INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Despite the ubiquity and importance of music in human culture, our understanding of the physiological bases of music perception is still in its infancy. A fundamental feature of music is harmony, which refers to characteristics of simultaneous note combinations or "vertical" musical structure (i.e., chords). It has been recognized since antiquity that certain chords sound more pleasant than others (Pythagoras, ca. 600 BC, in Apel 1972). Chords composed of tones related to each other by simple (small-integer) frequency ratios, e.g., octave (2:1) and perfect fifth (3:2), are typically judged to be harmonious, smooth, or consonant, whereas chords composed of tones related to each other by complex (large-integer) ratios, e.g., minor second (256:243) and major seventh (243:128), are considered unpleasant, rough, or dissonant.

In considering consonance and dissonance, it is important to distinguish between musical consonance/dissonance, i.e., of a given sound evaluated within a musical context, and psychoacoustic, or sensory consonance/dissonance, i.e., of a given sound evaluated in isolation (see Plomp and Levelt 1965; Terhardt 1974b, 1977). Musical consonance/dissonance is culturally determined, as evidenced by its variation across cultures and historical periods (see Apel 1972; Burns and Ward 1982). In contrast, judgments of sensory consonance/dissonance are culturally invariant and largely independent of musical training (Butler and Daston 1968). Moreover, rodents, birds, monkeys, and human infants discriminate isolated musical chords on the basis of sensory consonance and dissonance similarly to expert human listeners and experienced musicians (Fannin and Braud 1971; Hulse et al. 1995; Izumi 2000; Schellenberg and Trainor 1996; Zentner and Kagan 1996). These findings indicate that sensory consonance/dissonance is likely shaped by relatively basic auditory processing mechanisms that are not music specific and that can be studied in experimental animals.

Several psychoacoustic theories have been proposed to explain why musical intervals characterized by simple frequency ratios sound more consonant than intervals characterized by complex frequency ratios (see Plomp and Levelt 1965 for review). The most prominent of these theories, first promoted by Helmholtz (1954), states that dissonance is related to the sensation of "beats" and "roughness." These perceptual phenomena occur when two or more simultaneous components of a complex sound are separated from one another in frequency by less than the width of an auditory filter or "critical bandwidth" (10-20% of center frequency) (Zwicker et al. 1957) and are hence unresolved by the auditory system. Unresolved frequency components interact in the auditory periphery, producing fluctuations in the amplitude of their composite waveform envelope that are perceived as beats (fluctuations below 20 Hz) or roughness (fluctuations from 20 to 250 Hz) (Kameoka and Kuriyagawa 1969a,b; Plomp and Levelt 1965; Plomp and Steeneken 1968; Terhardt 1968a,b, 1974a,b, 1978). The rate of these amplitude fluctuations equals the difference in frequency between the components. The disappearance of roughness for stimuli with amplitude fluctuation rates exceeding ~250 Hz is thought to be due to the low-pass characteristic of the auditory nervous system (Plomp and Steeneken 1968; Terhardt 1974a, 1978).

The beats/roughness theory is impressive in its ability to predict the perceived dissonance of musical intervals on the basis of a relatively low-level psychoacoustic phenomenon. For intervals composed of harmonic complex tones, as produced by most musical instruments, dissonance depends on the ratio of the fundamental frequencies (f₀s) of the tones: dissonance is maximal when the f₀s of the complex tones form large-integer ratios and minimal when they form small-integer ratios (Kameoka and Kuriyagawa 1969b; Plomp and Levelt 1965). This pattern arises because chords composed of complex tones forming large-integer f₀ ratios have fewer harmonics in common and more harmonics lying within the same critical band than chords composed of complex tones forming small-integer f₀ ratios. Of these unresolved pairs of harmonics, the number with difference frequencies below 250 Hz is greater for intervals characterized by large-integer f₀ ratios than for intervals characterized by small-integer f₀ ratios. The summation of roughness contributed by each unresolved pair of frequencies separated by >20 Hz and by <250 Hz determines the overall perceived dissonance of musical intervals composed of complex tones (Kameoka and Kuriyagawa 1969b; Plomp and Levelt 1965; Terhardt 1974a, 1978). Consequently, musical intervals with large-integer f₀ ratios produce more roughness and therefore more dissonance.

The neurophysiological basis of sensory consonance/dissonance perception is largely unknown. Bilateral lesions of auditory cortical areas in humans and animals are associated with deficits in pitch perception (Whitfield 1980; Zatorre 1988) and a range of music perception impairments (e.g., Liegeois-Chauvel et al. 1998; Peretz et al. 1994), including aberrant consonance/dissonance perception (Peretz et al. 2001; Tramo et al. 1990). Several physiological studies have suggested that roughness may be represented in primary auditory cortex (A1) by neuronal responses phase-locked to the amplitude-modulated temporal envelope of complex sounds (Bieser and Muller-Preuss 1996; Schulze and Langner 1997; Steinschneider et al. 1998). This hypothesis is supported by the correlation found between the magnitude of neuronal ensemble phase-locking to the AM frequency (= difference frequency) of harmonic complex tones in A1 of the awake monkey and the degree of roughness perceived by human listeners. Specifically, phase-locking is maximal at stimulus modulation frequencies at which roughness is maximal and dissipates at stimulus modulation frequencies at which roughness disappears (Fishman et al. 2000a). Given the involvement of A1 in music perception and assuming the validity of Helmholtz's beats/roughness theory of sensory dissonance, it follows that if the hypothesized mechanism underlying the physiological representation of roughness is correct, then the perceived dissonance of musical chords should correlate with the magnitude of A1 activity phase-locked to the difference frequencies. The present study tests this hypothesis by examining phase-locked neuronal ensemble activity evoked by musical chords with varying degrees of consonance and dissonance in A1 of the awake macaque monkey. Macaques share similarities in basic auditory cortical anatomy and physiology with humans (Galaburda and Pandya 1983; Galaburda and Sanides 1980; Steinschneider et al. 1994, 1999) and are able to discriminate musical chords on the basis of sensory consonance/dissonance (Izumi 2000), making them appropriate animal models for investigating neural representation of sensory consonance and dissonance in the central auditory system.

