Introduction

To understand how the brain performs remarkable computational feats like speech perception, it is necessary to understand the cortical regions involved and the network architecture that connects them. The speech network in the human brain is postulated to have a largely hierarchical organization with information flow from subcortical structures to primary auditory cortex, then to lateral temporal cortex through successive dorsal and ventral pathway areas to reach apical targets in the frontal lobe1,2. Thus, activity in the frontal lobe during speech perception is thought to reflect downstream computational processes inherited from successive stimulus transformations in areas along the dorsal and ventral lateral cortex pathways (Fig. 1A, purple).

Fig. 1: Suprasylvian speech response characteristics.

A The classic hierarchical speech pathway from the medical geniculate body (MGB) of the thalamus to the front lobe is shown in the top panel. The orange arrows illustrate the hypothesized long-range parallel pathways from the medial geniculate body (MGB) and primary auditory cortex (PAC) to the frontal lobe. B Acoustic waveform of a single example sentence presented to participants. C Example responses to a single sentence from the right hemisphere participant in (D). D Representative example of two participants showing example responses in suprasylvian cortex and superior temporal gyrus (STG) to the single sentence shown in (B). The brain reconstructions show the spatial locations of responses to this sentence. This shows high similarity in response characteristics in the frontal lobe and STG. E Spatial distribution of significant responses to speech across all participants. This shows three areas of peak response probability in IFG, middle precentral gyrus (mPreCG), and ventral sensorimotor cortex (vSMC) (electrodes collapsed to the left hemisphere). F Spatial clustering using a Gaussian mixture model demonstrates three areas of speech responses in IFG, mPreCG, and vSMC. The number of clusters was determined by Bayesian information criterion and silhouette criterion values (both converged on 3 clusters)13. G Suprasylvian area versus STG responsivity. The proportion of electrodes with significant responses and the mean response amplitude is lower in IFG, mPreCG, and vSMC than in STG (p < 0.005, two-sided Wilcoxon rank-sum test, Bonferroni corrected, n proportion samples for left/right hem. = 9/8 IFG, 8/8 mPreCG, 9/8 vSMC, 9/8 STG, n response amplitudes for left/right hem. = 52/84 IFG, 120/180 mPreCG, 98/91 vSMC, 645/548 STG). Each area has no left-right hemispheric differences regarding the proportion or magnitude of significant responses (p > 0.05, two-sided Wilcoxon rank-sum test). All data are presented as mean values ± SEM. CS central sulcus; Hem hemisphere; IFG inferior frontal gyrus; IPL inferior parietal lobe; ITS inferior temporal sulcus; MGB medial geniculate body; mPreCG middle precentral gyrus; MTG middle temporal gyrus; PostC postcentral gyrus; PreC precentral gyrus; Spt Sylvian fissure at the parieto-temporal boundary; STG superior temporal gyrus; STS superior temporal sulcus; Sulc. sulcus; vSMC ventral sensorimotor cortex. A contains brain illustrations from Kenneth Probst, reproduced with permission, licensed under a Creative Commons Attribution 4.0 https://creativecommons.org/licenses/by/4.0/deed.en.

In addition to this hierarchical structure of the speech network, within nearby levels, there are short-range, local, parallel connections3,4. However, long-range parallel connections from the bottom of the network (i.e., sensory nuclei in the thalamus or primary auditory cortex) to apical regions in the frontal lobe have yet to be described (Fig. 1A, orange arrows). In support of such long-range parallel pathways, studies in primates found evidence for frontal lobe auditory responses that have short latencies in the range of those observed in lateral temporal cortex5,6. Additionally, auditory activity in these areas showed evidence of spectrotemporal representations7. These data, although in primates, raise the possibility of frontal cortex responses that may reflect parallel inputs from the auditory thalamus and primary auditory cortex, forming a long-range parallel pathway to the frontal cortex.

To test this hypothesis, we used electrocorticography (ECoG) to identify speech responses in the frontal and temporal lobes in awake participants while they passively listened to natural speech. We specifically asked whether there is evidence for: (1) short-latency speech-evoked responses that are synchronous between the frontal lobe and the earliest latency responses in the superior temporal gyrus (STG) and (2) whether such activity is explained by spectrotemporal encoding models as would be expected for direct inputs from low-level areas such as the primary auditory cortex or the medial geniculate body of the thalamus. Here, we identify three areas in the frontal lobe cortex that are activated by speech with latencies that occur simultaneously or even before the earliest latencies in the superior temporal gyrus (STG). Additionally, neural representations in these short-latency populations encode spectrotemporal content that is indistinguishable from populations in STG. Finally, to further establish evidence for anatomic connections that reflect parallel inputs from lower-level areas to the frontal lobe, we use white matter tractography and functional connectivity and find connections between frontal lobe regions and the auditory nucleus of the thalamus (MGB) and primary auditory cortex. Overall, these results demonstrate a fundamental divergence in the hierarchical architecture that is usually assumed to underlie cortical speech processing and shows multiple lines of evidence for long-range parallel inputs from low-level areas to frontal lobe cortical regions in the human speech network.

