Cortical representation of pitch perception in mice

Abstract

Pitch perception is fundamental to speech and music, yet the neural representation of pitch in auditory cortex remains poorly understood. We hypothesized that mice rely on temporal cues for pitch perception because of their wide auditory filters and high-frequency hearing. To investigate, we developed a mouse auditory filter bank model that estimates the temporal pitch salience of complex tones. We then trained mice to discriminate low vs. high fundamental frequencies (F0s) in complex tones and found that modeled temporal pitch salience was highly correlated with behavioral discrimination sensitivity. Widefield calcium imaging showed that the organization of periodotopy in auditory cortex predicts how well mice recognize temporal pitch cues. Mice trained on an F0 discrimination task had larger periodotopic maps with greater vector strengths than naïve mice, demonstrating experience-dependent expansion and refinement of temporal pitch maps in auditory cortex. Our findings establish mice as a robust model for studying the neural basis of temporal pitch perception.

Modeling, behavior, and imaging link auditory temporal cues to pitch perception and cortical maps in mice, revealing experience-dependent plasticity in pitch representations that predict how well mice perceive pitch.

Introduction

Pitch is the perception that sound is arranged from low to high on a scale derived from temporal and spectral acoustic cues. Pitch is central to speech and music, yet people with hearing loss or cochlear implants often struggle with pitch perception due, in part, to a greater reliance on temporal cues^1–5. For a pure tone, perceived pitch corresponds closely to its frequency. For complex tones, i.e., periodic broadband sounds, perceived pitch arises from central nervous system analyses of spectral and/or temporal cues encoded in the auditory nerve^6–27 (Fig. 1a). It is thought that the importance of temporal versus spectral cues in pitch perception depends on the bandwidths of “auditory filters”, which are a model of frequency selectivity in the auditory system (middle row in Fig. 1a)^28–33. Auditory filter bandwidths are largely determined by frequency selectivity along the basilar membrane^32–34. In harmonic complex tones (HCTs; top row Fig. 1a), harmonics that excite a single auditory filter (blue in Fig. 1a) are spectrally resolved. Spectral pitch cues are derived from the frequency distance between resolved harmonics, i.e., the fundamental frequency (F0). In contrast, if more than one harmonic excites each auditory filter (red in Fig. 1a), then the harmonics are spectrally unresolved. The combined harmonics cause periodic oscillations in auditory filter output at a rate of 1/F0, which is the temporal pitch cue.

Fig. 1. Temporal pitch salience predicts F0 discrimination sensitivity in mice.

a Illustration of spectral and temporal pitch cues. The top row shows the frequency spectrum of a harmonic complex tone (HCT). The distance between harmonically related frequencies is the spectral pitch cue (F0). In this example, there is no energy in the HCT at the pure-tone frequency corresponding to F0 (there is a “missing F0”). When auditory filters have narrow bandwidths (blue), individual harmonics are resolved and fall within distinct filters, enabling spectral pitch perception. When auditory filters have wide bandwidths (red), multiple harmonics are unresolved and fall within the same filter, causing periodic oscillations in filter output that modulate auditory nerve spike rates at 1/F0, giving rise to temporal pitch perception. [Adapted from Walker et al.¹²]. b Mouse auditory filter model (see Results and Methods for model details). Autocorrelations were used to quantify periodicity strength in the output of a gammatone auditory filter model. c Auditory filter model output. Model output for complex tones: pulse trains (PTs), HCTs, amplitude-modulated noise (AMN), and iterated ripple noise (IRN). d Behavioral testing of F0 discrimination, i.e., pitch discrimination. Left panel: mice were trained on a go/no-go task to discriminate low vs. high F0s in PTs, HCTs, AMNs, and IRNs. Middle panel: F0 discrimination sensitivity was highest for PTs and HCTs, intermediate for AMNs, and lowest for IRNs. Gray lines show behavioral data (d’, left axis) for individual mice (n = 11). The thick dark line shows the modeled periodicity strength (right axis) for each stimulus type. Right panel: modeled stimulus periodicity strength vs. behavioral d’. For each stimulus type, each dot shows an individual mouse from the middle panel. The black line and gray shading show the average ± 2 standard errors of the mean (SEMs) for linear fits to each mouse (mean R² = 0.97 ± 0.02). The annotated r_d’ value indicates the average ± 2 SEMs across mice for the correlation between stimulus periodicity strength vs. behavioral d’. e Phase manipulations for HCTs. f Sine- vs. alternating-phase F0 discrimination. Left panel: mice (n = 5) tended not to respond during catch trials with alternating-phase HCTs with temporal F0 = 110 Hz, matching non-target F0s. Right panel: mice (n = 6) tended to respond during catch trials with alternating-phase HCTs with temporal F0 = 110 Hz, matching target F0s. g Sine vs. random-phase F0 discrimination. Mice (n = 8) could not distinguish F0 = 55 vs. 330 Hz for random-phase HCTs. Bars show mean values. Symbols show individual mice. Filled and open symbols in Fig. 1g had targets F0 = 330 (n = 5) and 55 Hz (n = 3), respectively. The five filled-symbol mice in (g) were also included in the right panel of Fig. 1f. n.s. indicates that a difference was not statistically significant (bootstrap t-test, p > 0.05). Stars indicate statistical significance (bootstrap t-test, ***: p < 0.001).

