A distributed network of noise-resistant neurons in the central auditory system

Background noise strongly penalizes auditory perception of speech in humans or vocalizations in animals. Despite this, auditory neurons successfully detect and discriminate behaviorally salient sounds even when the signal-to-noise ratio is quite poor. Here, we collected neuronal recordings in cochlear nucleus, inferior colliculus, auditory thalamus, primary and secondary auditory cortex in response to vocalizations presented either against a stationary or a chorus noise. Using a clustering approach, we provide evidence that five behaviors exist at each level of the auditory system from neurons with high fidelity representations of the target, named target-specific neurons, mostly found in inferior colliculus and thalamus, to neurons with high fidelity representations of the noise, named masker-specific neurons mostly found in cochlear nucleus in stationary noise but in similar proportions in each structure in chorus noise. This indicates that the neural bases of auditory perception in noise rely on a distributed network along the auditory system.


Introduction
Here, we present evidence demonstrating that the target-specific neurons are in higher proportions in 92 inferior colliculus and thalamus in both noises, whereas the masker-specific neurons are found mostly 93 in the cochlear nucleus in stationary noise but in similar proportions in each structure in a noise com-94 posed of a mixture of conspecific vocalizations that we will name "chorus noise". We also provide 95 evidence that the noise-type sensitivity -that is the ability to switch category from a given background 96 noise to another -although present at each level of the auditory system in small proportions, is mostly 97 detected in the inferior colliculus and the thalamus. 98 99

100
From a database of 2334 multi-unit recordings collected in the five investigated auditory structures, 101 several criteria were used to include each neuronal recording in our analyses (see Table 1). A record-102 ing had to show significant responses to pure tones (see Methods section) and an evoked firing rate 103 significantly above spontaneous firing rate (200 ms before each original vocalization) for at least one 104 of the four original vocalizations (Fig. 1A illustrates their temporal envelopes and spectrograms). The-105 se four vocalizations were presented in quiet and embedded either in a vocalization-shaped stationary 106 noise (Fig. 1B) or in a chorus noise (Fig. 1C) using three SNRs. We selected neurons showing re-107 sponses at the three SNRs both in stationary and chorus noise in order to derive systematically six EI 108 values for each neuronal recording. To determine a significance level of the EI value, we computed an 109 EI Surrogate value for each recording (see Methods section) and included only the recordings with at least 110 one of the six EI values significantly higher than the EI Surrogate . Applying these criteria, we selected a 111 total of 1267 recordings (Table 1)   Quantification of the evoked firing rate and the response trial-to-trial temporal reliability in stationary 131 and chorus noise confirmed these observations (Fig. 2B-E). In both noises, the lower the SNR, the 132 lower the evoked firing rate and the trial-to-trial reliability. More precisely, in both noises, the de-133 crease in evoked firing rate was significant as early as the +10 dB SNR in all auditory structures ex-134 cept in VRB for which the decrease was significant at 0 dB SNR (Fig. 2B-2D; for the stationary noise 135 ( Fig. 2B): one-way ANOVA: F CN(4,1944) =315; F CNIC(4,1694) =265.5; F MGv(4,989) =174.9; F A1(4,1304) =95.8; 136 F VRB(4,399) =40.8, p<0.001; with post-hoc paired t tests, p<0.001; for the chorus noise (Fig. 2D): one-137 way ANOVA: F CN(4,1944) =108.7; F CNIC(4,1694) =92.7; F MGv(4,989) =93.8; F A1(4,1304) =74.7; F VRB(4,399) =24.2, 138 p<0.001; with post-hoc paired t tests, p<0.001). Similarly, in both noises, the trial-to-trial temporal 139 reliability (quantified by the CorrCoef index) was significantly decreased as early as the +10 dB SNR 140 in CN, MGv and A1 whereas in CNIC and VRB, the decrease was significant only at the 0 dB SNR 141   We found that the mean EI values were higher in the inferior colliculus and thalamus than in the coch-170 lear nucleus and cortex, except in chorus noise at -10 dB SNR, which strongly impacted all neuronal 171 responses at each stage (Fig. 3). Figure 3A displays  We next aimed at characterizing categories of neurons that display a particular behavior in noise in 188 relation to fidelity of neural representation either of the target or of the noise. Therefore, a clustering 189 analysis was performed on the entire database, i.e. the 1267 recordings obtained in the five structures, 190 separately for the stationary and for the chorus noise. 191

222
