Concurrent temporal channels for auditory processing: Oscillatory neural entrainment reveals segregation of function at different scales — TEI-tagged view

TengpersName, TianpersName, RowlandpersName, PoeppelpersName (2017date). “Concurrent temporal channels for auditory processing: Oscillatory neural entrainment reveals segregation of function at different scales.” PLoS Biologytitle·j.

Object of study — whether the auditoryterm system encodes temporal modulations equivalently across an intermediate range, or only within restricted bandsterm

Research question — whether the alphaterm bandterm, which a decade of studies ties to auditoryterm perceptionterm, reflects the processing of acoustic information in the same way as the bandsterm above and below it

Contribution — two timescalesterm track and the one between them does not - a hole in the middle of the range, which the paper reads as segregationterm of function in the time domainterm

Full text — tagged

The reference list is not part of the text we hold for this paper. Our copy was extracted from the publisher's page and the bibliography did not survive that extraction, so the article below ends before its references. Nothing has been removed by us.

Abstract Natural soundsterm convey perceptually relevant information over multiple timescalesterm, and the necessary extraction of multi-timescale information requires the auditoryterm system to work over distinct ranges. The simplest hypothesis suggests that temporal modulations are encoded in an equivalent manner within a reasonable intermediate range. We show that the human auditoryterm system selectively and preferentially tracks acoustic dynamics concurrently at 2 timescalesterm corresponding to the neurophysiological thetaterm bandterm (4–7 Hz) and gammaterm bandterm ranges (31–45 Hz) but, contrary to expectation, not at the timescaleterm corresponding to alphaterm (8–12 Hz), which has also been found to be related to auditoryterm perceptionterm. Listeners heard synthetic acoustic stimuli with temporally modulated structures at 3 timescalesterm (approximately 190-, approximately 100-, and approximately 30-ms modulation periods) and identified the stimuli while undergoing magnetoencephalography recording. There was strong intertrial phase coherenceterm in the thetaterm bandterm for stimuli of all modulation ratesterm and in the gammaterm bandterm for stimuli with corresponding modulation ratesterm. The alphaterm bandterm did not respond in a similar manner. Classificationterm·mth analyses also revealed that oscillatory phaseterm reliably tracked temporal dynamics but not equivalently across rates. Finally, mutual informationterm analyses quantifying the relation between phaseterm and cochlear-scaled correlationsterm·mth also showed preferential processing in 2 distinct regimes, with the alphaterm range again yielding different patterns. The results support the hypothesis that the human auditoryterm system employs (at least) a 2-timescale processing mode, in which lower and higher perceptual sampling scales are segregated by an intermediate temporal regime in the alphaterm bandterm that likely reflects different underlying computations.

Author summary Correctly perceiving behaviorally significant soundsterm—speechterm, music, and the acoustic environment—requires integrating acoustic information over time to extract relevant regularities. A fundamental question about this process is: How does the auditoryterm brain integrate information of continuously varying soundsterm, typical of many natural auditoryterm signals? How does the brain “sample” the input? To investigate this question, we measured how cortical activity is entrained by soundterm using the noninvasive technique magnetoencephalography. We presented soundsterm with temporal structure at different timescalesterm and examined how the brain encodes such signals. We found, unexpectedly, that the human auditoryterm system does not treat all rates equally but predominantly uses 2 nonoverlapping timescalesterm, the slower (thetaterm) and faster (gammaterm) bandsterm, to track acoustic dynamics, while the timescaleterm corresponding to an intermediate (alphaterm) timescaleterm is likely reserved for other cortical operations, perhaps relating to attentionterm and functional inhibition. The data support the hypothesis that the human auditoryterm system employs (at least) a 2-timescale processing mode and that the perceptionterm of natural soundsterm, which feels seamless and continuous, is underpinned by segregated and discontinuous neural processing.

Introduction Speechterm, music, and many natural soundsterm have a rich temporal structure over multiple timescalesterm [ 1 – 5 ]; such soundsterm contain perceptually critical information that is encoded over short periods (e.g., the identity and exact sequence of phonemes in a spoken word) and, concurrently, information encoded over longer periods (e.g., the intonation change over a word that signals intent or affect). Successful perceptual analysis of these signals requires the auditoryterm system to extract acoustic information at multiple scales. This presents a specific problem: how does the auditoryterm system process different, co-occurring rates of information across multiple timescalesterm? And, by extension, how can the requirements on temporal and spectral resolution simultaneously be met? To derive the appropriate perceptual representations, the auditoryterm system must extract rapidly varying information on a scale of milliseconds (approximately 10–50 ms), operating with high temporal resolution, and concurrently analyze more slowly varying signal attributes on a scale of hundreds of milliseconds (about 150–300 ms), enabling sufficient spectral resolution [ 6 ]. A strictly hierarchical model, starting, say, with short/small integration windows at more peripheral processing regions, which are then concatenated to build longer windows, cannot by itself meet the perceptual demands [ 7 – 9 ].

Behavioral research suggests that the human auditoryterm system may optimize processing by operating within separate temporal ranges instead of in a unitary way across a continuum of temporal variation [ 10 – 15 ]. Timescalesterm on the order of tens of milliseconds are argued to be optimized for rapid temporal integration, such as in modulation detection [ 16 ], gap detection [ 17 ], and nonsimultaneous masking [ 18 ]. On the other hand, previous models of temporal integration typically assuming leaky integration demonstrate timescalesterm above 150 ms, such as loudness summation [ 19 , 20 ], signal detection in noiseterm [ 21 , 22 ], and temporal integration at threshold [ 14 , 23 – 26 ].

Here, we investigate whether these very clear behavioral results arguing for distinct timescalesterm could be illuminated, especially in terms of their neural implementation, by considering cortical oscillationsterm. Cortical oscillationsterm reflect rhythmic activity of neural populations at different scales [ 27 , 28 ], and the frequencies of cortical oscillationsterm are thought to reveal the corresponding temporal characteristics of sensory processing [ 29 – 32 ]. Therefore, by hypothesis, neural activity in the human auditoryterm system may show a multiscale oscillatory pattern, which could reflect temporal scales of auditoryterm processing. A short processing timescaleterm, which guarantees high temporal resolution and may reflect the extraction or decodingterm of fine-grained acoustic information, could be reflected in the gammaterm bandterm of oscillationsterm that tracks fast acoustic dynamics, 30–50 Hz. A longer processing timescaleterm, which integrates acoustic information on a timescaleterm of 150–300 ms, could be reflected in the thetaterm bandterm oscillationterm, 4–7 Hz.

Of course, soundsterm contain information in ranges lying between these well-separated temporal regimes. Is there really such a compelling segregationterm of function in the time domainterm? Apart from the 2 timescalesterm mentioned above, cortical oscillationsterm within the alphaterm bandterm (8–12 Hz) have been found in many studies to critically correlate with auditoryterm attentionterm [ 33 – 35 ], auditoryterm working memory [ 36 , 37 ], and listening effort [ 34 , 38 ] and, furthermore, to predict speechterm intelligibility under challenging environments [ 39 ], which demonstrates a significant role for the alphaterm bandterm in auditoryterm processing. It is reasonable to hypothesize that the alphaterm bandterm is fundamental to auditoryterm perceptionterm, as demonstrated in a range of studies, but the timing (i.e., its intermediate position between lower thetaterm and higher gammaterm activity) also invites the conjecture that alphaterm range activity reflects aspects of auditoryterm processing on a timescaleterm of approximately 100 ms. Here, we investigate directly the most straightforward hypothesis: does alphaterm bandterm activity reflect the processing of acoustic information in a manner parallel to lower and higher neural frequencies?

Studies using amplitude-modulated soundsterm or click trains have shown a low-pass modulation transfer function (MTFterm) with a rebound above 30 Hz [ 40 – 45 ]. Although these findings demonstrate dominant auditoryterm responses in the low frequencyterm range (delta-theta and alphaterm) and the gammaterm bandterm, the cortical oscillationsterm entrained by stimuli of corresponding frequencies of the regular modulation are not sufficient to demonstrate the importance of thetaterm, alphaterm, and gammaterm bandsterm in auditoryterm processing, because the entrainmentterm could simply reflect a modulation-frequency following response which may have little to do with the auditoryterm system actively processing acoustic information in the thetaterm and gammaterm bandsterm.

In studies using speechterm stimuli with irregular and complex modulations, it has been found that thetaterm bandterm activity entrains to the envelope of speechterm (and other signals), and that phaseterm locking in the thetaterm bandterm is enhanced by increased speechterm intelligibility [ 46 – 54 ]. Different evidence suggests that the gammaterm bandterm also plays an important role in phonemic and syllabic processing and comprehension of speechterm [ 55 – 60 ]. The results from speechterm processing suggest that thetaterm bandterm oscillationsterm, instead of being simply entrained by modulations of the corresponding frequencyterm, may actively “chunkterm” complex acoustic signals at a timescaleterm corresponding to periods of the thetaterm bandterm [ 61 , 62 ], while the gammaterm bandterm is involved in processing detailed information, because the comprehension of speechterm and syllable processing requires access to acoustic information at the phonemic scale.

Here, we aimed to elucidate what kind of mechanism might form the basis of such multiscale hearing by using irregular temporal modulations of auditoryterm signals to ask whether neural oscillationsterm are entrained equally to auditoryterm stimuli of different irregular modulation ratesterm or, rather, only specific, restricted bandsterm; moreover, we asked whether different stimulusterm modulation ratesterm can be decoded from specific neural frequency bandsterm·var. We focus on cortical oscillationsterm in the thetaterm and gammaterm bandsterm as well as the alphaterm bandterm, in which strong effects related to auditoryterm processing have been reported [ 63 , 64 ].