The neural coding of temporal pitch cues by auditory nerve fibers has been described in detail^8,11,23, but little is known about how the central nervous system encodes temporal pitch. Three properties of neural responses to sound have been associated with pitch perception: (1) neural responses are tuned to F0, for both a pure-tone and complex tone with a missing F0, (2) neural responses show similar tuning to F0 for different complex tones and at varying intensities, and (3) neural responses depend on the salience of the pitch cue, i.e. the robustness of periodicity in the auditory filter output. Neurons with some of these properties emerge in the cochlear nucleus of guinea pigs^35–37 and chinchillas³⁸. In the rabbit inferior colliculus, neurons have rate-tuning to both HCTs and amplitude-modulated noise (AMN)³⁹. Gerbil inferior colliculus neurons show tuning to F0 for pulse trains (PTs), AMNs, and iterated ripple noises (IRNs), though the spatial patterns of F0-selective neurons may be heterogeneous across animals and stimulus-types^40–45. The human inferior colliculus shows activity that correlates with F0 discrimination task performance⁴⁶, and the size of the brainstem frequency following response in humans depends on the salience of temporal pitch cues⁴⁷. In the auditory thalamus of ferrets, focal lesioning of neurons that receive efferent input from AC impairs behavioral detection of mistuned harmonics⁴⁸. In dogs, auditory thalamic neurons respond more to IRN than to white noise⁴⁹, indicating the importance of periodic temporal cues in driving thalamic activity.

AC is thought to be important for the perception and discrimination of complex tones^50–52. Marmosets have neurons in an anterior lateral region of AC that are sensitive to the salience of temporal pitch cues in PTs, HCTs, and AMN^13,53. Humans and rhesus macaques show neural activity in AC that is tuned and sensitive to temporal pitch cues in IRN⁵⁴ and PTs^54,55. Experiments with cats⁵⁶ and humans^57–60 have shown that higher-order AC tends to encode temporal pitch cues more robustly than primary AC (A1) for HCTs, PTs, and IRN. Intrinsic imaging of AC in cats⁴¹, ferrets⁴², and gerbils⁴³, as well as magnetoencephalography⁴⁴ and functional magnetic resonance imaging⁴⁵ in humans, have shown spatial gradients mapping F0 across cortical space, i.e., “periodotopy”, reminiscent of tonotopy.

Mice present a useful model for studying temporal pitch perception because they have broad auditory filters and high-frequency hearing^61,62. Importantly, mice detect missing F0s in HCTs, may share some perceptual similarities with humans in processing multi-harmonic communication sounds^63,64, and, like humans, are most sensitive to F0s below a few hundred hertz^65,66. Despite the importance of pitch to hearing, it remains unclear how temporal pitch is represented in auditory cortex^{10,13,41–45,54–59}. We therefore asked if temporal pitch cues are salient to mice, under what conditions mice use temporal cues to discriminate F0s, and how cortical representation of F0 reflects temporal pitch perception.

Results

Modeled temporal pitch salience predicts F0 discrimination sensitivity in mice

To estimate the salience of temporal pitch cues in mice, we developed a gammatone filter bank model of the mouse auditory periphery in MATLAB software (see Materials and Methods; Fig. 1b). Four types of complex tones were synthesized for model analysis: PTs, HCTs, AMNs, and IRNs (Fig. 1c). All stimuli were generated at a 200 kHz sampling rate, tapered with 5 ms onset and offset ramps, normalized to unity magnitude, and band-limited between 4 and 48 kHz. F0s of 55 and 330 Hz were tested because mice are most sensitive to acoustic amplitude modulation rates less than several hundred hertz⁶⁵, which is similar to the upper limit of F0 sensitivity for temporal pitch perception in humans⁶⁶.

The cochlea was modeled by a bank of fourth-order, zero-phase gammatone filters with center frequencies spaced every 10 Hz from 4-48 kHz. Filter bandwidths were based on mouse behavioral critical band measurements reported by Ehret^61,67, using a linear fit to interpolate values across frequencies. Each filter’s impulse response was convolved with the input stimulus to produce a band-limited waveform for each cochlear channel. Filter outputs were half-wave rectified to reflect asymmetric inner hair cell transduction and auditory nerve fiber response bias, followed by a compressive nonlinearity (√x) to approximate cochlear level compression. Each channel was then low-pass filtered with a sixth-order Butterworth filter at 1 kHz to simulate the limit of phase-locking in the mouse auditory nerve. Model output represents pooled auditory nerve firing rates in response to sound¹¹.

For each stimulus input to the model, we computed autocorrelations of auditory filter outputs. The peak height at the lag corresponding to 1/F0 was taken as periodicity strength (PS), an index of temporal pitch salience^7,8,68,69 (Fig. 1b). Our model predicted the highest pitch salience for PTs (PS = 1.00) and HCTs (PS = 1.00), intermediate for AMN (PS = 0.92), and lowest for IRN (PS = 0.64) (Fig. 1d).

We tested model predictions for the temporal pitch salience of PTs, HCTs, AMNs, and IRNs by training mice on go/no-go F0 discrimination tasks. Mice were presented with randomly interleaved PTs, HCTs, AMNs, and IRNs, at either a low F0 (55 Hz) target or high F0 (330 Hz) non-target (Fig. 1d). Discrimination sensitivity was highest for PTs (d’=1.62 ± 0.20) and HCTs (d’=1.59 ± 0.16), intermediate for AMN (d’=1.25 ± 0.19), and lowest for IRN (d’=0.00 ± 0.09, p < 0.001, bootstrap t-test, Bonferroni correction). Task performance correlated strongly with periodicity strengths (r = 0.98 ± 0.01), indicating that temporal pitch salience predicts F0 discrimination sensitivity in mice.

