Vol 456 | 13 November 2008 | doi:10.1038/nature07448
ARTICLES
Using temperature to analyse temporal
dynamics in the songbird motor pathway
Michael A. Long1 & Michale S. Fee1
Many complex behaviours, like speech or music, have a hierarchical organization with structure on many timescales, but it is
not known how the brain controls the timing of behavioural sequences, or whether different circuits control different
timescales of the behaviour. Here we address these issues by using temperature to manipulate the biophysical dynamics in
different regions of the songbird forebrain involved in song production. We find that cooling the premotor nucleus HVC
(formerly known as the high vocal centre) slows song speed across all timescales by up to 45 per cent but only slightly alters
the acoustic structure, whereas cooling the downstream motor nucleus RA (robust nucleus of the arcopallium) has no
observable effect on song timing. Our observations suggest that dynamics within HVC are involved in the control of song
timing, perhaps through a chain-like organization. Local manipulation of brain temperature should be broadly applicable to
the identification of neural circuitry that controls the timing of behavioural sequences and, more generally, to the study of the
origin and role of oscillatory and other forms of brain dynamics in neural systems.
Motor behaviours are built out of a sequence of movements that
evolve through time. From the most basic, such as locomotion, to
the most complex, such as playing the piano, the timing of movements
is crucial. For some simple oscillatory behaviours, in which the movement evolves on a single timescale, it has been possible to identify the
particular neurons and biophysics that control the temporal dynamics
of the behaviour—for example pacemaker neurons in the stomatogastric ganglion1 or the oscillator network that controls swimming in
the leech2. However, it is not known what mechanisms underlie more
complex learned behaviours that have structure on many timescales.
Birdsong has a remarkably precise and hierarchically organized
temporal structure3,4 mediated by a number of distinct, well-studied
motor nuclei5,6 (Fig. 1a), allowing for an unprecedented view into the
central control of motor timing. Adult zebra finches generate a 0.5–
1.0-s song motif that is repeated a number of times during a bout of
singing7. The motif itself is made up of song syllables—individual
bursts of sound that are approximately 100 ms in length and occur
in a precise order. Syllables are highly stereotyped and often contain
complex acoustic structure that can evolve rapidly (10-ms timescale).
The duration of song elements at all timescales is stereotyped; trial-totrial fractional variations in song timing are roughly 1% (refs 8–10)z.
It is not known whether different brain regions are responsible for
the timing of motifs, syllables, and subsyllabic structure. Two forebrain nuclei in particular have been implicated in the control of the
temporal structure of song: HVC and RA. HVC projects to RA, which
in turn projects to the vocal motor neurons11 as well as midbrain vocal
control and brainstem respiratory areas12. Previous electrophysiological studies have found evidence that these brain regions contribute
to song structure in a hierarchical manner13,14 and have suggested that
the dynamics underlying the generation of different song timescales
may reside in different brain regions. For instance, syllable-timescale
dynamics have been suggested to occur in HVC15, whereas subsyllabletimescale dynamics may arise in RA14,16,17. Additionally, portions of
the midbrain and respiratory areas project back to HVC through
thalamic nucleus Uvaeformis (Uva)18,19, raising the further possibility
that syllables, which are tightly linked to respiratory patterning20, may
be timed by respiratory oscillator circuits9,21. With current techniques,
1
however, it has been difficult to test ideas about the origin of dynamics
that underlie the temporal control of song.
Localizing temporal dynamics with temperature
We set out to localize temporal dynamics within the song control
system, taking advantage of the fact that the speed of brain processes
is strongly temperature dependent22–24. The aim was to produce localized mild heating or cooling25,26, rather than inactivation, for which
cooling has also been used27. The basic logic of our experiments is as
follows. If the circuitry in a particular brain area is involved in controlling song timing, then cooling that area should slow the song.
Furthermore, if the neural control of song is organized with a dynamical hierarchy (that is, different song timescales are controlled by
biophysical dynamics in different brain areas), it should be possible
to differentially alter the behavioural timescales by individually
manipulating the temperature in these areas.
Dynamics in HVC
We started by bilaterally manipulating the temperature of nucleus
HVC. We designed a device, based on the Peltier effect, that is capable
of rapidly heating or cooling HVC in a spatially restricted manner
(Fig. 1a–c, Supplementary Fig. 1). Song timing was strongly affected
by changes in HVC temperature. At colder temperatures, song motifs
were produced more slowly than control songs (Fig. 1d,
Supplementary Fig. 2). All birds (n 5 10) showed a significant
increase in motif duration during cooling (ranging from 16.9 to
44.9%; Fig. 1e). Fractional change in motif duration (dilation) was
found to vary approximately linearly with temperature in the range
from 0 to 26.5 uC (0 to 1 A in terms of current; Fig. 1f). The slope of
this relation was used as a simple metric of temperature-dependent
song dilation, which we refer to as stretch (measured in per cent per
degree Celsius; see Supplementary Methods). The stretch metric in
different birds ranged from 21.89 to 23.97% uC21, with a mean of
22.83 6 0.22% uC21. Changes in song speed during cooling were
immediate and persisted for an hour or more (Supplementary Fig.
1d). Notably, temperature changes in HVC had only a small effect on
the acoustic structure of the song (Supplementary Fig. 4).
McGovern Institute for Brain Research, Department of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA.
189
©2008 Macmillan Publishers Limited. All rights reserved
ARTICLES
NATURE | Vol 456 | 13 November 2008
Subsyllabic
structure
5 ºC
0.5 A
t
−1 A
Heat sink
c
Dilation (%)
50 s
∆T ( ºC)
5
0
−5
0.5 mm
2 mm
4 mm
−10
−1.5 −1 −0.5 0
Current (A)
Respiratory
areas
0.5
d
30
30
20
20
20
10
10
10
0
0
0
−6
d
−4 −2
∆T (ºC)
5
−6
0
15 Syllable onset
Count
t
abcd
c
Syllable
Motif onset
−5
0
5
Stretch (% ºC−1)
e
−4 −2
∆T (ºC)
−2
−3
−4
−5
−5
−6
0
−4 −3 −2
Syllable onset
(% ºC−1)
f
abcd
−4 −2
∆T (ºC)
0
−2
−3
−4
−5
−5
−4 −3 −2
Motif onset
(% ºC−1)
Figure 2 | HVC cooling slows the song at all timescales. a, Dilation of
subsyllabic structure versus HVC temperature change for all five syllables of
bird no. 8. b, Dilation of syllable-onset intervals for the bird no. 8. c, Dilation
of motif-onset intervals for all seven birds that produced concatenated
motifs at all temperatures. All error bars, s.e.m. d, Distribution of stretch
metrics for the entire data set, including syllables (36 syllables, eight birds),
syllable onsets (43 syllables, nine birds) and motif onsets (seven birds).
e, f, Stretch of syllable-onset interval was strongly correlated with subsyllabic
stretch (e) and motif-onset stretch (f) (for further details, see Supplementary
Information).
100 ms
Control
(artificially
stretched)
e
f
40
Dilation (%)
Dilation (%)
b
30
0
20
0
−10
a
b
10
−1.9%
Motif
onsets
t
a
Peltier device
nXIIts
Syrinx
Syllable
onsets
Syllable onset (% ºC−1)
b
Power supply
Syllable (% ºC−1)
a
−5
0
∆T ( ºC)
5
40
20
0
−10
−5
∆T (ºC)
0
5
Figure 1 | Changes in HVC temperature affect song duration. a, The Peltier
device and relevant parts of the song production pathway. N and P,
semiconductor elements; DM, dorsomedial nucleus of the intercollicular
complex; nXIIts, tracheosyringeal part of the hypoglossal nucleus.
b, Temperature change in HVC as a function of time after onset (open circle)
of the indicated current through the Peltier device (top, heating; bottom,
cooling); current switched off at filled circles. c, Calibration curves for brain
temperature changes (DT) at various depths under the Peltier device (n 5 4).
d, Representative sonograms (frequency, 1–9 kHz) recorded from bird no. 3
with HVC heated (0.25 A) and cooled (0.25 to 1.5 A in 0.25-A steps), showing
percentage song dilation relative to control. Hotter colours represent greater
sound intensity. Bottom, spectrogram of the control motif shown artificially
stretched. e, Percentage change in duration (dilation) of song motif versus
change in temperature, relative to the pre-implantation song (n 5 10). Red,
bird no. 3. f, Motif dilation averaged over all ten birds. The shaded area
represents the range over which the song stretch metric (slope) was
calculated. All error bars, s.e.m.
To quantify whether HVC cooling slows the song even at the shortest
timescale of subsyllabic structure, we used a standard dynamic time
warping algorithm based on the correlation of sound features of the
control song with the cooled song (Supplementary Methods,
Supplementary Fig. 5), and also directly measured the duration of
subsyllabic elements. The average dilation of subsyllabic structure for
each song syllable was computed at each temperature condition (0 to
26.5 uC; Fig. 2a), and the slope (the stretch metric) of the dilation as a
function of temperature was computed for each syllable (Fig. 2d). The
mean stretch for subsyllabic structure was found to be
22.88 6 0.12% uC21, which differs significantly from zero (t-test,
P , 1026). This observation suggests that biophysical dynamics in
HVC are involved in controlling song timing on a fine timescale.
We then considered the control of syllable onsets. In principle,
respiratory circuits projecting to HVC (for example through Uva)
could act as a ‘clock’ that autonomously controls the initiation of
syllables, in which case cooling HVC should have little effect on
syllable onsets. Alternatively, the onset of a syllable may be linked
to the completion of the previous syllable, in which case the interval
between onsets should increase during cooling, as the duration of
each syllable increases. In fact, the intervals between syllable onsets
were significantly dilated by an average of 23.05 6 0.11% uC21
(t-test, P , 1026; Fig. 2b, d). This is not consistent with a model in
which syllable timing (or the timing of singing-related respiration) is
autonomously controlled by circuit dynamics in respiratory circuits
or any other area upstream or downstream of HVC.
The stretch of syllable-onset intervals was significantly correlated
with the stretch of the syllables within the intervals (r2 5 0.607, slope
of 0.92 6 0.16; Fig. 2e), and more weakly correlated with the stretch
of other syllables (r2 5 0.47; see Supplementary Materials). In other
words, for syllables that had a larger stretch than average, the onset
interval to the following syllable also had a larger stretch than average.
This is consistent with a model in which the onset of each syllable may
be causally linked to, or triggered by, the end of the previous syllable.