We also opted for cluster names using more symmetric terms. The neurons keeping a high-fidelity 223 representation of the vocalizations despite the presence of noise will be called the target-ultraspecific 224 or target-specific neurons, those keeping a high-fidelity representation of the noise across the SNRs 225 will be called masker-specific neurons and those showing no preference will be called non-specific 226 neurons. One last category of neurons is characterized by sensitivity to the signal-to-noise ratio and 227 therefore will be called SNR-dependent. This change in cluster names gives an equivalent role to tar-228 get-specific and noise-sensitive neurons, named here masker-specific neurons, since in some ethologi-229 cal conditions they could play a functional role as important as the target-specific neurons. noise. Figure 5C shows the mean EI values in stationary noise for these five clusters across the three 236 SNRs and the percentage of neurons in each cluster is displayed in figure 5D. Approximately 10% of 237 the neurons are target-ultraspecific characterized, on average, by EI values greater than 0.5 at +10 and 238 0 dB SNRs. More than 25% are target-specific characterized, on average, by EI values greater than 239 0.2 at +10 and 0 dB SNRs (Fig. 5C). About 5% of the neurons are SNR-dependent and more than 240 40% of the total population has EI values around 0 at all SNRs which corresponds to the non-specific 241 neurons. More than 10% of the auditory neurons have negative EI values at the three SNRs and corre-242 spond to masker-specific neurons. Figures 5H and 5I show the mean EI values for these five clusters 243 in the chorus noise with, roughly, similar proportions of the five clusters as in the stationary noise. 244 However, in the chorus noise there was a decrease in the proportions of neurons in the target-245 ultraspecific (from 10% to 7.5%) and target-specific (from 27% to 20%) clusters associated with an 246 increase in the proportion of SNR-dependent neurons (from 6.5% to 19.5%), whereas the proportion 247 of neurons in the non-specific cluster remained similar (42-39.5%). Note also that in the chorus noise, 248 the two clusters of target-specific neurons showed, on average, lower EI values at the 0 dB SNR than 249 in stationary noise (compared Fig. 5C and Fig. 5H). Based upon these quantifications, it is clear that, 250 in the entire auditory system, the chorus noise impacted more the neuronal responses than the station-251 ary noise. 252 Next, we determined the proportions of each cluster in a given structure. In stationary noise, target-253 ultraspecific and target-specific neurons were mostly present in the inferior colliculus and thalamus, 254 while the three other groups of neurons classified as SNR-dependent, non-specific, and masker-255 specific were mostly present in the cochlear nucleus and in the two cortical fields. For each auditory 256 structure, the percentage of neurons from each cluster is presented in the stationary noise (

285
Therefore, in both noises, the neurons with a high fidelity representation of the target were mostly 286 present in the inferior colliculus and thalamus. The non-specific neurons showing no preference either 287 for the target or the noise were found in majority in the cochlear nucleus and in the auditory cortex. 288 The SNR-dependent neurons represented a small fraction of neurons in stationary noise but were more 289 numerous in the chorus noise. One interesting feature is that, in both types of noise, the proportion of 290 these SNR-dependent neurons decreases progressively as one ascends in the auditory system. Finally, 291 In the marmoset auditory cortex, Ni and colleagues (2017) have pointed out that the neuronal behavior 298 in noise can be context-dependent: the behavior of a given neuron in a particular noise does not pre-299 dict its behavior in another noise. Is this a property that characterizes cortical neurons, or is it a prop-300 erty that exists at all levels of the auditory system? 301 Overall, we found that the neuronal response behaviors at all levels of the auditory pathway were part-302 ly, but not completely, preserved in different types of noise (Fig. 6A). On the whole database of 1267 303 recordings, about 50% of the target-ultraspecific and 40% of the target-specific neurons in the station-304 ary noise remained so in the chorus noise; most of the SNR-dependent (73.5%) and non-specific neu-305 rons (65.5%) in the stationary noise remained also in the same category in the chorus noise. Only the 306 masker-specific neurons tended to switch category to mostly became non-specific neurons. Consider-307 ing the proportions of neurons in each category in stationary noise, as indicated on the x-axis of figure 308 6A, around 45% of all recordings switched category in chorus noise (see Fig. 6B). Figure 6B shows 309 that the target-ultraspecific, target-specific and masker-specific neurons were the three categories with 310 the highest percentages of cluster changes (χ 2 p<0.001). We used a bootstrap procedure to have a bet-311 ter estimation of the percentage of cluster changes (see Methods section). Briefly, for each recording, 312 and from the 20 trials obtained for each stimulus, we resampled 20 trials (allowing repetitions), 313 recomputed the Extraction Index and reallocated each resampled recording in the closest cluster. This 314 entire procedure was performed 100 times for each recording. We assumed that in a given type of 315 noise, a recording could change category because of its response variability and/or because it was lo-316 cated very close to a border between two clusters, independently to the change in noise type. With the 317 resampled data obtained from the bootstrap procedure, we determined, for each cluster type (Fig. 6B), 318 and each structure (Fig. 6C), the percentage of cluster change averaged for the two types of noise. 