Next, we tested if mice rely on temporal or spectral cues to discriminate F0s. Sine-, alternating-, and random-phase HCTs share identical spectral profiles but differ in temporal waveforms (Fig. 1e). Alternating-phase HCTs introduce waveform peaks at half the sine-phase period, effectively doubling the temporal pitch. Mice trained to discriminate sine-phase HCTs (spectral F0 = 55 Hz targets, 110 Hz non-targets; d’=1.48 ± 0.30) withheld licking to alternating-phase catch trials (spectral F0 = 55 Hz), behaving as if they heard 110 Hz non-targets. Behavioral response rates were similar for sine-phase non-target trials (19.7 ± 11.0%) and alternating-phase catch trials (20.4 ± 7.7%; p = 0.81, bootstrap t-test) (Fig. 1f). With reversed behavioral meanings (spectral F0 = 110 Hz targets, 55 Hz non-targets; d’=1.37 ± 0.61), mice treated alternating-phase HCTs as 110 Hz targets, with alternating-phase catch trial response rates indistinguishable from sine-phase target hit rates (74.4 ± 8.9% vs. 73.8 ± 11.8%; p = 0.94). In contrast, random-phase perturbations, which largely disrupt acoustic waveform periodicity, abolished F0 discrimination (d’=0.24 ± 0.16) (Fig. 1g). Thus, mice used temporal, not spectral, acoustic cues to recognize low F0s.

Mouse auditory cortex exhibits periodotopic neural responses to F0

Having established that mice use temporal pitch cues, we examined the neural correlates of temporal pitch perception in the auditory cortex of awake CBA x Thy1-GCaMP6s mice^62,70. Widefield calcium imaging revealed both periodotopy (20–1280 Hz, 2 F0s per octave) and tonotopy (2–45 kHz, 2 tones per octave), i.e., maps of complex tone F0 and pure-tone frequency, respectively. We collected periodotopic maps from n = 8 mice. Examples of periodotopy are shown in Fig. 2a for two mice, from PTs, HCTs, AMNs, and IRNs, as well as for tonotopy. Table 1 shows the imaging experiment counts for each mouse and each stimulus type. Each experiment produced one whole-map for one stimulus-type, e.g., the examples shown in Fig. 2a.

Fig. 2. Periodotopic organization of neural responses to F0 in mouse auditory cortex (AC).

a Examples of periodotopy and tonotopy in two Thy1-GCaMP6s x CBA mice, mapped across AC using in vivo widefield imaging. The color bars show the range of tested F0s and pure-tone frequencies. b Periodotopic map features. We segmented individual periodotopic gradients, i.e., sub-maps, in each imaging field from experiments in eight mice. In the top panel, each colored dot shows the number (#) of sub-maps per experiment for each complex tone (total experiment counts: PT, n = 48; HCT, n = 50; AMN, n = 53; IRN, n = 49; see Table 1 for experiment counts per animal). The lower three panels quantify sub-map features using three metrics: vector strength (VS), size (% of AC), and angle (θ). Each colored dot shows the value from a single sub-map (PT, n = 96; HCT, n = 110; AMN, n = 145; IRN, n = 154). Black dots show average values. For each sub-map feature, the annotated r_d’ value indicates the average ± 2 SEMs across mice (n = 8), for the correlation between each mouse’s mean feature value per stimulus and the mean behavioral d′ per stimulus, computed from the average of the gray curves in Fig. 1d. Stars indicate statistical significance (#, VS, Size: bootstrap t-test; θ: Omnibus test; **p < 0.01, ***p < 0.001). n.s. indicates that IRN sub-map gradients did not have a significant direction (Omnibus test, p > 0.05).

Table 1.

Imaging experiment counts for each mouse and each stimulus-type

Mouse	Pulse Train (PT)	Harmonic Complex Tone (HCT)	Amplitude Modulated Noise (AMN)	Iterated Ripple Noise (IRN)
hs01	5	7	7	7
db01	5	4	5	4
hp01	6	7	7	7
op01	5	6	6	6
jz01	5	6	6	4
sa01	7	7	6	7
to01	6	7	7	6
je01	9	6	9	8

Underlined mouse names indicate mice that were trained on F0 discrimination before imaging experiments.

To quantify the structure of periodotopy (Fig. 2b), we segmented whole-maps into individual gradients, i.e., sub-maps (see Materials and Methods), counted the total number (#) of sub-maps per experiment (top right panel in Fig. 2b), and measured sub-map features including vector strength (VS), size (% of auditory cortex), and angle (θ). Periodotopic organization mirrored behavioral and model results: PTs (VS = 0.46 ± 0.02; size=40.8 ± 4.9%) and HCTs (VS = 0.44 ± 0.02; size=35.5 ± 4.6%) showed the greatest VS and size, AMN was intermediate (VS = 0.40 ± 0.02; size=24.2 ± 3.2%), and IRN was smallest (VS = 0.37 ± 0.02; size=20.5 ± 2.6%). Sub-map angles followed the same trend, with PTs (θ = 146.2 ± 10.3°) and HCTs (θ = 153.0 ± 9.8°) pointed rostral-dorsal, AMN pointed caudal-dorsal (θ = 29.7 ± 11.8°), and IRN showed no significant direction (θ = -68.6 ± 12.4°, p = 0.21, Omnibus test).

Periodotopic spatial overlap with secondary auditory cortex (A2) correlates with behavioral discrimination sensitivity for F0

Several studies have supported a specialized role for higher-order auditory cortex in processing spectrotemporally complex sounds^56,71–73. Here, we used the percent of spatial overlap between periodotopic and tonotopic sub-maps to determine the location of periodotopic sub-maps within four regions of mouse auditory cortex, including A1, A2, the anterior auditory field (AAF) and the dorsomedial field (DM). A1, A2, AAF, and DM were identified based on expected relative locations in cortical space, and their expected tonotopic gradient directions⁷⁴. We found that periodotopic sub-maps tended to locate in AAF and A2 (Fig. 3a). Moreover, only in A2 was the spatial overlap for sub-maps significantly correlated with mean behavioral discrimination sensitivity (d’) for F0 across stimulus-types (Fig. 3b).

