Neural dynamics in the rodent motor cortex enables flexible control of vocal timing

Abstract

Neocortical activity is thought to mediate voluntary control over vocal production, but the underlying neural mechanisms remain unclear. In a highly vocal rodent, the Alston’s singing mouse, we investigate neural dynamics in the orofacial motor cortex (OMC), a structure critical for vocal behavior. We first describe neural activity that is modulated by component notes (approx. 100 ms), likely representing sensory feedback. At longer timescales, however, OMC neurons exhibit diverse and often persistent premotor firing patterns that stretch or compress with song duration (approx. 10 s). Using computational modeling, we demonstrate that such temporal scaling, acting via downstream motor production circuits, can enable vocal flexibility. These results provide a framework for studying hierarchical control circuits, a common design principle across many natural and artificial systems.

Introduction

Many species exert voluntary control over vocal production, allowing rapid flexibility in response to conspecific partners or other environmental cues [1, 2]. Neocortical activity observed across a range of species [3–8] has been proposed to be important for executive control of vocalization [9–12]. For instance, cortical neurons are preferentially active when non-human primates vocalize in response to a conditioned cue [6]. In contrast, the primary vocal motor network consisting of evolutionarily conserved brain areas in the midbrain and brainstem [10–14] is sufficient to generate species-typical sounds. Pioneering work in squirrel monkeys [15] and cats [16] as well as recent studies in laboratory rodents [17–19] have identified many such areas, including the periaqueductal grey and specific pattern generator nuclei in the reticular formation. While these subcortical vocal production mechanisms have been well-characterized, much less is known about how cortical activity contributes to vocal production and flexibility.

To address this issue, we focus our attention on the highly tractable vocalizations of a Costa Rican rodent [20]: the Alston’s singing mouse (Scotinomys teguina, Fig. 1a). Singing mice produce a temporally patterned sequence of notes (approx. 20 to 200 ms) that become progressively longer over many seconds (e.g., Fig. 2a), henceforth referred to as a song. Singing mice can flexibly adjust their song duration in response to many internal [21] and external [22] factors, including social context [20]. Recently, we discovered that a specific forebrain region, the orofacial motor cortex (OMC), is crucial for vocal behavior in this species [20]. Electrical stimulation of OMC disrupted or paused ongoing singing, and its pharmacological inactivation abolished vocal interactions and significantly reduced variability in song duration [20]. A major gap in understanding, however, concerns the nature of the cortical activity that drives this ethologically relevant vocalization. We therefore performed the first electrophysiology recordings in singing mice to assess the impact of OMC dynamics on vocal production and flexibility.

Fig. 1. Reliable cortical population activity during singing in S. teguina.

(a) S. teguina singing (Photo credit: Christopher Auger-Dominguez). (b) Schematic of S. teguina brain highlighting the recording site (i.e., orofacial motor cortex, or OMC) as well as the positioning of electrodes (gray shaded region). (c) Spiking activity from 23 simultaneously recorded OMC neurons during song production. The sonogram at top depicts S. teguina song. Neurons with mean firing rates less than 1 spikes/s are excluded for visualization purposes. (d and e) Firing rates of OMC neural ensemble from (c) during three singing epochs (d) compared with equally timed epochs recorded outside of song (e). For plots (c) through (e), green and red dashed lines mark the beginning and end of the song, respectively. (f) and (g) For the example session, pairwise correlations of the joint activity of the OMC ensemble recorded across all singing (f) and nonsinging (g) epochs. Dimensions of this matrix reflect the total number of songs in this session (n = 29). (h) Correlation values across all songs are significantly higher during singing compared with nonsinging (one-sided Welch’s t-test, p = 3.0 × 10⁻¹³⁹). (i) Average correlation values for each recording session (mean ± S.E.M., n = 13 sessions, 4 mice). Red point refers to example session shown in (c)-(h).

Fig. 2. Note-related activity of OMC neurons.