Results

Frontal lobe areas are activated by passive speech listening

To characterize the extent to which neural populations in frontal lobe respond to speech, ECoG participants passively listened to 10–40 min of natural speech, which consisted of prerecorded sentences (2–4 s each) from the phonetically transcribed TIMIT speech corpus8. We extracted activity in the high-gamma (70–150 Hz) range9, which correlates with spiking activity10, spike-based tuning properties11, and reflects contributions from single neurons throughout the cortical depth12. At single electrodes, average responses to an example sentence (Fig. 1B) were highly robust and extended through the duration of the sentence in STG and three areas in the frontal lobe, including middle precentral gyrus (mPreCG; Fig. 1C; yellow), ventral sensorimotor cortex (vSMC; Fig. 1C; orange), and inferior frontal gyrus (IFG; Fig. 1C; blue). Figure 1D shows two example participants that illustrate typical electrode coverage in the left and right hemispheres and responses to a single sentence of speech that are qualitatively similar to responses in STG.

Across all 17 participants, we quantified the spatial distribution of responses to speech in the frontal and parietal cortex. We observed significant responses throughout the areas covered by ECoG grids, with three peaks of responsiveness in mPreCG, vSMC, and IFG (Fig. 1E, significant responses defined as electrodes with p < 0.05 for any response time bin, Bonferroni corrected for the number of time bins in each sentence, Wilcoxon rank-sum test). To further test whether these three peaks reflect spatially clustered responses, we used Gaussian mixture modeling to cluster the electrodes. The three areas that emerged reflect clusters of electrodes in mPreCG, vSMC, and IFG (Fig. 1F; cluster number = 3 determined using both Bayesian information criterion, p < 0.05, and Silhouette criterion13 p < 0.05). Thus, in addition to STG, there are three regions within the frontal lobe cortex that have significant responses to natural speech.

We compared the basic response properties in these three suprasylvian areas to STG, which is known to be a critical area that is central to speech perception14,15. To characterize the overall responsivity, we first quantified the proportion of all electrodes within an area that showed a significant response to speech (Fig. 1G, top row). Although all three regions had electrodes with significant speech-evoked activity in both hemispheres, there was a lower proportion of speech-responsive electrodes in frontal lobe areas compared to the proportion of speech-responsive electrodes in STG (p < 0.005, two-sided Wilcoxon rank-sum test, Bonferroni corrected; Fig. 1G, top row). Additionally, while significantly greater than zero, the average response magnitude in suprasylvian areas was lower than STG (p < 0.005, two-sided Wilcoxon rank-sum test, Bonferroni corrected; Fig. 1G, bottom row). For both the proportion of significant electrodes and the response magnitude, there were no significant differences between hemispheres (p > 0.05; two-sided Wilcoxon rank-sum test), suggesting that frontal lobe speech-evoked activity is similar between right and left hemispheres16,17.

Frontal lobe areas have short-onset latencies like the superior temporal gyrus

Having established the existence of robust speech-evoked activity throughout bilateral frontal lobe cortex during passive listening, we asked whether any of these neural populations had short-onset latencies that are synchronous or faster than the earliest onset latencies in lateral temporal cortex. If present, this would be consistent with parallel pathways from lower-order auditory areas to frontal and temporal lobes. To address this question, we were careful to only select participants with extensive coverage across STG and these frontal lobe areas within a single participant (Supplementary Fig. 1). The seventeen participants included in this study is a subset of all participants with grid-based coverage over a 10-year span in our center. Additionally, for higher spatial resolution, all participants have specialized high-density recording grids with 4 mm inter-electrode distance in contrast to 10 mm inter-electrode distance used at other centers18. Direct visual examination of speech-evoked activity showed near-synchronous shortest latency responses across frontal areas and STG (Fig. 2A, B, D). Furthermore, there were within-participant examples of frontal electrodes with onset latencies synchronous or earlier than the shortest onset latencies in the temporal lobe (Fig. 2C). To quantify this, we calculated the response onset latencies across all participants, similar to prior characterizations in humans and primates19,20 (Fig. 2A, D). This showed electrodes throughout the temporal, frontal, and parietal lobes with short response latencies, often less than 80 ms.