Fig. 3. Spatial overlap between periodotopic maps and tonotopically-defined regions of AC.

a Periodotopic map spatial overlap with primary AC (A1), secondary AC (A2), the anterior auditory field (AAF), and the dorsomedial field (DM). Bars show mean values across imaging experiments (see Table 1 for experiment counts per animal). Error bars show 2 SEMs. Stars indicate statistical significance (bootstrap t-test with Bonferroni correction, *: p < 0.05; **: p < 0.01; ***: p < 0.001). b. Periodotopic map spatial overlap with tonotopic regions, shown separately for A1, A2, AAF, and DM. Each dot shows the average value for a given animal (A1, AAF, DM: n = 8; A2: n = 5). For each cortical area, the annotated r_d’ value indicates the average ± 2 SEMs across mice, for the correlation between each mouse’s periodotopic map overlap per stimulus and the mean behavioral d′ per stimulus, computed from the average of the gray curves in Fig. 1d. Stars indicate statistical significance (bootstrap t-test, *: p < 0.05; **: p < 0.01; ***: p < 0.001). n.s. indicates that r_d’ was not significant, p > 0.05).

Periodotopic map gradient similarity and spatial overlap correlate with behavioral discrimination sensitivity for F0

We also compared the spatial overlap between periodotopic maps, both within and across stimulus-types (Fig. 4a). All comparisons were made within a given mouse. For both whole-maps and sub-maps, pairwise comparisons of spatial overlap tended to be greatest between maps within a stimulus-type (PT/PT, HCT/HCT, AMN/AMN, and IRN/IRN), though significant spatial overlap also occurred between different stimulus-types (left panels, Fig. 4b). In addition, the mean spatial overlap for all pairwise stimulus comparisons (right panels, Fig. 4b), both within and across stimulus-types, was significantly greater than zero (bootstrap t-test, p < 0.001). Similar stimulus-dependent patterns were observed for the similarity of whole-map and sub-map gradients, evaluated by first extracting F0 values from the shared pixels in two maps, and then computing Pearson’s correlation coefficient across paired F0 values (Fig. 5a). However, due to their more idiosyncratic periodotopic organization, AMN and IRN sub-maps each tended to have less gradient similarity compared to sub-maps from other stimulus-types (Fig. 5b). Also, AMN was the only stimulus whose gradient was sometimes inverted, i.e., the F0 spatial correlation was negative, e.g., for whole-map AMN/IRN and for sub-map AMN/PT. Importantly, when evaluated across the mouse population, the similarity of periodotopic maps across stimulus repeats tended to correlate well with behavioral discrimination sensitivity for F0 (Fig. 6).

Fig. 4. Periodotopic map similarity: spatial overlap.

a Pairwise comparisons.Top and bottom panels show pairwise comparisons for whole-maps and sub-maps, respectively. All comparisons were made within a given mouse and across multiple imaging experiments for each stimulus-type (see Table 1 for experiment counts per animal and stimulus-type). Whole-map overlap distributions comprised n = 127 (PT/PT), 302 (PT/HCT), 327 (PT/AMN), 304 (PT/IRN), 135 (HCT/HCT), 335 (HCT/AMN), 313 (HCT/IRN), 154 (AMN/AMN), 334 (AMN/IRN), and 133 (IRN/IRN) within-mouse pairwise comparisons. Sub-map overlap distributions comprised n = 510 (PT/PT), 1415 (PT/HCT), 1743 (PT/AMN), 1883 (PT/IRN), 684 (HCT/HCT), 2027 (HCT/AMN), 2182 (HCT/IRN), 1171 (AMN/AMN), 2789 (AMN/IRN), and 1265 (IRN/IRN) within-mouse pairwise comparisons. b Statistical tests for spatial overlap. Left panels show heatmaps of p-values from all non-identical stimulus-type bootstrap t-test comparisons of the distributions shown in (a), with Bonferroni corrections for multiple comparisons. Darker values indicate greater statistical significance between spatial overlap comparisons. Right panels show heatmaps for the mean value of spatial overlap for each distribution shown in panel a. For example, the upper left cell corresponds to the PT/PT distribution. Stars indicate statistical significance (bootstrap t-test, ***: p < 0.001). n.s. indicates that a comparison was not statistically significant (bootstrap t-test, p > 0.05).

Fig. 5. Periodotopic map similarity: spatial correlation.

a Pairwise comparisons. Top and bottom panels show pairwise comparisons for whole-maps and sub-maps, respectively. All comparisons were made within a given mouse and across multiple imaging experiments for each stimulus-type, matching the sample sizes described in Fig. 4. b Statistical tests for gradient similarity. Left panels show heatmaps of p-values from all non-identical stimulus-type bootstrap t-test comparisons of the distributions shown in (a), with Bonferroni corrections for multiple comparisons. Darker values indicate greater statistical significance between similarity comparisons. Right panels show heatmaps for the mean value of F0 spatial correlation for each distribution shown in (a). For example, the upper left cell corresponds to the PT/PT distribution. Stars indicate statistical significance (bootstrap t-test, ***: p < 0.001). n.s. indicates that a comparison was not statistically significant (bootstrap t-test, p > 0.05).

Fig. 6. Per-animal periodotopic map similarity for repeats of stimulus-types.

Each gray line shows mean values from an individual mouse (n = 8). The black line and dots show the mean values across mice ± 2 SEMs. The black lines are equivalent to the shade of the cells along the diagonal of the heatmaps in Figs. 4b and 5b. Annotated r_d’ values indicate the average ± 2 SEMs for the correlation between each mouse’s periodotopic map similarity per stimulus and the mean behavioral d′ per stimulus, computed from the average of the gray curves in Fig. 1d. Stars indicate statistical significance (bootstrap t-test, *: p < 0.05; **: p < 0.01; ***: p < 0.001). n.s. indicates that r_d’ was not significant, p > 0.05).

