Intracranial recordings from human auditory cortex reveal a neural population selective for song

Abstract

24What is the neural basis of the human capacity for music? Neuroimaging has suggested some 25 segregation between responses to music and other sounds, like speech. But it remains unclear 26 whether finer-grained neural organization exists within the domain of music. Here, using intracranial 27 recordings from the surface of the human brain, we demonstrate a selective response to music with 28 vocals, distinct from responses to speech and to music more generally. Song selectivity was evident 29 using both data-driven component modeling and single-electrode analyses, and could not be 30 explained by standard acoustic features. These results suggest that music is represented by multiple 31 neural populations selective for different aspects of music, at least one of which is specialized for 32 the analysis of song. 33 3 Music is a quintessentially human capacity that is present in some form in nearly every society 34 (Savage et al., 2015;Lomax, 2017; Mehr et al., 2018), and that differs substantially from its closest 35 analogues in non-human animals (Patel, 2019). Researchers have long debated whether the human 36 brain has neural mechanisms dedicated to music, and if so, what computations those mechanisms 37 perform (Patel, 2012;Peretz et al., 2015). These questions have important implications for 38 understanding the organization of auditory cortex (Leaver and Rauschecker, 2010; Norman-39 Haignere et al., 2015), the neural basis of sensory deficits such as amusia (Peterson and 40 Pennington, 2015;Norman-Haignere et al., 2016;Peretz, 2016), the consequences of auditory 41 expertise (Herholz and Zatorre, 2012), and the computational underpinnings of auditory behavior 42 (Casey, 2017;Kell et al., 2018). 43 44Neuroimaging studies have suggested that representations of music diverge from those of other 45 sound categories in non-primary human auditory cortex: some non-primary voxels show partial 46 selectivity for music compared with other categories (Leaver and Rauschecker, 2010; Fedorenko et 47 al., 2012;Angulo-Perkins et al., 2014), and a recent study from our lab, which modeled voxels as 48 weighted sums of multiple response profiles, inferred a component of the fMRI response with clear 49 selectivity for music (Norman-Haignere et al., 2015). However, there are few reports of finer-grained 50 organization within the domain of music (Casey, 2017), potentially due to the coarse resolution of 51 fMRI. As a consequence, we know little about the neural code for music. 52 53Here, we tested for finer-grained selectivity for music using intracranial recordings from the human 54 brain (electrocorticography or ECoG) ( Fig 1A). We measured ECoG responses to a diverse set of 55 165 natural sounds, and submitted these responses to a novel decomposition method that is well-56 suited to the statistical structure of ECoG to reveal dominant response components of auditory 57cortex. This data-driven method revealed multiple music-and speech-selective response 58 components. Our key finding is that one of these components re...

Multiscale integration organizes hierarchical computation in human auditory cortex

“…We ensured that differences in reliability could not explain our results in two ways: (1) we ensured that our model-estimated integration periods were unbiased by low data reliability (see Model-estimated integration periods and Simulations below) (2) we repeated our analyses using a denoising procedure that substantially increased the reliability of the electrode responses to a level well-above the point at which reliability might affect our integration period estimates. Our denoising procedure was motivated by the observation that iEEG responses are relatively low-dimensional and as a consequence much of the stimulusdriven response variation is shared across subjects 54 , in contrast with the noise which differs from subject to subject. We thus projected the electrode responses from one subject onto the responses from all other subjects (using regression), which has the effect of throwing out response variation that is not present in at least two subjects.…”

“…This projection is error prone because faraway points on the cortical surface can be nearby in space due to cortical folding. To minimize gross errors, we preferentially localized sound-responsive electrodes to regions where sound-driven responses are likely to occur 54 . Specifically, we calculated the likelihood of observing a significant response to sound using a recently collected fMRI dataset, where responses were measured to a large set of natural sounds across 20 subjects with whole-brain coverage 33 (p < 10 -5 , measured using a permutation test).…”

Music-selective neural populations arise without musical training

“…This approach makes it possible to disentangle the responses of neural populations that overlap within voxels, and has previously revealed a neural population with clear selectivity for music compared to both other realworld sounds (Norman-Haignere et al, 2015) and synthetic control stimuli matched to music in many acoustic properties (Norman-Haignere and McDermott, 2018). These results have recently been confirmed by intracranial recordings, which show individual electrodes with clear selectivity for music (Norman-Haignere et al, 2019). Although Norman-Haignere et al (2015) did not include actively practicing musicians, many of the participants had substantial musical training earlier in their lives.…”

“…Finally, our study is limited by resolution of fMRI. Voxel decomposition is intended to help overcome the spatial limitations of fMRI, and indeed appears to reveal responses that are not evident in raw voxel responses but can be seen with finer-grained measurement substrates such as electrocorticography (Norman-Haignere et al, 2019). But the spatial and temporal resolution of the BOLD signal inevitably constrain what is detectable, which likely explains why we have only been able to detect six reliable response components across all of auditory cortex.…”

EEG-based classification of natural sounds reveals specialized responses to speech and music

“…64 A few studies have used EEG to compare neural responses to speech and musical instrument 65 sounds (Cossy et al, 2014;Murray et al, 2006), and to our knowledge only one study has 66 compared time-varying differences in MEG responses to speech and instrument sounds to 67 differences in acoustic features (Ogg et al, 2019a). Still, it is unclear if these responses are 68 unique for speech and music (see Giordano et al, 2014;Norman-Haignere et al, 2015, 2019. 69…”