319 For each cluster type, except for the non-specific cluster, around 25% of resampled data switched cat-320 egory, which is indicated by the grey section (Fig. 6B). For the non-specific cluster, less than 20% 321 changed category with the resampled data (Fig. 6B). When subtracting these bootstrapped percent-   We performed the same analysis for each auditory structure (Figure 6 -figure supplement 1, Fig. 6C). 342 On average, between 21 and 31 out of 100 bootstrapped data per recording changed cluster in cortical 343 and subcortical structures in both noises (Figure 6 -figure supplement 1). Thus, with only the 344 resampled data, there is, on average, an important fraction of the recordings changed cluster in each 345 structure suggesting that the response variability and/or the proximity of the borders between two 346 clusters induced cluster changes. Figure 6C presents the mean percentage of cluster change obtained 347 from the physiological data for each auditory structure: for the two cortical areas about 38% of the 348 neurons changed categories, it was 57% in the MGv, 50% in the inferior colliculus and 43% in the 349 cochlear nucleus. With bootstrapped data, in all auditory structures, between 20 to 30% of the neurons 350 changed category, which is indicated by the grey section. When subtracting these bootstrapped per-351 centages, we obtained the bootstrap-corrected values of the percentages of cluster changes in each 352 structure, which dropped the percentage of cluster change obtained with physiological data to only 10-353 30% (Fig. 6C). The percentage of neurons changing category was higher in the inferior colliculus and 354 thalamus (31-22%) than in the cochlear nucleus and the auditory cortex (14% on average; Fig. 6C). 355 Therefore, the noise-type sensitivity is present at each stage of the auditory system but represents 356 small proportions. Furthermore, the inferior colliculus and the thalamus had the most noise-type sensi-357 tive neurons, the fewest were found in the auditory cortex.

366
One of the questions that should be addressed is whether or not the robustness to noise detected here 367 is correlated to the response characteristics obtained with the original vocalizations in quiet. For ex-368 ample, one can envision that the EI value (and its evolution across the three SNRs) is related either to 369  calizations. Therefore, in both noises, we looked for potential correlations between the EI values and 371 the response parameters to the original vocalizations. We focused only on the neurons exhibiting the 372 same neuronal behavior in the two noises (n Total =685, n CN =222, n CNIC =169, n MGv =81, n A1 =164, 373 n VRB =49) to look for correlations between stable EI values and other parameters obtained in quiet. For 374 these neurons, the evoked firing rate was significantly higher in the subcortical structures than in the 375 cortex (unpaired t-test, lowest p-value p<0.001; Fig. 7A). The CorrCoef values were significantly 376 higher in CNIC and MGv compared to A1 and VRB (Fig. 7B), and the MI Individual values obtained at 377 the subcortical level were also significantly higher than at the cortical level (unpaired t-test, highest 378 p<0.001 between the cortex and the subcortical structures; Fig. 7C). 379 Generally, both for the stationary noise (Fig. 7D-F) and for the chorus noise (Fig. 7G-I for the stationary noise). Except in one case, the same significant correlations were found in the cho-385 rus noise (see Table 2). Note that the strongest correlations between the EI and the CorrCoef values 386 were obtained in the inferior colliculus (0.65 and 0.63 in stationary and chorus noise respectively). 387 These results suggest that the trial-to-trial temporal reliability is the factor the more correlated with the 388 robustness to noise, especially in the inferior colliculus, the auditory structure where we detected the 389 largest number of target-specific neurons. 390

416