Building on experiments by Boemio et al. (2005) [ 65 ] and Luo and PoeppelpersName (2012) [ 66 ], we generated acoustic stimuli with modulation ratesterm that were centered at the typical periods of the neural thetaterm (4–7 Hz), alpha-beta (8–15 Hz), and gammaterm (31–45 Hz) frequencyterm bandsterm (See Fig 1a for illustration of the stimuli). We measured using magnetoencephalography (MEGterm·mth) the robustness of neural responses evoked by the stimuli and examined the correlations between the neural responses and the acoustic structure of the stimuli. We manipulated the signal-to-noise ratio (SNRterm·var) of the stimuli and evaluated the behavioral relevance of oscillationsterm at the different rates. Next, classificationterm·mth analyses of the MEGterm·mth data were performed to test which components of the neutrally elicited MEGterm·mth signal were informative about the auditoryterm signals. We show, contrary to expectation, that different temporal rates of the neurophysiological signal are differentially related to the behavioral, classificationterm·mth, and neural results. We further support the findings of processing at distinct timescalesterm by performing mutual informationterm analyses. Our combined neurophysiological and psychophysical results point to oscillatory neural mechanisms that underlie the segregated and discontinuous multiscale auditoryterm processing of signals which, subjectively, feel seamless and continuous. In particular, we argue that the neural computations reflected in thetaterm and gammaterm bandterm activity differ in a principled way from those reflected in the alphaterm bandterm and that alphaterm “splits” auditoryterm processing into paired low and high processing scales.

Fig 1. Stimuli and behavioral results.

(a) Cochleograms of 3 stimulusterm types. A Gammatone filterbank of 64 banks was used to decompose the stimuli. The prior distributions of segment duration for the 3 stimuli are shown on the right. (b) Behavioral performance. The blue, gray, and red lines show d-primeterm·var values for θ , α , and γ soundsterm across different signal-to-noise ratios (SNRsterm·var), respectively. The 3 separately plotted results at the top-right corner are d-primeterm·var values in clean stimulusterm conditions. Error bars: ±1 standard error of the mean. Data are deposited in the Dryad repository: http://dx.doi.org/10.5061/dryad.f357r [ 121 ].

Results Differential sensitivity to varying modulation ratesterm revealed by behavioral results We first examine how well listeners can recognize stimuli with varying frequencyterm modulation ratesterm, including especially rates that correspond to thetaterm, alphaterm, and gammaterm bandterm periods. The behavioral sensitivity to different modulation ratesterm provides a first indication whether specific rates are preferred by the human auditoryterm system.

We refer to the stimulusterm type with a modulation rateterm in thetaterm bandterm range (4–7 Hz) as a thetaterm (θ) soundterm, the stimulusterm type with modulation rateterm in the alphaterm bandterm range (8–12 Hz) as an alphaterm ( α ) soundterm, and the stimulusterm type with modulation rateterm in the gammaterm bandterm range (30–45 Hz) as a gammaterm (γ) soundterm. The cochleograms of 3 soundsterm are shown in Fig 1a . In the behavioral test, participants were asked to categorize the θ , α , and γ soundsterm by button press. Soundsterm were presented either as clean stimuli or presented in noiseterm at various SNRsterm·var. The results ( Fig 1b ) demonstrate that participants identified all clean stimuli well (all d′ > 2) and that performance deteriorated in all conditions as the SNRterm·var level decreased. Particularly noteworthy is the reduced performance in correctly identifying the α soundsterm across all levels compared to both γ and θ soundsterm.

The behavioral performance in identifying the 3 clean stimuli was analyzed using a 1-way repeated measures ANOVAterm·stat (rmANOVA) with Stimulus-type as the main factor. There is a main effect of Stimulus-type ( F (2,28) = 9.78, p = 0.001, η p 2 = 0.411). Planned comparisons using paired t tests reveal that the identification of α soundsterm was significantly worse than the identification of the other 2 stimuli (comparison to θ soundterm, t (14) = −3.16, p = 0.021, d = −0.82; comparison to γ soundterm, t (14) = −4.95, p = 0.001, d = −1.28; Bonferroni corrected). There was no difference in identification performance between the θ and γ soundsterm ( t (14) = 0.65, p > 0.05, d = 0.17).

For the masked stimuli, a Stimulus-type × SNRterm·var 2-way rmANOVA reveals main effects of Stimulus-type ( F (2,28) = 17.60, p < 0.001, η p 2 = 0.557) and SNRterm·var ( F (4,56) = 144.86, p < 0.001, η p 2 = 0.912) as well as a Stimulus-type × SNRterm·var interaction ( F (8,112) = 2.47, p = 0.017, η p 2 = 0.150). A quadratic trend analysisterm·stat indicates that identification of stimuli decreased as SNRterm·var decreases ( F (1,14) = 510.35, p < 0.001, η p 2 = 0.973).

Phase coherenceterm suggests differentiation between temporal regimes We computed the intertrial phase coherence (ITC)term·def across all frequencies from 2 Hz to 50 Hz and in a time range from 300 ms to 1,800 ms after the onset of the stimuli to measure how oscillatory cortical activity in each frequencyterm bandterm responds to different modulation ratesterm. ITCterm measures the robustness of neural responses to stimuli across trials. If cortical oscillationsterm at a certain frequencyterm bandterm reliably respond to a specific stimulusterm (and therefore, by hypothesis, probably encode information about this stimulusterm), we would obtain high ITCterm values at this frequencyterm bandterm. We created a distribution of ITCterm values by randomizing the onset time of each trial to normalize ITCterm in each subject across all frequencies and we converted the original ITCterm to z -scoresterm·stat of ITCterm (zITC). (See Materials and methods for details.) The zITC values for the clean stimuli across all frequencies are plotted in Fig 2a and are shown as topographies separated by each bandterm in Fig 2b . ITCterm, as a general measure of entrainmentterm, shows that the neural thetaterm bandterm (4–7 Hz) is strongly coherent to θ soundsterm and that the neural gammaterm bandterm (31–45 Hz) is strongly coherent to γ soundsterm. Unexpectedly, the alphaterm bandterm (8–12 Hz) reveals no selective response to the temporally corresponding stimuli. However, the α and γ soundsterm also evoked robust phase coherenceterm at a frequencyterm range below 8 Hz, in which zITC of the group means for α and γ soundsterm are above 1.64—the critical z -scoreterm·stat equivalent to an alphaterm level of 0.05 (1-tailed, Bonferroni corrected).

Fig 2. Z -scoresterm·stat of intertrial phase coherenceterm (zITC) to 3 clean stimuli and their response topographies.

(a) Top panel, spectra of zITC for θ , α , and γ soundsterm. The dashed line ( z -scoreterm·stat of 1.64) is equivalent to an alphaterm level of 0.05 (1-tailed, corrected). The shaded areas represent ±1 standard error of the mean. The bottom panel shows zITC for θ , α , and γ soundsterm in 4 frequencyterm bandsterm, thetaterm (4–7 Hz), alphaterm (8–12 Hz), betaterm (13–30 Hz), and gammaterm (31–45 Hz). The color scheme of blue, gray, and red in both panels represents zITC of θ , α , and γ soundsterm, respectively. The error bars represent ±1 standard error of the mean. (b) Topographies of zITC for each soundterm at each frequencyterm bandterm. Auditoryterm response patterns (compared to, for example, classic evoked M100 responses) are observed clearly in the thetaterm bandterm for θ soundsterm as well as for the α and γ soundsterm. In the gammaterm bandterm, the topographies show the auditoryterm response pattern only for the γ soundterm. No clear pattern is observed in other frequencyterm bandsterm. Data are deposited in the Dryad repository: http://dx.doi.org/10.5061/dryad.f357r [ 121 ].

Crucially, no robust phase coherenceterm was found in the alphaterm and betaterm bandsterm. This nonuniform phase-response pattern across temporal regimes does not support the simplest hypothesis of comparable neural trackingterm for all stimulusterm modulation ratesterm. Consistent results were obtained in the spatial analysis. The topographies of zITC, as shown in Fig 2b , reveal clear auditoryterm response patterns both for θ soundsterm in the thetaterm bandterm and for γ soundsterm in the gammaterm bandterm. Such auditoryterm response patterns, though weak, were also observed in the thetaterm bandterm for α and γ soundsterm.

To measure the effects of different stimuli on zITC, a Stimulus-type × Hemisphere × Frequencyterm bandterm 3-way rmANOVA was performed. This revealed main effects of Stimulus-type ( F (2,28) = 3.88, p = 0.033, η p 2 = 0.217) and Frequencyterm bandterm ( F (3,42) = 30.50, p < 0.001, η p 2 = 0.685) as well as an interaction between Stimulus-type and Frequencyterm bandterm ( F (6,84) = 3.67, p = 0.003, η p 2 = 0.208). The main effect of Hemisphere was not significant ( F (1,14) = 0.22, p = 0.648, η p 2 = 0.015).

Planned post hoc comparisons using paired t tests with Bonferroni correctionterm·stat on the main effect of Stimulus-type show that zITC of θ soundsterm is larger than zITC of α soundsterm ( t (14) = 3.36, p = 0.014, d = 0.87). Post hoc analysis of the main effect of Frequencyterm bandterm with Bonferroni correctionterm·stat shows that zITC at the thetaterm bandterm is larger than the alphaterm bandterm ( t (14) = 7.39, p < 0.001, d = 1.91), betaterm bandterm ( t (14) = 9.47, p < 0.001, d = 2.45), and gammaterm bandterm ( t (14) = 7.43, p < 0.001, d = 1.92) and that zITC at the alphaterm bandterm is larger than the betaterm bandterm ( t (14) = 3.52, p = 0.021, d = 0.91).