The silent gaps between syllables were also significantly dilated by
HVC cooling (23.70 6 0.32% uC21; t-test, P , 10212, median of
23.29% uC21), suggesting that biophysical dynamics in HVC are
involved in the timing of gaps. The cooling-related stretch of gaps
was slightly larger than that observed for syllables (paired t-test,
P , 0.05, median gap stretch was 12% larger than median syllable
stretch). Similarly, the average stretch of a syllable-onset interval was
slightly larger than the stretch of the syllable contained within that
interval (mean paired difference, 0.26 6 0.088% uC21; paired t-test,
P , 0.001, median interval stretch was 3.4% larger than median syllable stretch; Supplementary Fig. 6). These observations imply that
the circuit mechanisms involved in initiating song syllables may be
different from those involved in generating structure within song
190
©2008 Macmillan Publishers Limited. All rights reserved
ARTICLES
NATURE | Vol 456 | 13 November 2008
Dynamics in RA
Although the HVC cooling experiments strongly suggest the involvement of HVC in generating the fine temporal structure within syllables, they do not rule out some involvement of other brain areas. In
particular, circuit dynamics14,16,17 and connectivity28,29 within RA, as
well as reciprocal connections from RA to HVC30, have been implicated in the generation of these short timescales. In general, these
models would predict that song timing can be affected by manipulating circuit dynamics in RA. Here we directly test this prediction by
bilaterally cooling RA during singing. We use a Peltier device similar
to that used for HVC, but with attached gold probes (330-mm diameter) that were implanted into RA to facilitate thermal conduction
(Fig. 3a, Supplementary Figs 7, 8). At a distance of 200 mm from the
probe, the distance estimated to be the farthest extent of RA neurons,
we observed a temperature drop of 10 uC at the maximum current
used. We also found that the RA cooling device produced a slight
b
Cooling in RA
∆T (ºC)
10
HVC
RA
RA
HVC
0
−10
−2
5
0
HVC cooling
RA probe cooling
20
a+4.6 ºC
−3.5 ºC
0
−10
−5
0
Current (A)
10
10
5
0
−5
0
RA ∆T ( ºC)
−5
0
HVC ∆T ( ºC)
f
15
−5
−10
−10
5
5
b 20
0 ºC
Subsyllabic
structure
−8.6 ºC
−5
0
5
Stretch (% ºC−1)
Figure 3 | Effects of RA temperature change on song timing. a, X-ray image
of the implanted RA cooling device and approximate locations of HVC and
RA. b, Temperature in RA (200 mm from cooling probe) and HVC as a
function of RA probe current. We note that the RA probe produces some
cooling in HVC. c, Change in motif duration as a function of RA probe
current (n 5 4; red squares, mean). d, Average change in motif duration (red
squares) during RA probe cooling or heating, plotted as a function of HVC
temperature. Also plotted is the average change in motif duration (blue
circles) as a function of HVC temperature measured in the HVC cooling
experiment (Fig. 1f). e, Change in motif duration as a function of RA
temperature, corrected for the effect of HVC temperature change. All error
bars, s.e.m. f, Stretch of subsyllabic elements for the population of RA cooled
birds (n 5 4, 20 syllables), corrected for HVC temperature change.
10
0
−5
−10.1 ºC
0
250 ms
c
5
0
−6.4 ºC
5
Spikes per second
Dilation (%)
e
Dilation (%)
10
−5
1
d 40
Count
Dilation (%)
c 15
−1
0
Current (A)
20
10
0
5
Time (s)
10
d
30
Bursts per minute
a
temperature change in HVC of roughly 30% of the temperature
change in RA at each current level (Fig. 3b).
As expected, because of the residual effect of the RA cooling probe
on HVC temperature, we observed a slight increase in motif duration
at higher cooling currents (n 5 4; Fig. 3c, Supplementary Fig. 9a).
The effect of the residual HVC cooling on motif duration can be
accurately predicted by the results of the HVC cooling experiments
(Fig. 1f) and fully accounts for changes in motif duration produced
by the RA probe (Fig. 3d). Thus, after incorporating a correction for
HVC temperature changes, we find no evidence that changes in RA
temperature affect song motif duration (P . 0.20; Fig. 3e) or the
timing of subsyllabic structure (Fig. 3f), suggesting that dynamics
in RA may not contribute significantly to song timing, at least by
any mechanism that is sensitive to temperature changes in the range
we were able to achieve here.
The fact that RA cooling had so little effect on song structure led us
to wonder whether our temperature manipulation had any effect on
the neuronal properties in RA. In non-singing birds, RA neurons
spontaneously generate tonic, regular spiking14, possibly associated
with an intrinsic subthreshold membrane potential oscillation31. We
measured the spiking frequency of single units in RA in an anaesthetized preparation while changing the temperature using the RA cooling device. Cooling produced a rapid, roughly linear decrease in RA
neuron tonic spiking rate (19 cells from seven birds, slope of
0.85 Hz uC21; Fig. 4a–c and Supplementary Fig. 9b) that resulted in
a near cessation of spontaneous spiking at the coldest temperatures
(DT 5 210 uC). Our observation that cooling RA by 10 uC produces a
2.5-fold reduction in the intrinsic oscillation frequency of RA neurons, yet has no detectable effect on song structure or timing, implies
that these oscillations are not likely to be a source of dynamics underlying song production.
In contrast, the incidence of high-frequency spontaneous bursts in
RA (Fig. 4d, top), known to be driven by synaptic input from HVC
under anaesthesia and during sleep32,33, does not show a significant
trend with temperature (P . 0.6; Fig. 4d). The bursts exhibited only a
slight cooling-related decrease in firing rate (5.6 Hz uC21;
Supplementary Fig. 9c). RA is thus capable of a robust response to
burst input from HVC, even at temperatures low enough to substantially suppress tonic spiking.
Frequency (Hz)
syllables, as has been suggested by measurements of the variability in
timing of gaps and syllables in natural singing8,9.
An analogous argument can be made for the timing of motif
onsets; if there is a ‘motif clock’ outside HVC that independently
controls the intervals between motif onsets, then cooling HVC
should have little effect on motif-onset intervals. In fact, motif-onset
intervals were significantly dilated by an average of
23.19 6 0.24% uC21 (t-test, n 5 7, P , 1025; Fig. 2c, d, f), which is
not consistent with a model in which motif onsets are timed autonomously by circuit dynamics outside HVC.
−10
0
∆T (ºC)
10
60
200 ms
40
20
0
−10
0
∆T (ºC)
10
Figure 4 | Effects of RA temperature change on RA spiking activity. a, An
example of the tonic spiking activity of an RA neuron in an anaesthetized
bird for various temperature changes. b, Average firing rate response (25
trials) to the application of 1-A cooling current to the RA probe. c, Average
tonic spiking rate versus temperature for all recorded neurons (19 cells,
seven birds). Filled-red circles are from the example shown in a. d, Spike
train showing tonic spiking and spontaneous bursts (top) and incidence of
bursts (defined as an instantaneous firing rate greater than 100 Hz) for all
neurons (bottom).
191
©2008 Macmillan Publishers Limited. All rights reserved
ARTICLES
NATURE | Vol 456 | 13 November 2008
The control of song timing by HVC: lateralization
The HVC and RA cooling experiments highlight the centrality of
HVC in controlling song timing. HVC is a bilateral structure, and
it is natural to wonder how the two HVCs are coordinated during
singing. Bilateral multi-unit recordings in HVC have revealed brief
episodes of correlated activity across hemispheres that occur before
the onset of each syllable (and at some acoustic transitions within
long complex syllables)34,35, probably mediated by feedback pathways
from RA to midbrain areas and bilaterally back to HVC18,19. Do these
episodes reflect actual bilateral synchronization of the HVCs? If the
two HVCs were, hypothetically, synchronized only at the beginning
of the motif, after which they operated independently, cooling HVC
in only one hemisphere (Fig. 5a) should cause the two HVCs to
become misaligned in time by more than a whole syllable by the
end of the motif (compare control to bilaterally cooled song;
Fig. 5c), causing song degradation. However, we found that unilateral
cooling of HVC did not produce song degradation, but resulted in
a
b
c
20
Dilation (%)
T (ºC)
50
40
30
200 ms
10
0
20
−10
−1.5 −1 −0.5 0 0.5
Current (A)
d
−5
0
∆T ( ºC)
5
Cool right
e
Cool left
0 ºC
0 ºC
-5.2 ºC
-5.2 ºC
f
h
10
−4
0
10
0
15
Count
10
5
0
−5
−4
0
∆T (ºC)
−2
−4
j
k
15
Bilateral
5
Unilateral
10
Count
−8
i
0
Non-uniformity
(cool left)
−8
g
Stretch (% ºC−1)
Dilation (%)
0
0
5
Non-uniformity
5
0
−5
0
5
Non-uniformity
0
−5
−5
0
5
Non-uniformity
(cool right)
Figure 5 | Effect of unilateral HVC cooling on song timing. a, Simultaneous
temperature measurements from HVC in both hemispheres when the Peltier
device was configured for right HVC cooling. b, Change in motif duration as
a function of HVC temperature change during unilateral and bilateral
cooling in bird no. 11. Error bars, s.e.m. c, Spectrograms of song motif
during control, bilateral, left and right HVC cooling. d, Selective dilation of
subsyllabic element B, but not A, during left HVC cooling. e, Selective
dilation of element A, but not B, during right HVC cooling. f, g, Dilation of
subsyllabic element A (f) and B (g) during left (blue) and right (red) HVC
cooling. h, Stretch metric of identified song segments during cooling of left
(blue), right (red) or both (black) HVCs. i, j, Distributions of stretch nonuniformity values during bilateral (i) and unilateral (j) cooling. k, Nonuniformity values during left and right cooling show significant
anticorrelation (P 5 0.026). Solid line shows the first principal component
of the distribution.
slowed songs of normal acoustic structure (n 5 4; Fig. 5b, c,
Supplementary Fig. 10), ruling out any model in which hemispheric
synchronization occurs only at motif onsets.
Do both HVCs contribute to song timing? Given previous observations of hemispheric dominance in songbirds5,36, it is conceivable
that one HVC acts as a ‘master clock’ for song timing and that the
other follows as a slave. In fact, in all birds (n 5 4) we observed that
cooling either HVC alone caused song slowing intermediate to that
seen for bilateral cooling (Fig. 5b, Supplementary Fig. 10), ruling out
the possibility that song timing is controlled by dynamics in a single
hemisphere.