Here, we demonstrate that the processing of noisy vocalizations by neurons in the entire auditory sys-417 tem can be described by a limited number of neuronal behaviors found in different proportions de-418 pending on the auditory structure and the type of the masking noise. Target-specific neurons were 419 detected at each level of the auditory system but were in higher proportions at the collicular and tha-420 lamic level. For these neurons, the trial-to-trial temporal reliability of the responses in quiet is corre-421 lated to the robustness of responses in noise. In terms of proportions, the highest fidelity representa-422 tion of the target or noise was found at the subcortical level whereas at the cortical level, the majority 423 of neuronal responses showed no preference for the target or the noise suggesting that cortical neurons Our results do not indicate a progressive evolution from sensitivity to robustness to noise along the 462 central auditory system. However, based upon the proportion of SNR-dependent neurons, one inte-463 resting result is that, in both types of noise, these neurons decreased progressively as one ascends in 464 the auditory system which is in line with the idea of a progressive construction of an invariant repre-465 sentation of acoustic signals in noise described by Rabinowitz and colleagues (2013). 466 We also showed that the higher the trial-to-trial temporal reliability of the responses in quiet, the high-467 er the robustness of the neurons in noise, especially in the IC where we detected the highest propor-468 tion of target-specific neurons. In fact, it was previously reported that the firing rate and the temporal 469 reliability of IC neurons decreased when vocalizations were presented in natural stationary noise, but 470 they were still efficiently detected target stimuli in noise (Lesica and Groethe, 2008). Together, this 471 suggests that the more temporally precise are the synaptic inputs converging on a particular neuron, 472 the more the responses of this neuron are robust in background noise. 473 The subcortical robustness to noise described here might be surprising given that numerous studies 474 have pointed out the robustness of cortical representations in noise. However, Las and colleagues 475 (2005) reported that A1, MGB and IC neurons can detect low-intensity target tones in a louder fluctu-476 ating masking noise and display the so-called "phase-locking suppression", that is the interruption of 477 phase-locking to the temporal envelope of background noise. The last result indicates that, as cortical 478 neurons, both IC and MGB neurons have the ability to detect low-intensity target sounds in louder 479 background noise (even at -15 or -35 dB SNR) and the robustness of some of our subcortical neurons 480 may stem from this ability to detect target vocalizations even at SNR as low as the -10 dB SNR. 481 Robust perception of target sounds probably also requires a robust representation of competing sounds 482 (here, masking noise). This can be the functional role of the masker-specific neurons, which are po-483 tentially crucial to determine the characteristics of the noise type and to provide an accurate represen-er proportion in the CN in stationary noise, but they became more numerous and in equivalent propor-486 tion in all structures in chorus noise. Therefore, the noise representation can be based upon the neu-487 ronal activity in the cochlear nucleus in stationary noise, whereas this representation can be more dis-488 tributed in the chorus noise potentially because some of the target-ultraspecific or the target-specific 489 neurons in stationary noise became masked-specific neurons in chorus noise, due to its spectro-490 temporal acoustic richness. 491 In our results, the five categories rather form a continuum with no clear boundaries between clusters, 492 which inevitably led us to « impose » the clustering. Nonetheless, despite the lack of precise bounda-493 ries, 75% of the neurons remained assigned to the same cluster when the bootstrap procedure was per-494 formed ( Fig. 6B-C), suggesting a relatively good reliability of the classification. Also, these five cate-495 gories do represent distinct neuronal behaviors in the two types of noise, which have been previously 496 described at the cortical level in awake marmoset (Ni et al., 2017). Here, in the auditory cortex of 497 anesthetized animals, we found these same global behaviors, and our results show that these catego-498 ries also exist at the subcortical level. Thus, the cortical representation of noisy signals by different 499 neuronal categories characterized either by the preference of the target, the noise, a sensitivity to SNR 500 or an absence of these three acoustic features, is independent of the state of alertness of the animal. 501 We can wonder if choosing 7, 8, or 9 clusters, would have highlighted other neuron behaviors. A part 502 of the answer is provided by figure 4, which shows that with 6 clusters, similar behaviors re-appear 503 suggesting that a larger number of clusters would have been non-informative. As we collected multi-504 unit recordings composed of 2-6 shapes of action potentials, it is possible that more specific behaviors 505 might have been missed in our analyses. This is potentially the case at the cortical level where a large 506 number of cell types have been described (Ascoli et al., 2008;DeFelipe et al., 2013) and also in the 507 cochlear nucleus (Cant and Benson, 2003;Kuenzel, 2019). However, based on the output of small 508 groups of neurons, five neuronal behaviors seem to be present in noise at all the levels of the auditory 509 system. 