Fig. 2: Frontal lobe areas have short-onset latencies like the superior temporal gyrus.

A Spatial map of response latencies show short-onset latencies in STG, IFG, mPreCG, and vSMC (onset defined as first 1-ms time bin in which the response p-value is <0.05 for 15 consecutive 1-ms bins19,20; Wilcoxon rank-sum test). B All responses across participants with onset latencies less than 200 ms. C Two example participants showing the electrode sites with shortest onset latency in the STG and frontal lobe. As shown, the earliest onset latencies within a subject are similar between the frontal lobe and STG**. D** Latency distribution by area for onset latencies less than 160 ms (the proportion y-axis is relative to the full distribution of latencies up to 1600 ms). This shows the earliest onset latencies are similar across areas and is consistent with low-latency parallel inputs to each area. E Latency distribution boxplots for all onset latencies up to the temporal cutoff specified on the x-axis (80–720 ms). There is no significant difference between STG and IFG, mPreCG, or vSMC up to 240 ms in the right hemisphere and 200 ms in the left hemisphere (* = p < 0.05, two-sample Kolmogorov–Smirnov test, Bonferroni corrected, n latencies for left/right hem. = 52/84 IFG, 120/180 mPreCG, 98/91 vSMC, 330/274 STG). All box plots show the median (circle), 25th and 75th percentiles (box). These indistinguishable short-onset latency times are consistent with a subset of parallel inputs to the frontal cortex and STG rather than strictly hierarchical inputs from STG to the frontal cortex. CS central sulcus; IFG inferior frontal gyrus; mPreCG middle precentral gyrus; STG superior temporal gyrus; vSMC ventral sensorimotor cortex.

For each region, we quantified the latency distribution, which showed similar short-latency responses across all areas (Fig. 2D). To test for significant differences in short-onset latencies between areas, we used a temporal cutoff to define the left-sided tail of each distribution as the short-onset latencies of interest (Fig. 2E). The temporal cutoff ranged from 80 to 720 ms, with significance testing at each cutoff. As shown, onset latencies were not significantly different between frontal lobe areas (IFG, mPreCG, vSMC) and STG for latencies up to 200 ms in the left hemisphere and 240 ms in the right hemisphere (Fig. 2E, p > 0.05, two-sample Kolmogorov–Smirnov test, Bonferroni corrected). This data shows short-onset latencies within the same hemisphere are not significantly different between frontal lobe areas and STG. Thus, in response to natural speech, IFG, mPreCG, and vSMC are active beginning at the same time as neural populations in STG, consistent with parallel inputs to STG and frontal lobe areas.

Short-onset latency inputs to the frontal and temporal lobe encode the same spectrotemporal speech information

A feature of most functional-anatomical models of speech is frontal lobe areas receive inputs from the temporal lobe through dorsal and ventral pathways and may reflect qualitatively different (and perhaps higher-order) representations of speech1,2. Given the existence of parallel short-latency responses in the frontal and temporal lobes, we asked if the spectrotemporal information encoded in these responses fundamentally differed between lobes. To test this, we computed spectrotemporal receptive fields (STRFs, Fig. 3A) for short-onset latency sites (<200 ms) in STG, IFG, mPreCG, and vSMC. In all three frontal lobe areas, responses to speech were well-predicted by STRFs (mean r = 0.39 ± 24, Fig. 3B), with some electrodes reaching r = 0.75, demonstrating that frontal lobe neural population activity is well-explained by spectrotemporal models.

Fig. 3: Sites with short-onset latencies in frontal cortex encode the same spectrotemporal information as STG.

