simpler model might successfully classify natural vocalizations, applying the model to manipulated stimuli reveals the shortcomings of the spectrum-based model in capturing GP behavioral trends, further highlighting the need for using spectrotemporal features for classification.

Comparing models with different training procedures yields insight into GP behavioral strategy

The high explanatory power of the feature-based model could be leveraged to gain further insight into what information the GPs were using or learning to accomplish these categorization tasks. On the one hand, because GPs are exposed to these call categories from birth, the GPs may simply be employing the features that they have already acquired for call categorization over their lifetimes to solve our specific categorization tasks. The model presented so far is aligned with this possibility—we trained features to categorize one call type from all other call types (one vs. many categorization) and used a large number of call exemplars for training. Alternatively, GPs could be de-novo learning stimulus features that distinguished between the particular Go and No-go exemplars we presented during training. To test this possibility, we re-trained the model only using the eight exemplars of the targets and distractors that we used to train GPs for one versus one categorization. When tested on manipulated calls, the one versus one model typically performed poorly compared to the original one versus many model. Compared to the one versus many model, the one versus one model was less consistent with behavior as indicated by lower R² (Figure 11B, E) and higher MAE values (Figure 11C, F). One explanation for this result could be that eight exemplars of each category are insufficient to adequately train the model. But the one versus one model achieved suprathreshold (d′>1) performance on both the training and generalization call exemplars, suggesting that these training data were, to some degree, sufficient for classifying natural calls. Furthermore, if GPs were indeed learning stimulus features from the training data set, they would also face the same constraints on training data volume as the model. A second explanation is that the one versus many model better matches GP behavior because rather than re-learning new task-specific features, GPs might be using call features that they had acquired previously over their lifespan to solve our call categorization task. These results also suggest that training a feature-based categorization system (in-silico or in-vivo) on exemplars that capture within-category variability is critical to obtain a system that can flexibly adapt and maintain robust performance to unheard stimuli that exhibit large natural or artificial variations.

The effect of training our model on the one versus many categorization task using a large number of call exemplars for training was that the model learned features that truly captured the within-class and outside-class variability of calls. This resulted in a model that accurately predicted GP performance across a range of stimulus manipulations. To understand how the model was able to achieve robustness to stimulus variations, and to gain insight into how GPs may flexibly weight features differently across the various stimulus manipulations, we examined the relative detection rates of various model MIFs across different stimulus paradigms in which we observed strong behavioral effects (Figure 12, Figure 12—figure supplement 1).

Figure 12 with 1 supplement see all

Download asset Open asset

Figure 12—source data 1

Characteristics of most informative features (CF, bandwidth, and duration) and their relative detection rates for the various stimulus manipulations.

https://cdn.elifesciences.org/articles/78278/elife-78278-fig12-data1-v2.xlsx

Download elife-78278-fig12-data1-v2.xlsx

In Figure 12, we plot the relative detection rates of MIFs (discs, ~20 MIFs per instantiation) from different model instantiations (colors, 5 instantiations) for the different call types used. That is, we computed the difference between the rate of detection of each MIF (for each call category) in response to within-category and outside-category stimuli, and plotted this difference (disc areas) as a function of MIF tuning properties (CF, bandwidth, and duration). For the generalization stimuli, that is, new natural calls on which the model had not been trained, almost all MIFs showed relatively large net detection rates which resulted in plots (Figure 12A–D, left panels) with discs of about equal area. For example, all chut MIFs were detected at high rates in novel chut calls (within-category calls), and detected at low rates in non-chut call types (outside-category calls). Note, however, that the learned MIFs are spread out across a range of CFs, bandwidths, and durations. Given these data alone, one might argue that learning ~20 MIFs per call category is highly redundant, and that high performance could be achieved using only a subset of these MIFs. But examining which features maintain high relative detection rates in other stimulus paradigms underscores the utility of learning this wide feature set. When we added white noise to the stimulus, low-CF features showed higher relative detection rates (Figure 12A, B, right top) and thus contributed more towards categorization. This could likely be attributed to GP calls having high power at low frequencies, resulting in more favorable local SNRs at lower frequencies. But when we altered stimulus F0, high-CF features contributed more towards categorization (Figure 12A, B, right bottom). Similarly, low-duration, high-bandwidth features contributed more when categorizing time-restricted calls, whereas high-duration, low-bandwidth features contributed more when categorizing F0-shifted calls (Figure 12C, D). That the model quantitatively matched GP behavior suggests that a similar strategy might be employed by GPs as well. Note that our contention is not that the precise MIFs obtained in our model are also the precise MIFs used by the GPs—indeed, we were able to train several distinct MIF sets that were equally proficient at categorizing calls. Rather, we are proposing a framework in which GPs learn intermediate-complexity features that account for within-category variability and best contrast a call category from all other categories, and similar to the model, recruit different subsets of these features to solve different categorization tasks.