(a) At top, singing behavior in a single S. teguina example song. At bottom, an expanded view of 7 notes from the above example. Horizontal lines represent the timing of notes, and the durations for each note (in ms) are provided below. (b) Histogram of note (n = 30,540) and song (n = 305) durations plotted on a logarithmic axis across all recorded mice in this study (n = 4). (c) Spiking activity corresponding to note timing for an example neuron. For visualization, the spike raster plot was restricted to notes within a range of 55 to 65 ms (full range: 31.4 to 175.9 ms). Green and red ticks indicate the onset and offset of notes, respectively. (d) Spiking activity of an example neuron linearly warped to a common note duration (onsets indicated by dashed green lines). Rasters (top) and spike probability density plots (bottom) are provided for the recorded spike trains (i) and after imposing a ‘sensory’ (− 30 ms) (ii) or ‘motor’ (+ 30 ms) (iii) offset. (e) Modulation strength and offset values for three example neurons. Gray circles and roman numerals in plot for Cell 36 refer to the corresponding panels depicted in (d) (see Methods). (f) Summary plot showing the best-fit latency (restricted to ± 50 ms) corresponding to the maximum note modulation strength for 96 neurons. Gray symbols represent cases that are not significantly different from zero, and red (n = 15) and blue (n = 2) symbols represent points with sensory and motor off-sets, respectively. The three example cells depicted in (e) are indicated.

Results

High-density silicon probe recordings in freely behaving singing mice

We recorded OMC neural activity during vocal production in four adult male S. teguina using high-density silicon probes (Cambridge NeuroTech or Diagnostic Biochips) (Fig. 1b, c). Electrodes were inserted to a final depth of 600–1000 µm, such that most recording sites were in the ventral portion (i.e., motor output layers) of OMC. We used this approach to monitor neural activity continuously over 3 to 20 days, and 13 sessions with robust vocal behavior (duration: 10.4 ± 5.7 hours, mean ± SD) were analyzed further. During these recording sessions, singing mice produced songs both spontaneously (n = 226) and in response to the playback of a conspecific vocalization (n = 79). For this study, which focuses on vocal production, we combined data across these conditions, yielding a total of 23 ± 17 (mean ± SD) songs per session (range: 8 to 72). In total, we recorded from 375 neurons (29 ± 11 per session, mean ± SD) whose spiking was stably monitored throughout those recording sessions (see Methods).

OMC spiking is modulated during vocal production

We began by examining whether OMC neural activity was related to singing behavior. Although song-related spiking patterns often differed across neurons (e.g., Fig. 1c), we found that the ensemble activity of simultaneously recorded OMC neurons was similar across song epochs compared to nonsinging periods (Fig. 1d, e). Since each session consisted of multiple songs, we calculated the correlation values of OMC ensemble activity across all pairs of songs and found them to be significantly greater compared to nonsinging epochs in the example session (Fig. 1f-h) as well as across all recording sessions (Corr_singing = 0.61 ± 0.11, Corr_nonsinging = 0.44 ± 0.12, p = 2.76 × 10⁻⁶, paired t-test) (Fig. 1i). Taken together, we find that OMC population activity is consistently modulated during song production.

Since OMC ensemble activity displayed reliable neural dynamics during singing, we next proceeded to characterize song-related spiking in individual OMC neurons. Each song is composed of a series of notes (Fig. 2a, b); therefore, neural activity could a priori be related to the production of each note at a fast timescale (approx. 100 ms), or it could follow slower dynamics at timescales comparable to the entire song (approx. 10 s). By statistically comparing neural activity during vocal production (versus nonsinging epochs), we found that 29.6 percent of neurons (111 out of 375) were correlated with notes (Extended Data Fig. 1a-c, Fig. 2c) while 35.5 percent (133 out of 375) of neurons displayed dynamics spanning the entire song (Extended Data Fig. 1d-f, see Methods), and 13.1 percent were active at both timescales. Therefore, more than half of individual OMC neurons were significantly modulated with some aspect of singing behavior.