A Example middle precentral gyrus STRF with the predicted and actual response to a single sentence. Predicted responses are obtained by convolving the spectrogram with the STRF and are proportional to the similarity between the stimulus’s spectrotemporal content and the STRF structure. B STRF prediction value distributions for all short-latency electrodes with a significant response to speech in mPreCG, VSMC, and IFG (suprasylvian cortex global mean Pearson correlation coefficient 0.39 ± 0.24; tested on held-out data). STRFs, which are predictive of neural responses, characterize spectrotemporal encoding at each site and are consistent with the presence of spectrotemporal representations in these areas. All box plots show the median, 25th and 75th percentiles. Notches approximate the 95% confidence interval of the median. n site prediction values = 46 IFG, 115 mPreCG, 110 vSMC, and 389 for STG. C Examples showing the high degree of STRF similarity between frontal lobe areas and STG. D There is no significant difference in short-onset (<200 ms) STRF tuning parameters across frontal lobe sites and STG, including: best frequency tuning (p > 0.05 Kruskal–Wallis test), spectral, or temporal tuning (p > 0.05 cosine similarity permutation test of mean spectral and temporal tuning vectors). Spectral and temporal tuning plots show the mean ± s.e.m for each area in each hemisphere; the total contralateral hemisphere mean tuning is plotted as a dotted line for reference. E Ensemble modulation transfer functions by area. The average modulation tuning for each area has a high degree of similarity. F There is no significant difference in temporal or spectral modulation tuning between frontal lobe areas and STG (p > 0.05, cosine similarity permutation test of mean spectral and temporal modulation tuning vectors26). BF best frequency; CC Pearson correlation coefficient; c/o cycles per octave; Hem. hemisphere; IFG inferior frontal gyrus; mPreCG middle precentral gyrus; Spec. Mod. spectral modulation; STRF spectrotemporal receptive field; Temp. Mod. Temporal Modulation; vSMC ventral sensorimotor cortex.

Next, we asked whether the spectrotemporal content encoded in short-onset (<200 ms) frontal lobe electrodes is similar to STG. Based on visual examination of short-onset STRFs from each frontal lobe area and STG, there was a high degree of similarity in spectrotemporal encoding (Fig. 3C). Additionally, the STRF features correlated with spectrotemporal structure seen in phonetic features similar to STG21. For example, the bottom row of Fig. 3C shows broadband excitatory tuning followed shortly by broadband inhibitory tuning. This is the canonical spectrotemporal structure that is tuned for plosives in phonetic feature space21. To quantify this we examined the STRF weights and computed three key parameters (Fig. 3D): (1) the best frequency (first column)22,23, (2) spectral tuning (second column)24, and (3) temporal tuning (third column)24,25. For each of these parameters, there was no significant difference across regions (p > 0.05 Kruskal–Wallis test and cosine similarity permutation test of mean spectral and temporal tuning vectors26).

While individual STRF tuning properties were not different between STG and frontal lobe areas, the joint combination of these tuning properties in each STRF characterizes the spectrotemporal processing at each site. To jointly characterize spectrotemporal processing for each electrode across regions, we transformed the STRFs into their modulation transfer functions (MTFs)23,27,28. MTFs are the modulation domain representation of spectrotemporal processing characterized by STRFs28,29. They summarize neural processing in terms of spectral and temporal modulation tuning and are frequently used to characterize processing within the auditory system. As shown in Fig. 3E, the ensemble (mean) modulation transfer function across frontal lobe areas and STG is similar. To quantify this, the mean temporal modulation tuning and spectral modulation tuning distributions derived from the MTFs were calculated (Fig. 3F), and no significant differences in temporal or spectral modulation tuning were observed across frontal lobe areas and STG (p > 0.05, cosine similarity permutation test of mean spectral and temporal modulation tuning vectors26). Together, these results indicate that the short-latency responses in frontal lobe areas and STG encode the same spectrotemporal representations of speech, consistent with parallel spectrotemporal speech representations in the temporal and frontal lobes.

While short-onset (<200 ms) electrodes in the frontal lobe are well-predicted by spectrotemporal speech representations, it is possible that they reflect higher-order linguistic features that are correlated with the speech spectrogram. To test this alternative, we compared encoding performance for STRFs to semantic-based encoding models generated with word vector representations of speech derived from the FASTTEXT data set30. We expected that both spectrotemporal and semantic encoding models would perform well in each region, however, they would differ from each other as a function of short or long onset latency. Specifically, we hypothesized that the STRF would be a better model for electrodes with short-onset latencies since these populations reflect direct inputs from low-level areas such as the medial geniculate body or primary auditory cortex.

Confirming our hypothesis, we found that electrodes with short-onset latencies (<200 ms) in IFG, mPreCG, and vSMC show significantly higher STRF model predictions than semantic model predictions, consistent with dominant spectrotemporal representations at those sites (Fig. 4A, p < 0.01, Wilcoxon rank-sum test). At longer latencies (>200 ms), spectrotemporal and semantic encoding models were not significantly different, further supporting the notion that these short latency electrodes reflect spectrotemporal representations. Overall, these data demonstrate that neural populations in the frontal lobe with short-onset responses encode spectrotemporal representations of sound similar to populations in more low-level auditory areas and STG.