510 511 Noise-type sensitivity is present, in small proportions, at each stage of the auditory system but 512 mostly in the inferior colliculus and thalamus 513 514 Ni and colleagues (2017) found about two-thirds of cortical neurons switching category from one 515 background noise to another, suggesting that the majority of cortical neurons have a behavior specific 516 to the type of noise. Although we initially found between 40 and 60% of such neurons in the different 517 auditory structures, the bootstrap procedure indicated that more realistic percentages should be much 518 lower, potentially between 10-30%. Also, the response variability, which is probably much larger in 519 awake than in anesthetized animals, can explain the difference between our results and those of Ni and 520 colleagues (2017). Here, these neurons were detected in auditory cortex but were found in higher pro-521 portions in subcortical structures. This indicates that only a small fraction of neurons display a behav-522 ior specific to a particular noise. We preferred to call this phenomenon noise-type sensitivity rather 523 context-dependence (proposed by Ni and colleagues, 2017) because the latter refers to situations 524 where the same stimulus is presented in different contexts; whereas here inserting target stimuli in two 525 types of noise generated different auditory streams. As mentioned by Ni  larly changes, probably because neural adaptation suppresses the representation of noise features, a 543 mechanism that seems to be independent of the attentional focus of the listeners. 544 Here, we propose that the noise-robustness observed in many studies at the cortical level stems, at 545 least partially, from subcortical mechanisms ( In the MGB, IC and CN, the recordings were obtained using 16 channel multi-electrode arrays (Neu-615 roNexus) composed of one shank (10 mm) of 16 electrodes spaced by 110 µm and with conductive 616 site areas of 177µm 2 . The electrodes were advanced vertically (for MGB and IC) or with a 40° angle 617 (for CN) until evoked responses to pure tones could be detected on at least 10 electrodes. 618 All thalamic recordings were from the ventral part of MGB (see above surgical procedures) and all 619 displayed latencies < 9ms. At the collicular level, we distinguished the lemniscal and non-lemniscal 620 divisions of IC based on depth and on the latencies of pure tone responses. We excluded the most su-select recordings from the central nucleus of IC (CNIC). At the level of the cochlear nucleus, the re-623 cordings were collected from both the dorsal and ventral divisions. 624 The raw signal was amplified 10,000 times (TDT Medusa). It was then processed by an RX5 multi-625 channel data acquisition system (TDT). The signal collected from each electrode was filtered (610-626 10000 Hz) to extract multi-unit activity (MUA). The trigger level was set for each electrode to select 627 the largest action potentials from the signal. On-line and off-line examination of the waveforms sug-628 gests that the MUA collected here was made of action potentials generated by a few neurons at the 629 vicinity of the electrode. However, as we did not used tetrodes, the result of several clustering algo- Acoustic stimuli were generated using MatLab, transferred to a RP2.1-based sound delivery system 641 (TDT) and sent to a Fostex speaker (FE87E). The speaker was placed at 2 cm from the guinea pig's 642 right ear, a distance at which the speaker produced a flat spectrum (± 3 dB) between 140 Hz and 36 643 kHz. Calibration of the speaker was made using noise and pure tones recorded by a Bruel and Kjaer 644 microphone 4133 coupled to a preamplifier BandK 2169 and a digital recorder Marantz PMD671. 645 The Time-Frequency Response Profiles (TFRP) were determined using 129 pure-tones frequencies 646 covering eight octaves (0.14-36 kHz) and presented at 75 dB SPL. The tones had a gamma envelop 647 given by ( ) = ( ! ! whistle were selected. As shown in figure 1a (lower panels), despite the fact the maximal energy of 657 the four selected whistle was in the same frequency range (typically between 4 and 26 kHz), these 658 calls displayed slight differences in their spectrograms. In addition, their temporal (amplitude) enve-659 lopes clearly differed as shown by their waveforms (Fig. 1a, upper panels). 660 The four whistles were also presented in two frozen noises ranging from 10 to 24,000 Hz. To generate 661 these noises, recordings were performed in the colony room where a large group of guinea pigs were 662 housed (30-40; 2-4 animals/cage). Several 4-seconds of audio recordings were added up to generate 663 the "chorus noise", which power spectrum was computed using the Fourier transform. This spectrum 664 was then used to shape the spectrum of a white Gaussian noise. The resulting vocalization-shaped 665 stationary noise therefore matched the "chorus-noise" audio spectrum, which explains why some fre-666 quency bands were over-represented in the vocalization-shaped stationary noise. Figures 1b et 1c dis-667 play the spectrograms of the four whistles in the vocalization-shaped stationary noise (1b) and in the 668 chorus noise (1c) with a SNR of +10 dB SPL, 0 dB SPL, -10 dB SPL. The last spectrograms of these 669 two figures represent the noises only. 670