Correlation between patterns of cortical activity in an animal model and psychoacoustical features of consonance/dissonance perception leaves in question, however, whether these neural response patterns are applicable to the human brain. A stronger argument for the relevance of these physiological responses could be made if physiological findings similar to those obtained in the animal model are observed in human neural responses. Therefore, in parallel to the studies in monkeys, auditory-evoked potentials (AEPs) evoked by musical chords were also recorded directly from the auditory cortex of two patients undergoing surgical evaluation for medically intractable epilepsy. This cross-species approach has already been used to advantage in the study of auditory cortical representation of the voice onset time phonetic feature (Steinschneider et al. 1999) and offers several significant benefits. Clearly, it bolsters the relevance of the animal results by testing the suitability of the macaque as a model in which to examine neural correlates of higher perceptual processes. Furthermore, if a similarity between human and animal physiological response patterns can be demonstrated, the more refined sampling and analysis inherent in animal physiological studies can help to characterize the detailed mechanisms underlying the neural representation of the perceptual process under study.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Monkey surgery and electrophysiological recordings

Three adult male monkeys (Macaca fascicularis) were studied using previously reported methods (Steinschneider et al. 1992, 1994, 1998). Animals were housed in our Association for Assessment and Accreditation of Laboratory Animal Care-accredited Animal Institute under daily supervision by veterinary staff. All experiments were conducted in accordance with institutional and federal guidelines governing the experimental use of primates. Briefly, using aseptic surgical techniques under general anesthesia (pentobarbital, initial and supplementary doses of 20 and 5 mg/kg iv, respectively), holes were drilled in the exposed skull to accommodate epidural matrices consisting of adjacent 18-gauge stainless steel tubes. Matrices were stereotaxically positioned to target A1 and were oriented at an angle of 30° from normal to approximate the anterior-posterior tilt of the superior temporal plane. This orientation guided electrode penetrations roughly perpendicular to the cortical surface, thereby fulfilling one of the major technical requirements of one-dimensional current-source density (CSD) analysis (Vaughan and Arezzo 1988). Matrices and Plexiglas bars used for painless head immobilization during the recording sessions were held in place by a pedestal of dental acrylic fixed to the skull by inverted screws keyed into the bone. Animals were given peri- and postoperative analgesic, antibiotic, and anti-inflammatory medications. Recordings began 2 weeks after surgery and were conducted in an electrically shielded, sound-attenuated chamber with the animals awake and comfortably restrained.

Intracortical recordings were obtained using linear-array multi-contact electrodes containing 14 recording contacts, evenly spaced at 150-µm intervals (Barna et al. 1981). Individual contacts were constructed from 25-µm-diameter stainless steel wires, each with an impedance of ~200 k. An epidural stainless steel guide tube positioned over the occipital cortex served as a reference electrode. Field potentials were recorded using unity-gain headstage preamplifiers, and amplified 5,000 times by differential amplifiers with a frequency response down 3 dB at 3 Hz and 3 kHz. Signals were digitized at a sampling rate between 2 and 4 kHz (depending on the analysis time used) and averaged by computer (Neuroscan software and hardware, Neurosoft) to yield AEPs. To derive multiunit activity (MUA), signals were simultaneously high-pass filtered above 500 Hz, amplified an additional eight times, and full-wave rectified prior to digitization and averaging. MUA is a measure of the summed action potential activity of neuronal aggregates within a sphere of about 50-100 µm in diameter surrounding each recording contact (Brosch et al. 1997; Vaughan and Arezzo 1988). For some electrode penetrations, raw data were stored on a 16-channel digital tape recorder (Model DT-1600, MicroData Instrument; sample rate: 6 kHz) for off-line analyses. Due to limitations of the acquisition computer, the sampling rates used were below the Nyquist frequency corresponding to the 3 kHz upper cutoff of the amplifiers. However, empirical testing revealed negligible signal distortion due to aliasing, as most of the spectral energy in the MUA lies below 1 kHz. Using shorter analysis windows and fewer channels, raw data re-digitized at 6 kHz, yielded nearly identical averaged waveforms as data sampled at the lower rate. Absence of aliasing was also confirmed by low-pass filtering the MUA at 800 Hz (96 dB/octave roll-off) following rectification and prior to digitization at 2 kHz, using digital filters (RP2 modules, Tucker Davis Technologies) acquired after the completion of this study. Differences between unfiltered and low-pass filtered MUA signals were negligible (see Fig. 2). To further confirm the validity of MUA measures, off-line multi-unit cluster analyses of unrectified high-pass filtered data were performed for some sites. Peristimulus time histograms (PSTHs) were constructed with a binwidth of 1 ms. Triggers for spike acquisition were set at 2.5 times the amplitude of the background "hash" of lower-amplitude, high-frequency activity.

One-dimensional CSD analyses characterized the laminar pattern of net current sources and sinks within A1 generating the AEPs. CSD was calculated using a three-point algorithm that approximates the second spatial derivative of voltage recorded at each recording contact (Freeman and Nicholson 1975; Nicholson and Freeman 1975). Current sinks represent net inward transmembrane current flow associated with local depolarizing excitatory postsynaptic potentials or passive, circuit-completing current flow associated with hyperpolarizing potentials at adjacent sites. Current sources represent net outward transmembrane currents associated with active hyperpolarization or passive current return associated with adjacent depolarizing potentials. The corresponding MUA profile is used to help distinguish these possibilities: current sinks coincident with increases in MUA likely reflect depolarizing synaptic activity, whereas current sources associated with concurrent reductions in MUA from baseline levels likely reflect hyperpolarizing events rather than passive current return for adjacent synaptic depolarization.

Electrodes were manipulated with a microdrive and positioned using on-line examination of click-evoked potentials as a guide. Pure tone and chord stimuli were delivered when the electrode channels bracketed the inversion of early AEP components and the largest MUA, typically occurring during the first 50 ms within lamina IV (LIV) and lower lamina III (LLIII), was situated in the middle channels. Evoked responses to 75 presentations of the stimuli were averaged with an analysis window (including a 25-ms prestimulus baseline interval) of 300 ms for pure tones and 520 ms for musical chord stimuli.