Fig. 4: Spectrotemporal encoding models are better than semantic encoding models for short-onset latency sites in the frontal lobe and STG.

A Short-onset latency sites in each area encode low-level spectrotemporal information. Sites with onset latencies less than 200 ms show significantly higher STRF model predictions than semantic model predictions consistent with dominant spectrotemporal over semantic coding at these short-onset latency sites. This is consistent with short-onset latency sites encoding spectrotemporal speech information in parallel (p < 0.001, Wilcoxon rank-sum test, two-sided). Longer onset latency sites in the same areas show no difference between spectrotemporal or semantic encoding models. All data are presented as mean values ± SEM. n site CC prediction values for early (<200 ms)/late (>200 ms) onset latencies = 46/31 IFG, 115/68 mPreCG, 110/43 vSMC, 389/72 STG. CC Pearson correlation coefficient; STRF spectrotemporal receptive field.

White matter tractography and functional connectivity show connections from thalamus and primary auditory cortex to frontal lobe areas

To establish whether the short-latency responses with STG-like spectrotemporal encoding observed in frontal lobe cortex reflects a direct anatomical pathway from lower-level auditory areas, we used diffusion tensor imaging (DTI) tractography. Based on the known connectivity patterns with STG, we hypothesized that the medial geniculate body (MGB) within the thalamus and the primary auditory cortex within Heschl’s gyrus are key areas that might have parallel projections to the frontal lobe cortex.

Using data from 842 individuals from the Human Connectome project31, we calculated white matter tractography in each hemisphere with the MGB and Heschl’s gyrus as seed regions. To localize the MGB, we used MNI coordinates in conjunction with the terminal point of white matter tracts from the inferior colliculus to the thalamus (Fig. 5A). We restricted the suprasylvian target region of interest to the total area spanned by the lateral frontal and parietal cortex. We identified tracts from MGB that projected to IFG, mPreCG, and right vSMC (Fig. 5A). We also identified tracts from Heschl’s gyrus that projected to vSMC and left IFG.

Fig. 5: White matter tractography and functional connectivity show connections from thalamus and primary auditory cortex to frontal lobe areas.

A White matter tractography between the medial geniculate body (MGB), Heschl’s gyrus (HG), and suprasylvian cortex (IFG, middle precentral gyrus (mPreCG), and ventral sensorimotor cortex (vSMC)). White matter tracts are color-coded by the suprasylvian site of termination. This analysis demonstrates tractography data consistent with direct MGB-to-frontal cortex and HG-to-frontal cortex white matter tracts. B Resting-state functional connectivity of the medial geniculate body and C Heschl’s gyrus showing functional connectivity to frontal lobe areas, middle precentral gyrus, ventral sensorimotor cortex, and inferior frontal gyrus (connectivity maps derived by performing one-sample t-tests, one-sided, with a statistical threshold of p < 5 × 10−5 after peak-level family wise error (FWE) correction for multiple comparisons). Colorbar represents T-score. Overall, these data provide additional structural and functional evidence for parallel pathways to frontal lobe areas from low-level areas, such as Heschl’s gyrus and the medial geniculate body. CS central sulcus; FC functional connectivity; HG Heschl’s gyrus; IFG inferior frontal gyrus; IC inferior colliculus; MGB medial geniculate body; mPreCG middle precentral gyrus; vSMC ventral sensorimotor cortex.

To test whether identified tracts relate to functional measures, we also calculated resting-state functional connectivity between the medial geniculate body or Heschl’s gyrus to all of the cortex using fMRI. For the MGB seed region, there was significant functional connectivity with IFG, mPreCG, and vSMC (Fig. 5B, p < 5 × 10−5 after peak-level family-wise error (FWE) correction for multiple comparisons). For the Heschl’s gyrus seed region, there was significant functional connectivity to mPreCG and vSMC (Fig. 5C, p < 5 × 10−5 after peak-level family-wise error (FWE) correction for multiple comparisons).

Finally, although limited to the supplementary material due to the scarcity of data (Supplementary Fig. 2), we show single pulse stimulation in the primary auditory cortex with cortico-cortical evoked potentials in these frontal lobe areas consistent with the DTI results above. Together, white matter tractography, resting-state functional connectivity, and single-pulse stimulation demonstrate that short-latency evoked activity that encodes speech-relevant spectrotemporal content in frontal lobe cortex is consistent with a parallel auditory pathway that functions alongside the classical hierarchical speech network mediated by STG.