In this study, we trained GPs to report call categories using an appetitive Go/No-go task. We then tested GP call categorization when we challenged them with spectrally and temporally manipulated calls. We found that GPs maintained their call categorization across a wide range of gross temporal manipulations such as changes to tempo and altered ISI distributions. In contrast, GP behavior was strongly affected by altering the F0 of calls. In parallel, to determine which features GPs might be using to extract knowledge of call category from stimuli, we extended a previously developed feature-based model of auditory categorization by adding a WTA feature-integration stage that enabled us to obtain a categorical decision from the model on a trial-by-trial basis. We trained the model using natural GP calls. When we challenged the model with the identical stimuli used in the GP experiments, we obtained model responses that explained ~50% of the overall variance in GP behavior. This result suggests that learning to generalize over the variability in natural call categories resulted in the model being able to generalize over some artificial call manipulations. That the model tracks GP performance on these manipulated stimuli is a key observation that supports our reasoning that GPs learn model-like features to generalize over call variability and accomplish call categorization. We had previously reported electrophysiological support for the feature-based model by demonstrating that a large fraction of neurons in the superficial layers of auditory cortex exhibited feature-selective responses, resembling the FD stage of the model (Montes-Lourido et al., 2021a). The results described in the present manuscript lend further support to the model at a behavioral level. Taken together, these studies strongly suggest how a spectral content-based representation of sounds at lower levels of auditory processing may be transformed into a goal-directed representation at higher processing stages by extracting and integrating task-relevant features.

The feature-based model was highly predictive of GP behavior, although it was conceptualized from purely theoretical considerations, trained only using natural GP calls, and implemented without access to any behavioral data. We developed the feature-based model to capture how knowledge of auditory categories is learned by encoding informative features, but did not account for how this knowledge is expressed during behavior. Rather, we made minimal assumptions and used a WTA framework with a static decision criterion (with a small amount of error). Insights from our behavioral observations could be used to further refine the model by adding a realistic behavioral expression stage. For example, our data indicated that GPs altered their behavioral strategy over the course of multiple sessions within a given day. This could possibly reflect an early impulsivity in their decision-making brought on by food deprivation (evidenced by a FA rate) that gradually switches to a punishment-avoidance strategy with increasing satiation (although d′ remains consistent across sessions). Additional arousal-dependent effects on behavior are likely. Nevertheless, the fact that the current model could explain ~50% of the variance in behavioral trends we observed suggests that the fundamental strategy employed by the model—that of detecting features of intermediate complexity to generalize over within-category variability—also lies at the core of GP behavior.

Furthermore, we could leverage the model to gain insight into possible behavioral strategies used by GPs in performing the tasks. For example, we could compare models trained to categorize calls in one versus many or one versus one conditions to ask which scenario was more consistent with GP behavior: (1) whether the GPs used prior features that they acquired over their lifetimes to categorize a given call type from all other calls, or (2) whether GPs were de-novo learning new features to solve the particular categorization task on which they were trained. While the relatively lower volume of data used to train the one vs. one model is a potential confound, the model trained on call features that distinguish a particular call from all other calls was more closely aligned with GP behavioral data, supporting the first possibility. Examining how different subsets of features could be employed to solve different categorization tasks revealed possible strategies that GPs might use to flexibly recruit different feature representations to solve our tasks. While we have used GPs as an animal model for call categorization in this study, we have previously shown that the feature-based model shows high performance across species (GPs, marmosets, and macaques), and that neurons in the primary auditory cortex (A1) of marmosets also exhibit feature-selective responses (Liu et al., 2019). Thus, it is likely that our model reflects general principles that are applicable across species, and offers a powerful new approach to deconstruct complex auditory behaviors.