Human electrophysiological recordings

Intracranial AEPs were recorded in one man (subject 1) and one woman (subject 2). Both subjects had medically intractable epilepsy, were right-handed, and required placement of multiple temporal lobe electrodes to determine the location of seizure onsets. Experimental procedures were approved by the University of Iowa Human Subjects Review Board and the National Institutes of Health. Informed consent was obtained from the subjects prior to their participation. Subjects underwent surgical implantation of intracranial electrodes (Radionics, Burlington, MA) to acquire diagnostic electroencephalographic (EEG) data required for planning subsequent surgical treatment. Subjects did not undergo any additional risk by participating in this study.

Subject 2 had depth electrodes (Howard et al. 1996a,b) implanted in the right Heschl's gyrus and planum temporale. Data from this subject using different stimulus protocols have been reported (Steinschneider et al. 1999). Bipolar recordings at three locations were obtained from closely spaced recording contacts (impedance, 200 k, 2.5-4.2 mm inter-contact distance) placed sterotaxically along the long axis of Heschl's gyrus. Spectral sensitivity of two of these sites, site 1 (the most posteromedial site) and site 3 (the most anterolateral site), was assessed via independent analysis of multiple unit responses. Maximal tone responses of units at sites 1 and 3 were 2,125 ± 252 and 736 ± 91 (SD) Hz, respectively, consistent with findings that higher frequencies are represented at more posteromedial locations in human A1 (Howard et al. 1996a; Steinschneider et al. 1999). Subject 1 had three depth electrodes implanted in the right superior temporal gyrus: the first in Heschl's gyrus, the second in the planum temporale, and the third in a more posterior location within the planum temporale. Click-evoked responses recorded at the location of the most posterior electrode were of low amplitude, and, consequently, musical chord-evoked responses were not recorded at this electrode. Responses at the Heschl's gyrus and planum temporale electrodes were recorded from two higher-impedance (200 k) and one lower-impedance (30 k) recording contacts (2.5-4.2 mm inter-contact distance). Spectral sensitivities of sites in subject 1 were not determined. The reference electrode was a subdural electrode located on the ventral surface of the ipsilateral, anterior temporal lobe.

Recording sessions took place in a quiet room in the Epilepsy Monitoring Unit of the University of Iowa Hospitals and Clinics with the subjects lying comfortably in their hospital beds. Subjects were awake and alert throughout the recordings. For both subjects, sweeps exhibiting high-amplitude epileptic spikes at any time point within the analysis window were rejected by the acquisition computer or discarded following visual inspection of the data.

AEPs were recorded at a gain of 5,000 using headstage amplification followed by differential amplification (BAK Electronics). Field potentials were filtered (band-pass, 2-500 Hz; roll-off, 6 dB/octave), digitized (1.0- or 2.050-kHz sampling rate), and averaged, with an analysis window of 500 ms (including a 25-ms prestimulus baseline interval) in the case of subject 2 and 1,000 ms (including a 325-ms prestimulus baseline interval) in the case of subject 1. Averages were generated from 50 to 75 stimulus presentations. Raw EEG and timing pulses were stored on a multi-channel tape recorder (Racal) for off-line analysis.

Stimuli

MONKEY RECORDINGS. Frequency response functions (FRFs), based on pure tone responses, were used to characterize the frequency tuning of the cortical sites. Pure tones ranging from 0.2 to 17.0 kHz were generated and delivered at a sampling rate of 100 kHz by a PC-based system using SigGen and SigPlay (Tucker Davis Technologies). Pure tones were 175 ms in duration with 10-ms linear ramps. Stimulus onset asynchrony (SOA) for pure tone presentation was 658 ms. All stimuli were monaurally delivered at 60 dB SPL via a dynamic headphone to the ear contralateral to the recorded hemisphere. Sounds were introduced to the ear through a 3-in-long, 60-ml plastic tube attached to the headphone. Sound intensity was measured with a Bruel and Kjaer sound level meter (type 2236) positioned at the opening of the plastic tube. The frequency response of the headphone was flattened (±3 dB) from 0.2 to 17.0 kHz by a graphic equalizer (Rane).

View larger version (49K): [in this window] [in a new window]   Fig. 1. Waveforms and spectral content of the 8 musical interval stimuli presented in the study (Pythagorean tuning, only 150 ms displayed). Stimuli with base tones of 256 Hz are shown; all frequencies are doubled for the 512-Hz base tone stimuli. Each stimulus is composed of 2 simultaneous harmonic complex tones. Each complex tone contains the fundamental frequency (f0) and the 2nd-10th harmonic, all at equal amplitude.  and , frequencies comprising the lower base tone and upper interval-defining tone, respectively. Corresponding musical notation is shown in the top left corner of each stimulus box. Ratio of the f0 of the higher tone to that of the lower tone is indicated above each stimulus box.

HUMAN RECORDINGS. In the case of subject 1, stimuli were delivered to the left ear (contralateral to the recorded hemisphere) by an insert earphone (Etymotic Research). In the case of subject 2, stimuli were delivered to the left ear by an external headphone (Koss, model K240DF) coupled to a 4-cm cushion. Stimuli were presented at a comfortable listening level (60-70 dB SPL). In the case of subject 1, musical interval stimuli were identical to those used in the monkey recordings and were presented in pseudorandom order with a SOA of 2,000 ms. Due to time constraints, only a subset of the chords presented in the monkey studies was presented in the human studies. In the case of subject 2, two-tone chords were generated using a keyboard synthesizer (Roland, model JV-35) in organ mode. Keyboard-generated sounds were edited and presented in the same manner as sounds created by addition of frequency components, except that their total duration was 375 ms. Spectral analyses of the organ sounds indicated the presence of multiple harmonic components (see Fig. 18). In contrast to the chords constructed from sine wave addition, keyboard generated chords were based on equal temperament tuning (the tuning system conventionally used in modern Western music), thereby allowing qualitative comparison between neural responses evoked by intervals derived from different systems of tuning (Pythagorean vs. equal temperament). Keyboard-generated chords were presented in pseudorandom order with a SOA of 658 ms. Subjects were informally asked to relate their impression of the musical chords following the acquisition of a block of electrophysiological responses (e.g., "Did the chord sound pleasant or unpleasant?"). Patients' subjective evaluations of the chords were consistent with those reported in psychoacoustic studies on consonance and dissonance (Butler and Daston 1968; Kameoka and Kuriyagawa 1969b; Malmberg 1918).