Post hoc analyses on Stimulus-type × Frequencyterm bandterm interactions using adjusted false discovery rate (FDR) correction [ 67 , 68 ] show that, in the thetaterm bandterm, zITC of θ soundsterm is larger than both zITC of α soundsterm ( t (14) = 4.00, p = 0.007, d = 1.14) and γ soundsterm ( t (14) = 3.86, p = 0.010, d = 1.00); in the gammaterm bandterm, zITC of γ soundsterm is larger than zITC of α soundsterm ( t (14) = 2.93, p = 0.044, d = 0.76). Before adjusted FDRterm·stat correction, zITC of γ soundsterm is significantly larger than θ soundsterm at the gammaterm bandterm ( t (14) = 2.20, p = 0.022, d = 0.57). There are no significant differences of zITC across different soundsterm found in the alphaterm and betaterm bandsterm after adjusted FDRterm·stat correction.

To investigate whether θ , α , and γ soundsterm evoke robust phase coherenceterm in the frequencyterm bandsterm other than their corresponding frequencyterm bandsterm, a 1-sample t testterm·stat of zITC in comparison with a baseline of 1.64 was conducted in each frequencyterm bandterm for each soundterm. zITC above 1.64 means that the phase coherenceterm observed is above the 95th percentile of ITCterm distribution over trials of randomized onset time. After adjusted FDRterm·stat correction, we found robust phase coherenceterm in the thetaterm bandterm for θ soundsterm ( t (14) = 38.06, p < 0.001, d = 9.82) and for γ soundsterm ( t (14) = 3.01, p = 0.028, d = 0.78) but not for α soundsterm ( t (14) = 1.66, p = 0.240, d = 0.43). There are no significant results in alphaterm, betaterm, and gammaterm bandsterm. zITC of γ soundsterm in the gammaterm bandterm is not significant ( t (14) = −0.15, p = 1.00, d = −0.03), likely because robust phase coherenceterm peaks within a narrow frequencyterm range centered around 37 Hz, as shown in the spectrum of zITC in Fig 1a . Averaging zITC of γ soundterm from 30 to 45 Hz decreases the mean of zITC in the gammaterm bandterm.

To summarize this first set of analyses, we showed that cortical oscillationsterm in the thetaterm and gammaterm bandsterm, but not in the alphaterm bandterm, robustly entrained to soundsterm with modulation ratesterm in the corresponding frequencyterm ranges. The further analysis of zITC indicates that thetaterm bandterm oscillationsterm reliably respond to γ soundsterm whose modulation rateterm is in the gammaterm bandterm range. zITC for α soundterm in the thetaterm bandterm also shows phase coherenceterm, although it is not significantly above threshold.

This preferential 2-scale (thetaterm and gammaterm) response pattern of the cortical auditoryterm system shown by phase coherenceterm results aligns with the findings in neurophysiology related to speechterm perceptionterm and production [ 69 ] (although nonspeech stimuli were used here). These consistent results across different stimulusterm types motivate the hypothesis that 2 discrete timescalesterm, 150–300 ms (thetaterm bandterm) and approximately 30 ms (gammaterm bandterm), play an important role in general auditoryterm processing. The robust phase coherenceterm observed in the thetaterm bandterm for all 3 soundsterm further suggests that the phase coherenceterm reflects more than passive entrainmentterm to modulated soundsterm. Although the alphaterm bandterm has been found in many auditoryterm tasks to correlate with auditoryterm perceptionterm, we did not find robust and distinct entrainmentterm in the alphaterm bandterm here. It is possible that the stimulus-evoked activity in the alphaterm bandterm does not manifest in phaseterm but in powerterm, or that alphaterm does not show preference to any soundsterm and can be equally entrained by all 3 soundsterm. Next, we therefore analyzed powerterm responses to 3 soundsterm. In the classificationterm·mth analysis shown subsequently, we investigated whether phaseterm patterns in each frequencyterm bandterm provide critical information for processing θ , α , and γ soundsterm.

Analysis of evoked powerterm responses shows robust but differential entrainmentterm Having first quantified phase coherenceterm patterns, we next tested whether neural response powerterm reveals patterns that support the observed segregationterm across bandsterm, as the powerterm response may be differentially modulated by different soundsterm in specific frequencyterm bandsterm, which could reflect powerterm coding for temporal information. We analyzed evoked powerterm and induced powerterm separately, because evoked powerterm is conceived as a stimulus-locked response while induced powerterm is often argued to be generated by nonstimulus-locked processes [ 70 , 71 ].

The spectrograms of evoked powerterm for each clean stimulusterm are plotted in Fig 3 . The θ and γ soundsterm show corresponding responses in the MEGterm·mth signal, but the α soundterm does not: the powerterm evoked by θ soundsterm largely distributes in the thetaterm bandterm; the powerterm evoked by γ soundsterm distributes in the gammaterm bandterm. In contrast, the powerterm evoked by α soundsterm is not well observed in the alphaterm or any other frequencyterm bandterm.

Fig 3. Time-frequency analyses of evoked powerterm.

Evoked powerterm responses to θ , α , and γ soundsterm, respectively. The vertical dashed lines indicate the onset of the auditoryterm stimuli. The dashed boxes in the panels mark the frequencyterm bandsterm of evoked powerterm responses corresponding to the stimulusterm modulation ratesterm. θ soundsterm evoke powerterm increase in the thetaterm bandterm and γ soundsterm in the gammaterm bandterm. Data are deposited in the Dryad repository: http://dx.doi.org/10.5061/dryad.f357r [ 121 ].

A Stimulus-type × Hemisphere × Frequencyterm bandterm 3-way rmANOVA was performed and revealed main effects of Stimulusterm type ( F (2,28) = 4.29, p = 0.024, η p 2 = 0.234) and Frequencyterm bandterm ( F (3,42) = 28.84, p < 0.001, η p 2 = 0.673) as well as an interaction between Stimulus-type and Frequencyterm bandterm ( F (6,84) = 2.88, p = 0.013, η p 2 = 0.171). The main effect of Hemisphere was not significant ( F (1,14) = 0.33, p = 0.575, η p 2 = 0.023).

Post hoc paired t tests with Bonferroni correctionterm·stat on the main effect of Stimulus-type show that the evoked powerterm of θ soundsterm was larger than that of α soundsterm ( t (14) = 3.58, p = 0.009, d = 0.92). Post hoc analysis of the main effect of Frequencyterm bandterm showed that the evoked powerterm in the thetaterm bandterm is larger than that in the alphaterm bandterm ( t (14) = 3.88, p = 0.010, d = 1.00), in the betaterm bandterm ( t (14) = 8.84, p < 0.001, d = 2.28), and in the gammaterm bandterm ( t (14) = 5.14, p < 0.001, d = 1.33). Evoked powerterm in the alphaterm bandterm is larger than the betaterm bandterm ( t (14) = 5.23, p < 0.001, d = 1.35), and evoked powerterm in the gammaterm bandterm is larger than betaterm bandterm ( t (14) = 6.32, p < .001, d = 1.64).

A post hoc analysis of the Stimulus-type × Frequencyterm bandterm interaction using adjusted FDRterm·stat correction shows that, in the thetaterm bandterm, the evoked powerterm of θ soundsterm is larger than the evoked powerterm of α soundsterm ( t (14) = 4.77, p < 0.001, d = 1.23) and γ soundsterm ( t (14) = 3.32, p = 0.020, d = 0.86). In the gammaterm bandterm, the evoked powerterm of γ soundsterm is larger than the evoked powerterm of α soundsterm ( t (14) = 3.94, p = 0.006, d = 1.02). There is no significant difference between evoked powerterm across different soundsterm found in alphaterm and betaterm bandsterm after adjusted FDRterm·stat correction.

Powerterm responses induced by the different stimuli were explored by a Stimulus-type × Hemisphere × Frequencyterm bandterm 3-way rmANOVA. The main effect of Frequencyterm bandterm was significant ( F (3,42) = 15.36, p < 0.001, η p 2 = 0.523). A post hoc analysis with Bonferroni correctionterm·stat showed that the powerterm at the betaterm bandterm was less than the thetaterm bandterm ( t (14) = 6.88, p < 0.001, d = 1.78), alphaterm bandterm ( t (14) = 5.26, p = 0.001, d = 1.36), and gammaterm bandterm ( t (14) = 5.77, p < 0.001, d = 1.49). There is no difference in powerterm between thetaterm, alphaterm, and gammaterm bandsterm. The main effect of Hemisphere was marginally significant ( F (1,14) = 4.51, p = 0.052, η p 2 = 0.244), with the powerterm in the right hemisphere larger than in the left hemisphere.

Neural responses in the thetaterm and gammaterm bandsterm lock to the temporal structure of θ and γ soundsterm, respectively The selective phase coherenceterm observed in the thetaterm and gammaterm bandsterm that we show in Fig 2 may be a result of reliable auditoryterm responses to any soundsterm, but not necessarily caused by the specific temporal structure of the stimuli. We therefore asked next whether phaseterm patterns of cortical oscillationsterm actually correlate with the temporal structure of the stimuli. We first used a measure, cochlear-scaled correlationterm·mth, inspired by the concept of cochlear-scaled entropy, to extract salient acoustic changes that may reset the phaseterm of cortical oscillationsterm and therefore lead to robust phase coherenceterm across trials [ 51 ]. The cochlear-scaled correlationterm·mth was calculated using a moving temporal window and represents acoustic changes along time (see Materials and methods for details). Next, we computed mutual informationterm between the phaseterm series of cortical oscillationsterm at each frequencyterm and the cochlear-scaled correlationsterm·mth. Mutual informationterm can quantify how much information in the temporal structure of the stimuli can be explained by the phaseterm patterns of cortical oscillationsterm and indicate whether the robust phase coherenceterm across trials observed in the neural frequency bandsterm·var is evoked by the temporal structure of a specific soundterm.