One possible explanation for the observation of intermediate slowing is that the left HVC may control the timing of some parts of the
song and the right HVC may control the timing of other parts of the
song. In fact, we found that cooling the right HVC or left HVC produced less uniform stretching of the song in comparison with bilateral
cooling, as predicted by this model (bilateral non-uniformity s.d. of
0.67% uC21, unilateral non-uniformity s.d. of 1.33% uC21; Fig. 5d–j).
This inhomogeneity in stretch during unilateral cooling is comparable
to the mean stretch (,1.2% uC21), suggesting that a large fraction of
song elements were not stretched by unilateral cooling (bilateral cooling, 14% not significantly stretched; left only, 71%; right only, 43%).
Furthermore, some elements that were not stretched by right HVC
cooling were strongly stretched by left HVC cooling, and vice versa
(Fig. 5d–h, Supplementary Fig. 10). Consistent with this, stretch during left cooling was significantly anticorrelated with stretch during
right cooling (P 5 0.026; Fig. 5k; see Supplementary Methods).
Thus, it appears that there may be some switching of the control of
song timing between the two HVCs on the timescale of song syllables
or long subsyllabic elements.
A chain model of song dynamics
The results of our HVC and RA cooling experiments do not support a
view in which the dynamics underlying song timing are divided at
different timescales between different brain areas. One possible
model of dynamics in the song control system that is consistent with
all of our observations is strongly anticipated by the singing-related
firing patterns of HVC neurons that project to RA. These neurons
burst extremely sparsely during singing, each generating a single brief
(,6-ms) burst of spikes at a particular moment in every repetition of
the song motif 37. In addition, different HVC neurons burst at different times throughout the song. We have proposed that, as a population, these HVC neurons form domino-like chains of activity that
control the timing of the song38. The HVC cooling experiments suggest that the dynamics underlying the sparse sequential activation of
HVC neurons reside at least partly within HVC; if the sparse HVC
bursts were driven by an independent upstream sequence generating
circuit, cooling HVC should not have affected the speed of the song.
An interesting possibility is that such sequential chains of activity
arise, in part, from a chain-like synaptic organization within
HVC39–43. In this case, cooling HVC may simply increase the time
it takes each neuron to burst following activity in a preceding neuron,
thus introducing an accumulating delay that slows down the chain.
Although the cooling results rule out models in which brain areas
outside HVC independently or autonomously control song timing
on any timescale, they do not exclude the involvement of other brain
areas important for song production, in particular feedback projections from RA, through the brainstem/midbrain and Uva back to
HVC19,21. These projections could form an integral part of the connectivity that generates sequential bursts in HVC. Consider the timescale on which this feedback might operate. In principle, every burst
in HVC could be driven by fast, rapidly cycling feedback through this
loop, rather than by intrinsic connections within HVC. Cooling anywhere in this feedback loop, including RA, should introduce accumulating delays as well. The fact that we do not observe song slowing
from cooling in RA suggests that the feedback circuitry may not
192
©2008 Macmillan Publishers Limited. All rights reserved
ARTICLES
NATURE | Vol 456 | 13 November 2008
operate at this rapidly cycling timescale, but less frequently during the
song.
Thus, one interesting possibility is that HVC may contain multiple
independent chains8 (or modules), which may be associated with
syllables or long subsyllabic elements34. These modules, each of which
could run autonomously by virtue of circuit dynamics within HVC,
may then be linked together in time by the feedback connections
from RA through the thalamus and back to HVC (Supplementary
Fig. 11). The feedback circuitry may act to detect the end of one
syllable and rapidly and bilaterally initiate the next gap and syllable,
thus simultaneously continuing the song sequence and resynchronizing the two HVCs. In this case, cooling HVC would slow the production of a song syllable, because such production is generated by
dynamics within HVC, and would also delay the onset of the next
syllable, because this onset is linked to the termination of the previous
syllable by feedback circuitry.
A module of chain-like activity in HVC may produce a song syllable as follows. During singing, RA neurons generate a complex but
highly stereotyped sequence of spike bursts14,44. We have previously
suggested that these RA bursts are driven rapidly and on a momentto-moment basis by bursting inputs from HVC37,38. In this view,
because RA burst patterns simply follow the timing set by HVC, we
would expect RA cooling to have a minimal effect on the timing of RA
activity, whereas slowing the chain in HVC would necessarily slow
the sequence of bursts in RA. Furthermore, if structures downstream
of RA (brainstem motor neurons and syringeal muscles) respond
rapidly to descending drive from RA, perhaps on the timescale of
the fastest song modulations (10–20 ms)38,we would expect that
slowing the sequence of bursts in RA might slow the song yet have
a minimal effect on song acoustic structure. Thus, a simple model of
chain-like dynamics in HVC that drives a fast response in RA and
downstream structures is consistent with the electrophysiological
data and the HVC and RA cooling experiments.
Here we have used local manipulation of brain temperature to
identify components within the avian song system that control the
timing of a complex behavioural sequence. This approach may be
broadly useful in localizing specialized brain circuits that control the
timing of other behaviours. We have also used temperature changes
to test ideas about the contribution to song production of oscillatory
dynamics in the song control pathway, an approach that should be
generally applicable to localizing the biophysical origin of oscillatory
and other forms of brain dynamics, and for studying their role in
brain function.
thermocouples were secured with dental acrylic and the bird was placed in a cage
and allowed to awaken. At each Peltier current level, three minutes were allowed
for the brain to reach a steady-state temperature before measurements were
taken from all thermocouple locations.
Received 22 May; accepted 23 September 2008.
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
METHODS SUMMARY
Subjects. Subjects were adult zebra finches (.120 days post-hatch) obtained
either from our colony or from an outside distributor (Preferred Birds). All
animal procedures were approved by the committee on animal care at the
Massachusetts Institute of Technology.
Cooling devices. We used a small (0.7-g) custom-built thermoelectric device
based on the Peltier effect to cool HVC and RA. The HVC cooling device was
constructed from two 1 mm 3 2 mm gold cooling elements that bilaterally contacted the surface of the dura overlying the left and right HVCs. The temperature
change in HVC was spatially restricted (Fig. 1c), producing a maximal change of
only 0.5 uC in RA, the nearest brain region known to be involved in song production in the zebra finch5,45. The current could be switched to flow bilaterally or
unilaterally through only the left or only the right cooling element. The RA
cooling devices were equipped with a gold spike implanted into RA to facilitate
heat transfer.
RA electrophysiology. A craniotomy was made over RA under isoflurane anaesthesia (1.5%). The borders of RA were identified electrophysiologically and the
cooling device implanted at the centre. Single neurons were isolated using carbon
fibre electrodes (Carbostar-1, Kation Scientific), with a signal-to-noise ratio of
greater than 10:1.
Temperature measurements. Temperatures were measured using small thermocouples (5SRTC-TT-K-40-36, Omega). For HVC calibration, three thermocouples were inserted in one hemisphere, under anaesthesia, at respective depths
of 0.5, 2.0 and 4.0 mm beneath the gold pad. In some birds, an additional probe
was placed in the contralateral HVC (at a depth of 0.5 mm). Once inserted, the
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
Marder, E. & Bucher, D. Understanding circuit dynamics using the stomatogastric
nervous system of lobsters and crabs. Annu. Rev. Physiol. 69, 291–316 (2007).
Stent, G. S. et al. Neuronal generation of the leech swimming movement. Science
200, 1348–1357 (1978).
Williams, H. Birdsong and singing behavior. Ann. NY Acad. Sci. 1016, 1–30 (2004).
Konishi, M. Birdsong: from behavior to neuron. Annu. Rev. Neurosci. 8, 125–170
(1985).
Nottebohm, F., Stokes, T. M. & Leonard, C. M. Central control of song in the
canary, Serinus canarius. J. Comp. Neurol. 165, 457–486 (1976).
Nottebohm, F., Kelley, D. B. & Paton, J. A. Connections of vocal control nuclei in
the canary telencephalon. J. Comp. Neurol. 207, 344–357 (1982).
Zann, R. A. The Zebra Finch: A Synthesis of Field and Laboratory Studies 196–246
(Oxford Univ. Press, 1996).
Glaze, C. M. & Troyer, T. W. Temporal structure in zebra finch song: implications
for motor coding. J. Neurosci. 26, 991–1005 (2006).
Cooper, B. G. & Goller, F. Physiological insights into the social-context-dependent
changes in the rhythm of the song motor program. J. Neurophysiol. 95,
3798–3809 (2006).
Sossinka, R. & Böhner, J. Song types in the zebra finch (Poephila guttata
castanotis). Z. Tierpsychol. 53, 123–132 (1980).
Vicario, D. S. Organization of the zebra finch song control system: II. Functional
organization of outputs from nucleus Robustus archistriatalis. J. Comp. Neurol.
309, 486–494 (1991).
Wild, J. M. Descending projections of the songbird nucleus robustus
archistriatalis. J. Comp. Neurol. 338, 225–241 (1993).
Vu, E. T., Mazurek, M. E. & Kuo, Y. C. Identification of a forebrain motor
programming network for the learned song of zebra finches. J. Neurosci. 14,
6924–6934 (1994).
Yu, A. C. & Margoliash, D. Temporal hierarchical control of singing in birds.
Science 273, 1871–1875 (1996).
Solis, M. M. & Perkel, D. J. Rhythmic activity in a forebrain vocal control nucleus in
vitro. J. Neurosci. 25, 2811–2822 (2005).
Chi, Z. & Margoliash, D. Temporal precision and temporal drift in brain and
behavior of zebra finch song. Neuron 32, 899–910 (2001).
Laje, R. & Mindlin, G. B. Diversity within a birdsong. Phys. Rev. Lett. 89, 288102
(2002).
Striedter, G. F. & Vu, E. T. Bilateral feedback projections to the forebrain in the
premotor network for singing in zebra finches. J. Neurobiol. 34, 27–40 (1998).
Ashmore, R. C., Renk, J. A. & Schmidt, M. F. Bottom-up activation of the vocal
motor forebrain by the respiratory brainstem. J. Neurosci. 28, 2613–2623 (2008).
Goller, F. & Cooper, B. G. Peripheral motor dynamics of song production in the
zebra finch. Ann. NY Acad. Sci. 1016, 130–152 (2004).