Monkey histology

At the end of the recording period, monkeys were deeply anesthetized with pentobarbital sodium and transcardially perfused with 10% buffered formalin. Tissue was sectioned (80 µm thickness) and stained for acetylcholinesterase and Nissl substance to reconstruct the electrode tracks and to identify A1 according to previously published criteria (Hackett et al. 1998; Merzenich and Brugge 1973; Morel et al. 1993; Wallace et al. 1991a). Field R was demarcated from A1 by a reversal in the best frequency gradient (Merzenich and Brugge 1973; Morel et al. 1993). The earliest sink/source configuration was used to locate LIV (Steinschneider et al. 1992). Other laminar locations were then determined by their relationship to LIV and the measured widths of laminae within A1 for each electrode penetration histologically identified.

Data analysis

MONKEY RECORDINGS. The best frequency (BF) of the cortical site sampled in a given electrode penetration was defined as the pure tone frequency eliciting the largest peak amplitude MUA within LLIII during the first 50 ms following stimulus onset. Determination of BF was generally based on MUA averaged across two to three LLIII electrode contacts. Use of peak amplitude initial MUA as a measure of BF yielded the expected anterolateral to posteromedial topographic gradients of low- to high-frequency representation in all animals (Merzenich and Brugge 1973; Morel et al. 1993; Recanzone et al. 2000).

HUMAN RECORDINGS. Spectral analysis (FFT) was also used to quantify phase-locked activity in AEPs recorded in human auditory cortex. For subject 1, the FFT was applied to the same window as that examined in the analysis of electrophysiological data obtained from monkey A1 (175-445 ms). This analysis window excluded ON and OFF response components. A shorter FFT analysis window (175-370 ms), consistent with shorter stimuli, was applied to the data from subject 2. Phase-locked activity was quantified using two complementary measures. The first measure was similar to that used to quantify monkey data: the peak of the amplitude spectrum from 10 to 300 Hz (no significant oscillatory activity was observed at frequencies >150 Hz). The second measure was the area under the amplitude spectrum from 10 to 150 Hz, which thereby includes spectral peaks corresponding to multiple difference frequencies or harmonics of a single difference frequency. Accordingly, an increase in oscillatory phase-locked activity at multiple frequencies, and potentially relevant for the encoding of roughness and sensory dissonance, would thus be represented by an increase in the integral of the amplitude spectrum of the "steady-state" response within the 10- to 150-Hz frequency range. As in the analysis of monkey data, statistical significance of spectral peaks was assessed by comparing mean spectra of AEPs evoked by non-octave intervals with mean spectra of octave-evoked AEPs.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Neural ensemble activity evoked by musical chords in monkey A1

Results are based on 32 perpendicularly oriented (error, <20%) electrode penetrations into A1. 256 and 512 Hz base tone stimulus sets (each comprised of 8 intervals) were each presented in 17 penetrations. Due to the comparatively low number of sampled cortical sites with BFs <1,280 Hz (the highest frequency component of the base tone in the 128-Hz octave stimulus), chords with base tones of 128 Hz were presented in only four electrode penetrations. As this small sample size precluded meaningful interpretation of statistical measures, data based on 128 Hz base tone interval-evoked responses are not discussed.

Figure 2 shows representative CSD and MUA laminar response profiles evoked by the two most consonant and the two most dissonant musical chords presented: octave and perfect fifth, minor second and major second, respectively (base tone = 256 Hz; BF = 1,600 Hz). CSD and MUA waveforms in each quadrant of the figure represent neuronal activity recorded simultaneously at 150-µm interval depths within A1. The dashed boxes superimposed on the LLIII responses delineate the temporal window subjected to spectral analysis.

View larger version (42K): [in this window] [in a new window]   Fig. 2. Representative examples of current-source density (CSD) and multiunit activity (MUA) laminar response profiles evoked by the 2 most dissonant chords (left: minor and major 2nd) and by the 2 most consonant chords (right: perfect 5th and octave) presented in the study [best frequency (BF) = 1,600 Hz, base tone = 256 Hz]. Approximate laminar boundaries are shown on the left of the figure. Stimulus duration is represented by the black bar above the time axes. Consonant and dissonant stimuli evoke similar early response components reflecting initial cortical activation in lower lamina III (LLIII) and lamina IV (LIV; initial sink, MUA ON) and delayed activation in upper lamina III (ULIII; supragranular sink). Later activity differs between consonant and dissonant interval-evoked responses: dissonant stimuli evoke oscillatory activity phase-locked to the difference frequencies, whereas consonant stimuli evoke little or no oscillatory activity. The present study examines the degree of phase-locking in LLIII MUA, LLIII CSD, and ULIII CSD from 175 to 445 ms poststimulus onset (portion of the response enclosed by the dashed box) as a function of musical interval. Waveforms of LLIII MUA low-pass filtered at 800 Hz (96 dB/octave roll-off) prior to digitization are shown superimposed on waveforms of unfiltered MUA data to illustrate absence of significant signal aliasing, as demonstrated by the nearly flat difference waveforms shown below the superimposed waveforms.

All of the musical chords elicit a stereotypical laminar pattern of activity characterized by short-latency current sinks (below-baseline deflections in the CSD) located in the thalamorecipient zone (lamina IV and LLIII, Initial sink), and slightly later supragranular sinks located in mid- and ULIII (supragranular sink). These sinks are coincident with above-baseline bursts of MUA in lamina III and IV (MUA ON), indicating that they primarily represent current flow associated with depolarizing synaptic potentials. The LLIII and ULIII sinks are balanced by deeper and more superficial current sources (above-baseline deflections in the CSD, e.g., P28 source) that, together with the sinks, form current dipole configurations consistent with initial activation of pyramidal cells in LLIII and delayed polysynaptic activation of pyramidal cell elements in ULIII.