Note-related responses of OMC neurons

Cortical activity has been shown to represent relevant kinematic features (e.g., velocity and force of effector muscles) for many movements [23]. Applying this framework to vocal production, we would expect OMC neurons to show phasic activity patterns preceding each note. To determine the relationship of OMC firing and note production, we linearly warped spiking activity to both the onset and offset of notes (Fig. 2d). A close inspection of note-related neurons revealed a diverse relationship between spike timing and note duration. For instance, in some cases, there appeared to be a systematic shift in the spike timing as note durations increased (e.g., Fig. 2di), which may arise from systematic offsets between neural activity and note production. Specifically, if this shift were due to a motor delay, or the timing needed for premotor signals to result in a behavioral change, activity would precede the production of notes [24]. Conversely, if the timing shift were due to sensory feedback, spiking activity would lag note production [25].

To explore these possibilities, we systematically varied the timing of spikes with respect to the audio recordings (Fig. 2d, Extended Data Fig. 2a, b) and determined the time lag that resulted in the most consistent alignment with notes (Fig. 2e, see Methods). Among the population of note modulated neurons, shifts resulted in significantly better alignment between neural activity and note phase in 25 cases (Fig. 2e, f, bootstrap p < 0.01, see Methods). Of these, 23 were consistent with sensory shifts and only 2 with motor offsets (Fig. 2f, Extended Data Fig. 2e). Based on the relative timing of neural activity and behavior, less than 1 percent (2 out of 375) of all recorded OMC neurons have a response profile consistent with a motor command for note production. Therefore, while we find phasic note-related activity in OMC, it is unlikely to be directly involved in the production of individual notes.

Precise temporal scaling of OMC neural dynamics with song duration

We next explored an alternative schema based on hierarchical control in which OMC population dynamics is dominated by a set of motor primitives (i.e., distinct patterns of neural activity) which do not directly represent movement kinematics [26]. In this view, motor commands for note production are determined by downstream vocal pattern generators driven by time-varying OMC activity spanning the duration of the song, a dynamical systems framework that has been proposed in other motor control studies [27, 28]. Therefore, we broadened our view to examine the extent to which neural activity relates to the structure of the produced song at timescales comprising the entire song duration (approx. 10 s).

We tested how OMC neural dynamics covaries with song duration, which can substantially differ across renditions (Fig. 3a). The activity of individual neurons may evolve with identical timing regardless of song duration and therefore be correlated with ‘Absolute Time’ (Fig. 3b). Consequently, dynamics associated with shorter songs would simply look like truncated versions of those observed during longer songs. Alternatively, OMC neurons could reflect ‘Relative Time’ (Fig. 3c), in which neural activity expands and contracts to track the progression through longer and shorter songs, respectively. To test these models, we analyzed trial-to-trial differences in song duration across renditions (average variation: 139.9 percent, n = 13 sessions, e.g., Fig. 3a) and used a similarity analysis to compare the firing patterns of each modulated neuron after the timing of activity had been linearly warped to align the onset and offset of song (Fig. 3d, e, Extended Data Fig. 3). The Absolute Time model would predict a higher degree of correlation when maintaining original timing and comparing initial portions of longer songs to shorter songs, while the Relative Time model suggests the opposite (i.e., higher correlation after warping). We therefore directly compared these two scenarios and found that the explained variance of single trial firing rates was significantly greater in the warped condition compared with the unwarped condition (p = 7.5 × 10⁻⁷, one-sided paired t-test) (Fig. 3f), supporting the Relative Time model of OMC neural dynamics.

Fig. 3. Scaling of neural activity with song duration.