671
As inserting an array of 16 electrodes in a brain structure almost systematically induces a deformation 672 of this structure, a 30-minutes recovering time lapse was allowed for the structure to return to its ini-673 tial shape, then the array was slowly lowered. Tests based on measures of time-frequency response 674 profiles (TFRPs) were used to assess the quality of our recordings and to adjust electrodes' depth. For 675 auditory cortex recordings (A1 and VRB), the recording depth was 500-1000 µm, which corresponds 676 to layer III and the upper part of layer IV according to Wallace and Palmer (2008). For thalamic re-677 cordings, the NeuroNexus probe was lowered about 7mm below pia before the first responses to pure 678 tones were detected. 679 When a clear frequency tuning was obtained for at least 10 of the 16 electrodes, the stability of the 680 tuning was assessed: we required that the recorded neurons displayed at least three successive similar 681 TFRPs (each lasting 6 minutes) before starting the protocol. When the stability was satisfactory, the 682 protocol was started by presenting the acoustic stimuli in the following order: We first presented the 683 four whistles at 75 dB SPL in their natural versions (in quiet), followed by their masked versions pre-684 sented against the chorus and the vocalization-shaped stationary noise at 65, 75 and 85 dB SPL. Thus, 685 the level of the original vocalizations was kept constant (75 dB SPL), and the noise level was in-686 creased (65, 75 and 85 dB SPL). In all cases, each vocalization was repeated 20 times. Presentation of 687 this entire stimulus set lasted 45 minutes. The protocol was re-started either after moving the electrode

Data analysis 690
Quantification of responses to pure tones 691 The TFRP were obtained by constructing post-stimulus time histograms for each frequency with 1 ms 692 time bins. The firing rate evoked by each frequency was quantified by summing all the action poten-693 tials from the tone onset up to 100 ms after this onset. Thus, TFRP were matrices of 100 bins in ab-694 scissa (time) multiplied by 129 bins in ordinate (frequency). All TFRPs were smoothed with a uni-695 form 5x5 bin window. 696 For each TFRP, the Best Frequency (BF) was defined as the frequency at which the highest firing rate 697 was recorded. Peaks of significant response were automatically identified using the following proce-698 dure: A positive peak in the TFRP was defined as a contour of firing rate above the average level of 699 the baseline activity plus six times the standard deviation of the baseline activity. Recordings without 700 significant peak of responses or with inhibitory responses were excluded from the data analyses. 701

Quantification of responses evoked by vocalizations 702
The responses to vocalizations were quantified using two parameters: (i) The firing rate of the evoked 703 response, which corresponds to the total number of action potentials occurring during the presentation 704 of the stimulus minus spontaneous activity; (ii) the trial-to-trial temporal reliability coefficient (Cor-705 rCoef) which quantifies the trial-to-trial reliability of the response over the 20 repetitions of the same 706 stimulus. This index was computed for each vocalization: it corresponds to the normalized covariance 707 between each pair of spike trains recorded at presentation of this vocalization and was calculated as 708 follows: 709 710 where N is the number of trials and σx i x j is the normalized covariance at zero lag between spike trains 711 x i and x j where i and j are the trial numbers. Spike trains x i and x j were previously convolved with a 712 10-ms width Gaussian window. Based upon computer simulations, we have previously shown that 713 this CorrCoef index is not a function of the neurons' firing rate (Gaucher et al., 2013a). 714 We have computed the CorrCoef index with a Gaussian window ranging from 1 to 50 ms to determine 715 if the selection of a particular value for the Gaussian window influences the difference in CorrCoef 716 mean values obtained in the different auditory structures. Based upon the responses to the original 717 vocalizations, we observed that the relative ranking between auditory structures remained unchanged 718 whatever the size of the Gaussian window was. Therefore, we kept the value of 10 ms for the Gaussi- tion at a particular SNR. EI is bounded between -1 and 1: a positive value indicates that the neural 774 response to noisy vocalization is more vocalization-like, and a negative value implies that the neural 775 response is more noise-like. The EI profile for each recording was determined by computing EI at 776 every SNR level. The normalized inner product was used to compute distance between