Periodotopy in AC reflects temporal pitch sensitivity in mice

We quantified the relationship between periodotopy and temporal pitch perception by regressing z-scored F0 discrimination (d’) against z-scored sub-map metrics including VS, Size (% AC), gradient angle (θ), gradient similarity, and A2 spatial overlap, separately for each mouse (Fig. 7). VS showed the strongest explanatory power (R² = 0.73 ± 0.14), followed by gradient similarity (R² = 0.62 ± 0.23), size (R² = 0.58 ± 0.17), angle (R² = 0.55 ± 0.28), and finally A2 overlap (R² = 0.39 ± 0.19). All predictors were significant (p < 0.001, Bonferroni correction), suggesting that temporal pitch salience is encoded by periodotopic organization in mouse auditory cortex.

Fig. 7. Periodotopic map features predict temporal pitch sensitivity in mice.

The dotted black trace shows the mean z-scored behavioral sensitivity for F0 discrimination (d′) across stimulus-types (PT, HCT, AMN, and IRN) in the behavioral cohort of mice (n = 11; Fig. 1d). Shading indicates ±2 SEMs. Colored traces show the mean regression-predicted behavioral z-score obtained from five periodotopic map features measured across stimuli: vector strength (VS), gradient size, gradient angle, map similarity (F0 spatial correlation), and A2 spatial overlap. For each predictor, regressions were fit separately for each imaged mouse and then averaged across mice. Shading indicates ±2 SEMs of the regression-predicted behavioral z-score across mice (n = 8 for VS, Size, Angle, and Similarity; n = 5 for A2 overlap). All five predictors had mouse-averaged R² values significantly greater than zero (Bonferroni-corrected, ***p < 0.001 for all predictors).

Auditory training on F0 discrimination increases periodotopic map size and vector strength

Previous studies have shown that auditory perceptual learning drives task-specific reorganization of cortical maps^75–79. To examine the effect of auditory perceptual learning on periodotopic maps, four of the eight imaged mice were trained on the F0 discrimination task prior to imaging experiments (see Table 1). Periodotopic sub-maps in trained animals were not significantly different compared to naïve mice (Fig. 8) for θ, gradient similarity, and A2 spatial overlap (bootstrap t-test, p > 0.05). However, in trained mice, VS and Size were significantly greater than in naïve mice (bootstrap t-test, VS: p < 0.05; Size: p < 0.01), indicating that experience with F0 discrimination enhanced cortical representation of F0.

Fig. 8. Experience-dependent periodotopy.

Mice trained on the F0 discrimination task shown in Fig. 1d showed significantly greater vector strength (VS; trained n = 248 sub-maps, naïve n = 257 sub-maps) and size (% AC; trained n = 248 sub-maps, naïve n = 257 sub-maps) of periodotopic maps compared with naïve mice. Stars indicate statistical significance (bootstrap t-test, *: p < 0.05; **: p < 0.01). n.s. indicates that a comparison was not statistically significant (bootstrap t-test, p > 0.05).

Discussion

Our results establish mice as a useful animal model for investigating the neural basis of temporal pitch perception. Across cortical imaging, behavior, and auditory modeling, the same general pattern emerged: sounds with stronger modeled temporal pitch salience were also the sounds that mice discriminated more accurately and that evoked larger and more organized periodotopic maps in auditory cortex. Together, our findings suggest that temporal pitch salience in mice is linked closely to both auditory peripheral coding and cortical functional organization.

A central conclusion of our study is that mice rely primarily on temporal acoustic cues to discriminate low F0s. Our phase manipulation experiments showed that mice categorized alternating-phase HCTs according to their temporal periodicity rather than their spectral F0, whereas randomizing harmonic phase abolished discrimination. This pattern is consistent with long-standing theories and on-going work indicating that species with broad auditory filters rely strongly on temporal cues for low-F0 pitch judgments^1–5,16.

Our widefield imaging results show that temporal pitch perception in mice is represented across the cortical surface by orderly gradients of F0 preference, i.e., periodotopic maps. Importantly, periodotopy was not equally well-organized for all periodic stimuli. Instead, periodotopic organization depended on stimulus-type and closely paralleled both modeled temporal pitch salience and behavioral F0 discrimination. Stimuli that were predicted to generate stronger temporal pitch signals produced larger and more coherent maps, whereas stimuli with weaker temporal pitch salience, especially IRN, produced smaller and less structured maps. At the same time, periodotopic organization was repeatable within a given stimulus-type, but comparisons across different stimulus-types showed weaker overlap and F0 spatial correlations. Thus, the mouse cortex may not contain a fully stimulus-invariant map of temporal pitch. Rather, different classes of periodic sounds recruit partially distinct, yet stable, patterns of cortical organization. This is broadly consistent with previous work, in which pitch- or periodicity-related organization has been described across species but differed in stimulus-dependence and map organization^40–45.

We found that periodotopy was behaviorally predictive, suggesting that the coherence of periodotopic organization may reflect temporal pitch salience. Several features of periodotopy, including vector strength, size, angle, F0 spatial correlation, and A2 spatial overlap, covaried with behavioral discrimination sensitivity for F0. Moreover, periodotopic maps showed perceptual experience-dependent plasticity. Periodotopic maps in trained animals were larger and had greater vector strength than naïve animals. This is consistent with prior work showing that auditory perceptual learning drives plasticity in cortical maps^75–79. Our results therefore extend the literature on auditory cortical plasticity to temporal pitch coding.

Our study also places mice within the broader comparative framework for temporal pitch coding in auditory cortex. In marmosets, neurons in an anterolateral region of auditory cortex are sensitive to the salience of temporal pitch cues¹³. In macaques, neurons in auditory cortex are tuned to temporal pitch cues⁵⁴. In humans, neurons encoding temporal pitch cues have been identified throughout and beyond Heschl’s gyrus, with stronger representations often observed in higher-order auditory cortical fields^54–60,80. Indeed, A2 in the mouse and higher-order auditory cortical fields in other species may be specialized for processing spectrotemporally complex sounds^56,71–73. This pattern parallels our observation that periodotopy in mice was strongest in A2 and AAF. More broadly, periodotopy appears to occur across multiple species, including cats, ferrets, gerbils, and humans, although its precise topography and anatomical locus vary across species^40–45. These comparative findings support the idea that spatial representations of periodicity are a broadly conserved feature of auditory systems, while also suggesting that the detailed organization of periodotopy is species dependent.