On the behavioral side, previous GP studies have largely used conditioned avoidance (e.g., Heffner et al., 1971) or classical conditioning using aversive stimuli (e.g., Edeline et al., 1993) to study simple tone detection and discrimination behaviors. Studies probing GP behaviors using more complex stimuli are rare (e.g., Ojima et al., 2012; Ojima and Horikawa, 2015). Our study of GP call categorization behavior using multiple spectrotemporally rich call types and parametric manipulations of spectral and temporal features offers comprehensive insight into cues that are critical for call categorization and builds significantly on previous studies. First, we showed that GPs can categorize calls in challenging SNRs, and that thresholds vary depending on the call types to be categorized. We demonstrated that information for GP call categorization was available in short-duration segments of calls, and consistent with some previous studies in other species (Holfoth et al., 2014; Knudsen and Gentner, 2010; Marslen-Wilson and Zwitserlood, 1989; Pitcher et al., 2012), GPs could extract call category information soon after call onset. GP call perception was robust to large temporal manipulations, such as reversal and larger changes to tempo than have been previously tested (Neilans et al., 2014). These results are also consistent with the resilience of human word identification to large tempo shifts (Janse et al., 2003). Our finding that GP call categorization performance is robust to ISI manipulations is also not necessarily inconsistent with results from mice (Perrodin et al., 2020); in that study, while female mice were found to strongly prefer natural calls compared to calls with ISI manipulations, it is possible that mice still identified the call category correctly. For gross spectral manipulations, we found that GP call categorization was robust to a larger range of F0-shifts than have been previously tested (Neilans et al., 2014). Critically, for all but one of these manipulations, the feature-based model captured GP behavioral trends with surprising accuracy both qualitatively and quantitatively.

An analysis of model deviation from behavior could suggest a roadmap for future improvements to our model that could yield further insight into auditory categorization. The one paradigm where we observed a systematic under-performance of the model compared to GP behavior was when we presented call segments of varying lengths from call onset. While the GPs were able to accomplish categorization by extracting information from as little as 75 ms segments, the model required considerably more information (~150 ms). This is likely because the model was based on the detection of informative features that were on average of ~110 ms duration, which were identified from an initial random set of features that could be up to 200 ms in duration. We set this initial limit based on an upper limit estimated from electrophysiological data recorded from A1 ( Montes-Lourido et al., 2021a). We consciously did not impose further restrictions on feature duration or bandwidth to ensure that the model did not make any assumptions based on observed behavior. Upon observing this deviation from behavior, to test whether restricting feature lengths to shorter durations would further improve model performance, we repeated the modeling by constraining the initial feature length to a maximum of 75 ms (the lowest duration for which GPs show above-threshold performance). We found that the constrained MIF model better matched GP behavior on the segment-length task (R² of 0.62 and 0.58 for the chuts vs. purrs and wheeks vs. whines tasks) with the model exceeding d′=1 for 75 ms segments for most tested cases. The constrained MIF model maintained similarity to behavior for the other manipulations as well (Figure 11—figure supplement 2), and yielded higher overall R² values (0.66 for chuts vs. purrs, 0.51 for wheeks vs. whines), explaining ~59% of the variance in GP behavior. This result illustrates how behavioral experiments can provide constraints for the further development of theoretical models in the future.

We also observed the over-performance of the model compared to behavior in some paradigms. Some of this over-performance might be explained by the fact that the model does not exhibit motivation changes, and so forth, as outlined above. A second source of this over-performance might arise from the fact that our model integrates evidence from the FD stage perfectly, that is, we take the total evidence for the presence of a call category to be the weighted sum of the log-likelihoods of all detected features (counting detected features only once) over the entire stimulus, and do not explicitly model a leaky integration of feature evidence over time, as is the case in evidence-accumulation-to-threshold models (Cheadle et al., 2014; Keung et al., 2020). Future improvements to the model could include a realistic feature-integration stage, where evidence for a call category is generated when a feature is detected and degrades with a biologically realistic time constant. In this case, a decision threshold could be reached before the entire stimulus is heard, but model parameters would need to be derived from or fit to observed behavioral data (Glaze et al., 2015).

How do the proposed model stages map onto the auditory system? In an earlier study, we provided evidence that feature detection likely occurs in the superficial layers of A1, in that a large fraction of neurons in this stage exhibits highly selective call responses and complex spectrotemporal receptive fields (Montes-Lourido et al., 2021a). How and at what stage these features are combined to encode a call category remains an open question. Neurons in A1 can acquire categorical or task-relevant responses to simple categories, for example, low versus high tone frequencies, or low versus high temporal modulation rates, with training (Bao et al., 2004; Fritz et al., 2005). In contrast, categorical responses to more complex sounds or non-compact categories only seem to arise at the level of secondary or higher cortical areas or the prefrontal cortex (Russ et al., 2008; Yin et al., 2020), which may then modulate A1 via descending connections. These results, taken together with studies that demonstrate enhanced decodability of call identity from the activity of neurons in higher cortical areas (Fukushima et al., 2014; Grimsley et al., 2012; Grimsley et al., 2011), suggest that secondary ventral-stream cortical areas, such as the ventral-rostral belt in GPs, are promising candidates as the site of evidence integration from call features. The WTA stage may be implemented via lateral inhibition at the same level using similar mechanisms as has been suggested in the primary visual cortex (Chettih and Harvey, 2019) or may require a further upstream layer. Further experiments are necessary to explore these questions.