(a) Duration of all songs (n = 143) produced from one example mouse (left). Raw waveforms for three example songs of different durations (right). (b and c) Hypothetical time-varying neural activity from a single neuron as predicted by the Absolute Time (b) and Relative Time (c) models for three songs of varying durations. (d) and (e) Spiking responses for a single neuron across 29 songs aligned to the start of the song (d) or temporally warped to the beginning and end of the song (e). (f) Comparison of explained variance for 133 song-modulated neurons across trials using recorded song times (Model 1, x-axis) and following temporal warping (Model 2, y-axis). Data are better fit by Model 2 (one-sided paired t-test, p = 3.95 × 10⁻⁷). (g) Peri-song time histograms (PSTHs) for three example neurons. Each trace represents an average of 4–21 similarly timed trials. Cell 19 is the same neuron shown in (d) and (e). Song blocks used to calculate consensus firing rate profiles are indicated by numbers and vertical lines. (h) Two example pairwise comparisons of the instantaneous firing plots from (g). For each pair, the black arrow indicates scaling factor with maximum correlation (Sneural), and the gray arrow shows the ratio of song times (Sbehavioral). (i) and (j) All pairwise comparisons (n = 10) of Sneural and Sbehavioral for the example neuron (i) (colored circles refer to panels in (h)) and for the entire population (n = 105 neurons, x-axis bin size: 0.05) (j) The error bars refer to the standard error of the median estimated by bootstrapping.

To further quantify the magnitude of time scaling for each neuron, we generated a consensus neural activity profile for songs with similar durations (Fig. 3g, Extended Data Fig. 3a-c, see Methods). For each pair of blocks, we compared the neural activity profiles to determine the scaling factor that maximized the pairwise correlation (e.g., Fig. 3h), which we call the neural scaling factor (S_neural). If the optimal neural scaling (i.e., the ratio of activity profiles leading to the highest correlation value) matched the relative ratio of associated song durations (S_behavioral), then the S_neural/S_behavioral slope is expected to be 1 (equivalent to the Relative Time model). When S_neural was plotted against the behavioral scaling factor (i.e., ratio of the associated song durations, S_behavioral), we found them to be linearly proportional (Fig. 3i, j). Across all the neurons, the neural scaling/behavioral slope was 1.01 ± 0.01 (n = 659 pairs, 105 neurons, Fig. 3j, see Methods). For comparison, the Absolute Time model would predict a slope of 0. This result demonstrates that activity of individual OMC neurons linearly stretches or compresses by a magnitude determined by the ratio of the song durations, enabling OMC activity to precisely track the proportion of elapsed song.

Diverse individual neuron dynamics in OMC

What are the motor primitives observed in OMC during vocalization? Since OMC circuit activity precisely scales with song duration, we linearly warped the firing rates of song-modulated neurons to both the onset and offset of song. Using this strategy, we observed diverse firing patterns within the OMC during vocalization (Fig. 4). To quantify this heterogeneity, we performed hierarchical clustering (Fig. 4a, see Methods) and found that 28.6 percent of neurons increased firing during song production while the remainder were suppressed. Further analyses of their response profiles revealed 8 distinct clusters of neurons (Fig. 4a, b). We observed that some neurons exhibited transient responses coincident with song onset (Cluster 7), song offset (Cluster 6), or both (Cluster 2), and other neurons showed more persistent increases (Clusters 1, 3) or decreases (Clusters 4, 5, 8) in neural activity during singing. Overall, neurons were responsive throughout the duration of the song and not just at song initiation and termination, consistent with moment-by-moment control of ongoing song production. We conclude that the population of OMC neurons that keep track of Relative Time (i.e., phase) shows diverse firing patterns during song production.

Fig. 4. Diverse categories of OMC firing patterns during singing.

(a) A hierarchical clustering plot describing the response profiles of OMC neurons whose activity was modulated during singing (see Methods, n = 105 neurons). Individual clusters are indicated by colored bars on the right. (b) Spiking responses for each cluster displayed as average firing rate plots. The mean activity profile of each neuron is represented with gray lines, and colored lines are average waveforms for each cluster corresponding to categories from (a). Black vertical bars indicate a normalized firing rate (z-score) of 1. Gray shaded blocks denote song epochs, with green and red arrows marking song starts and stops respectively.