Taken together, our findings show that mice use temporal acoustic cues to discriminate low F0s, auditory cortex contains orderly and behaviorally relevant temporal pitch maps, and auditory training on F0 discrimination is associated with more robust cortical representation of temporal pitch. Our results position the mouse as a tractable animal model for investigating the neural basis of temporal pitch perception, laying the groundwork for future studies of how cortical circuits transform sensory information into perceptual experience.

Materials and Methods

Auditory filter model

To simulate cochlear and auditory nerve processing in mice, we implemented a gammatone auditory filter bank model in MATLAB. Our approach captures key auditory transformations relevant for temporal pitch processing in mice, but omits detailed biophysical mechanisms contained in models such as the Zilany et al. (2009) model of the auditory periphery⁸¹. Four types of periodic sounds were synthesized for model analysis: pulse trains (PTs), harmonic complex tones (HCTs), amplitude-modulated noises (AMNs), and iterated ripple noises (IRNs). All stimuli were generated using a 200 kHz sampling rate, tapered with 5 ms onset and offset ramps, normalized to unity magnitude, and band‑limited between 4 and 48 kHz. Fundamental frequencies (F0s) were tested at 55 and 330 Hz. HCTs were synthesized by summing sine-phase pure-tones with harmonic frequencies spaced at integer multiples of the F0, beginning at F0. PTs were synthesized by concatenating impulses across silent temporal intervals of 1/F0. AMNs were synthesized by 1/F0 temporal envelope modulation of broadband noise at 100% depth. IRNs were synthesized using the delay-and-add method⁶⁹ for a broadband noise token, using 5 iterations, and the delay set to 1/F0.

The model cochlea was represented by a bank of fourth‑order zero-phase gammatone filters with center frequencies spaced every 10 Hz from 4 to 48 kHz. Filter bandwidths were set according to mouse behavioral critical band (CB) measurements reported by Ehret⁶¹ (CBs: 3.9, 4.4, 6.4, 7.9, 15.7, 17.7, and 21.9 kHz for characteristic frequencies of 5, 10, 15, 20, 30, 40, and 50 kHz, respectively), using a linear fit to interpolate values across frequencies. The impulse response of each gammatone filter was convolved with the input stimulus, yielding a band‑limited waveform at each cochlear channel. The output of each gammatone filter was half‑wave rectified to capture only positive‑going responses, reflecting the asymmetric transduction of inner hair cells and because auditory nerve fibers fire only to positive deflections of the basilar membrane. A compressive non‑linearity ($\sqrt{x}$) was then applied to the rectified signal to approximate the saturation and level compression introduced by cochlear mechanics and hair cell transduction. Finally, each channel was low‑pass filtered with a 6th‑order Butterworth filter at 1 kHz to simulate the limits of phase locking in the mouse auditory nerve. The spectral profile of the auditory response was obtained by taking the maximum amplitude across time for each channel and normalizing by the channel’s peak response.

To quantify temporal pitch salience, we computed the autocorrelation function of the filter output within each channel. For each stimulus, the periodicity strength (PS) was defined as the autocorrelation function peak located at the lag corresponding to the fundamental period (1/F0). We then averaged across channels to obtain a single PS value per stimulus.

Animals

We have complied with all relevant ethical regulations for animal use, and all procedures were approved by the University of Maryland Institutional Animal Care and Use Committee. We used a total of N = 29 mice (14 male, 15 female, aged 13–35 weeks) in our experiments. The mice were F1 offspring of CBA/CaJ mice (The Jackson Laboratory; stock #000654) crossed with transgenic C57BL/6J-Tg (Thy1GCaMP6s)GP4.3Dkim/J mice⁷⁰ (The Jackson Laboratory; stock #024275). We used the F1 generation of crossed mice because they have both GCaMP6s expression in excitatory neurons and healthy hearing at least 1 year into adulthood⁶². 25 mice were used for behavioral experiments. 8 mice were used for imaging, with 4 of the 8 imaged mice being naïve. The other 4 imaged mice were initially in the group of 25 behaviorally trained mice, before continuing on to imaging. All mice were housed under a reversed 12 h-light/12 h-dark light cycle.

Acoustic stimulus presentation

During behavioral and imaging experiments, we presented awake mice with sounds from an ES1 free-field speaker (Tucker-Davis Technologies) connected to an ED1 amplifier (Tucker-Davis Technologies). The speaker was calibrated in situ to give a flat frequency response in the 1–80 kHz range. Our acoustic stimulus set included PTs, HCTs, AMNs, and IRNs. Stimuli were synthesized in MATLAB software (Mathworks) and output at 75 dB SPL from a National Instruments board (NI-6211, 200 kHz sampling rate) to the ED1 amplifier. Each stimulus had 5 and 495 ms raised-cosine attack and decay ramps, respectively. These ramp parameters were chosen to minimize off-responses to sounds⁸². HCTs were synthesized by summing pure tones spaced at integer multiples of the F0, beginning at F0. All components were typically in sine-phase, except during phase manipulation experiments, in which components were either (1) all in sine-phase, (2) alternating between sine- and cosine-phase for even and odd harmonics, respectively, or (3) set to a random phase uniformly distributed between -pi and pi. PTs were synthesized by concatenating impulses across silent temporal intervals of 1/F0. AMNs were synthesized by 1/F0 temporal envelope modulation of broadband noise at 100% depth. IRNs were synthesized using the delay-and-add method for a broadband noise token, using five iterations, and the delay set to 1/F0. All broadband sounds were bandpass filtered between 4 and 45 kHz.