To compute the cochlear-scaled correlationterm·mth, we decomposed the stimuli using a Gammatone filterbank with 64 bandsterm, then averaged the amplitude of the envelope in each cochlear bandterm using a moving temporal window of 10 ms, generating 64 total values (1 per bandterm). A Pearson’s correlationterm·stat was then calculated for these values between each adjacent time point. The cochlear-scaled correlationterm·mth for each soundterm is shown in Fig 4a . We computed mutual informationterm between the cochlear-scaled correlationterm·mth of each soundterm and the phaseterm series of the neural oscillationterm to all 3 soundsterm. For example, we computed mutual informationterm between the cochlear-scaled correlationterm·mth of the θ soundsterm and the 3 phaseterm series evoked by θ , α , and γ soundsterm. We used the phaseterm series evoked by α and γ soundsterm as controls to examine whether mutual informationterm between the cochlear-scaled correlationterm·mth of θ soundsterm and the phaseterm series evoked by θ soundsterm is significant. The results of mutual informationterm analysis are shown in Fig 4b .

Fig 4. Mutual informationterm between cochlear-scaled correlationterm·mth and phaseterm of neural oscillationsterm.

(a) Cochlear-scaled correlationterm·mth of θ , α , and γ soundsterm. (b) Mutual informationterm results. The colors represent mutual informationterm computed using the phaseterm of the neural oscillationterm to different soundsterm: blue, θ soundterm; gray, α soundterm; red, γ soundterm. The inset bar graphs show the averaged mutual informationterm within frequencyterm ranges that show significant main effects of the stimulusterm type. The robust phase coherenceterm observed in the thetaterm bandterm for θ soundterm and in the gammaterm bandterm for γ soundterm is caused by trackingterm acoustic structures instead of simply being evoked by general acoustic stimuli. The shaded areas represent ±1 standard error of the mean. Data are deposited in the Dryad repository: http://dx.doi.org/10.5061/dryad.f357r [ 121 ].

We ran a 1-way rmANOVA on mutual informationterm with the factor of phaseterm responses evoked by soundterm type (phaseterm soundterm type). After adjusted FDR correction across frequenciesterm·stat, we found significant main effects of the phaseterm soundterm type between 4 and 6 Hz using the cochlear-scaled correlationterm·mth of θ soundsterm ( p < 0.05) and significant main effects of the phaseterm soundterm type between 33 and 39 Hz using the cochlear-scaled correlationterm·mth of γ soundsterm. Importantly, no significant main effects of the phaseterm soundterm type were found for mutual informationterm computed using the cochlear-scaled correlationterm·mth of α soundsterm. In the post hoc comparison with Bonferroni correctionterm·stat, we averaged mutual informationterm within the frequencyterm ranges in which significant main effects were observed and found that, when the cochlear-scaled correlationterm·mth of θ soundsterm was used, the mutual informationterm computed using phaseterm series of θ soundterm is significantly larger than that using phaseterm series of α soundterm ( t (14) = 5.29, p < 0.001, d = 1.37) and that using phaseterm series of γ soundterm ( t (14) = 3.97, p = 0.004, d = 1.03). When the cochlear-scaled correlationterm·mth of γ soundsterm was used, the mutual informationterm computed using phaseterm series of γ soundterm is significantly larger than that using phaseterm series of θ soundsterm ( t (14) = 5.05, p < 0.001, d = 1.30) and that using phaseterm series of α soundsterm ( t (14) = 5.71, p < 0.001, d = 1.47).

The results of the mutual informationterm analyses demonstrate that phaseterm patterns in the thetaterm and gammaterm bandsterm track the temporal structure of the stimuli, as quantified by the cochlear-scaled correlationterm·mth. The robust phase coherenceterm observed in the thetaterm bandterm for θ soundsterm and that in the gammaterm bandterm for γ soundsterm is indeed caused by trackingterm specific acoustic structures, rather than simply being evoked by generic acoustic stimuli. The fact that no significant results were found for α soundsterm suggests, again, that the alphaterm bandterm may play a different role in processing soundsterm, and especially their temporal structure. In contrast, the thetaterm bandterm and the gammaterm bandterm may be central to auditoryterm processing and the construction of neural representations underlying perceptual analysis.

We did not find specialized trackingterm for γ soundsterm in the thetaterm bandterm, although γ soundsterm evoked robust phase coherenceterm in the thetaterm bandterm ( Fig 2a ). This could be because the thetaterm bandterm, instead of faithfully coding the temporal structure of γ soundsterm, chunksterm acoustic information and forms a perceptual unit on a timescaleterm of approximately 200 ms (the thetaterm bandterm range). We further explored the contribution of the thetaterm bandterm to coding temporal information of γ soundsterm in the following classificationterm·mth analyses.

Phaseterm series: Classificationterm·mth analyses and the primacy of the neural thetaterm bandterm We next performed classificationterm·mth analysis to investigate whether information in each frequencyterm bandterm of cortical oscillationsterm can be used to classify different stimulusterm types. If a frequencyterm bandterm, for example, the thetaterm bandterm, reflects sufficient information to classify θ , α , and γ soundsterm, it implies that the thetaterm bandterm plays an important role in processing θ , α , and γ soundsterm. On the other hand, if a frequencyterm bandterm does not contribute to classifying any soundsterm, it may indicate that this frequencyterm bandterm is not a key component to processing the soundsterm.

We first use phaseterm and powerterm responses in a frequencyterm range of 4–45 Hz and a time range of 300–1,900 ms (i.e., after the onset of stimuli) for classificationterm·mth to test whether phaseterm or powerterm provide information related to the temporal structure of each soundterm. Second, we measure the contribution of each frequencyterm bandterm in classificationterm·mth to determine which frequencyterm bandterm is by hypothesis critical to auditoryterm processing. Finally, we use a new method to classify stimulusterm type at each time point to investigate temporal progression of the classificationterm·mth performance.

Classificationterm·mth performance was first computed for each stimulusterm type using the phaseterm and powerterm response profile of all frequencyterm bandsterm (4–45 Hz). Confusion matrices of phaseterm classificationterm·mth and powerterm classificationterm·mth are plotted in Fig 5a . d-primeterm·var values computed based on the confusion matrices, indicated by D′ (to differentiate it from d′ in the behavioral results), are shown in Fig 5b . A Stimulus-type × Classificationterm·mth source (phaseterm classificationterm·mth or powerterm classificationterm·mth) 2-way rmANOVA reveals the main effect of Classificationterm·mth source ( F (1,14) = 30.44, p < 0.001, η p 2 = 0.685), with the D′ of phase-based classificationterm·mth significantly larger than the D′ of power-based classificationterm·mth. To determine whether classificationterm·mth performance is better than chance (D′ = 0), a 1-sample t testterm·stat with Bonferroni correctionterm·stat was applied on each stimulusterm type and each classificationterm·mth source. For phaseterm classificationterm·mth, the performance of all stimulusterm types was better than chance (for θ soundsterm, t (14) = 6.19, p < 0.001, d = 1.60; for α soundsterm, t (14) = 6.40, p < 0.001, d = 1.65; for γ soundsterm, t (14) = 6.44, p < 0.001, d = 1.66). For powerterm classificationterm·mth, classificationterm·mth performance of α soundsterm was significant ( t (14) = 4.73, p = 0.002, d = 1.22) as well as θ soundsterm ( t (14) = 3.10, p = 0.046, d = 0.80), but performance of powerterm classificationterm·mth was only slightly above chance. These results demonstrate that phaseterm patterns of cortical oscillationsterm reliably encode the temporal dynamics of stimuli.

Fig 5. Classificationterm·mth results.

(a) Confusion matrices for phase-based and power-based classificationsterm·mth. “Stimulusterm label” represents the actual stimulusterm type and “classified label” represents the classified bin. Color bar codes the percentage of trials classified into each bin. (b) Classificationterm·mth performance for each soundterm using phaseterm and powerterm. Phase-based classificationterm·mth performs significantly better than power-based classificationterm·mth. The blue, gray, and red bars represent θ , α , and γ soundsterm, respectively. (c) Frequencyterm bandterm contributions to phase-based classificationterm·mth of each soundterm. “N” indicates classificationterm·mth without the given frequencyterm bandterm; “Y” indicates classificationterm·mth with the given frequencyterm bandterm. Phaseterm in the thetaterm bandterm improves performance for classifying all stimuli and phaseterm in the gammaterm bandterm contributes to classifying γ soundterm. The color scheme is as in (b). The error bars represent ±1 standard error of the mean. Data are deposited in the Dryad repository: http://dx.doi.org/10.5061/dryad.f357r [ 121 ].