The feature-based model we developed offers a trade-off between performance and biological interpretability. Modern deep neural network (DNN) based models can attain human-level performance (e.g., in vision: Rajalingham et al., 2015, in audition: Kell et al., 2018) but what features are encoded at the intermediate network layers remain somewhat hard to interpret. These models also typically require vast quantities of training data. In contrast, our model is based on an earlier model for visual categorization (Ullman et al., 2002) that is specifically trained to detect characteristic features that contrast the members of a category from non-members. Thus, we can develop biological interpretations for what features are preferably encoded and more importantly, why certain features are more advantageous to encode. Because the features used in the model are the most informative parts of the calls themselves, they can be identified without a parametric search. This approach is especially well-suited for natural sounds such as calls that are high-dimensional and difficult to parameterize. We are restricted, however, in that we do not know all possible categorization problems that are relevant to the animal. By choosing well-defined categorization tasks that are ethologically critical for an animal’s natural behavior (such as call categorization in the present study), we can maximize the insight that we can derive from these experiments as it pertains to a range of natural behaviors. In the visual domain, the concept of feature-based object recognition has yielded insight into how human visual recognition differs from modern machine vision algorithms (Ullman et al., 2016). Our results lay the foundation for pursuing an analogous approach for understanding auditory recognition in animals and humans.

All experimental procedures conformed to the NIH Guide for the use and care of laboratory animals and were approved by the Institutional Animal Care and Use Committee of the University of Pittsburgh (protocol number 21069431).

All data generated or analyzed during this study are included in the manuscript and supporting file; Source Data files have been provided for Figures 3 - 12.

1. 1. Aushana Y
   2. Souffi S
   3. Edeline JM
   4. Lorenzi C
   5. Huetz C

   (2018) Robust neuronal discrimination in primary auditory cortex despite degradations of spectro-temporal acoustic details: comparison between guinea pigs with normal hearing and mild age-related hearing loss

   Journal of the Association for Research in Otolaryngology 19:163–180.

   https://doi.org/10.1007/s10162-017-0649-1

   • PubMed
   • Google Scholar
2. 1. Bao S
   2. Chang EF
   3. Woods J
   4. Merzenich MM

   (2004) Temporal plasticity in the primary auditory cortex induced by operant perceptual learning

   Nature Neuroscience 7:974–981.

   https://doi.org/10.1038/nn1293

   • PubMed
   • Google Scholar
3. Preprint

   1. Bates D
   2. Machler M
   3. Bolker B
   4. Walker S

   (2014) Fitting Linear Mixed-Effects Models Using Lme4

   arXiv.

   https://www.jstatsoft.org/article/view/v067i01

   • Google Scholar
4. 1. Berditchevskaia A
   2. Cazé RD
   3. Schultz SR

   (2016) Performance in a GO/NOGO perceptual task reflects a balance between impulsive and instrumental components of behaviour

   Scientific Reports 6:1–5.

   https://doi.org/10.1038/srep27389

   • PubMed
   • Google Scholar
5. 1. Boinski S
   2. Mitchell CL

   (1997) Chuck vocalizations of wild female squirrel monkeys (Saimiri sciureus) contain information on caller identity and foraging activity

   International Journal of Primatology 18:975–993.

   https://doi.org/10.1023/A:1026300314739

   • Google Scholar
6. 1. Chabout J
   2. Sarkar A
   3. Dunson DB
   4. Jarvis ED

   (2015) Male mice song SYNTAX depends on social contexts and influences female preferences

   Frontiers in Behavioral Neuroscience 9:76.

   https://doi.org/10.3389/fnbeh.2015.00076

   • PubMed
   • Google Scholar
7. 1. Cheadle S
   2. Wyart V
   3. Tsetsos K
   4. Myers N
   5. de Gardelle V
   6. Herce Castañón S
   7. Summerfield C

   (2014) Adaptive gain control during human perceptual choice

   Neuron 81:1429–1441.