While initial components of responses evoked by the consonant and dissonant stimuli are qualitatively very similar, later portions of the responses differ considerably. Both of the dissonant intervals evoke prominent oscillations in the MUA and CSD that are phase-locked to the predicted difference frequencies (minor 2nd: 13.6 Hz; major 2nd: 32 Hz). Neuronal beating patterns are evident both in the thalamorecipient zone and in ULIII, with maximal phase-locked activity typically occurring in LLIII. This laminar distribution is consistent with that observed in previous investigations of phase-locked neural ensemble activity in macaque A1 (Fishman et al. 2000a; Steinschneider et al. 1998). In contrast, little or no oscillatory activity is evoked by the consonant intervals. Waveforms of LLIII MUA low-pass filtered at 800 Hz (96 dB/octave roll-off) prior to digitization are superimposed on those of unfiltered data. MUA waveforms generated under these two conditions are virtually identical, confirming the absence of significant signal aliasing. This is further demonstrated by the nearly flat difference waveforms, shown below the superimposed waveforms.

Figure 3 shows representative chord-evoked responses (base tone = 256 Hz) recorded at a site with a BF of 1,000 Hz. LLIII MUA and CSD waveforms of chord-evoked responses (from 175 to 445 ms poststimulus onset) and associated amplitude spectra are depicted in Fig. 3, A and B, respectively. The most dissonant intervals (minor and major 2nd) evoke robust phase-locked oscillations in both the MUA and CSD, which are manifested as prominent peaks in the associated amplitude spectra at predicted difference frequencies (indicated by arrows). In contrast, responses evoked by the most consonant intervals (octave and perfect 5th) display little or no oscillatory activity and are characterized by comparatively flat amplitude spectra. Intervals of intermediate dissonance (e.g., major 7th, minor 7th, and augmented 4th) also evoke oscillatory phase-locked responses. Even the perfect fourth evokes oscillatory activity, consistent with the theoretical prediction that perfect fourths become disproportionately more dissonant, compared with octaves and perfect fifths, when the base tone of the chord is lower than ~300 Hz (see Fig. 12 in Plomp and Levelt 1965).

View larger version (47K): [in this window] [in a new window]   Fig. 3. Representative waveforms (A) and corresponding amplitude spectra (B) of LLIII MUA and CSD (175-445 ms poststimulus onset) evoked by musical chords with base tones of 256 Hz. For clarity, only the frequency range from 10 to 200 Hz is displayed because no peaks in the amplitude spectra were evident above 200 Hz. C: frequency response functions (FRFs) based on peak amplitude of LLIII MUA and CSD within the first 50 ms poststimulus onset (BF = 1,000 Hz). Frequency components of the smallest interval (minor 2nd) and of the largest interval (octave) are schematically represented above the FRFs to illustrate overlap between stimulus spectra and the excitatory frequency response area of the neuronal ensemble. Dissonant intervals (e.g., minor and major 2nd) evoke oscillatory phase-locked responses, manifested as peaks in amplitude spectra. Arrows in the spectra indicate major peaks corresponding to predicted difference frequencies (values, in Hz, next to arrows). In contrast, the most consonant intervals (e.g., octave and perfect 5th) evoke little or no phase-locked activity, leading to comparatively flat amplitude spectra.

Figure 4 shows representative chord-evoked responses (base tone = 512 Hz) recorded at a site with a BF of 4,000 Hz. Similar to the pattern of responses evoked by chords with base tones of 256 Hz, the most dissonant 512-Hz base tone chords (minor and major 2nd) evoke phase-locked oscillations in both the MUA and CSD, which are represented as prominent peaks in the associated amplitude spectra at predicted difference frequencies (indicated by arrows). In contrast, responses evoked by the most consonant chords (octave and perfect 5th) are characterized by a virtual absence of rapid oscillations and by comparatively flat amplitude spectra. Intervals of intermediate dissonance also evoke oscillatory responses phase-locked to predicted difference frequencies.

View larger version (50K): [in this window] [in a new window]   Fig. 4. Representative waveforms (A) and corresponding amplitude spectra (B) of LLIII MUA and CSD responses (175-445 ms poststimulus onset) evoked by musical chords with base tones of 512 Hz recorded at a different site from that shown in Fig. 3. C: LLIII MUA and CSD FRFs (BF = 4,000 Hz). Same conventions as in Fig. 3. Dissonant intervals evoke oscillatory phase-locked responses manifested as peaks in amplitude spectra. In contrast, consonant intervals evoke little or no phase-locked activity, leading to comparatively flat amplitude spectra.

Oscillatory phase-locked activity is visible not only in the averaged responses but also in responses evoked by individual stimulus presentations. Statistical significance of phase-locked activity in the dissonant-chord-evoked responses relative to octave-evoked responses is demonstrated in Figs. 5 and 6 for two representative A1 sites (the same as those shown in Figs. 2 and 4, respectively). The figures show mean (±SE) LLIII MUA and CSD waveforms and corresponding mean (±SE) amplitude spectra of responses evoked by the two most dissonant chords (minor and major 2nd) and by the two most consonant chords (octave and perfect 5th). Major peaks in the mean spectrum of responses evoked by the minor second and the major second occur at predicted difference frequencies (). Means at the peaks are significantly larger than means at corresponding frequencies in the spectrum of octave-evoked responses (one-tailed t-test; t and P values are shown in the figures). In contrast, the mean spectrum of perfect fifth-evoked responses (above 10 Hz) is not significantly different from that of octave-evoked responses (P > 0.05). No significant differences between mean spectra are observed at frequencies >150 Hz. Peaks at 60 Hz are present in the mean spectra of Fig. 5 due to the fact that mean spectra of responses to individual stimulus presentations include 60-Hz line noise, which disappears with time domain averaging.

View larger version (36K): [in this window] [in a new window]   Fig. 5. Mean (±SE; n = 70) LLIII MUA and CSD waveforms and corresponding mean (±SE) amplitude spectra of responses (175- to 445-ms poststimulus onset) evoked by the two most dissonant chords (minor and major 2nd) and by the two most consonant chords (octave and perfect 5th) at the same site as that shown in Fig. 2 (base tone = 256 Hz). Major peaks in the mean spectra of responses evoked by the minor 2nd and the major 2nd occur at predicted difference frequencies (). Means at the peaks are significantly larger than means at corresponding frequencies in the spectrum of octave-evoked responses (one-tailed t-test; t and P values are shown in the figures). In contrast, the mean spectrum of perfect 5th-evoked responses (above 10 Hz) is not significantly different from that of octave-evoked responses (P > 0.05). No significant differences between mean spectra are observed at frequencies >150 Hz. A peak at 60 Hz corresponds to line noise, which disappears with time domain averaging.