Ashmore, R. C., Wild, J. M. & Schmidt, M. F. Brainstem and forebrain contributions
to the generation of learned motor behaviors for song. J. Neurosci. 25, 8543–8554
(2005).
Thompson, S. M., Masukawa, L. M. & Prince, D. A. Temperature dependence of
intrinsic membrane properties and synaptic potentials in hippocampal CA1
neurons in vitro. J. Neurosci. 5, 817–824 (1985).
Volgushev, M., Vidyasagar, T. R., Chistiakova, M. & Eysel, U. T. Synaptic
transmission in the neocortex during reversible cooling. Neuroscience 98, 9–22
(2000).
Volgushev, M., Vidyasagar, T. R., Chistiakova, M., Yousef, T. & Eysel, U. T.
Membrane properties and spike generation in rat visual cortical cells during
reversible cooling. J. Physiol. (Lond.) 522, 59–76 (2000).
Arbas, E. A. & Calabrese, R. L. Rate modification in the heartbeat central pattern
generator of the medicinal leech. J. Comp. Physiol. A 155, 783–794 (1984).
Katz, P. S., Sakurai, A., Clemens, S. & Davis, D. Cycle period of a network oscillator
is independent of membrane potential and spiking activity in individual central
pattern generator neurons. J. Neurophysiol. 92, 1904–1917 (2004).
Ferster, D., Chung, S. & Wheat, H. Orientation selectivity of thalamic input to
simple cells of cat visual cortex. Nature 380, 249–252 (1996).
Herrmann, K. & Arnold, A. P. The development of afferent projections to the
robust archistriatal nucleus in male zebra finches: a quantitative electron
microscopic study. J. Neurosci. 11, 2063–2074 (1991).
Spiro, J. E., Dalva, M. B. & Mooney, R. Long-range inhibition within the zebra finch
song nucleus RA can coordinate the firing of multiple projection neurons. J.
Neurophysiol. 81, 3007–3020 (1999).
Roberts, T. F., Klein, M. E., Kubke, M. F., Wild, J. M. & Mooney, R. Telencephalic
neurons monosynaptically link brainstem and forebrain premotor networks
necessary for song. J. Neurosci. 28, 3479–3489 (2008).
Mooney, R. Synaptic basis for developmental plasticity in a birdsong nucleus. J.
Neurosci. 12, 2464–2477 (1992).
Janata, P. & Margoliash, D. Gradual emergence of song selectivity in sensorimotor
structures of the male zebra finch song system. J. Neurosci. 19, 5108–5118 (1999).
193
©2008 Macmillan Publishers Limited. All rights reserved
ARTICLES
NATURE | Vol 456 | 13 November 2008
33. Hahnloser, R. H., Kozhevnikov, A. A. & Fee, M. S. Sleep-related neural activity in a
premotor and a basal-ganglia pathway of the songbird. J. Neurophysiol. 96,
794–812 (2006).
34. Schmidt, M. F. Pattern of interhemispheric synchronization in HVc during singing
correlates with key transitions in the song pattern. J. Neurophysiol. 90, 3931–3949
(2003).
35. Schmidt, M. F., Ashmore, R. C. & Vu, E. T. Bilateral control and interhemispheric
coordination in the avian song motor system. Ann. NY Acad. Sci. 1016, 171–186
(2004).
36. Williams, H., Crane, L. A., Hale, T. K., Esposito, M. A. & Nottebohm, F. Rightside dominance for song control in the zebra finch. J. Neurobiol. 23, 1006–1020
(1992).
37. Hahnloser, R. H., Kozhevnikov, A. A. & Fee, M. S. An ultra-sparse code underlies
the generation of neural sequences in a songbird. Nature 419, 65–70 (2002).
38. Fee, M. S., Kozhevnikov, A. A. & Hahnloser, R. H. Neural mechanisms of vocal
sequence generation in the songbird. Ann. NY Acad. Sci. 1016, 153–170 (2004).
39. James, W. Psychology: The Briefer Course 1–17 (Dover, 2001).
40. Li, M. & Greenside, H. Stable propagation of a burst through a one-dimensional
homogeneous excitatory chain model of songbird nucleus HVC. Phys. Rev. E 74,
011918 (2006).
41. Jin, D. Z., Ramazanoglu, F. M. & Seung, H. S. Intrinsic bursting enhances the
robustness of a neural network model of sequence generation by avian brain area
HVC. J. Comput. Neurosci. 23, 283–299 (2007).
42. Glaze, C. M. & Troyer, T. W. Behavioral measurements of a temporally precise
motor code for birdsong. J. Neurosci. 27, 7631–7639 (2007).
43. Abeles, M. Coriconics: Neural Circuits of the Cerebral Cortex 208–226 (Cambridge
Univ. Press, 1991).
44. Leonardo, A. & Fee, M. S. Ensemble coding of vocal control in birdsong. J. Neurosci.
25, 652–661 (2005).
45. Cardin, J. A., Raksin, J. N. & Schmidt, M. F. Sensorimotor nucleus NIf is necessary
for auditory processing but not vocal motor output in the avian song system. J.
Neurophysiol. 93, 2157–2166 (2005).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements We thank D. Aronov, T. Gardner, J. Goldberg, L. Las,
B. Ölveczky, S. Seung and M. Wilson for their comments on earlier versions of this
manuscript. This work is supported by funding from the US National Institutes of
Health to M.S.F. (MH067105) and to M.A.L. (DC009280) as well as funding from
the Human Frontiers Science Project.
Author Contributions M.A.L. and M.S.F. both contributed to all aspects of this
work.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. Correspondence and requests for materials should be
addressed to M.S.F. ([email protected]).
194
©2008 Macmillan Publishers Limited. All rights reserved
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
METHODS
Peltier device for cooling HVC. The HVC cooling device (Supplementary Fig 1) was constructed with
bilaterally placed cooling elements. At the tip of each cooling element was soldered a 1mm x 2mm gold pad
to make thermal contact with the dura directly above HVC. The gold pads on each element were tilted at a 15
degree angle to match the curvature of the brain, improving the uniformity of thermal transfer. Each element
was constructed using a p-doped and an n-doped semiconductor block, obtained from a commercially
available thermoelectric element (Melcor, CP1.0-127-05L-1-W6). The cooling elements were constructed on
a ceramic base plate (separation of 4.5 mm) so as to fit bilaterally over both HVCs. The two elements were
wired in series electrically for bilateral cooling, but the wires were placed such that the left or right elements
could easily be disconnected from the circuit for unilateral cooling. Current to the cooling device was
provided by a programmable current source (Kepco, Model # ATE 15-15M) via fine braided copper wire
allowing freedom of movement for the bird (Cooner Wire, Inc. #CZ-1187).
The series resistance of the cooling elements (0.115 ) generates approximately 0.5 W of waste heat
at our maximum Peltier current (2A) that rapidly heats the cooling device if not removed. We constructed a
water-cooled heat sink using copper sheet shaped into a pouch and soldered together with inlet and outlet
tubes. The heat sink was attached to the back of the ceramic base plate with thermally conductive epoxy
(Melcor, EG7658). During the cooling experiments, water was flowed through the heat sink using smalldiameter silicone tubing (A-M Systems, #806400) at a rate of 15 mL/min. The source of water to the heat sink
was immersed in a water bath maintained at 47 ºC such that the temperature of the cooling device, in the
absence of electrical current flow, was held near normal avian body temperature (~40 ºC).
Because RA is not located near the brain surface, we penetrated RA with a gold spike to facilitate
thermal transfer. We used the same Peltier cooling elements described for the HVC cooling device, but at the
www.nature.com/nature
1
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
tip of each cooling element was soldered a 330μm gold wire, 3mm in length and sharpened to facilitate
insertion into RA (Supplementary Fig 7a). The optimal access angle for the RA probe had a 10 degree medial
tilt, requiring us to implant the probes separately. Thus, the cooling element for each RA was constructed on a
separate ceramic base, with separate water-cooled heat sinks. A thin layer of a bio-inert silicone elastomer
(Kwik-Cast, WPI) was placed around the proximal part of the gold spike for thermal insulation, improving
heat transfer out of RA (Supplementary Fig 7a). The last 1.0-1.2 mm of the gold spike was left uninsulated.
Before implantation, Peltier devices were treated with EndoZime AW (Ruhof Corp) and 80% ethanol.
Surgical procedure. During the implantation of the Peltier device, birds were anesthetized with 1.5-2.0%
isoflurane. For the HVC device, craniotomies matching the dimensions of the gold pads (plus 100 m on
each axis) were prepared. The device was then fitted into place and secured with dental acrylic. For RA
cooling, the center of that nucleus was first localized electrophysiologically using a carbon fiber electrode
(Kation Scientific). The RA cooling probe was then lowered slowly (100 m/min) at a 10º angle from
normal into the center of RA to a depth of 3 mm, and then retracted to a final depth of 2.5 mm.
Behavioral recordings. Subjects were adult male zebra finches selected for singing prolificacy. After
surgery, birds were housed in a ventilated custom-built sound chamber. Songs were recorded using a lapel
microphone (Audio Technica, AT803B) and an audio amplifier (M-Audio, DMP3). Songs were recorded
before and after surgical implantation of HVC and RA cooling devices. Before recording the song, electrical
current was applied across the Peltier device for at least three minutes to establish a stable temperature
change. A female bird was then presented to elicit directed song. The male was allowed to sing for up to
three minutes before stopping the current flow to the Peltier device. Birds were then given 5-10 minutes to
equilibrate before further testing. The Peltier current was monitored by measuring the voltage drop across a 1
Ohm resistor in series with the Peltier elements. Both the acoustic microphone signal and the Peltier current
were sampled (at 40 kHz) using a National Instruments acquisition board and custom MATLAB software (A.
Andalman).
www.nature.com/nature
2
SUPPLEMENTARY INFORMATION
doi: 10.1038/nature07448
RA temperature calibration and electrophysiology. For the RA electrophysiology experiments, the firing
properties of RA neurons were measured at different Peltier currents. In order to relate the observed
physiological effects to brain temperature, we calibrated the temperature change as a function of Peltier
current and distance from the probe, under anesthesia. Temperature measurements were made with a small
thermocouple (Omega, 5SRTC-TT-K-40-36) implanted at various distances from the probe tip. Furthermore,
we estimated the distance of every recorded RA neurons from the probe in order to estimate the local
temperature change at the site of each neuron.