   https://doi.org/10.1016/j.neuron.2014.01.020

   • PubMed
   • Google Scholar
8. 1. Chettih SN
   2. Harvey CD

   (2019) Single-Neuron perturbations reveal feature-specific competition in V1

   Nature 567:334–340.

   https://doi.org/10.1038/s41586-019-0997-6

   • PubMed
   • Google Scholar
9. 1. Coye C
   2. Zuberbühler K
   3. Lemasson A

   (2016) Morphologically structured vocalizations in female diana monkeys

   Animal Behaviour 115:97–105.

   https://doi.org/10.1016/j.anbehav.2016.03.010

   • Google Scholar
10. 1. de Gee JW
    2. Tsetsos K
    3. Schwabe L
    4. Urai AE
    5. McCormick D
    6. McGinley MJ
    7. Donner TH

    (2020) Pupil-linked phasic arousal predicts a reduction of choice bias across species and decision domains

    eLife 9:e54014.

    https://doi.org/10.7554/eLife.54014

    • PubMed
    • Google Scholar
11. 1. Edeline JM
    2. Pham P
    3. Weinberger NM

    (1993) Rapid development of Learning-induced receptive field plasticity in the auditory cortex

    Behavioral Neuroscience 107:539–551.

    https://doi.org/10.1037//0735-7044.107.4.539

    • PubMed
    • Google Scholar
12. 1. Fritz JB
    2. Elhilali M
    3. Shamma SA

    (2005) Differential dynamic plasticity of A1 receptive fields during multiple spectral tasks

    The Journal of Neuroscience 25:7623–7635.

    https://doi.org/10.1523/JNEUROSCI.1318-05.2005

    • PubMed
    • Google Scholar
13. 1. Fukushima M
    2. Saunders RC
    3. Leopold DA
    4. Mishkin M
    5. Averbeck BB

    (2014) Differential coding of conspecific vocalizations in the ventral auditory cortical stream

    The Journal of Neuroscience 34:4665–4676.

    https://doi.org/10.1523/JNEUROSCI.3969-13.2014

    • PubMed
    • Google Scholar
14. 1. Fukushima M
    2. Doyle AM
    3. Mullarkey MP
    4. Mishkin M
    5. Averbeck BB

    (2015) Distributed acoustic cues for caller identity in macaque vocalization

    Royal Society Open Science 2:150432.

    https://doi.org/10.1098/rsos.150432

    • PubMed
    • Google Scholar
15. 1. Gamba M
    2. Colombo C
    3. Giacoma C

    (2012) Acoustic cues to caller identity in lemurs: a case study

    Journal of Ethology 30:191–196.

    https://doi.org/10.1007/s10164-011-0291-z

    • Google Scholar
16. 1. Glaze CM
    2. Kable JW
    3. Gold JI

    (2015) Normative evidence accumulation in unpredictable environments

    eLife 4:08825.

    https://doi.org/10.7554/eLife.08825

    • PubMed
    • Google Scholar
17. Book

    1. Green DM
    2. Swets JA

    (1966)

    Signal Detection Theory and Psychophysics

    Wiley.

    • Google Scholar
18. 1. Grimsley JMS
    2. Palmer AR
    3. Wallace MN

    (2011) Different representations of tooth chatter and purR call in guinea pig auditory cortex

    Neuroreport 22:613–616.

    https://doi.org/10.1097/WNR.0b013e3283495ae9

    • PubMed
    • Google Scholar
19. 1. Grimsley JMS
    2. Shanbhag SJ
    3. Palmer AR
    4. Wallace MN

    (2012) Processing of communication calls in guinea pig auditory cortex

    PLOS ONE 7:e51646.

    https://doi.org/10.1371/journal.pone.0051646

    • PubMed
    • Google Scholar
20. 1. Heffner R
    2. Heffner H
    3. Masterton B

    (1971) Behavioral measurements of absolute and frequency-difference thresholds in guinea pig

    The Journal of the Acoustical Society of America 49:1888–1895.

    https://doi.org/10.1121/1.1912596

    • PubMed
    • Google Scholar
21. 1. Holfoth DP
    2. Neilans EG
    3. Dent ML

    (2014) Discrimination of partial from whole ultrasonic vocalizations using a go/no-go task in mice

    The Journal of the Acoustical Society of America 136:3401–3409.

    https://doi.org/10.1121/1.4900564

    • PubMed
    • Google Scholar
22. 1. Janse E
    2. Nooteboom S
    3. Quené H

    (2003) Word-level intelligibility of time-compressed speech: prosodic and segmental factors

    Speech Communication 41:287–301.