Computational model of vocal motor control

To understand how motor commands for note timing can be generated from the motor primitives described above (Fig. 4), we next constructed a data-driven hierarchical model that makes experimentally testable behavioral predictions. In this model, OMC does not determine note timing directly (consistent with a lack of ‘premotor’ timing in Fig. 2), but vocal motor control is instead shared by cortical and downstream circuits. Inspired by our data, we posit that cortex dictates the moment-by-moment song phase and overall duration (Fig. 3), while the motor command for individual notes is generated by midbrain/brainstem areas comprising the primary vocal motor network (Fig. 5a, Extended Data Fig. 4). In the model, OMC activity provides descending synaptic drive, which influences the rate of note production in the subcortical song pattern generator (Fig. 5b). To account for the decreasing rate of note production with time, the synaptic drive onto the downstream note pattern generator may decrease throughout the song. We accomplish this in our model through linear weighting of OMC activity profiles directly measured in our recordings (Extended Data Fig. 4a) which sum up to produce synaptic drives with varying slopes (Fig. 5b). We model the workings of the note pattern generator such that individual notes are produced upon reaching a fixed firing rate threshold (see Methods), akin to an integrate-and-fire module. Appropriate time-scaling of cortical activity will thus result in songs of different durations without the need for modifying the note-generating mechanism (Fig. 5b). Importantly, this role of OMC is robust to the choice of the precise means by which note generation is implemented in the note pattern generator, either via postsynaptic adaptation mechanisms or synaptic drive from another brain region (Extended Data Fig. 4b, c).

Fig. 5. Hierarchical model of vocal motor control.

(a) Schematic depicting shared control of vocal production, where OMC controls song duration and rate of progression while individual notes are produced by a downstream note pattern generator. The synaptic drive to the note pattern generator is derived from OMC neural activity (see Extended Data Fig. 4). (b) Activity profiles of four model OMC neurons for a long song (purple) compared to a short song (cyan). Linear summation of neural activity creates the synaptic drive to the note pattern generator. The note pattern generator is modeled as an integrate-and-fire module, such that the rate of note production depends upon the strength of the OMC synaptic input. (c) Model output using seven different values of time-scaling, leading to a prediction in which the number of notes linearly covaries with song duration. Cyan and purple indicate examples from (b). (d) The number of notes scales with song duration in an example mouse (n = 144 songs, left) as well as across the population (n = 13 mice, right). Diagonal lines at right represent linear regression fits for each individual animal. Red line indicates data from the example at left. (e) Cooling OMC results in a shift in both song duration and number of notes in one example animal (Mouse 4, left). The average change in song duration and number of notes as the result of cooling for each animal (n = 9 mice, right). Across all animals, OMC cooling significantly increased average song durations (control: 8.0 ± 0.3 s, cooled: 9.3 ± 0.4 s, p = 0.002, paired t-test) as well as average number of notes (control: 92.8 ± 3.2, cooled: 103.9 ± 3.5, p = 0.004, paired t-test). Red line indicates data from Mouse 4 (left). (f) Hierarchical model of vocal motor control, wherein OMC confers flexibility to a downstream song pattern generator.

We next test a specific behavioral prediction of our hierarchical model to assess its validity. Our model predicts that songs become longer by incorporating more notes and not by increasing the duration of individual notes (Fig. 5b, c). Alternately, if note timing were directly triggered by note-modulated OMC activity (Fig. 2), longer songs would have the same number of notes with their durations proportionately stretched, as observed in the songbird [29, 30]. We tested these predictions by examining the structure of songs produced with different durations and found that the number of notes systematically increased as a function of song duration (n = 13 animals, 4 from this study and an additional 9 from a published data set [20]) (Fig. 5d), a finding that strongly agrees with our hierarchical model.