The distance of the tip of the recording electrode (or thermocouple, for temperature measurements)
from the edge of the cooling probe was estimated as follows: The coordinates of the recording electrode,
relative to a fixed fiducial point on the gold probe near the insertion point into the brain, were determined
using the digital readout of the micromanipulator (MyNeuroLab.com, Benchmark Digital). The distance of
the recording electrode from the edge of the probe was estimated trigonometrically, taking into account the
20º angle between the RA cooling probe and the electrode axis. In some cases the distance estimates were
confirmed by taking an x-ray (Rad-icon, Shad-o-Snap) of the bird with the implanted cooling probe with a
.005 tungsten wire placed at the site of the recorded cell (Fig 4a).
For the temperature calibration, thermocouple measurements were made at several Peltier current
levels (0.5A to -2A) with the thermocouple placed within RA at a fixed distance from the probe (0.2 mm).
The temperature change exhibited the expected quadratic dependence on Peltier current (Supplementary Fig
8a). In addition, measurements were made at a fixed current (1A) but at several distances from the probe
(0.2mm to 1.0mm, Supplementary Fig 8b). Temperature change as a function of distance was well fit by an
exponential (Supplementary Fig 8c). Temperature change as a function of distance (D) and current (I) was
modeled as a separable function in these two variables: T = e
(D D0 )
(a + bI + cI ). The parameters , a, b,
2
and c were fit to the temperature changes measured at various distances and currents.
www.nature.com/nature
3
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Temperature calibration in RA was also carried out in awake birds (n=3). A thermocouple was
implanted 0.2mm from the cooling probe tip, under isofluorane anesthesia. Once inserted, the thermocouple
was secured with dental acrylic and the bird was placed in a cage and allowed to awaken. At each Peltier
current level, three minutes were allowed for the brain to reach a steady-state temperature before
measurements were taken. Although the baseline brain temperature was somewhat lower in the anesthetized
bird, the temperature changes measured in RA as a function of Peltier current were not significantly different
(1A cooling, -5.1±0.4ºC anesthetized, n=3 ; -6.0±1.5ºC awake, n=3).
The spontaneous spike rate of RA neurons was calculated as the inverse of the median value of all
ISIs greater than 20 ms for each neuron in each condition. Bursts were defined as events in which one or
more consecutive interspike intervals (ISIs) were less than 10 ms (i.e. 100 Hz threshold).
Localization of RA cooling probes. The RA cooling probe was placed in the center of RA as defined by
electrophysiological mapping of the RA borders. In a subset of implanted birds, we also histologically
confirmed the placement of the RA cooling probe. Following intercardial perfusion with paraformaldehyde
(Sigma), sections were cut on a vibrating microtome (Vibratome), stained with NeuN (conjugated to
AlexaFluor 488, Chemicon), and visualized using an Axioplan-2 florescence microscope (Carl Zeiss, Inc). In
both birds examined, the cooling probes penetrated the center of RA in both hemispheres.
Analysis of anticorrelation of unlateral stretch. We used two methods to analyze the extent to which the
stretch nonuniformities during left-only and right-only cooling were anticorrelated (Fig 5k). First, we
calculated the principal components of the distribution of these stretches. The first principal component was
found to be nearly parallel to the anticorrelated (-1,1) direction (solid line shown in Fig 5k). The ratio of the
variance along the two principal components was found to be 2.2. We also calculated the ratio of variance in
the anticorrelated (-1,1) direction to that in the correlated (1,1) direction. This ratio was found to be 2.19. We
tested the statistical significance of this anticorrelation using a Monte-Carlo test on shuffled matchings of the
left-right pairs. For each set of randomly shuffled pairs, we computed the variance in the correlated direction
www.nature.com/nature
4
SUPPLEMENTARY INFORMATION
doi: 10.1038/nature07448
and in the anticorrelated direction. By repeating the shuffling process 100,000 times, we determined the
distribution of the ratio of variances in these directions. This analysis showed that the observed ratio, 2.19,
was exceeded in only 2.6% of the shuffled samples. Thus, the stretch exhibited during cooling of left versus
right hemispheres showed a significant anti-correlation (p=0.026).
Stretch analysis I: motif duration. Motif onsets and offset times were determined using a threshold
crossing of song power, in which the threshold was held constant for all temperature conditions within each
bird. From the onset and offset times, the motif duration d m (T ) was calculated as a function of HVC
temperature change, T. The fractional change in motif duration (referred to as ‘dilation’, Fig 1e) was
0
0
calculated as d m (T ) d m / d m , where d m0 is the pre-surgical motif duration. The dilation was calculated
relative to pre-surgical motif duration to make explicit the difference between the control (zero current)
condition and the presurgical condition. We found that motif dilation varied quite linearly with brain
temperature changes (Fig 1d). Thus,
d m (T ) d m0
= mT + const ,
d m0
[1]
where m is the linear stretch (units of inverse temperature, C 1 ). Linear motif stretch, m , was extracted
for each bird by fitting a line to the motif dilation as a function of temperature change T over the range -6.5ºC
to 0ºC. The stretch values plotted in Fig 1g are multiplied by 100 to give units of %/ºC. The offset constant
from the linear fit in Eq 1 was 3.82±0.95%, consistent with an HVC temperature in the control (zero current)
condition that was 1.35±0.33ºC below pre-surgical HVC temperature.
Stretch analysis II: syllable onset interval and motif onset interval. The analysis of syllable onset
intervals and motif onset intervals shown in Fig 2d,e and Fig 2f,g respectively was carried out using an
approach similar to that used to analyze motif durations. Syllable onsets were determined using a threshold
crossing of song power. Syllable onset interval was calculated as the time difference (in seconds) between the
www.nature.com/nature
5
SUPPLEMENTARY INFORMATION
doi: 10.1038/nature07448
onset of each syllable and the onset of the following syllable (n=43 syllables in 9 birds). Likewise, motif
onset interval was defined as the time difference between the onset of the first and second motifs in the song
bout. Only a subset of birds sang bouts of multiple motif renditions at all temperatures tested (n=7 birds). The
syllable onset and motif onset intervals were measured from the entirety of our dataset, which included a
average of 22 motifs per temperature condition for each bird (median=23).
The temperature-dependent dilation of syllable onset intervals and motif onset intervals was also
found to vary linearly with brain temperature changes over the range 0 to -6.5ºC (Fig 2d,f). Thus, syllable and
motif interval stretches, I and M , were extracted for each syllable of each bird by fitting a line to the
dilation of interval durations as a function of temperature change T over the range -6.5ºC to 0ºC. The stretch
values plotted in Fig 2 were multiplied by 100 to give units of %/ºC.
Stretch analysis III: subsyllabic structure. Our principal aim in analyzing the temperature-dependent
dilation of subsyllabic song structure was to estimate a linear stretching factor for each syllable. However,
rather than just measuring the onset and offset of each syllable, we employed a different approach in which all
of the sybsyllabic structure is used to estimate the syllable stretch (Figs 2b,c, Fig 3f). We used a crosscorrelation method in which the fine structure of control and cooled motifs could be directly compared. The
cross-correlation matrix (Supplementary Fig 5a) was calculated as the normalized outer product of an array of
standard sound features extracted from the song motifs, including gravity center, amplitude, frequency
modulation, amplitude modulation, entropy, pitch goodness, pitch, pitch choose, and pitch weight (Sound
Analysis Pro, O. Tchernichovski et al). We also included the acoustic power in seven Gaussian bands of
width 660 Hz at center frequencies from 1.1 kHz through 7.7 kHz in 1.1 kHz increments. Each feature was
demeaned and normalized by its largest value. The 16 acoustic features of the control song were each
resampled at regularly spaced times
ti = it ,
www.nature.com/nature
[2a]
6
SUPPLEMENTARY INFORMATION
doi: 10.1038/nature07448
within the motif, and features of the cooled song were resampled at times
t j = jt ,
[2b]
where i=1,2,…1000 and j=1,2,…1000. Sampling interval ( t ) was chosen such that the vector of features for
the control song was of length 1000 samples and spanned the duration ( d m (0) ) of the control motif (not the
presurgery motif): t = d m (0) /1000 . Similarly the sampling interval ( t ) of the cooled motif was chosen
such that the vector of features for the cooled song had 1000 samples and spanned the duration ( d m (T ) ) of
the motif recorded at a temperature change T: t = d m (T )/ 1000 . At each temperature condition, the motif
duration was estimated as the average duration of several sample motifs. The number of sample motifs used
varied between 3 and 60 (average =14), depending on how many complete motifs were obtained that were
free of contaminating female calls.
Using the array of feature vectors of the control song ( R (i, q ) ) and the cooled song ( R( j , q ) , where
the index q spans the 16 song features), the time-time cross-correlation C (i, j ) of the vectors was computed
using the MATLAB xcorr function, producing a square matrix of size [1000, 1000] (Supplementary Fig 5a).
Pairwise cross-correlations of all cooled motifs (up to 6.5ºC) against control motifs were performed. Using a
[
]
dynamic time warping algorithm, an optimal central path Iˆ(k ), Jˆ (k ) through the correlation matrix was
determined, where k is the index for the path steps (D. Ellis (2003), Dynamic Time Warp (DTW) in Matlab;
web resource, available at http://www.ee.columbia.edu/~dpwe/resources/matlab/dtw/). For each sample
motif within a cooling condition, deviations of the central path from the matrix diagonal were calculated in
the j-direction: J Iˆ( k ) = Jˆ (k ) Iˆ( k ) , to produce an estimate of the sampled ‘time point’ in the cooled
( )
song, J (i ) , that best matches any sampled ‘time point’, i, in the control song. Mean values of these
deviations were calculated over all of the motif samples within each temperature condition. Supplementary
Fig 5b shows the average path deviations for all five temperature conditions in one bird.
www.nature.com/nature
7
SUPPLEMENTARY INFORMATION
doi: 10.1038/nature07448
The slope of the path deviation at any point in the song is a measure of the local time dilation of the
resampled cooled song relative to the resampled control song. We can see from Supplementary Fig 5b, the
path deviations were small, and do not show a clear trend with temperature, indicating that subsyllabic
structure was stretched uniformly with motif duration. To quantify this, we calculated the average slope of the
path deviation function during each syllable by fitting a line to the values of J (i ) between the beginning
and end of the syllable ( asn < i bsn where asn = tsonn / t and bsn = tsoffn / t are the resampled onset and
offset times of the syllable under control conditions, and the index n refers to the values for the nth syllable).