Next, to explore how each frequencyterm bandterm might contribute to phase-based classificationterm·mth of different stimulusterm types, classificationterm·mth was conducted using different combinations of frequencyterm bandsterm (e.g., thetaterm bandterm plus gammaterm bandterm, etc.). We then compared the classificationterm·mth performance that was obtained with or without a particular frequencyterm bandterm included in the analysis. For example, to quantify the thetaterm bandterm contribution, we obtained 2 values by either averaging D′ values across frequencyterm bandterm combinations that included the thetaterm frequencyterm bandterm, or without the thetaterm frequencyterm bandterm. The contributions of each frequencyterm bandterm to different stimuli are plotted in Fig 5c . A Stimulus-type × Frequencyterm bandterm × Inclusion (with or without a given frequencyterm bandterm) 3-way rmANOVA shows main effects for Frequencyterm bandterm ( F (3,42) = 6.28, p = 0.001, η p 2 = 0.310) and Inclusion ( F (1,14) = 16.34, p = 0.001, η p 2 = 0.539). The 2-way interactions between Frequencyterm bandterm and Inclusion ( F (3,42) = 6.28, p = 0.001, η p 2 = 0.310) and between Stimulus-type and Frequencyterm bandterm ( F (6,84) = 5.42, p < 0.001, η p 2 = 0.279) were significant as well as the 3-way Stimulusterm type × Frequencyterm bandterm × Inclusion interaction ( F (6,84) = 5.42, p < 0.001, η p 2 = 0.279). To further examine how each frequencyterm bandterm contributes to classificationterm·mth for each stimulusterm type, paired t tests with adjusted FDRterm·stat correction were performed on each frequencyterm bandterm and each stimulusterm type. The thetaterm bandterm contributes to the decodingterm of all stimulusterm types (for θ soundsterm, t (14) = 8.16, p < 0.001, d = 2.11; for α soundsterm, t (14) = 3.68, p = 0.010, d = 0.95; for γ soundsterm, t (14) = 3.16, p = 0.021, d = 0.82). The betaterm bandterm deteriorated decodingterm of γ soundsterm ( t (14) = −4.66, p = 0.002, d = 1.20). Crucially, the alphaterm bandterm did not contribute significantly to decodingterm any stimuli. Before adjusted FDRterm·stat correction, the gammaterm bandterm shows a contribution to decodingterm γ soundsterm ( t (14) = 2.07, p = 0.057, d = 0.53). After removing 1 subject who showed abnormal decodingterm performance, we found that the gammaterm bandterm significantly contributes to decodingterm γ soundsterm after FDR correction ( t (13) = 3.35, p = 0.017, d = 0.90).

Finally, because we found that the thetaterm and gammaterm bandsterm provided the main contributions to classificationterm·mth, we examined how classificationterm·mth performance in the thetaterm bandterm and gammaterm bandterm progress temporally by using each time point of a phaseterm series to classify a stimulusterm type. Classificationterm·mth was conducted in the thetaterm bandterm and the gammaterm bandterm separately by combining the MEGterm·mth channels selected from 500 ms before the onset of the stimuli to 2,000 ms after. We used a cluster-based permutation testterm·stat to quantify significance of classificationterm·mth performance (see Materials and methods for details). The results are shown in Fig 6a . We then averaged classificationterm·mth performance for each soundterm on each bandterm from 300 to 1,900 ms after the onset of stimuli ( Fig 6b ) and found that in the thetaterm bandterm, classificationterm·mth performance for θ soundsterm is significantly larger than that for α soundsterm ( t (14) = 3.06, p = 0.027, d = 0.79) and γ soundsterm ( t (14) = 4.04, p = 0.006, d = 1.04) after Bonferroni correctionterm·stat. In the gammaterm bandterm, classificationterm·mth performance for γ soundsterm is significantly larger than for θ soundsterm ( t (14) = 2.74, p = 0.048, d = 0.71) and α soundsterm ( t (14) = 4.79, p < 0.001, d = 1.24) after Bonferroni correctionterm·stat.

Fig 6. Temporal progression of classificationterm·mth performance in the thetaterm and the gammaterm bandsterm.

(a) Classificationterm·mth performance along each time point. Upper panel: classificationterm·mth results using phaseterm series of the thetaterm and the gammaterm bandsterm for θ soundsterm, from 500 ms before the onset of stimuli to 2,000 ms after the onset of stimuli; middle panel shows results for α soundsterm; lower panel shows results for γ soundsterm. Classificationterm·mth results using the thetaterm bandterm in black and the gammaterm bandterm in green. Lines represent the mean classificationterm·mth results averaged across subjects. The bolded parts represent clusters significantly larger than baseline, D′ = 0. Significant clusters are observed in the thetaterm bandterm after the onset of stimuli for all 3 soundsterm. In the gammaterm bandterm, the significant clusters are found mainly for γ soundsterm. (b) Classificationterm·mth results averaged from 300 to 1,900 ms after the onset of stimuli for each bandterm. The upper panel represents the classificationterm·mth results using the thetaterm bandterm for θ , α , and γ soundterm and the low panel the classificationterm·mth results using the gammaterm bandterm. Classificationterm·mth performance for θ soundsterm using the thetaterm bandterm is significantly larger than that for α and γ soundsterm, whereas classificationterm·mth performance for γ soundterm using the gammaterm bandterm is larger than that for θ and α soundsterm. The error bars and shaded areas represent ±1 standard error of the mean. Data are deposited in the Dryad repository: http://dx.doi.org/10.5061/dryad.f357r [ 121 ].

The significant classificationterm·mth performance in the thetaterm bandterm after the onset of stimuli for all stimulusterm types demonstrates that the thetaterm bandterm not only entrains to soundsterm with corresponding modulation ratesterm but also provides critical information for classifying stimuli of all modulation ratesterm. The gammaterm bandterm showed significant classificationterm·mth performance for γ soundsterm; by averaging time points from 300 to 1,900 ms, we see significantly higher classificationterm·mth performance in the gammaterm bandterm for γ soundsterm than for θ and α soundsterm. This argues for a higher degree of specificity for gammaterm trackingterm.

To summarize the classificationterm·mth results: phaseterm series in the thetaterm bandterm temporally track acoustic dynamics across all modulation ratesterm (used in our study), which suggests that the thetaterm bandterm is not only entrained by modulation ratesterm with corresponding frequencyterm range but also chunksterm soundsterm with modulation ratesterm outside of the thetaterm bandterm range into acoustic segments at a timescaleterm corresponding to the thetaterm period range. The gammaterm bandterm specifically locks to modulation ratesterm with a timescaleterm corresponding to gammaterm bandterm, approximately 30 ms. The contribution of the alphaterm bandterm must be seen as functionally separate from the other bandsterm.

If, broadly speaking, the (paired) activity of the thetaterm and gammaterm bandsterm is associated with the construction of perceptual objects in audition, contrary to the alphaterm bandterm, it stands to reason that the localization of the thetaterm and gammaterm neural activity should be associated. The supporting information (see S1 Text ) provides additional new data to verify that the neural sources are overlapping by localizing the MEG-recorded activity in source space based on individual participants’ structural MRIs.

Correlation between behavioral results and neural markers Connecting back to the behavioral data ( Fig 1b ), we tested how the degraded behavioral performance induced by noiseterm is correlated with the neural markers (ITCterm and powerterm response) to indicate which neural marker may account for the behavioral results. The positive correlation between d′ and ITCterm was significant in the thetaterm bandterm for 3 stimuli (for θ soundsterm, r = 0.815, t (14) = 8.50, p < 0.001, d = 2.19; for α soundsterm, r = 0.378, t (14) = 3.50, p = 0.014, d = 0.90; for γ soundsterm, r = 0.417, t (14) = 3.64, p = 0.014, d = 0.94). The positive correlation between d′ and evoked powerterm was significant in the thetaterm bandterm for θ soundsterm (r = 0.622, t (14) = 8.20, p < 0.001, d = 2.12). Analysis of the correlation between d′ and induced powerterm response showed a negative correlation. A significant negative correlation was found in the alphaterm bandterm for α soundsterm (r = −0.370, t (14) = −3.01, p = 0.049, d = −0.78) and in the gammaterm bandterm for γ soundsterm (r = −0.347, t (14) = −3.22, p = 0.048, d = 0.83). Adjusted FDRterm·stat correction was applied to all tests.

The ITCterm in the thetaterm bandterm showed significant correlation with behavioral performance on recognizing all 3 soundsterm. These results echo the classificationterm·mth results and demonstrate that the phaseterm series in the thetaterm bandterm provides critical information for auditoryterm processing.

Discussion In this MEG-based neurophysiological experiment, we investigate temporal codingterm at different scales by exploring the entrainmentterm of auditoryterm cortical oscillationsterm to soundsterm with different modulation ratesterm. Because healthy listeners (appear to) perceive soundsterm that contain modulation ratesterm over various temporal scales in a manner that reflects a continuous MTFterm with a low-pass filter shape [ 16 , 72 ], the most straightforward hypothesis suggests that different auditoryterm stimulusterm rates are tracked in a comparable manner across modulation ratesterm.

Contrary to this hypothesis, we find that oscillationsterm in the (slower) thetaterm and (faster) gammaterm bandsterm reliably track acoustic dynamics—but not in the (intermediate) alphaterm and betaterm bandsterm. Subsequent analyses showed that the information carried in the neural thetaterm bandterm contributes to the decodingterm of all modulation ratesterm used in this study, whereas the gammaterm bandterm mainly contributes to decodingterm only gamma-modulated soundsterm. Moreover, intertrial coherence in the thetaterm bandterm correlates with identification performance across all stimuli, underscoring that there is a clear perceptual consequence of the entrained oscillatory activity.

Our results are consistent with previous work showing entrainmentterm of auditoryterm cortical activity to low acoustic modulation ratesterm [ 31 , 32 , 45 – 47 , 50 , 51 , 53 , 54 , 57 , 73 – 81 ]. The present results lend support from a new perspective to those studies using modulated soundsterm with amplitude modulation ratesterm less than 10 Hz that have found strong entrainmentterm in the delta and thetaterm bandsterm. Importantly, the failure to observe entrainmentterm in the alphaterm and betaterm bandsterm in the present study also aligns with previous electrophysiological data of monkey primary auditoryterm cortex: cortical oscillationsterm could entrain to modulation ratesterm in the delta-theta bandsterm but not at 12 Hz [ 73 ]. Preliminary MEGterm·mth data have also revealed such a response pattern [ 66 ].