    https://doi.org/10.1016/S0167-6393(02)00130-9

    • Google Scholar
23. 1. Kell AJE
    2. Yamins DLK
    3. Shook EN
    4. Norman-Haignere SV
    5. McDermott JH

    (2018) A task-optimized neural network replicates human auditory behavior, predicts brain responses, and reveals a cortical processing hierarchy

    Neuron 98:630–644.

    https://doi.org/10.1016/j.neuron.2018.03.044

    • PubMed
    • Google Scholar
24. 1. Keung W
    2. Hagen TA
    3. Wilson RC

    (2020) A divisive model of evidence accumulation explains uneven weighting of evidence over time

    Nature Communications 11:2160.

    https://doi.org/10.1038/s41467-020-15630-0

    • PubMed
    • Google Scholar
25. 1. Knudsen DP
    2. Gentner TQ

    (2010) Mechanisms of song perception in oscine birds

    Brain and Language 115:59–68.

    https://doi.org/10.1016/j.bandl.2009.09.008

    • PubMed
    • Google Scholar
26. 1. Kuchibhotla KV
    2. Hindmarsh Sten T
    3. Papadoyannis ES
    4. Elnozahy S
    5. Fogelson KA
    6. Kumar R
    7. Boubenec Y
    8. Holland PC
    9. Ostojic S
    10. Froemke RC

    (2019) Dissociating task acquisition from expression during learning reveals latent knowledge

    Nature Communications 10:1–3.

    https://doi.org/10.1038/s41467-019-10089-0

    • PubMed
    • Google Scholar
27. 1. Liu ST
    2. Montes-Lourido P
    3. Wang X
    4. Sadagopan S

    (2019) Optimal features for auditory categorization

    Nature Communications 10:1302.

    https://doi.org/10.1038/s41467-019-09115-y

    • PubMed
    • Google Scholar
28. 1. Marslen-Wilson W
    2. Zwitserlood P

    (1989) Accessing spoken words: the importance of word onsets

    Journal of Experimental Psychology 15:576–585.

    https://doi.org/10.1037/0096-1523.15.3.576

    • Google Scholar
29. 1. Miller CT
    2. Mandel K
    3. Wang X

    (2010) The communicative content of the common marmoset phee call during antiphonal calling

    American Journal of Primatology 72:974–980.

    https://doi.org/10.1002/ajp.20854

    • PubMed
    • Google Scholar
30. 1. Montes-Lourido P
    2. Kar M
    3. David SV
    4. Sadagopan S

    (2021a) Neuronal selectivity to complex vocalization features emerges in the superficial layers of primary auditory cortex

    PLOS Biology 19:e3001299.

    https://doi.org/10.1371/journal.pbio.3001299

    • PubMed
    • Google Scholar
31. 1. Montes-Lourido P
    2. Kar M
    3. Kumbam I
    4. Sadagopan S

    (2021b) Pupillometry as a reliable metric of auditory detection and discrimination across diverse stimulus paradigms in animal models

    Scientific Reports 11:3108.

    https://doi.org/10.1038/s41598-021-82340-y

    • PubMed
    • Google Scholar
32. 1. Moore S
    2. Kuchibhotla KV

    (2022) Slow or sudden: re-interpreting the learning curve for modern systems neuroscience

    IBRO Neuroscience Reports 13:9–14.

    https://doi.org/10.1016/j.ibneur.2022.05.006

    • PubMed
    • Google Scholar
33. 1. Neilans EG
    2. Holfoth DP
    3. Radziwon KE
    4. Portfors CV
    5. Dent ML

    (2014) Discrimination of ultrasonic vocalizations by CBA/caj mice (Mus musculus) is related to spectrotemporal dissimilarity of vocalizations

    PLOS ONE 9:e85405.

    https://doi.org/10.1371/journal.pone.0085405

    • PubMed
    • Google Scholar
34. 1. Ojima H
    2. Taira M
    3. Kubota M
    4. Horikawa J

    (2012) Recognition of non-harmonic natural sounds by small mammals using competitive training

    PLOS ONE 7:e51318.

    https://doi.org/10.1371/journal.pone.0051318

    • PubMed
    • Google Scholar
35. 1. Ojima H
    2. Horikawa J

    (2015) Recognition of modified conditioning sounds by competitively trained guinea pigs

    Frontiers in Behavioral Neuroscience 9:373.

    https://doi.org/10.3389/fnbeh.2015.00373

    • PubMed
    • Google Scholar
36. 1. Paraouty N
    2. Charbonneau JA
    3. Sanes DH