Because the cooled and control motifs are resampled to the same length, the slope of the path deviation
function ( sn ) gives an estimate of the dilation of that syllable relative to the dilation of the song motif (see
Appendix). We found that syllable dilation was approximately linear with temperature (see Fig 2b), thus we
have expressed the temperature dependence of syllable stretch as a linear coefficient sn , where:
sn sn T ,
[3]
We then extracted the linear coefficient sn by fitting a line to the measured values of the slope of the path
deviation, s n , as a function of temperature change. Because s n is the difference between the linear stretch
of the nth syllable and the linear stretch of the motif, s n would be zero for all syllables if the song were
stretched perfectly uniformly. Thus, we refer to s n as the ‘deviation from uniformity’ or ‘nonuniformity’ of
the syllable stretch. We can approximate the linear stretch of the syllable, s n , by adding the measured stretch
of the motif for that bird (see Appendix): s n = s n + m . This was done to generate the stretch and dilation
values shown in Fig 2b,c and Fig 3f. Finally, the stretch values plotted in Fig 2 and 3were multiplied by 100
to give units of %/ºC.
www.nature.com/nature
8
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
As a demonstration of the sensitivity of this technique, we used the above method to analyze data in
which gaps had been artificially stretched (Supplementary Fig 5c) in one bird. After feature vectors were
resampled to 1000 points, we increased the duration of gaps by 0, 25.1, 61.7, 75.4, and 97.9% for the different
current conditions (0, 0.25, 0.5, 0.75, and 1A, respectively). This dilation of the gaps reproduces the observed
dilation of the motif, because the gaps represented only 21.2% of the motif duration in that bird (original
motif stretch values: 0, 5.3, 13.1, 16.0, and 20.8%). In other words, this corresponds to the situation in which
the stretch of the motif is generated entirely by a temperature-dependent stretch of the gaps, with no syllable
stretch. The motifs were again resampled to 1000 points, and our algorithm was used to analyze the mean
residual values. We find that the gaps show a very large increase in slope (positive dilation) at more negative
temperatures. Syllables show a large decrease in slope (negative dilation) at more negative temperatures
(Supplementary Fig 5c). Thus, gaps show a large negative nonuniformity and syllables show a large positive
nonuniformity. This analysis demonstrated that the dynamic time warping approach is effective at identifying
local changes in song stretch when such changes are artificially introduced into the data.
Stretch analysis IV: unilateral cooling. A primary aim for the analysis of the unilateral cooling data was to
measure the non-uniformity of the stretch of different syllables and subsyllabic elements in the motif. For this
analysis, we took the approach of identifying the times of syllable onsets and offsets, and of rapid acoustic
transitions between sybsyllabic elements. Two acoustic features in particular permitted the reliable
identification of acoustic transitions within complex syllables: the first derivative of song power (taken from
1.5 to 7.1 kHz) and amplitude modulation. As described above, features of the control song and cooled song
were resampled at regularly spaced times within the motif, ti = it and t j = jt , respectively, where
i=1,2,…1000 and j=1,2,…1000. Sampling intervals were also chosen as above: t = d m (0) /1000 and
t = d m (T ) /1000 . The resampled feature vectors for the control and cooled song, R(i, q ) and R( j , q ) ,
respectively, were then averaged over several (3-60, mean=14, median=10) clean samples of the song motif
under each condition. Transitions in song structure appeared as sharp features in the average feature vectors,
www.nature.com/nature
9
SUPPLEMENTARY INFORMATION
doi: 10.1038/nature07448
particularly in the amplitude modulation feature, and a threshold was set to determine the onset and offset
times, in units of samples, for each syllable (or subsyllabic element) in both the control and cooled conditions.
From these we calculated the durations Ds n (T ) of each cooled syllable at every temperature condition T (in
units of samples). From these values, we can directly compute the duration ratio of each syllable
sn = Dsn (T ) Dsn (0) , which has a simple relationship to s n , the stretch nonuniformity: sn = 1 + sn T
(see Appendix). Thus, after calculating the values of sn as a function of T, we used a linear fit to extract the
slope s n . The value of s n was calculated separately for bilateral, and for left and right unilateral cooling
(Fig 5).
Stretch analysis V: intervals, syllables and gaps. The time interval ( din ) between the onset of the nth
syllable and the following syllable is, by construction, the sum of the duration of the enclosed syllable ( d sn )
and the duration of the enclosed gap ( d gn ), at all temperature conditions:
d in (T ) = d s n (T )+ d g n (T ).
[4]
Thus, we expect that cooling-related variations in these three quantities may be related. For example, if the
cooling-related stretch of a particular syllable onset interval is larger than average, this could be the result of a
large syllable stretch, or a large gap stretch, or both. Our aim here is to derive a unified measure of the
temperature-related change in the duration of gaps, syllables and intervals that captures the inherent constraint
between these elements.
If we express the temperature dependence of these durations as a linear stretch, for example
(
)
d sn (T ) = d sn (0 ) 1 + sn T , then we can rewrite Eq 4 as:
(
)
(
)
(
)
din (0) 1 + in T = d sn (0) 1 + sn T + d gn (0) 1 + gn T .
www.nature.com/nature
[5]
10
SUPPLEMENTARY INFORMATION
doi: 10.1038/nature07448
Because the control durations satisfy the additive relation, din (0) = d sn (0) + d gn (0) , we can rewrite Eq 5 as
din (0) in = d sn (0) sn + d gn (0) gn . Note that d sn (0) sn is simply the absolute change in duration of the
syllable in units of seconds/ºC. Thus, while the temperature-dependent stretch, , of gaps, syllables and
intervals do not have an additive relation ( in s n + g n ), the absolute stretch measures, d (0) , do.
Further algebraic manipulation of Eq 5 shows that
(
)
(
)
(
)
din (0) in m = d sn (0) sn m + d gn (0) gn m ,
which can also be written din (0) in = d sn (0) sn + d gn (0) gn where g n , s n , and in are the stretch
nonuniformities of gaps, syllables and intervals, respectively. Note that d sn (0) m is the change in syllable
duration expected if the syllable exhibited the same fractional stretch as the song motif. Thus,
since s n = s n m , we see that d sn (0) sn gives the difference between the actual change in syllable
duration compared to that expected for a uniform motif stretch, also in units of seconds/ºC. We refer to the
quantity sn = sn d s0n as the absolute stretch nonuniformity, in contrast to the fractional stretch
nonuniformity, s n .
We see then that the absolute stretch nonuniformity of gaps, syllables, and syllable intervals exhibits
the same additive relation, in = s n + g n , as does the durations of these elements. Thus we have derived a
measure of the temperature-related change in the duration of gaps, syllables and intervals that captures the
inherent constraint between durations of these song elements.
The absolute stretch nonuniformities, g , s , and i , were calculated using an approach similar to
that used for finding the stretch nonuniformity in the unilateral cooling analysis. The acoustic features of
control and cooled song motifs were resampled at 1000 points during the motif. The feature vectors of
multiple example motifs in each condition were averaged and the onsets and offsets of song syllables were
www.nature.com/nature
11
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
easily identified as sharp peaks or valleys in the amplitude modulation feature. A threshold was set to
automatically determine syllable onset and offset times, in units of samples, for both the control and cooled
conditions. From these we calculated the durations of each gap ( Dg n (T )), syllable ( Ds n (T )) and syllable
onset interval ( Din (T )), in each temperature condition, where n is the index over each gap, syllable and
syllable interval triplet. As described for the unilateral cooling analysis, we then found the stretch
nonuniformities g n , s n , and in . Each of these is multiplied by the durations under control condition,
d gn (0) , d sn (0) , and din (0) , respectively, to obtain the absolute stretch nonuniformities g n , s n , and in .
Note that the control durations, in units of time, are easily determined as d sn (0) = Dsn (0)d m (0) /1000 .
SUPPLEMENTARY RESULTS AND DISCUSSION
I. Homeostasis – Does HVC exhibit slow recovery to normal temporal dynamics during cooling, either by
compensatory blood-flow or other biophysical changes? To address this question, in three birds we cooled
HVC for either an hour (n = 2 birds) or 255 minutes (n = 1 additional bird) and measured song speed at
regular intervals during this period. We observed no return of motif duration (Supplementary Fig 1d) during
the cooling period. In addition, we saw no evidence of an “overshoot” once cooling was removed, which
might have resulted from a compensatory mechanism in HVC dynamics during the prolonged cooling period.
From these data, we determine that the Peltier device is capable of generating stable changes in brain
temperature over long periods of time, and that slow homeostatic modification to the circuit does not occur
over the timescale of several hours.
II. Introductory notes and distance calls – In addition to song motifs, zebra finches produce at least two
other vocalizations thought to be HVC dependent. First, song bouts typically begin with a number of
repeated syllable-like vocalizations called introductory notes. Second, male zebra finches produce a learned
www.nature.com/nature
12
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
distance call (also known as a long call). We wondered whether the timing of these vocalizations was also
dependent upon HVC. We measured the duration of introductory notes and distance calls at different HVC
temperatures. In all cases, both introductory notes and distance calls exhibited a significant cooling-related
stretch (Supplementary Fig 3, intro notes: -2.14±0.60 %/ºC, n = 6 birds; calls: -2.30±0.55 %/ºC, n = 4 birds).
On average, the extent of stretching of introductory notes and distance calls was 72.5% and 70.4% of that
seen for singing, respectively. From these data, we can conclude that these non-song vocalizations are
influenced to some extent by dynamics within HVC.
III. Effects of HVC temperature on song acoustic structure. As one measure of changes in song acoustic
structure, we quantified the pitch of harmonic stack syllables or notes. We analyzed 17 such elements across
9 birds at control (average of 16.2 renditions of each stack) and cooled (-6.5ºC, average of 10.6 renditions of
each stack) temperatures. At a cooling of 6.5ºC there was an average 2.7±1.3% (S.D.) decrease in pitch
(Supplementary Fig 4a). Furthermore, there was no relationship between pitch change and the baseline pitch
of the harmonic stack (slope not statistically different from zero, p > 0.3).
We measured sound amplitude by calculating the standard deviation of the mean-subtracted audio
signal, averaged over each song motif for 9 birds. At 6.5ºC cooling, song amplitude decreased by
17.9±12.0% (S.D.) At 5.0ºC heating, song amplitude increased 3.6±16.4% (S.D., Supplementary Fig 4b).
These results suggest that a change in HVC temperature alters the drive nucleus to drive downstream targets.