View larger version (40K): [in this window] [in a new window]   Fig. 6. Mean (±SE; n = 70) LLIII MUA and CSD waveforms and corresponding mean (±SE) amplitude spectra of responses (175-445 ms poststimulus onset) evoked by the two most dissonant chords (minor and major 2nd) and by the two most consonant chords (octave and perfect 5th) at the same site as that shown in Fig. 4 (base tone = 512 Hz). Major peaks in the mean spectra of responses evoked by the minor 2nd and the major 2nd occur at predicted difference frequencies (indicated by arrows). Means at the peaks are significantly larger than means at corresponding frequencies in the spectrum of octave-evoked responses (one-tailed t-test; t and P values are shown in the figures). In contrast, the mean spectrum of perfect 5th-evoked responses (above 10 Hz) is not significantly different from that of octave-evoked responses (P > 0.05). No significant differences between mean spectra are observed at frequencies >150 Hz.

Similar chord-evoked oscillatory response patterns are observed when responses are analyzed using more conventional neurophysiological techniques. Figure 7A shows PSTHs based on multiunit spike activity recorded in LLIII at three representative sites in A1. Data from two of these sites are represented in Figs. 2-6. Similarly to MUA results, PSTHs of the minor- and the major-second-evoked responses display periodic oscillations that are absent in the PSTHs of the octave- and perfect-fifth-evoked responses. Dissonant-chord-evoked oscillations are manifested as peaks in corresponding amplitude spectra at predicted difference frequencies or their harmonics (Fig. 7B). In contrast, spectra of consonant-chord-evoked responses are characterized by a general absence of significant peaks at frequencies >10 Hz.

View larger version (40K): [in this window] [in a new window]   Fig. 7. A: peristimulus time histograms (PSTHs) of multiunit cluster activity recorded in LLIII at 3 A1 sites (binwidth = 1 ms). BFs of the sites shown in the 1st and 3rd rows are 1,600 and 4,000 Hz, respectively. The site shown in the second row did not display a clear BF and exhibited broad frequency tuning ranging from 200 to 15,000 Hz. Black bars above the PSTHs indicate stimulus duration. Note that chords presented at the site shown in the 3rd row had base tones of 512 Hz. B: amplitude spectra of PSTHs in A from 175 to 445 ms poststimulus onset. Spectra of minor- and major-2nd-evoked responses display peaks at predicted difference frequencies and their harmonics, whereas major peaks are absent in spectra of perfect 5th- and octave-evoked responses. No significant peaks are observed at frequencies >150 Hz.

Figures 2-7 illustrate a general pattern of musical chord-evoked responses in A1: dissonant intervals evoke oscillatory phase-locked activity, whereas consonant intervals evoke comparatively little or no phase-locked activity. However, the relative magnitude of phase-locked activity differs across sites. According to the roughness theory of sensory dissonance, the total dissonance of a musical chord reflects the sum of roughness contributed by each pair of unresolved frequency components (Kameoka and Kuriyagawa 1969b; Plomp and Levelt 1965; Terhardt 1974a, 1978). Accordingly, the overall dissonance of a musical chord should be represented by response patterns averaged across the tonotopic map in A1.

To quantify the average relative phase-locked activity across the cortical sites sampled, the peak of the amplitude spectrum of each chord-evoked response was first expressed as a percentage of the minimum peak amplitude of the eight chord-evoked response spectra obtained in each electrode penetration. Normalized peak spectrum amplitudes were subsequently averaged across penetrations. The resultant mean normalized amplitudes, plotted as a function of musical interval (ordered from left to right according to interval width) for each of the three response measures examined are shown in Fig. 8. On average, the octave and perfect fifth evoke comparatively little phase-locked activity, whereas the minor and major second generally evoke the highest amplitude phase-locked responses. The perfect fourth, the third most consonant interval, also yields comparatively little phase-locked activity when presented within the octave above middle C (i.e., with a base tone of 512 Hz) but becomes physiologically more "dissonant" when presented within the octave of middle C. Differences among mean normalized amplitudes across interval conditions are statistically significant (repeated-measures ANOVA: all F > 8.5, P < 0.00001).

View larger version (31K): [in this window] [in a new window]   Fig. 8. Mean normalized peak spectrum amplitude (±SE) from 10 to 300 Hz as a function of musical interval (ordered from small to large). LLIII MUA and CSD and ULIII CSD data for 256- and 512-Hz base tone intervals are represented in separate histograms as indicated. Means are based on data from 17 electrode penetrations. Data corresponding to the 3 most consonant intervals (octave and perfect 5th and 4th) are represented by the white bars. Differences in mean peak spectrum amplitude across stimulus conditions are statistically significant (one-way ANOVA: F > 8.5, P < 0.00001).

To examine the extent to which the magnitude of phase-locked neuronal ensemble activity in A1 correlates with the perceived dissonance of the musical intervals, spectra of the eight musical interval-evoked responses from each electrode penetration were ranked according to their peak amplitude (1 = lowest amplitude, 8 = highest amplitude). Physiological ranks from each penetration were then compared with perceptual ranks of the intervals (1 = least dissonant, 8 = most dissonant; see details in METHODS). For all three of the response measures examined, mean rank of spectral amplitude tends to increase with the perceived dissonance of the chords (Fig. 9). This relationship, quantified by Spearman rank-order correlation analysis based on raw data (n = 17 penetrations) and emphasized by the superimposed linear regression lines, is statistically significant (r values are indicated in the figure; P < 0.00001) for all response components and octave ranges examined. The strongest correlation between neural and perceptual measures is seen for LLIII CSD, while the weakest association is seen for ULIII CSD.

View larger version (39K): [in this window] [in a new window]   Fig. 9. Maxima of amplitude spectra of chord-evoked responses (175-445 ms poststimulus onset) in each electrode penetration were ranked (from lowest to highest amplitude). Mean ranks (n = 17 penetrations) are shown as a function of the perceived dissonance of the musical chords (ordered from least dissonant to most dissonant). Error bars indicate SE. LLIII MUA and CSD and ULIII CSD data for 256- and 512-Hz base tone intervals are represented in separate histograms as indicated. For all response measures and for both chord octave ranges, physiological ranks are significantly correlated with perceptual ranks (Spearman correlation analysis: r and one-tailed P values are indicated in the figure). A linear regression line superimposed on the histograms emphasizes this relationship.