    (2020) Social learning exploits the available auditory or visual cues

    Scientific Reports 10:14117.

    https://doi.org/10.1038/s41598-020-71005-x

    • PubMed
    • Google Scholar
37. Preprint

    1. Perrodin C
    2. Verzat C
    3. Bendor D

    (2020) Courtship Behaviour Reveals Temporal Regularity Is a Critical Social Cue in Mouse Communication

    bioRxiv.

    https://doi.org/10.1101/2020.01.28.922773

    • Google Scholar
38. 1. Phatak SA
    2. Allen JB

    (2007) Consonant and vowel confusions in speech-weighted noise

    The Journal of the Acoustical Society of America 121:2312–2326.

    https://doi.org/10.1121/1.2642397

    • PubMed
    • Google Scholar
39. 1. Pisupati S
    2. Chartarifsky-Lynn L
    3. Khanal A
    4. Churchland AK

    (2021) Lapses in perceptual decisions reflect exploration

    eLife 10:e55490.

    https://doi.org/10.7554/eLife.55490

    • PubMed
    • Google Scholar
40. 1. Pitcher BJ
    2. Harcourt RG
    3. Charrier I

    (2012) Individual identity encoding and environmental constraints in vocal recognition of pups by Australian sea lion mothers

    Animal Behaviour 83:681–690.

    https://doi.org/10.1016/j.anbehav.2011.12.012

    • Google Scholar
41. 1. Rajalingham R
    2. Schmidt K
    3. DiCarlo JJ

    (2015) Comparison of object recognition behavior in human and monkey

    The Journal of Neuroscience 35:12127–12136.

    https://doi.org/10.1523/JNEUROSCI.0573-15.2015

    • PubMed
    • Google Scholar
42. Book

    1. Rescorla RA
    2. Wagner AR

    (1972)

    A theory of pavlovian conditioning: variations in the effectiveness of reinforcement and nonreinforcement

    In: Black AH, Prokasy WF, editors. Classical Conditioning II. Appleton-Century-Crofts. pp. 64–99.

    • Google Scholar
43. 1. Russ BE
    2. Orr LE
    3. Cohen YE

    (2008) Prefrontal neurons predict choices during an auditory same-different task

    Current Biology 18:1483–1488.

    https://doi.org/10.1016/j.cub.2008.08.054

    • PubMed
    • Google Scholar
44. 1. Salasoo A
    2. Pisoni DB

    (1985) Interaction of knowledge sources in spoken word identification

    Journal of Memory and Language 24:210–231.

    https://doi.org/10.1016/0749-596X(85)90025-7

    • PubMed
    • Google Scholar
45. 1. Screven LA
    2. Dent ML

    (2016) Discrimination of frequency modulated sweeps by mice

    The Journal of the Acoustical Society of America 140:1481–1487.

    https://doi.org/10.1121/1.4962223

    • PubMed
    • Google Scholar
46. 1. Seyfarth RM
    2. Cheney DL

    (2003) Meaning and emotion in animal vocalizations

    Annals of the New York Academy of Sciences 1000:32–55.

    https://doi.org/10.1196/annals.1280.004

    • PubMed
    • Google Scholar
47. 1. Shannon RV
    2. Zeng FG
    3. Kamath V
    4. Wygonski J
    5. Ekelid M

    (1995) Speech recognition with primarily temporal cues

    Science 270:303–304.

    https://doi.org/10.1126/science.270.5234.303

    • PubMed
    • Google Scholar
48. 1. Smith ZM
    2. Delgutte B
    3. Oxenham AJ

    (2002) Chimaeric sounds reveal dichotomies in auditory perception

    Nature 416:87–90.

    https://doi.org/10.1038/416087a

    • PubMed
    • Google Scholar
49. 1. Souffi S
    2. Lorenzi C
    3. Varnet L
    4. Huetz C
    5. Edeline JM

    (2020) Noise-sensitive but more precise subcortical representations coexist with robust cortical encoding of natural vocalizations

    The Journal of Neuroscience 40:5228–5246.

    https://doi.org/10.1523/JNEUROSCI.2731-19.2020

    • PubMed
    • Google Scholar
50. 1. Ter-Mikaelian M
    2. Semple MN
    3. Sanes DH

    (2013) Effects of spectral and temporal disruption on cortical encoding of gerbil vocalizations

    Journal of Neurophysiology 110:1190–1204.