This could, in principle, result from a decrease (or increase) in the spike rate of the HVC bursts, similar to
what we observe in RA (Supplementary Fig 9c).
IV. Direct analysis of subsyllabic stretch. In order to directly address the effect of temperature changes on
subsyllabic elements, we examined long, complex syllables (n = 9 syllables from 6 birds, average of 3.9
elements per syllable) that could be broken into distinct elements, in a manner similar to that used for the
unilateral cooling analysis (see Stretch Analysis III, also see Fig 5, Supplementary Fig 10). The mean stretch
of subsyllabic elements was -2.75±0.17%/ºC. The standard deviation of the stretch nonuniformity was
www.nature.com/nature
13
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
0.89%/ºC. To assess the significance of stretching, we took the linear regression of the stretch for each
segment versus temperature. We then calculated the probability that this relationship exhibited a slope that
was different from zero. Segments with a p-value lower than 0.05 were defined as significantly stretched.
V. Correlation of syllable-onset interval stretch versus syllable stretch. In the following discussion, we
will use the following notation: for onset interval ‘b-c’, the preceding syllable is ‘a’, the enclosed syllable is
‘b’, and the following syllable is ‘c’. We wanted to determine whether the correlation between the stretch of
syllables and syllable onsets was in some way specific to the enclosed syllable. In other words, is the
correlation between stretch of a syllable-onset intervals and the stretch of the enclosed syllable stronger than
the correlation to the preceding syllable or the following syllable? We found that the correlation between the
stretch of syllable onsets to the preceding syllable was r2 = 0.48. The correlation between the stretch of
syllable onsets to the following syllable was r2 = 0.44. Thus, we find that the correlation was strongest for the
enclosed syllable (r2 = 0.607), suggesting that variations in the stretch of a syllable onset interval are more
closely related to the stretch of the enclosed syllable than to other syllables in the motif. Furthermore, the
slope of this correlation (0.92±0.16 s.e.) is not significantly different from one. Thus, the amount by which a
syllable stretches is highly predictive of the amount by which the associated onset interval stretches, and these
two measures deviate from their mean values by roughly the same amount.
These findings imply that the onset of a syllable is delayed until after the offset of the previous
syllable. We argue that this is consistent with a model in which the onset of a syllable is causally linked to the
offset of the previous syllable. Of course the cooling experiments shown here cannot prove such a causal link,
but they do suggest this possibility.
An analogous argument can be made for motif-onset intervals: The stretch of motif-onset intervals
was well predicted by the duration of the enclosed motif (r2 = 0.97, slope =0.97 ±0.08), consistent with the
view that motif onsets may be causally linked to the termination of the previous motif. Furthermore, the
stretch of motif-onset intervals was correlated with the stretch of enclosed syllable-onsets (Fig 2f, r2 = 0.56),
and the slope of this relation was 0.87 ±0.15 s.e., not significantly different from one.
www.nature.com/nature
14
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
VI. Intervals, Syllables, and Gaps - We examined the effect of HVC temperature change on gap duration,
syllable duration, and syllable onset intervals (n = 8 birds). Gaps, syllables, and syllable onset intervals
stretched by similar amounts (median values of stretch, g , s , and i , were -3.29 %/ºC, -2.93 %/ºC, -3.03
%/ºC, respectively (Supplementary Fig 6a). Note, for self-consistency, that the mean value of syllable onset
interval stretch found here (-3.14±0.16 %/ºC) agrees well with that determined by direct measurement of
syllable onset times in each sample motif (-3.05±0.11 %/ºC, see Stretch Analysis II). Also, the mean value of
syllable stretch measured here (-2.91±0.13 %/ºC) using syllable onsets and offsets, agrees well with that
measured using the Dynamic Time Warping method (-2.88±0.12 %/ºC, see Stretch Analysis III).
We found that gaps stretched significantly more than syllables (p<0.05, paired t-test) and that
syllable-onset intervals stretched significantly more than syllables (p<0.001, paired t-test). Note that the size
of these differences is relatively small. Comparing the medians given in the preceding paragraph, gaps stretch
only 12% more than syllables, and syllable onsets stretch only 3.4% more than syllables. Thus, while these
differences are statistically significant, and give insight into the mechanisms underlying song production, the
stretch of song during HVC cooling is remarkably uniform.
We find that gap stretches were considerably more variable than the stretches of syllables or intervals
(SD = 1.71 %/ºC, 0.72 %/ºC, 0.82 %/ºC, respectively). These differences were also evident in the
distributions of stretch nonuniformity g , s , and i , which represent the difference of the stretch of the
song element relative to the stretch of the entire motif (Supplementary Fig 6b). The larger variability of gap
stretches is not very surprising since these numbers represent fractional variations, and gaps have a much
shorter duration than syllables (roughly 22% of the average syllable duration). Thus, a small absolute
variability in gap duration has five times larger effect on gap stretch than an equal variability in syllable
duration would have on syllable stretch.
www.nature.com/nature
15
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Nevertheless, it is interesting to examine the source of variability in gap durations. The time interval
between the onsets of two adjacent syllables ( d i ) is, by construction, the sum of the duration of the enclosed
syllable ( d s ) and the duration of the enclosed gap ( d g ), at all temperature conditions:
d in (T ) = d s n (T )+ d g n (T ). Thus, we expect that variations in the stretch of these three quantities will be
related. For example, if the cooling-related stretch of a particular syllable onset interval is larger than average,
this could be the result of a large syllable stretch, or a large gap stretch, or both. In other words, variations in
interval stretch may be correlated with variations in syllable stretch, or perhaps correlated with variations in
gap stretch. As we argue in the Supplementary Methods, an appropriate measure with which to examine these
correlations is the absolute stretch nonuniformity of gaps, syllables, and intervals ( g s , and i ,
respectively) because, for any particular syllable, these quantities exhibit an additive constraint:
in = s n + g n . Note that absolute stretch nonuniformity of a syllable, for example, represents the linear
increase in duration of the syllable as a function of temperature (ms/ºC) relative to the increase in duration
expected if the syllable stretched by the same fractional amount as the song motif. In the rest of this section,
we will use the term ‘stretch’ when referring to the quantities g s , and i .
The stretch values g s , and i were calculated for all triplets of gaps, syllables and intervals in our
data set (28 syllables in 8 birds, Supplementary Fig 6). The mean and standard deviation of g s , and i
were (in units of ms/ºC) -0.25±0.59, 0.23±0.44, -0.02±0.60, respectively (Supplementary Fig 6c). We found
that gap stretch was significantly correlated with interval stretch (p<0.0001, slope=0.72, Supplementary Fig
6f), but not correlated with syllable stretch (p=0.08, Supplementary Fig 6d). This suggests that if the cooling
related stretch of a gap was particularly large, it was due to a larger-than-expected stretch of the syllable onset
interval rather than a smaller-than-expected stretch of the syllable. One implication of this result is that the
large variability in gap stretch is not simply due to a systematic (i.e. temperature-dependent) error in
www.nature.com/nature
16
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
determining the offset of certain syllables, which would produce a negative correlation between gap stretch
and syllable stretch.
We also found that interval stretch was significantly but weakly correlated with syllable stretch
(p<0.04 , slope=0.28, Supplementary Fig 6e), suggesting that a larger-than-expected stretch of a syllable onset
interval was only weakly associated with a larger-than-expected stretch of the enclosed syllable. From the
slopes of these correlations, we estimate that 72% variance in the stretch of syllable onset intervals is
explained by variance in gap stretch, and 28% is explained by variance in syllable stretch. These results are
consistent with the view that gaps exhibit a more variable temperature-dependent stretch, and that this is
related to variations in the onset time of the following syllable.
VII. Interhemispheric synchronization and lateralization of timing control in HVC - Bilateral multiunit
recordings in HVC (Schmidt, 2003) and unilateral single-unit studies in right HVC and RA (Hahnloser et al,
2002; Yu and Margoliash, 1996; Leonardo and Fee, 2005) suggest that, in zebra finches, the song motor
system is active bilaterally at all points in the song. Thus, the question of how the song system maintains
synchrony throughout the song is highly relevant. Schmidt (2003) showed that right and left HVC exhibit
brief episodes of correlated activity prior to the onset of each syllable, and also at some acoustic transitions
within long complex syllables. As a result, he suggested that the two HVCs are bilaterally synchronized by
projections from Uva, prior to each syllable onset and possibly at some subsyllabic transitions.
Our unilateral cooling experiments strongly support the view that the two HVCs are synchronized
multiple times within the motif, and are consistent with the idea that such synchronization occurs prior to each
syllable onset. The HVC cooling experiments address the question of the origin of these synchronization
signals. Since syllable onsets are tightly linked to the offset of the previous syllable, we have suggested that
feedback circuitry may act to detect the end of one syllable and rapidly and bilaterally initiate the next
syllable, thus simultaneously continuing the song sequence and re-synchronizing the two HVCs.
How does the feedback circuitry detect the end of a syllable? Again, an interesting question is raised
by the bilateral nature of HVC. If there are two HVC clocks running, which one determines when the next
www.nature.com/nature
17
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
syllable is triggered? Let us first consider the hypothetical case that the next syllable is triggered by whichever
HVC finishes its syllable sequence first. What would we predict as the outcome of unilateral cooling in this
case? Cooling the right HVC, for example, would slow the right HVC chain, so the left HVC would always
finish first. Thus, the next syllable in the motif would always be triggered by the uncooled left HVC, and we
would not expect to see any cooling-related stretch of syllable onset intervals. Cooling the left HVC would
have the same result. However, this is completely inconsistent with the results of the unilateral cooling
experiment, which show partial stretch for unilateral cooling.
Let us next consider another hypothetical case in which the next syllable in the motif is triggered by
whichever HVC finishes its syllable sequence last. In this case, cooling the right HVC would slow the right
HVC chain, and the right HVC would always finish last. Thus, the next syllable in the motif would always be
triggered by the cooled right HVC, and we would expect to see the same stretch for unilateral cooling as
observed for bilateral cooling. Again, this is inconsistent with the results of unilateral cooling.