The neural gammaterm bandterm, in addition to the delta and thetaterm bandsterm, also reliably codes temporal information. The finding that the neural gammaterm bandterm is entrained by soundsterm with a corresponding temporal modulation rateterm—and that this alignment contributes to the classificationterm·mth of γ soundsterm—indicates that the auditoryterm system can track acoustic dynamics over short timescalesterm, approximately 30 ms. This observation is consistent with studies using amplitude modulation created by binaural beats; in that work, strong entrainmentterm both in the thetaterm and gammaterm bandsterm is found—but, again, not in the alphaterm or betaterm bandsterm [ 82 ]. Recordings in the primary auditoryterm cortex of monkeys also show a phase-locked response using amplitude modulation at 30 Hz [ 83 , 84 ]. The data we show also confirm the contribution of gammaterm bandterm entrainmentterm to speechterm separation found in the multiple talker environments [ 57 ]. These studies, complemented by the data shown here, support the emerging view that the auditoryterm system extracts precise temporal information mainly on 2 discrete, segregated timescalesterm.

The classificationterm·mth results we report show that the phaseterm information of thetaterm bandterm oscillationsterm contributes to decodingterm soundsterm with modulation ratesterm not only at the thetaterm bandterm timescaleterm but also at the alphaterm and gammaterm scales. One possible explanation is that thetaterm bandterm oscillationsterm (i) track acoustic dynamics at that specific temporal scale and (ii) at the same time actively chunkterm (at a scale of the mean thetaterm bandterm period) soundsterm with faster modulation ratesterm, so that acoustic properties over larger scales can be further extracted [ 85 ]. The gammaterm bandterm exclusively contributes to the classificationterm·mth of soundsterm with corresponding temporal dynamics. This may indicate specific processing at a fine-grained scale for acoustic temporal details. Overall, thetaterm bandterm oscillationsterm may be necessary, although not sufficient, for processing soundsterm with temporal variations across different scales, and gammaterm bandterm oscillationsterm may be needed for fine-grained processing.

The classificationterm·mth results also reveal data patterns that have not been shown in previous findings on the MTFterm. As the MTFterm shows decreased neural responses with increasing modulation rateterm, it is plausible to conjecture that temporal codingterm for acoustic dynamics would also show such a pattern. However, as observed in the classificationterm·mth analysis, the temporal codingterm capability does not correspond to the magnitude of the MTFterm—large ITCterm values for the alphaterm bandterm do not indicate high temporal codingterm capability. Some previous findings, though, do show a tendency for a rebound of neural activity in the gammaterm bandterm [ 40 , 41 , 44 ], although the magnitude of gammaterm bandterm activity is small compared with lower frequencyterm bandsterm [ 45 ]. The most relevant finding is from Wang et al. [ 45 ], in which a robust response to amplitude-modulated soundsterm of 31.5 Hz was indicated by the percentage of subjects that showed robust auditoryterm steady-state response (aSSR). These results were suggestive, although the study did not test temporal codingterm of different frequencyterm bandsterm and concluded “the MTFterm of the low-frequency aSSR generally has a low-pass pattern and only weakly depends on the carrier bandwidth.” Therefore, although we built our current study on assumptions about the MTFterm, the previous findings could not resolve the question of auditoryterm temporal codingterm on different timescalesterm or demonstrate that the powerterm of alphaterm bandterm is preserved (and “reserved”) not for temporal codingterm but for other auditoryterm cognitive process.

The absence or marked reduction of trackingterm acoustic dynamics in the alphaterm bandterm suggests that neural activity reflected in the alphaterm bandterm may play a different role in audition. In the auditoryterm system, the neural computations reflected in the alphaterm bandterm signal may be more explicitly involved in auditoryterm attentionterm, working memory, listening effort, or functional inhibition [ 34 , 36 – 39 , 86 ], and the alphaterm bandterm may as such be more related to aspects of auditoryterm perceptionterm that differ from constructing the elementary perceptual representations. The alphaterm bandterm is associated with suppressing activity of cortical areas that are irrelevant for ongoing sensory processing according to tasks [ 87 ], which suggests (for the current context) that alphaterm bandterm oscillationsterm co-occur with and segregate 2 temporal codingterm regimes (thetaterm and gammaterm bandsterm) and modulate auditoryterm processing as a top-down process [ 88 ]. The alphaterm bandterm is involved in processing in other sensory systems and has been well established in visual and somatosensory perceptual analysis [ 89 – 92 ]. It will be relevant to further investigate in the auditoryterm domain how the alphaterm bandterm interacts with thetaterm and gammaterm domains to comprehend auditoryterm analysis more fully.

The finding that, like alphaterm bandterm activity, the betaterm bandterm also does not track acoustic dynamics may reflect that neural oscillationsterm on that scale are reflective of different operations as well. Betaterm bandterm oscillationsterm have been argued to play a role in predictive coding [ 93 , 94 ]; the task of the present study does not require active prediction. The random phaseterm in the betaterm bandterm adds more noiseterm to the classificationterm·mth process, so removing the phaseterm information of the betaterm bandterm actually results in better classificationterm·mth performance.

Our finding of entrainmentterm, and specifically concurrent parallel processing at different scales of the thetaterm and gammaterm bandsterm, converges with the 2 perceptual time constants often found in behavioral studies [ 13 ]. Experiments on temporal integration frequently report a time constant in the (few) hundreds of milliseconds [ 14 , 21 – 26 ], while studies examining the high temporal resolution of the auditoryterm system show a time constant less than 30 ms [ 16 – 18 ]. The behavioral results we show also point to higher perceptual sensitivity for thetaterm and gammaterm soundsterm compared with alphaterm soundsterm. However, the behavioral method used in the present study cannot circumvent a concern with a selection biasterm to different soundsterm, because participants may identify θ soundsterm and γ soundsterm more easily simply because the modulation ratesterm of these 2 soundsterm are located at the perceptual extremes in this experimental design. We acknowledge that this behavioral task is suboptimal—and primarily employed to ensure attentionterm during neurophysiological recording—so we scrupulously refrain from overinterpreting these data, beyond pointing out that the pattern is consistent with our hypothesis. Notwithstanding this potential concern, results from a recent psychophysical study dovetail with the view that the auditoryterm system works concurrently on a short timescaleterm (about 30 ms) to extract fine-grained acoustic temporal detail while processing more global acoustic patterns on a longer timescaleterm (>200 ms) [ 95 ].

Importantly, our results, based on nonspeech stimuli, suggest that such dual-scale entrainmentterm is not speechterm specific but may rather reflect an intrinsic auditoryterm processing property. The auditoryterm cortex tunes to both thetaterm frequencyterm and gammaterm frequencyterm acoustic dynamics [ 65 ]. The alphaterm bandterm reflects different operations. This segregated, dual tuning of the auditoryterm system at different scales may facilitate the extraction of information of different types in speechterm, such as featural, segmental, or phonemic information versus syllabic scale information [ 1 ]. We suggest that the measured oscillatory patterns at different timescalesterm encode acoustic information over multiple scales, which leads to a temporal multiplexingterm of sensory information [ 59 , 96 ].

Mounting evidence shows that human perceptual systems employ a discrete process in which continuous signal information is broken up into segments [ 29 , 30 , 97 – 99 ]. As natural soundsterm contain information at multiple scales, the auditoryterm system may chunkterm continuous soundsterm using temporal windows of different sizes to sample information at different timescalesterm, instead of processing acoustic information on a unitary scale. This multiplexingterm strategy solves the requirement in auditoryterm processing that both fine resolution and integration over time are needed for perceiving soundsterm with regularities both at large and small scales. One model proposes that, although a very high resolution is represented in subcortical areas, in the auditoryterm cortex, there are 2 main temporal windows used for processing acoustic information: one centered around 200 ms and the other around 30 ms [ 6 , 100 ]. On this view, acoustic information is analyzed and integrated using 2 temporal windows at these scales so that perceptual information at such “global” and “local” scales, whether in speechterm or nonspeech, can be abstracted concurrently to form a unitary percept that forms the basis for perceptual decision-making, lexical access, memory encoding, and other cognitive operations building on elementary perceptual representations. This design, however, builds in a hole in processing, a segregationterm of function between low and high processing rates—perhaps optimized for sensory sampling—by an intermediate rate, perhaps optimized for allocating attentional and memory resources and functionally inhibiting task- or stimulus-irrelevant actions. Whereas we typically address segregationterm of function in the spatial domainterm, i.e., different regions are specialized for different operations, here, we provide a compelling example of cortical segregationterm of function in the time domainterm.

Materials and methods Ethics statement The study was approved by the New York University Institutional Review Board (IRB# 10–7277) and conducted in conformity with the 45 Code of Federal Regulations (CFR) part 46 and the principles of the Belmont Report.

Participants Sixteen right-handed volunteers (9 females; mean age: 24.8; standard deviation: 3.2) participated in this experiment. All participants provided informed written consent and received monetary compensation for their participation. Handedness was determined using the Edinburgh Handedness Inventory [ 101 ]. All participants had normal hearing and no neurological deficits. We excluded the data from 1 participant because of noiseterm issues during neurophysiological recording. Therefore, the analysis included the data from 15 participants (8 females; mean age: 25.2; standard deviation: 3.0).