We considered a directed circuit perturbation to assess whether the relationship between notes and song duration relies upon activity within OMC. We reanalyzed a data set in which OMC was focally cooled in 9 mice [20]. Previous experimental [29, 31–33] and theoretical [34] work predicts that mild focal cooling should dilate the temporal profile of OMC neural activity thereby slowing the progression of subcortically controlled note production. For each animal, OMC cooling resulted in an increase in both song duration (control: 8.0 ± 0.3 s, cooled: 9.3 ± 0.4 s, p = 0.002, paired t-test) as well as the number of notes (control: 92.8 ± 3.2, cooled: 103.9 ± 3.5, p = 0.004, paired t-test) (Fig. 5e). Therefore, OMC-cooled songs became longer by incorporating more notes, further supporting the role of OMC activity in our hierarchical model. In sum, these results suggest that cortical activity can generate the necessary vocal motor commands to account for natural variability in behavior.

Discussion

In this study, we observed robust modulation of motor cortical activity during vocalization corresponding to two behaviorally relevant timescales: (1) phasic responses during note production (approx. 100 ms) and (2) persistent song-related dynamics (approx. 10 sec). We found that many neurons modulated at the faster timescale exhibited a delay between note timing and spiking that could represent either sensory feedback or efference copy signals (Fig. 5f). Sensory feedback is known to be important in animal and human vocal motor control [35–38], and a systematic perturbation of sensory streams (e.g., auditory, proprioceptive) [39] could test whether these signals are important in similar control processes in the singing mouse. Nevertheless, our time-shift analysis, modeling, and perturbation results confirm that these fast-varying responses in OMC do not reflect vocal motor commands to produce individual notes. At the slow timescale, responses were heterogeneous (e.g., transient at song onsets, ramping responses, etc.) and appear to reflect a set of motor primitives related to the control of song duration and the rate of note production. Future work will determine whether these spiking profiles map onto specific neuronal cell types in the OMC defined by critical circuit features, such as their output targets, as seen in motor cortical circuits in the laboratory mouse [40–42].

These results provide a striking example of how motor cortical dynamics can modulate song production, perhaps reflecting a voluntary mechanism of generating adaptive vocal flexibility. To accomplish this moment-to-moment control, our cortical recordings support a model in which OMC acts hierarchically via downstream song pattern-generator circuits (Fig. 5f, Extended Data Fig. 4b, c), likely corresponding to regions that have been recently characterized in the laboratory mouse [17–19] and appear to be highly conserved across vocalizing species [10, 11]. The hierarchical model proposed here is consistent with our previous work, where we found that OMC inactivation did not abolish singing but significantly reduced the variability in song durations [20], suggesting that activity in OMC is providing necessary input to the brainstem to generate socially appropriate vocalizations (Extended Data Fig. 4b, c). Future work is needed to determine the full song circuit in the singing mouse and elucidate the synaptic mechanisms by which OMC influences downstream vocal production circuits.

The singing mouse vocal control network appears to operate in a partially autonomous hierarchical configuration – a successful design principle for biological and artificial systems – wherein a higher order modulator (i.e., OMC) extends the capabilities of lower-level motor controllers (i.e., note production circuitry) without being necessary for generating the basic motor program [43–45]. Such an arrangement enables behavioral flexibility without relying upon synaptic plasticity in downstream motor patterning circuits. Similar mechanisms have been observed when animals are trained to keep track of time [46–51] or in primate cortex during cycling tasks at different speeds [52]. Our results extend the scope of this temporal scaling algorithm over an expanded time window (approx. 10 s) and to a new domain: controlling vocal flexibility in mammals. Despite its ubiquity, the neural mechanisms contributing to temporal scaling are not well-understood, though several ideas have been proposed, including feedback loops [46, 51] and neuromodulatory gain control [53]. The OMC circuit in the singing mouse offers a valuable opportunity to examine these and other circuit features for generating motor flexibility in the context of an ethologically-relevant behavior.