Another alternative is that the left and right HVC take turns triggering the next song syllable. For
example, the end of the syllable A sequence in the left HVC could bilaterally trigger the onset of the syllable
B sequence in both HVCs. Then, the end of the syllable B sequence in the right HVC could bilaterally trigger
the onset of the syllable C sequence in both HVCs, and so on, with responsibility for initiating the next
syllable in the sequence alternating between the two hemispheres. In this view, cooling the left HVC would
delay the onset of syllable B and D, etc, while cooling the right HVC would delay the onset of syllable C and
E, etc. , thus producing alternating patterns of stretch. Furthermore, because roughly half of the syllable onset
intervals would be stretched during left HVC cooling and the other half of the syllable onset intervals would
be stretched during right HVC cooling, the cooling-related stretch of the entire motif would be roughly half
that seen for bilateral cooling, similar to what we observe (Fig 5b, Supplementary Fig 10). Clearly the results
of our unilateral cooling experiments do not show such a simple alternating pattern. However, our data show
some aspects of this behavior, suggesting that some form of switching of control of timing between left and
right hemispheres may be occurring.
www.nature.com/nature
18
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Note that in this hypothetical model, both HVCs will be operating and controlling syringeal muscles
throughout the song. In this model, the only thing that alternates between the two sides is the control over
initiating the next syllable in the sequence. This is, in principle, entirely separate from the question of
whether one hemisphere dominates in the production of vocal output. That said, it is interesting that right
cooling consistently resulted in a slightly larger song slowing than left cooling (average ratio of right to left
stretch was 1.67±0.29 S.D.), consistent with reports of right hemispheric dominance in zebra finches
(Williams et al, 1992).
VIII. A possible model of song sequence generation - We have proposed a model (Supplementary Fig 11)
in which intrinsic circuitry within HVC is broken up into multiple syllable-length chains (or modules), each
of which can run autonomously by virtue of circuit dynamics within HVC. These modules may then be linked
together in time by the feedback connections from RA through the thalamus and back to HVC. The feedback
circuitry may act to detect the end of one syllable and rapidly and bilaterally initiate the next module, thus
simultaneously continuing the song sequence and re-synchronizing the two HVCs. Bursts of light or sound
cause the interruption of song selectively at the ends of syllables or at acoustic transitions within complex
multi-note syllables (Cynx, 1990), suggesting that the links betweens these HVC modules (mediated in this
model by feedback connections) are less robust than the links within the modules.
In this model, silent gaps between syllables are explicitly timed by HVC, consistent with the fact that
gaps are stretched by HVC cooling, and the fact that single-unit recordings in HVC and RA show sequential
bursting during gaps similar to that seen during syllables. The association between these hypothetical modules
and acoustic features (gaps, syllables, and notes) is not clear. Brief bursts of bilaterally correlated activity
(revealed by bilateral multi-unit recording in HVC, Schmidt, 2003) occur roughly 45ms before the onset each
syllable Thus, these bilateral feedback interactions may actually be related to the onset of gaps. In other
words, one possibility, shown in Supplementary Fig 11, is that silent gaps may be explicitly encoded in HVC
as the initial part of a syllable module. In general, the relation between the timing of feedback connections
and acoustic structure will need to be elucidated by other experiments, such as single-unit recordings in Uva.
www.nature.com/nature
19
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Cooling in HVC would slow the production of the modules, thereby increasing the duration of song
syllables. This would also stretch the syllable-onset intervals, since each syllable is triggered by the end of the
previous syllable. In this way cooling of HVC would produce a fairly uniform stretch of the entire song. Note
that cooling HVC could also have the effect of reducing the drive to the feedback circuitry, thus producing an
increased and more variable latency to triggering the next syllable. This increased latency would likely be
temperature dependent, and thus could explain the fact that gap and syllable-onset stretches exhibit slightly
more stretch than observed for syllables.
www.nature.com/nature
20
SUPPLEMENTARY INFORMATION
doi: 10.1038/nature07448
APPENDIX
The slope ( s ) of the optimal time-warping path through the correlation matrix, during any segment
of sound, gives a measure of the dilation of that segment. In other words, the slope of the optimal path is
given roughly by the ratio of the duration Ds (in units of samples) of the cooled segment to the duration
( Ds0 = bs as ) of the control segment: s = Ds / Ds0 . Using Equation 2a and 2b, we can express D 0s and
n
Ds n in units of time as Ds0n = d sn (0) / t and Ds n = d s n (T )/ t . Thus we have
sn =
Dsn
0
Ds
n
=
t d sn (T )
.
t d sn (0 )
Using Eq 1 and the definitions of t and t , we find that t t = d m (0) / d m (T ) . If we assume a linear
(
)
model of motif and syllable stretch, then we have d sn (T ) d sn (0 ) = 1 + sn T and
(
1
)
d m (T ) d m (0 ) = (1 + mT ) . Thus we find that sn = 1 + sn T (1 + mT ) . With the Taylor series
1
(
)
expansion (1 + x ) = 1 x + x 2 ... for x << 1, we have: sn 1 + sn m T , where we neglect
terms of order m2 T 2 . These higher order terms are negligible for linear stretches of the magnitude we
measure ( mT 0.2 ). In our calculations, we used the slope ( sn ) of the path deviation function, which
differs from the slope of the optimal path by 1. That is, = sn + 1 . (For a perfectly uniform stretch the
slope of the optimal path is one, while the slope of the path deviation function is zero.) Thus,
sn (
sn
)
m T . If we define sn = sn m then we obtain Eq 3, and show that the method of
measuring the slopes of the path deviation function allows us to extract the linear stretch of the song syllables.
www.nature.com/nature
21
doi: 10.1038/nature07448
Supplementary Fig 1. The HVC cooling device. The Peltier device used to cool HVC is shown from both the bottom view (showing the part of the
device that faces the bird) and the top view. a, The bottom view shows the gold pads (at top) that contact the dura over HVC and the semiconductor
elements of the Peltier device. These are mounted on a ceramic base cut from the commercial cooling device. Also shown is the current path for
bilateral cooling (yellow arrows). b, The top view shows the copper heat sink with the cooling-water inlet and outlet tubes. c, A schematic showing the
placement of the cooling device with relation to HVC. The view is from above the bird’s head. d, Motif dilation measured during an hour of constant
Peltier current (mean values from 2 birds). For further clarification, see Supplementary Results and Discussion, section I.
www.nature.com/nature
22
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Supplementary Fig 2. Changes in HVC temperature affect motif duration. Representative sonograms recorded with HVC heated (0.25A, 0.5A) and
cooled (0.25A steps) in four additional birds.
www.nature.com/nature
23
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Supplementary Fig 3. Introductory notes and distance calls are affected by HVC
temperature. a, Examples of an introductory note and distance call from one bird recorded
at control temperature and at - 8.2ºC. b, Introductory note and distance call stretch plotted
against motif stretch. For further clarification, see Supplementary Results and Discussion,
section II.
www.nature.com/nature
24
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Supplementary Fig 4. Amplitude and pitch are affected by HVC temperature. a, The change in pitch of identified harmonic
stacks (17 stacks, 9 birds) during 1A (-6.5ºC) HVC cooling plotted against the pitch of the element under control conditions.
b, A histogram showing the range of pitch changes under control conditions (relative to individual averages), with a dashed
line showing the mean pitch change during 6.5ºC cooling. c, Average amplitude of the song motif plotted against HVC
temperature (n = 8 birds). d, A histogram showing the range of amplitude changes under control conditions (relative to
individual averages), with a dashed line showing the mean amplitude change during 6.5ºC cooling. For further clarification,
see Supplementary Results and Discussion, section III.
www.nature.com/nature
25
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Supplementary Fig 5. Dynamic time warping. a, A cross-correlation of sound features from single examples of a control
and cooled (-6.5ºC) motif for bird #8. Also shown is the line from the dynamic time warping algorithm (blue) and a diagonal
reference (red). b, A representative motif and the average path deviation function for each temperature condition. c, The
average path deviation function for the same bird, but with gaps artificially stretched to account for all motif stretch. This
corresponds to the situation in which syllables do not show any cooling-related stretch, and all motif stretch is produced by a
temperature-dependent dilation of gaps. For further clarification, see Supplementary Methods.
www.nature.com/nature
26
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Supplementary Fig 6. Syllables and gaps. Distributions of (a) stretch, (b) fractional nonuniformity, and (c) absolute nonuniformity for
syllable onset intervals, syllable durations, and gap durations. The median values for each distribution are shown with dashed lines. Scatterplots
show the relation of absolute nonuniformity between (d) gaps and preceding syllables, (e) syllables and syllable onsets, and (f) gaps and
syllable onset intervals. For further clarification, see Supplementary Results and Discussion, sections V and VI.
www.nature.com/nature
27
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Supplementary Fig 7. The RA cooling device. a, A photograph of the device. b, Histological
sections showing the track of the RA cooling probe through RA. c, Examples of song motifs
before (left) and after (right) implantation of the RA cooling device for six birds. The
highlighted example corresponds to the histology shown in b.
www.nature.com/nature
28
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Supplementary Fig 8. Temperature calibration of the RA cooling probe. a, The
temperature measured in RA at a distance of 0.2 mm from the edge of the gold
cooling probe. b, Temperature change resulting from 1A current through the RA
cooling device as a function of distance from the probe. c, The data in b, replotted
as ln(∆T/∆To) versus distance and fit with a line. All data are averages from 3
birds. For further clarification, see Supplementary Methods.
www.nature.com/nature
29
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Supplementary Fig 9. Examples of songs during RA cooling. a, Representative examples of a song
motif following RA heating (0.5A) and cooling (0.5A steps). b, Average ISI distributions of 19 RA
neurons (from 7 birds) recorded at six different values of RA probe current, showing a monotonic shift
toward longer ISIs at colder temperatures. c, Spike frequency during spontaneous bursts as a function of
temperature change in RA.
www.nature.com/nature
30
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Supplementary Fig 10. Unilateral cooling. Results of unilateral HVC cooling in three more birds. All panels are analogous to those shown in Fig 5.
Birds #3 and #10 were cooled unilaterally on the left and right sides; Bird #4 was cooled only on the left side. For further clarification, see Supplementary
Results and Discussion , section VII.
www.nature.com/nature
31
doi: 10.1038/nature07448
SUPPLEMENTARY INFORMATION
Supplementary Fig 11. One possible model of song sequence generation. In this model, intrinsic circuitry within HVC is broken up into multiple
syllable-length chains (or modules), each of which can run autonomously by virtue of circuit dynamics within HVC. These modules are then be
linked together in time by the feedback connections from RA through the thalamus and back to HVC. For further clarification, see Supplementary
Results and Discussion, section VIII.
www.nature.com/nature
32