    https://doi.org/10.1152/jn.00645.2012

    • PubMed
    • Google Scholar
51. Book

    1. Toarmino C
    2. Neilans EG
    3. Dent ML

    (2011) Identification of Conspecific Calls by Budgerigars

    Melopsittacus Undulatus.

    https://doi.org/10.1037/e525792013-005

    • Google Scholar
52. 1. Ullman S
    2. Vidal-Naquet M
    3. Sali E

    (2002) Visual features of intermediate complexity and their use in classification

    Nature Neuroscience 5:682–687.

    https://doi.org/10.1038/nn870

    • PubMed
    • Google Scholar
53. 1. Ullman S
    2. Assif L
    3. Fetaya E
    4. Harari D

    (2016) Atoms of recognition in human and computer vision

    PNAS 113:2744–2749.

    https://doi.org/10.1073/pnas.1513198113

    • PubMed
    • Google Scholar
54. 1. Wichmann FA
    2. Hill NJ

    (2001) The psychometric function: I. fitting, sampling, and goodness of fit

    Perception & Psychophysics 63:1293–1313.

    https://doi.org/10.3758/bf03194544

    • PubMed
    • Google Scholar
55. 1. Yin P
    2. Strait DL
    3. Radtke-Schuller S
    4. Fritz JB
    5. Shamma SA

    (2020) Dynamics and hierarchical encoding of non-compact acoustic categories in auditory and frontal cortex

    Current Biology 30:1649–1663.

    https://doi.org/10.1016/j.cub.2020.02.047

    • PubMed
    • Google Scholar
56. 1. Zilany MSA
    2. Bruce IC
    3. Carney LH

    (2014) Updated parameters and expanded simulation options for a model of the auditory periphery

    The Journal of the Acoustical Society of America 135:283–286.

    https://doi.org/10.1121/1.4837815

    • PubMed
    • Google Scholar

Author details

1. Manaswini Kar

   1. Center for Neuroscience at the University of Pittsburgh, Pittsburgh, United States
   2. Center for the Neural Basis of Cognition, Pittsburgh, United States
   3. Department of Neurobiology, University of Pittsburgh, Pittsburgh, United States

   Contribution

   Conceptualization, Data curation, Formal analysis, Investigation, Visualization, Methodology, Writing – review and editing

   Competing interests

   No competing interests declared

2. Marianny Pernia

   Department of Neurobiology, University of Pittsburgh, Pittsburgh, United States

   Contribution

   Investigation, Methodology, Writing – review and editing

   Contributed equally with

   Kayla Williams

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-9889-3577

3. Kayla Williams

   Department of Neurobiology, University of Pittsburgh, Pittsburgh, United States

   Contribution

   Investigation, Writing – review and editing

   Contributed equally with

   Marianny Pernia

   Competing interests

   No competing interests declared

4. Satyabrata Parida

   Department of Neurobiology, University of Pittsburgh, Pittsburgh, United States

   Contribution

   Data curation, Software, Formal analysis, Investigation, Visualization, Methodology, Writing – original draft, Writing – review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-2896-2522

5. Nathan Alan Schneider

   1. Center for Neuroscience at the University of Pittsburgh, Pittsburgh, United States
   2. Center for the Neural Basis of Cognition, Pittsburgh, United States

   Contribution

   Software, Investigation, Methodology

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-9145-5427

6. Madelyn McAndrew

   1. Center for the Neural Basis of Cognition, Pittsburgh, United States
   2. Department of Neurobiology, University of Pittsburgh, Pittsburgh, United States

   Contribution

   Investigation

   Contributed equally with

   Isha Kumbam

   Competing interests

   No competing interests declared

7. Isha Kumbam

   Department of Neurobiology, University of Pittsburgh, Pittsburgh, United States

   Contribution

   Investigation

   Contributed equally with

   Madelyn McAndrew

   Competing interests

   No competing interests declared

8. Srivatsun Sadagopan

   1. Center for Neuroscience at the University of Pittsburgh, Pittsburgh, United States
   2. Center for the Neural Basis of Cognition, Pittsburgh, United States
   3. Department of Neurobiology, University of Pittsburgh, Pittsburgh, United States
   4. Department of Bioengineering, University of Pittsburgh, Pittsburgh, United States
   5. Department of Communication Science and Disorders, University of Pittsburgh, Pittsburgh, United States

   Contribution

   Conceptualization, Resources, Software, Formal analysis, Supervision, Funding acquisition, Validation, Methodology, Writing – original draft, Project administration, Writing – review and editing

   For correspondence

   vatsun@pitt.edu

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-1116-8728

• 677

  views

• 140

  downloads

• 5

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.