Discussion

We examined the nature of short-latency, speech-evoked activity in frontal lobe areas that are not typically associated with auditory processing. Using high-density ECoG, we found that a subset of neural populations in the frontal lobe cortex exhibited response latencies that were synchronous with or preceded the earliest response latencies in STG. Spectrotemporal representations in these populations were largely indistinguishable from those found in STG, encoding key spectral and temporal modulation rates for speech. Finally, we identified white matter tracts that connect early auditory regions like MGB and Heschl’s gyrus with frontal lobe areas and were associated with resting-state functional connectivity. Together, these results indicate the existence of long-range pathways between low-level areas (medial geniculate body, primary auditory cortex) and neural populations in the frontal cortex, which work in parallel to the more well-established hierarchical pathway via lateral cortex areas.

Sensory systems, including vision and hearing, are known to have hierarchical structure, with serial feedforward connections from low-level areas to mid-level and then to apical areas within the frontal cortex1,2,32. In addition to this hierarchical structure, locally, within nearby levels, there are short-range parallel connections3,4. However, the presence of long-range parallel connections in humans from low-level areas, such as the primary auditory cortex or MGB directly to apical areas in frontal cortex have not been described. The present results demonstrate that, in addition to short-range parallel connections, long-range parallel connections exist from the thalamus and primary auditory cortex to areas in the frontal lobe. To what extent these different proposed parallel pathways exist is an important question for further work and could be examined further in participants with depth electrode coverage or tracer studies in primates.

It is important to emphasize the exploratory nature of the tractography results. While focusing on the robust ECoG findings in combination with DTI data, we can present broad interpretation that respects the complexity of the neural architecture involved and produces a model that can be tested with more robust structural analysis of these pathways. DTI-based methods have inherent limitations, including issues of resolution, crossing fibers, and challenges in differentiating close anatomical structures such as the MGB and adjacent thalamic nuclei33,34. Although steps within the scope of the study were taken to mitigate these limitations, these findings require further validation via post-mortem ultra-high-resolution MRI, virtual lesioning35, tracer injections in primates, and MGB single pulse stimulation in primates36. Additionally, future work in which preoperative high-resolution DTI and fMRI recordings in study participants could show that the individual electrode responses are predicted by single-subject connectivity for additional support of the model.

Perhaps a surprising result was how similar spectrotemporal representations encoded by early onset responses are between the frontal lobe and STG. Based on the model of long-range parallel pathways proposed here, frontal neuronal populations and STG would receive the same low-level spectrotemporal representations in MGB or primary auditory cortex to serve as raw snapshot of the of incoming speech signal. This is analogous to work from our group showing primary auditory cortex in Heschl’s gyrus, and STG receive rapid parallel inputs that convey spectrotemporal representations of speech4. Rapid snapshots of raw spectrotemporal information to frontal lobe areas may help with top-down modulatory functions carried out by the frontal lobe37,38,39,40 or real-time spectrotemporal feedback during speech production41,42. Further comparisons between the information encoded in fast frontal lobe responses and longer latency frontal lobe responses (reflecting inputs from canonical hierarchical pathways) is one future direction to investigate the function of these rapid parallel inputs to frontal lobe areas. A second future direction is specialized orthogonal stimuli designed to test which lower-level speech representation is dominant (spectrotemporal versus phonemic versus phonetic feature representations)21,23,43. Lastly, due to experimental time constraints, our standard protocol is focused on presenting speech sounds. It is less common to collect data that could speak to questions of whether these early frontal lobe responses are specific to speech. Future work using simple tones, clicks, or white noise bursts could investigate this important question.

It may be surprising to find such long-range connections, particularly as part of a network involved in processing such a complex stimulus. However, three important frontal lobe characteristics suggest a long-range parallel auditory pathway. First, these frontal areas are highly heterogeneous, and sub-populations of neurons in primates appear to be responsive to the acoustic properties of sound, similar to the tuning properties seen in lower-level areas7. Second, prior work in primates has suggested the presence of short-onset latencies in the frontal lobe to sound5,6, and in humans, electrical stimulation of primary auditory cortex elicits short-latency responses in the inferior frontal gyrus44. Lastly, in rhesus macaques, non-primary auditory belt regions have tracer-defined anatomic projections to the frontal lobe45. The present work shows short-latency neural populations with acoustic representations that are indistinguishable from STG, consi