Stimuli and experimental procedure We created 3 stimulusterm types following the methods used in Boemio et al. (2005) and Luo and PoeppelpersName (2012). Each stimulusterm was 2 s long and generated by concatenating narrow-band frequency-modulated segments. The mean starting frequencyterm of each segment was randomly drawn from 2 frequencies, 1,000 Hz and 1,500 Hz. If the mean starting frequencyterm is 1,000 Hz, the frequency-modulated segment could sweep up to 1,500 Hz. If the mean starting frequencyterm is 1,500 Hz, the frequency-modulated segment could sweep down to 1,000 Hz. The bandwidth of segmentsterm·var was 100 Hz (within a critical bandterm at the center frequencies used). We generated each segment by adding up 100 frequency-modulated sinusoids with randomized amplitude and phaseterm. To create a segment that sweeps down, the starting frequencyterm of 100 sinusoid is randomly distributed between 1,450 Hz and 1,550 Hz and the end frequencyterm is distributed between 950 Hz and 1,050 Hz. To create a segment that sweeps up, the starting frequencyterm of 100 sinusoid is randomly distributed between 950 Hz and 1,050 Hz and the end frequencyterm is distributed between 1,450 Hz and 1,550 Hz.

The duration of the segments for each of the 3 stimulusterm types was drawn from a Gaussian distribution with means of 190 ms, 100 ms, and 27 ms, with standard deviations of 30 ms, 15 ms, and 3 ms, respectively. The distribution of the segment durations of the stimuli aligned with the range of periods typical of thetaterm (4–7 Hz), alphaterm/low betaterm (8–15 Hz), and low gammaterm (30–45 Hz) bandterm neural oscillationsterm. We refer to the stimulusterm type with mean segment durationterm·var of 190 ms as a thetaterm ( θ ) soundterm, the stimulusterm type with mean segment durationterm·var of 100 ms as an alphaterm ( α ) soundterm, and the stimulusterm type with mean segment durationterm·var of 27 ms as a gammaterm ( γ ) soundterm. The cochleograms of the 3 stimuli were created using a Gammatone filterbank with 64 banks to decompose the stimuli from 50 to 22,050 Hz [ 102 , 103 ] and are shown in Fig 1a , with the corresponding prior distributions of segment duration for each stimulusterm type.

We generated white noiseterm segments of 4 s using the random number generator, the function “randn,” in Matlab R2014a (The MathWorks, Natick, MA). Then, we embedded the 3 types of clean stimuli into white noiseterm to create noise-masked stimuli at 5 levels of SNRterm·var: −9, −13, −17, −21, and −25 dB. The onset of white noiseterm preceded the onset of the embedded clean stimulusterm for a random interval uniformly distributed from 1 s to 1.5 s. Thirty stimuli for each SNRterm·var level and each stimulusterm type were created using individually generated noiseterm. As only 1 sample of each stimulusterm type was generated, the θ , α , and γ soundsterm were the same across all conditions. Therefore, there were 18 total conditions that included 3 clean stimuli and 15 (3 types × 5 SNRterm·var levels) noise-masked stimuli. In total, 540 trials (18 conditions × 30 trials per condition) were presented. The order of all stimuli was pseudorandomized for each participant. After each stimulusterm was presented, participants were required to push 1 of 3 buttons to indicate the type of stimulusterm. Between 1 and 2 s after participants responded, the next stimulusterm was presented, so that all stimuli were presented at random onset points. Participants were required to keep their eyes open and to focus on a white fixation cross in the center of a black screen.

All stimuli were normalized to about 65 dB SPL and delivered through plastic air tubes connected to foam ear pieces (E-A-R Tone Gold 3A Insert earphones, Aearo Technologies Auditoryterm Systems).

MEGterm·mth recording and channel selection MEGterm·mth signals were measured with participants in a supine position and in a magnetically shielded room using a 157-channel whole-head axial gradiometer system (KIT, Kanazawa Institute of Technology, Japan). A sampling rate of 1,000 Hz was used with an online 1–200 Hz analog band-pass filter and a notch filter centered around 60 Hz. After the main experiment, participants were presented with 1-kHz tone beeps of 50 ms duration as a localizer to determine their M100 evoked responses, which is a canonical auditoryterm response [ 104 ]. Ten channels in each hemisphere, selected based on the peak of M100 response between 60 ms and 120 ms, were used as auditoryterm channels for each participant individually. A layout of channels that are selected based on the peak of M100 response across 15 subjects is shown in Fig 7 .

Fig 7. Selected channels pooled across subjects.

The plot shows a layout of channels that are selected based on the peak of M100 response across 15 subjects. Twenty channels are selected for each subject (10 in each hemisphere). The channels selected for analysis are indicated by black circles. The contours indicate the extent of overlap across subjects. Data are deposited in the Dryad repository: http://dx.doi.org/10.5061/dryad.f357r [ 121 ].

Behavioral data analysis Behavioral data analysis was conducted in MATLAB using the Palamedes toolbox 1.50 [ 105 ]. For each SNRterm·var level as well as the 3 clean stimuli, a 3-by-3 confusion matrix was created and then was collapsed into three 2-by-2 tables by treating 1 stimulusterm as “target” and pooling the observations on the other 2 stimuli as “noiseterm.” Correct identification of the target stimulusterm was counted as a “hit” while misidentification of the other 2 stimuli as the target stimulusterm was counted as “false alarm;” d-primeterm·var values were computed based on hit rates and false alarm rates of each table. A half artificial incorrect trial was added to the table with all correct trials [ 106 ].

MEGterm·mth data preprocessing and analysis MEGterm·mth data analysis was conducted in MATLAB using the Fieldtrip toolbox 20140619 [ 107 ] and wavelet toolbox. Raw MEGterm·mth data were noiseterm reduced offline using the time-shifted principle component analysis [ 108 ] and sensor noiseterm suppression [ 109 ]. Trials were visually inspected, and those with artifacts such as signal jumps and large fluctuations were discarded. An independent component analysis was used to correct for eye blink, eye movement, heartbeat-related and system-related artifacts. Twenty-five trials were included in the analysis for each condition. Each trial was divided into 5-s epochs (1-s prestimulus period and 4-s stimulusterm period). Baseline was corrected for each trial by subtracting out the mean of the whole trial before further analysis.

To extract time-frequency information, single-trial data for each condition in each MEGterm·mth channel were transformed using functions of the Morlet wavelets embedded in the Fieldtrip toolbox, with a frequencyterm ranging from 1 to 50 Hz in steps of 1 Hz. To balance spectral and temporal resolution of the time-frequency transformation, from 1 to 20 Hz, the window length increased linearly from 1.5 circles to 7 circles and was kept constant at 7 circles above 20 Hz. Phaseterm and powerterm responses (squared absolute value) were extracted from the wavelet transform output at each time-frequency point.

The ITCterm, a measure of consistency of phase-locked neural activity entrained by stimuli across trials, was calculated for each time-frequency point (details as in [ 110 ]). ITCterm in different frequencyterm bandsterm reflects phaseterm trackingterm of cortical oscillationsterm to temporally modulated stimuli. As the baselines of phase-locking may be different across frequencyterm bandsterm, which may be affected by powerterm distributions with a 1/f characteristic, a shuffling method of onset time was used to avoid this confound. Within each condition, the onset time of the stimulusterm was randomly chosen on each trial and a new dataset for each condition was created. The same analysis of ITCterm was applied on this new dataset. To create a distribution of shuffled ITCsterm, this shuffling procedure was repeated 1,000 times. The z -scoreterm·stat of ITCterm was computed using the percentile of the original ITCterm in the distribution.

Induced powerterm was normalized by dividing the mean powerterm value in the baseline range (−0.6 to −0.1 s) and taking logarithms with base 10 and then was converted into values with the unit of decibel by multiplying by 10. The evoked powerterm response was computed by applying the time-frequency transform on averaged temporal responses across all trials. The baseline correction was the same as that used in computing the induced powerterm.

The ITCterm and powerterm data were averaged from 0.3 s to 1.8 s poststimulus onset to minimize the effects of stimulus-evoked onsets and offsets and within 4 frequencyterm bandsterm: thetaterm (4–7 Hz), alphaterm (8–12 Hz), betaterm (13–30 Hz), and gammaterm (31–45 Hz).

To examine whether phaseterm and powerterm in different frequencyterm bandsterm can explain behavioral performance, correlations between behavioral performance and neural measurements were tested by calculating correlation coefficients between d′ and either ITCterm or powerterm across the 5 SNRsterm·var, and then a 1-sample t testterm·stat was performed for each frequencyterm bandterm and each stimulusterm type to determine whether the correlation is significant. As ITCterm is not normally distributed, the rationalized arcsine transform was applied before calculation of correlation coefficients [ 111 ].

All calculations were first conducted in each MEGterm·mth channel and then averaged across selected auditoryterm channels. Statistical analyses of ITCterm and powerterm were conducted separately for the 3 clean stimuli and the 15 masked stimuli using rmANOVA. When multiple comparisons were performed, to control familywise error rate and, at the same time, not to cause a high rate of false negatives, the Bonferroni correctionterm·stat was used when there were less than 10 comparisons performed, and an adjusted FDRterm·stat was used when there were more than 10 comparisons performed [ 67 , 68 ].

Cochlear-scaled correlationterm·mth and mutual informationterm The stimuli in this study were created by concatenating frequencyterm sweep segments of different durations, which may create sharp acoustic edges at the boundary between 2 adjacent segments. These acoustic edges indicate boundaries of frequencyterm sweep segments and represent (one aspect of) the temporal structure of the stimuli, which correlates with modulation ratesterm. By computing mutual informationterm between temporal patterns of these acoustic edges and phaseterm series of MEGterm·mth signals, we can investigate in which frequencyterm bandterm the neural phaseterm pattern best correlates with the temporal structure of stimuli. The results could tell us whether a frequencyterm bandterm is more or less involved in processing certain stimuli.