AEPs evoked by musical chords in human auditory cortex

AEPs recorded directly from human auditory cortex display strikingly similar response patterns to those observed in monkey A1. Figure 10 shows musical chord-evoked AEPs recorded at a single site within Heschl's gyrus (subject 1; base tone = 256 Hz). AEPs evoked by the most dissonant chords (minor 2nd, major 2nd, and major 7th) display prominent oscillations phase-locked to the predicted difference frequencies, which are manifested as peaks (indicated by arrows) in the corresponding amplitude spectra (middle). In contrast, responses evoked by the octave and by the perfect 5th are characterized by an absence of rapid oscillations and by comparatively flat, low-amplitude spectra (above ~10 Hz).

View larger version (33K): [in this window] [in a new window]   Fig. 10. Waveforms and corresponding amplitude spectra of chord-evoked AEPs recorded at a single site in Heschl's gyrus of subject 1 (base tone = 256 Hz). Far left: musical notation representation of the chords. Stimulus duration is represented by the black bar above the time axis. The "P70" component of the AEP is indicated in the octave-evoked response. AEPs evoked by dissonant chords (e.g., minor 2nd) display oscillations phase-locked to the predicted difference frequencies, whereas AEPs evoked by consonant chords (e.g., octave and perfect 5th) display little or no oscillatory activity. Arrows in the amplitude spectra indicate major peaks occurring at predicted difference frequencies (values, in Hz, next to arrows). Mean spectra of AEPs evoked by non-octave chords in individual stimulus presentations are shown superimposed on mean spectra of octave-evoked AEPs in the right-hand side of the figure. Error bars indicate SE. Means at peaks are significantly larger than means at corresponding frequencies in the spectrum of octave-evoked responses (P values indicated in figure). No significant differences between mean spectra are observed at frequencies >50 Hz.

Statistical significance of the spectral peaks was assessed by comparing the mean spectrum of AEPs evoked by the non-octave chords with that of AEPs evoked by octaves (shown superimposed in Fig. 10, right; error bars represent SE). Means at the peaks of the dissonant chord-evoked response spectra are significantly larger than means at corresponding frequencies in the mean spectrum of octave-evoked responses (one-tailed t-test; P values are indicated in the figure). Mean spectra of AEPs evoked by the perfect fifth and the augmented fourth are not significantly different from the mean spectrum of octave-evoked AEPs (P > 0.05). No significant differences between the mean spectrum of non-octave chord-evoked responses and that of octave-evoked responses were observed at frequencies >50 Hz. Similar phase-locked response patterns are displayed by AEPs averaged across the three recording sites in Heschl's gyrus of subject 1 (Fig. 11), indicating that oscillatory activity is synchronized (i.e., displays phase coherence) over a considerable distance across the cortical tissue. As quantified in Fig. 12, the magnitude of oscillatory phase-locked activity at each of the three sites and in the averaged data tends to increase with increasing dissonance of the chords. The only major deviation from this trend is the greater oscillatory activity evoked by the major seventh relative to that evoked by the major second.

View larger version (24K): [in this window] [in a new window]   Fig. 11. Waveforms and amplitude spectra of chord-evoked AEPs averaged across 3 Heschl's gyrus recording sites in subject 1 (base tone = 256 Hz). Same conventions as in Fig. 10.

View larger version (28K): [in this window] [in a new window]   Fig. 12. Normalized (percent maximum) peak and area measurements of amplitude spectra (10-150 Hz) of chord-evoked AEPs recorded in Heschl's gyrus of subject 1 as a function of the dissonance of the chords (base tone = 256 Hz). Data from each of the 3 recording sites are represented in separate histograms as indicated. Spectrum peaks and areas tend to increase with increasing dissonance of the chords. This trend is also apparent for AEP waveforms averaged across the 3 recording sites (right-most histogram).

Figure 13 depicts AEPs evoked by chords with base tones of 128 Hz recorded at the same site in Heschl's gyrus as that shown in Fig. 10. In this case, with the exception of the octave, all intervals, including the perfect fifth, evoke oscillatory responses. This pattern is consistent with the observation that in octave ranges below middle C, all intervals, except octaves with base tones >100 Hz, sound rougher and more dissonant than their higher octave counterparts (Plomp and Levelt 1965). This may explain why, in lower octave ranges, intervals smaller than octaves, including perfect fifths, tend to be avoided in music composition (Plomp and Levelt 1965). Oscillatory phase-locked responses are manifested as peaks in corresponding amplitude spectra at predicted difference frequencies (indicated by arrows). Many of these spectral peaks are statistically significant relative to the mean spectrum of octave-evoked responses (right-hand column; one-tailed t-test; P values are indicated in the figure; same conventions as in Fig. 10). No significant differences between mean spectra are observed at frequencies >75 Hz (P > 0.05).

View larger version (37K): [in this window] [in a new window]   Fig. 13. Waveforms and amplitude spectra of chord-evoked AEPs recorded at a single site in Heschl's gyrus of subject 1 (base tone = 128 Hz). Same site and conventions as in Fig. 10. AEPs evoked by all chords, except for the octave, display oscillations phase-locked to the predicted difference frequencies. Arrows in the amplitude spectra indicate major peaks corresponding to predicted difference frequencies (values, in Hz, next to arrows). Most of the spectral peaks are statistically significant, relative to mean spectra of octave-evoked AEPs (P values indicated in the figure). Peaks at 60 Hz correspond to line noise, which disappears with time domain averaging. No significant differences between mean spectra are observed at frequencies >75 Hz.

As in the case of responses evoked by intervals with base tones of 256 Hz, similar phase-locked response patterns are evident when AEPs evoked by intervals with base tones of 128 Hz are averaged across the three Heschl's gyrus recording sites (Fig. 14), indicating that oscillatory activity is synchronized over a considerable distance across the cortical tissue. Responses are quantified in Fig. 15, which again shows that, with the exception of the comparatively high values obtained for augmented fourth-evoked responses, the magnitude of oscillatory activity tends to increase with increasing dissonance of the chords. This trend is not apparent, however, for AEPs recorded at the most medial location in Heschl's gyrus, Site 3. 