Behavioral testing

F0 discrimination was tested using a custom training system based on a previous design⁸³. Each operant chamber contained a TDT ES1 speaker calibrated in situ and a lick-sensitive waterspout. Water-restricted mice were trained to lick a waterspout in response to hearing a target F0, and to avoid licking the waterspout after hearing a non-target F0. PT, HCT, AMN, and IRN presentations were randomly interleaved, and each type of complex tone was presented at target and non-target F0s. Each trial of the task was 4 s long with a randomized 5–9 s inter-trial interval. To initiate trials, the mouse refrained from licking for 5 s. Trials began with a 1 s silent period, followed by the 1 s stimulus, and ended with a 2 s silent period. If the mouse licked the waterspout between the sound onset and the end of the trial, i.e., a Hit (H), then the mouse received a water droplet. A false alarm (FA) occurred when the mouse licked the waterspout after a non-target, resulting in a 10 or 30 s time-out. Discrimination sensitivity was quantified as d’ = z(H)-z(FA). Each behavioral session lasted 30–60 min. Analysis of behavioral data was done using custom MATLAB software.

Behavioral testing was conducted in four cohorts whose training differed only by the stimulus types used in a given trial. Stimulus types used for a given cohort were randomly interleaved across trials. In the initial experiment, 11 mice were trained to discriminate F0 = 55 Hz vs. 330 Hz in PTs, HCTs, AMNs, and IRNs. A second cohort of five mice was then trained to discriminate target F0 = 55 Hz vs. non-target F0 = 110 Hz in alternating-phase HCTs, and a third cohort of six mice was trained to discriminate target F0 = 110 Hz vs. non-target F0 = 55 Hz in alternating-phase HCTs. The subsequent random-phase HCT experiment included 5 of the mice from the third cohort and three additional naïve mice. Thus, we used 25 unique mice.

Widefield calcium imaging

Chronic window implantation

We performed widefield calcium imaging in eight mice. Four mice were naïve, and 4 were first trained in the initial behavioral cohort before continuing on to imaging. Mice were given an intraperitoneal injection of dexamethasone (5 mg/kg) at least 1 h prior to surgery to prevent inflammation and edema. Mice were deeply anesthetized using isoflurane (5% induction, 0.5–2% for maintenance) and given a subcutaneous injection of cefazolin (500 mg/kg). Internal body temperature was maintained at 37.5 °C using a feedback-controlled heating blanket. Scalp fur was trimmed using scissors, and any remaining fur was removed using Nair. The scalp was disinfected with alternating swabs of 70% ethanol and betadine. A patch of skin over the temporal bone was removed and the underlying bone cleared of connective tissue using a scalpel. The temporal muscle was detached from the skull, and the skull was cleaned and dried. A thin layer of cyanoacrylate glue (VetBond) was applied to the exposed skull surface, and a 3D printed stainless steel head-plate was affixed to the midline of the skull. A circular craniotomy (3 mm diameter) was made over the auditory cortex where the glass imaging window was implanted. The craniotomy was located in the posterior portion of the thin lateral skull, just dorsal to the zygomatic arch and anterior to the lambdoid suture, near the squamosal-parietal junction and partially overlapping the parietal-temporal ridge. The space between the glass and the brain was sealed with an optically transparent silicone elastomer (Kwik-Sil). Dental cement (C&B Metabond) was used to cover the entire head-plate and the edges of the glass, and the skull was then sealed. Finally, the entire implant except for the imaging window was coated with black dental cement created by mixing methyl methacrylate with iron oxide powder to reduce optical reflections. Meloxicam (0.5 mg/kg) was given subcutaneously as a post-operative analgesic.

Image acquisition

Widefield images were acquired as previously described^84,85. To examine periodotopy and tonotopy, we performed widefield imaging experiments in the auditory cortex of Thy1-GCaMP6s x CBA mice. Awake mice were placed into a 3D-printed plastic tube and head-restraint system. Blue excitation light was shone by an LED (473 nm) through an excitation filter (470 nm) and directed into the cranial window. Emitted fluorescence from neurons was collected through a 4x objective (Thorlabs), passed through a longpass filter (cutoff: 505 nm), followed by a bandpass emission filter (531 nm) attached to a CMOS camera. Images were acquired at a rate of 5 Hz and resolution of 512×512 pixels using ThorCam software (Thorlabs). After acquiring an image of the cortical surface, the focal plane was advanced to approximately 750 μm below the surface for imaging neural activity.

Acoustic stimuli

Acoustic stimuli were presented to mice from an ES1 free-field speaker, as described above. To measure tonotopy, the frequency of each pure-tone was randomly selected from ten equiprobable values (2–45 kHz, 2 tones per octave). To measure periodotopy, the F0 of PTs, HCTs, AMNs, and IRNs were randomly selected from 13 equiprobable values (20-1280 Hz, 2 F0s per octave). Trials began with a 1 s silent period, followed by a 0.5 s stimulus, and ended with a 2 s silent period. Each stimulus was repeated 20 times per experiment, with inter-stimulus intervals randomized at 6, 7, or 8 s. Mice were imaged once per day, and only one type of stimulus was presented per ~20-minute imaging session.

Image registration

Widefield calcium imaging data were preprocessed in MATLAB. For each session, cortical surface images were averaged to provide a high‑contrast template for alignment. Fluorescence movies were registered frame‑by‑frame to the surface reference using rigid translation-based alignment (imregtform, Image Processing Toolbox, MATLAB). Registered movies were processed with a 16‑pixel spatial filter, downsampled by a factor of four, half‑wave rectified, and a variance mask was applied to suppress low‑signal regions.