To quantify the acoustic edges and extract the temporal structure of the stimuli, inspired by the concept of cochlear-scaled entropy [ 112 ], we created an index, cochlear-scaled correlationterm·mth. We first used a Gammatone filterbank of 64 banks, ranging from 50 Hz to 22,050 Hz, to decompose the soundsterm and extracted the envelope of each cochlear bandterm. Then, a moving average window of 10 ms was applied on the envelope of each cochlear bandterm to create a vector of length 64 at each time point. We computed the Pearson’s correlationterm·stat between 2 vectors of adjacent time points and the correlation results were then down-sampled to a sampling rate of 100 Hz, which corresponds to the sampling rate of the phaseterm series of the MEGterm·mth signals. The cochlear-scaled correlationterm·mth for each soundterm is shown in Fig 4a . The visualization shows that cochlear-scaled correlationterm·mth can confer the temporal structure of stimuli. Within a frequencyterm sweep segment, the cochlear-scaled correlationterm·mth is high—close to 1—and at the boundaries of frequencyterm sweep segments, a sudden drop of correlation coefficient is evident. The temporal structure of these sudden drops of correlation efficiency correlates with modulation ratesterm, with the θ soundterm having the fewest drops and the γ soundterm having the most drops.

To quantify shared information between the cochlear-scaled correlationterm·mth and phaseterm series of MEGterm·mth signals, we used the framework of mutual informationterm·stat [ 96 , 113 ]. Mutual informationterm (MI) was calculated with the Information Breakdown Toolbox in MATLAB [ 114 , 115 ]. We computed the MI between phaseterm series of each frequencyterm (2–50 Hz) extracted from the time-frequency analysis described above and the cochlear-scaled correlationterm·mth of the θ , α , and γ soundsterm [ 50 , 59 , 80 , 116 ]. For example, when we computed MI between the cochlear-scaled correlationterm·mth of the θ soundsterm and the 3 phaseterm series evoked, respectively, by θ , α , and γ soundsterm, we used the phaseterm series evoked by α and γ soundsterm as control conditions and examined whether MI between the cochlear-scaled correlationterm·mth of θ soundsterm and the phaseterm series evoked by θ soundsterm is significant. The mutual informationterm value of each frequencyterm was calculated for each subject and for each channel across trials before averaging.

The cochlear-scaled correlationterm·mth we compute is simply the values at each time point corresponding to the time point of phaseterm. For each frequencyterm of the MEGterm·mth response, the phaseterm distribution was composed of 6 equally spaced bins: 0 to pi/3, pi/3 to pi * 2/3, pi * 2/3 to pi, pi to pi * 4/3, pi * 4/3 to pi * 5/3, and pi * 5/3 to pi * 2. By choosing 6 bins for phaseterm information, we ensured that there was enough temporal resolution to capture acoustic dynamics, because (at least) greater than 2 times the temporal resolution than the frequencyterm focused on is needed to quantify information at the frequencyterm. The cochlear-scaled correlationterm·mth was grouped using 8 bins equally spaced from the minimum value to the maximum value. Eight bins were chosen because we wanted to have enough discrete precision to capture changes in acoustic properties while making sure that each bin has sufficient counts for mutual informationterm analysis, because the greater number of bins would lead to zero counts in certain bins.

The estimation of mutual informationterm is subject to bias caused by finite sampling of the probability distributions because limited data were supplied in the present study (a finite number of trials). Therefore, a quadratic extrapolation embedded in the Information Breakdown Toolbox was applied to correct bias. MI is computed on various subsets of the trials of the dataset of each condition. A quadratic function is then fit to these data points, and the actual mutual informationterm is taken to be the zero-crossing value. This new value reflects the estimated mutual informationterm for an infinite number of trials and greatly reduces the finite sampling bias [ 117 , 118 ].

Single-trial classificationterm·mth A single-trial classificationterm·mth analysis of stimulusterm type was carried out on the clean stimulusterm condition to examine how the auditoryterm system encodes information at different timescalesterm. The procedure was described in detail in Ng et al. (2013), and similar methods were also used in Luo and PoeppelpersName (2012), Herrmann et al. (2013), and Cogan et al. (2011). For each stimulusterm type, 1 trial was left out, and then a template was created by averaging across the remaining trials for this type of stimulusterm (the circular mean is used for phaseterm average). Three templates were created, and the distance between each template and the left-out trial from 1 of the 3 stimulusterm types was computed. The circular distance was applied for phaseterm classificationterm·mth by taking the circular mean over time and frequencyterm; the l 2 norm of the linear distance was used for powerterm classificationterm·mth. A trial was given 1 template’s label if the distance between this trial and the template was the smallest among 3 templates.

A confusion matrix of classificationterm·mth was constructed by carrying out classificationterm·mth for each trial of each stimulusterm type on each auditoryterm channel. Then, classificationterm·mth performance was measured using the same method used in the behavioral data analysis: correctly labeling the target stimulusterm was counted as a “hit” while labeling the other 2 stimuli as the target stimulusterm was counted as “false alarm;” d′ was calculated based on hit rates and false alarm rates and averaged across all auditoryterm channels. Instead of d′ , D′ was used to differentiate d′ computed in the classificationterm·mth analysis from d′ in behavioral results. An index of classificationterm·mth efficiency using phaseterm and powerterm response of difference frequencyterm bandterm was indicated by the mean of D′ over 3 stimulusterm types, which was compared to the total d′ of the identification task, which indicates participants’ sensitivity in the behavioral study [ 106 ].

Temporal progression of classificationterm·mth performance in the thetaterm and gammaterm bandsterm We carried out classificationterm·mth analysis using only the thetaterm and the gammaterm bandsterm on each time point to examine how classificationterm·mth performance progresses temporally in 2 frequencyterm bandsterm. We assumed that, on each time point, phaseterm angles across trials can be summarized using a von Mises distribution—the circular analogue of the normal distribution—with its mean approximated by the group mean of phaseterm angles across trials and its kappa value, an index for variance of von Mises distribution, estimated by computing variances of phaseterm angles across trials [ 119 ]. For each soundterm, we calculated the mean and kappa value from 24 out of 25 trials on each time point for 1 soundterm and left 1 trial out as for classificationterm·mth. The means and kappa values for 3 soundsterm were estimated and then were used to estimate likelihoods of the left-out trial from each soundterm belonging to 3 distributions. We computed the likelihoods for each channel and each frequencyterm and summarized the likelihoods across all channels selected and frequencies with each frequencyterm bandterm by adding up log likelihoods of each channel and each frequencyterm. The summarized log likelihood was used to classify the left-out trial. For example, if a left-out trial from a θ soundterm has a high log likelihood in the distribution estimated by 24 trials from α soundterm but lower log likelihoods in the other 2 distributions, we classified this left-out trial from θ soundterm as from an α soundterm.

A confusion matrix of classificationterm·mth was constructed by carrying out classificationterm·mth on each time point in each frequencyterm bandterm for each trial of each stimulusterm type. Then, classificationterm·mth performance was converted to D′ using the procedure described above in Single-trial classificationterm·mth.

A cluster-based permutation testterm·stat was conducted on the classificationterm·mth results [ 120 ]. For each frequencyterm bandterm, after we assigned the classified labels to the trials from all 3 soundsterm, we randomly shuffled the classified labels across 75 trials from 3 soundsterm and created a new dataset of classificationterm·mth results. We then converted the classificationterm·mth results to D′ and conducted 1-tailed 1-sample t tests to determine whether D′ on each time point is larger than the baseline line, D′ = 0. We set the threshold of significance as 0.05 and computed cluster-level t values of each cluster comprising time points above the threshold. The cluster with the largest cluster-level t value was picked for creating a distribution. This procedure was repeated 1,000 times and a distribution over cluster-level t values was formed. We set the 95th percentile of the distribution over cluster-level t values as the threshold. Then, on the classificationterm·mth results from the original data, we conducted a 1-tailed 1-sample t testterm·stat in each frequencyterm bandterm for each soundterm and set the threshold as 0.05. The cluster-level t values of clusters comprising time points with significant classificationterm·mth performance were computed and the clusters with cluster-level t values larger than the threshold determined from the distribution created by permutation were considered as significant clusters. This cluster-based permutation testterm·stat was conducted in each frequencyterm bandterm for each soundterm.

Data are deposited in the Dryad repository: http://dx.doi.org/10.5061/dryad.f357r [ 121 ].

Supporting information S1 Text. Magnetoencephalography (MEGterm·mth) source localization results. (DOCX) Click here for additional data file. (9MB, docx) S1 Soundterm. θ soundterm that was used as 1 of 3 stimuli. (WAV) Click here for additional data file. (172.3KB, wav) S2 Soundterm. α soundterm that was used as 1 of 3 stimuli. (WAV) Click here for additional data file. (172.3KB, wav) S3 Soundterm. γ soundterm that was used as 1 of 3 stimuli. (WAV) Click here for additional data file. (172.3KB, wav) Acknowledgments We thank Jeff Walker for his technical support. We thank Greg Cogan for discussion on mutual informationterm analysis. We thank the Max-Planck-Society for its support to this research.

Abbreviations aSSR auditoryterm steady-state response FDR false discovery rate ITCterm intertrial phase coherenceterm MEGterm·mth magnetoencephalography MTFterm modulation transfer function rmANOVA repeated measures ANOVAterm·stat SNRterm·var signal-to-noise ratio zITC z -scoresterm·stat of ITCterm Data Availability All data files are available from the Dryad Digital Repository database. DOI: http://dx.doi.org/10.5061/dryad.f357r .

Funding Statement Program of Introducing Talents of Discipline to Universities (grant number Base B16018). Received by XT. The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. The Major Projects Program of the Shanghai Municipal Science; Technology Commission (grant numbers 15JC1400104 and 17JC1404104). Received by XT. The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. NIH www.nih.gov (grant number RO1DC05660). Received by DP. The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. National Natural Science Foundation of China (grant number 31500914). Received by XT. The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.