View larger version (25K): [in this window] [in a new window]   Fig. 14. Waveforms and corresponding amplitude spectra of chord-evoked AEPs averaged across the 3 Heschl's gyrus recording sites in subject 1 (base tone = 128 Hz). Same conventions as in Fig. 11. AEPs evoked by all chords, except for the octave, display oscillations phase-locked to the predicted difference frequencies. Arrows in the amplitude spectra indicate major peaks corresponding to predicted difference frequencies (values, in Hz, next to arrows).

View larger version (32K): [in this window] [in a new window]   Fig. 15. Normalized (percent maximum) peak and area measurements of amplitude spectra (10-150 Hz) of chord-evoked AEPs recorded in Heschl's gyrus of subject 1 as a function of the dissonance of the chords (base tone = 128 Hz). Same conventions as in Fig. 12.

Chord-evoked AEPs recorded simultaneously at posterior electrode contacts located in the planum temporale display little or no oscillatory activity, even when elicited by the most dissonant chords. This is illustrated in Fig. 16, which shows representative AEPs recorded at a single site in the planum temporale (base tone = 128 Hz). Phase-locked activity is largely absent except for low-amplitude (but statistically significant, relative to octave-evoked responses; P < 0.005) oscillations at 64 Hz in the perfect fifth-evoked response. Above 10 Hz, mean spectra of AEPs evoked by all non-octave intervals are not significantly different from the mean spectrum of octave-evoked AEPs (P > 0.05), except for the mean spectrum of the perfect fifth-evoked response. AEPs evoked by chords with base tones of 256 Hz (whose difference frequencies are double those of the 128-Hz base tone chords) in the planum temporale are characterized by a similar absence of oscillatory activity, even in the case of perfect fifth-evoked responses (data not shown). Correspondingly, mean spectra of all non-octave 256-Hz base tone chord-evoked responses are not significantly different from mean spectra of octave-evoked responses (P > 0.05; data not shown). This markedly diminished sensitivity to temporal features of the stimuli is remarkable given the comparatively large amplitude of the AEPs (e.g., the P70 componentindicated in the octave responsesrecorded in the planum temporale is approximately twice the amplitude of that recorded in Heschl's gyrus).

View larger version (31K): [in this window] [in a new window]   Fig. 16. Representative waveforms and corresponding amplitude spectra of chord-evoked-AEPs recorded in the planum temporale of subject 1 (base tone = 128 Hz). Same conventions as in Fig. 10. Note the larger amplitude of planum temporale responses (compared with Heschl's gyrus responses) and the absence of significant oscillatory activity in AEPs evoked by even the most dissonant chords. The only exception is the low-amplitude (but statistically significant) oscillatory activity at 64 Hz in the perfect 5th-evoked response. A complete absence of statistically significant phase-locked activity is observed for AEPs evoked by chords with base tones of 256 Hz (data not shown).

In both Heschl's gyrus and the planum temporale, the amplitude of the P70 component tends to increase with increasing width of the musical intervals (i.e., as interval ratio increases from the minor 2nd, the smallest interval, to the octave, the largest interval), as illustrated in Fig. 17A, left. Previous physiological studies in monkey A1 have shown that the amplitude of intracortical AEP components increases with increasing frequency separation between the harmonics of a complex tone spectrally centered at the BF, consistent with a manifestation of critical band masking phenomena (Fishman et al. 2000b). Because the number of pairs of resolved harmonics in the musical chords increases with interval width (Fig. 17A, right), we hypothesized that the enhancement in P70 amplitude with increases in interval width may reflect similar critical band masking effects. In support of this hypothesis, P70 amplitudes in both Heschl's gyrus and the planum temporale are correlated with the number of pairs of spectrally resolved harmonics in the chords (Fig. 17B). Linear regression lines superimposed on the scatter-plots emphasize this relationship. Spearman correlation coefficients for AEPs recorded at each of the three Heschl's gyrus and three planum temporale electrode contacts are indicated in the scatter-plot insets. Coefficients >0.83 are statistically significant (n = 6, P < 0.05). Amplitudes of chord-evoked responses in A1 of the monkey are unrelated to interval width (data not shown). A possible explanation for why such effects in monkey A1 are not observed in the present study is considered in the DISCUSSION.

View larger version (34K): [in this window] [in a new window]   Fig. 17. A, left: amplitude of the P70 component of human intracranial AEPs (subject 1) as a function of musical interval width (only data for AEPs averaged across 3 recording sites are shown). Symbols representing data corresponding to AEPs evoked by chords presented in the 2 octave ranges and recorded in Heschl's gyrus and in the planum temporale are identified in the legend below. Amplitudes tend to increase with increasing interval width. Right: number of pairs of resolved harmonics in the chords as a function of interval width (- - - , intervals with base tones of 128 Hz; , intervals with base tones of 256 Hz). B: normalized amplitude (percent maximum) of P70 as a function of the number of pairs of resolved harmonics in the chords. Heschl's gyrus and planum temporale data for the 2 octave ranges are represented in separate scatter plots, as indicated. Symbols representing data corresponding to AEPs recorded at each of the 3 electrode contacts located in Heschl's gyrus and the planum temporale are identified in the legend at the bottom of the figure. , data corresponding to AEPs averaged across the 3 recording sites. P70 amplitude tends to increase with the number of pairs of resolved harmonics comprising the chords. Spearman correlation coefficients are shown in the insets. Coefficients >0.83 are statistically significant (P < 0.05). Superimposed linear regression lines emphasize this relationship.

The relative perceived consonance/dissonance of musical intervals tuned according to the Pythagorean or pure fifth tuning system does not differ substantially from that of intervals tuned according to the equal temperament system, which divides the octave into 12 equal semitones. Accordingly, patterns of oscillatory activity evoked by equal temperament intervals were similar to those evoked by Pythagorean intervals. Figure 18B shows AEPs evoked by chords played an octave below middle C on an electronic keyboard (tuned in equal temperament) at three sites within Heschl's gyrus of subject 2. AEPs evoked by the minor and major second display oscillations that are mani