Analysis of periodotopic gradients

For each stimulus condition, imaging frames were aligned to stimulus onset and averaged across repetitions to obtain a mean time series for every pixel. For each pixel, ΔF/F traces were calculated by finding the average fluorescence during the silent pre‑stimulus baseline period (F), subtracting that value from subsequent time points until 1 s after onset, and then dividing all time points by the baseline fluorescence. To identify auditory responses, we retained traces within 90% of the maximum ΔF/F observed at each pixel (ΔF/F₉₀). Each pixel was assigned a preferred frequency (for pure-tones) or preferred F0 (for complex tones), defined as the median of the F0 or frequency that evoked ΔF/F₉₀ responses. The median was chosen instead of the maximum to provide a stable estimate when several adjacent stimuli produced similar responses. Only one stimulus-type (PTs, HCTs, AMNs, or IRNs) was presented per experiment (see Table 1 for experiment and stimulus counts per animal). This procedure yielded either one tonotopic or periodotopic map per experiment. Maps in Fig. 2a show unsmoothed periodotopy and tonotopy.

To segment individual gradients within periodotopic and tonotopic maps⁸⁶, first the ΔF/F₉₀ images were smoothed with a 5×5 moving average filter, and local maxima within 10‑pixel neighborhoods were identified as candidate gradient centers. The corresponding tonotopic or periodotopic values of these local maxima were then subtracted from all other values in the map. A value of 1 was then added to all pixels to ensure positive values, which facilitates finding reversal points in gradients. Then the tonotopic and periodotopic maps were smoothed with a 2×2 moving average filter to regularize maps. From each center, radial trajectories were traced using Bresenham’s line algorithm⁸⁷, with angles discretized in 0.25° steps. Along each trajectory, reversals in preferred frequency or F0 values were defined as local peaks, and the first reversal from the center was taken as a gradient boundary. Reversal points that were more than 2 standard deviations (STDs) from the mean boundary distance were removed. Boundaries were extended to include pixels within three pixels of a reversal and converted into polygonal masks. Masks were further smoothed with a 5-pixel moving average. Smoothing changes the binary mask into a continuous distribution that tapers to zero beyond the original boundary. To account for this blurring, we removed regions of the smoothed mask that fell below 1 STD from the mean value of the smoothed mask. To reduce redundancy, we took the intersection of the masks with >75% overlap.

Within each segmented gradient, gradient vectors were computed on a per-pixel basis as the change in preferred frequency or F0 (on a log2 scale) divided by cortical distance, oriented along the unit vector between points⁸⁶. Vector strength (VS) was defined as the mean resultant length of these vectors, and gradient angle (θ) was computed from the summed vector components. Gradient size (Size) was defined as the percentage of auditory cortex occupied by a given gradient. Tonotopic and periodotopic vectors were calculated as the normalized mean of gradient vectors. Segmented tonotopic gradients were used to identify auditory cortical fields, based on expected field locations and gradient directions in cortical space⁷⁴.

For each pair of maps, similarity was quantified using spatial overlap and F0 spatial correlation, with comparisons performed both within and across stimulus types. All comparisons were performed within each mouse using all available imaging experiment pairs for each stimulus-type (Table 1). F0 spatial correlation was measured by identifying the overlapping region between two maps, extracting F0 values from the shared pixels in each map, and computing the Pearson correlation across paired F0 values. Spatial overlap was quantified as percent regional overlap, defined as the intersection area divided by the area of the smaller region.

Statistics and reproducibility

Non-parametric bootstrap t-test

To determine significantly different values between experimental conditions, we used a non-parametric bootstrap t-test, as previously described⁸⁴. Given 2 datasets, A and B, having sample sizes of n and m, respectively, we tested A and B against the null hypothesis that they were drawn from a common distribution. The hypothesis test began by taking the absolute value of the observed difference of means, Δµ, between A and B. Next, we created the null distribution by pooling the individual values of A and B. Two sample sets, A* and B*, of size min(n,m), were randomly selected (with replacement) from the null distribution. The test statistic, Δµ*, was computed from the absolute value of the difference of the means obtained from the A* and B* sample sets. We repeated the random selection of A* and B* from the null distribution and the calculation of Δµ*, 10,000 times, to form a bootstrap distribution of Δµ*. A was taken to have a statistically significantly different mean than B if Δµ* was greater than Δµ in less than 5%, 1%, or 0.01% of the 10,000 bootstrapped values. This would mean that the probability was <5%, 1%, or 0.01% that samples in A and B came from a common distribution. For one-sample mean comparisons against zero, we used a null-centered bootstrap procedure. All mean values are reported with standard errors of the mean (SEMs). Where noted in the text, multiple comparisons across groups were corrected with a Bonferroni procedure.

Circular statistics

Gradient angles were defined relative to the cortical rostrocaudal and dorsoventral axes. An Omnibus test was used to determine whether angle distributions deviated from uniformity.

Linear regression of behavioral pitch discrimination on periodotopic map features

To assess the relationship between behavioral F0 discrimination performance and structural properties of periodotopy, we performed standardized linear regressions across stimulus types. For each imaged mouse, we formed a 4-point predictor vector for each periodotopic map feature, with one value for each of the four stimulus types (PT, HCT, AMN, and IRN). The predictors were vector strength (VS), size, angle, A2 spatial overlap, and gradient similarity (F0 spatial correlation). Because behavioral sensitivity (d′) and periodotopic features were measured in separate cohorts of mice, the behavioral outcome was the mean z-scored behavioral d′ template across stimulus types, obtained by first z-scoring d′ within each behavioral mouse and then averaging across mice. Predictor vectors were z-scored within each imaged mouse, and regressions were fit separately for each imaged mouse and each predictor. This yielded one coefficient of determination (R²) per imaged mouse for each predictor, which were then averaged across mice. Statistical significance of R² across mice was assessed against zero using the bootstrap procedure described above, with Bonferroni correction for the five predictors.

Peer review

Peer review information

Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editor: Benjamin Bessieres.

Data availability

Source data for figures (10.13016/bojv-i5ug) and code for the auditory filter model (10.5281/zenodo.19616783) are available in online repositories.

Competing interests

The authors declare no competing interests.