Académique Documents
Professionnel Documents
Culture Documents
The spectral density offluctuations inthaudio power ofmany musical selections and ofEnglish speech varies approximately as 1/f(f isthe frequency) down toafrequency of5X10 -4Hz. This result implies that the audio-power fluctuations are correlated over alltimes inthe same manner as"l/f noise" in
down tothe inverse ofthe length ofthe piece ofmusic. The frequency fluctuations ofEnglish speech have
electronic components. The frequency fluctuations ofmusic also have a 1/fspectral density atfrequencies
a quite different behavior, with a single characteristic time ofabout 0.1s,the average length ofa syllable. The observations on music suggest that 1/fnoise isa good choice for stochastic composition. Compositions inwhich the frequency and duration ofeach note were determined by1/f noise sources sounded pleasing. Those generated by white-noise sources sounded too random, while those generated by
1/f 2 noise sounded toocorrelated.
PACSnumbers: 43.75.Wx, 43.60.Cg, 43.75.--z,43.70.-h
INTRODUCTION
A second characterization
as 1/f , wheref is the frequencyand0.5 <T <1.5, over manydecades. Thusvacuum tubes, t carbonresistors, ' semiconducting devices, s continuous 4'5or discontinuous 6 metal films, ionic solutions, ? films at the superconducting transition, Josephson junctions, 9 nerve membranes, sunspot activity, ll andthe floodlevels of the river Nilel
V(t) is the autocorrelationfunction, <V(t)V(t+r)>. <V(t)V(t+ r)> is a measure of how the fluctuatingquantities at times ! and t+ are related. For a stationary process<V(t)V(t+ )> is independent of t and depends only on the time difference. Sv(f) and<V(t)V(t + )>
are not independent, but are related by the Wiener-
Khintchine relations 4
havespectral densitiesthat vary as 1/f. The I/f behavior implies some correlation in these fluctuating quantities over all times corresponding to the frequency range
Sv(f) = 4
(2)
by
for whichthe spectral densityis 1/f. The observation of the 1/f spectral densityin music has implicationsfor music compositional procedures. We haveuseda 1/f
noise source in a simple computer algorithm to produce
V(t) is independent (f<< 1/2r,); andis a rapidly decreasing function of frequency, usually1/f ', in thefrequency rangeover whichV(t) is correlated(f >>1/2r,). A
quantity with a 1If spectral densitycannot,therefore, be
characterized by a single correlation time. In fact, the
spectral density Sv(f) of a quantity V(t) fluctuatingwith time t is a measure of the mean squaredvariation
in a unit bandwidth centered on the frequency f. The
average is taken over a time that is long compared with the period; in practice, we usually average over at
least 30 periods. Sv(f) may be measuredby passing V(t) througha tunedfilter of frequencyf andbandwidth 5f. Sv(f) is then the average of the squaredoutputof the filter dividedby 5f. Thus, if V(t) is a voltage,Sv(f)
is in units of volt squared per hertz.
aPresent address: IBM Thomas J. Watson Research Center, Yorktown Heights, NY 10598.
258 J. Acoust. Soc. Am. 63(1), Jan. 1978
changes. The 1If ' noiseis the most correlatedshowing only slow changes. The 1/f noise is intermediate.
258
Downloaded 12 Jul 2013 to 128.237.233.75. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
259
259
(a)
v(t)
WHITE
NOISE
t
t
1/f NOISE
versus
1/f 2 NOISE
t
(b)
II.
tuatingquantity of interest was convertedto a voltage whosespectral densitywas measured by an interfaced PDP-11 computerusing a Fast Fourier Transform algorithm that simulates a bank of filters. The most famil-
iar fluctuating quantity associatedwith music is the audiosignal V(t) suchas the voltageused to drive a
I
-3
I
-Z
I
-I
I
0
I
I
LOGlo(f/1Hz)
FIG. 3. Bach'sFirst Brandenburg Concerto(log scales). (a) Sv.(f)rs f; (b) $ve(f) vsf.
speakersystem. Figure 2(a) shows a plot of the spectral density S,(f) of the audiosignalfrom J. S. Bach's
First Brandenburg Concerto averaged over the entire
concerto. The spectral density consists of a series of sharp peaks in the frequency range 100 Hz to 2 kHz correspondingto the individual notes in the concerto
trum contains much useful information, our primary interest is in more slowly varying quantities.
o IOOO 2o 50
f (Hz)
squared, andfilteredwith a 20-Hzlow-pass filter. This process produced a slowly varying signal,V'(t),that
i
I II I I
I F'ILTER I I OH= I
was proportionalto the "instantaneous" audiopower of the music. Correlations of V'(t) represent correlations
of the audio power of successive notes. The spectral density of the audio power fluctuations of the First BrandenburgConcerto, Sv.(f), averagedover the entire concerto is shownin Fig. 2(b). On this linear-linear plot, the audiopower fluctuationsappear as a peak close
to zero frequency.
i,,
f (Hz)
(a) Spectral density of audio signal,$.(f) vsf; (b) spectral densityof audiopowerfluctuations, Sv2(f)vsf.
J. Acoust. Soc. Am.,Vol.63, No. 1, January 1978
Figure 3 is a log-log plot of the same spectra as in Fig. 2. In Fig. 3(a), the spectral density of the audio signal, S(f), is distributedover the audiorange. In Fig. 3(b), however, the spectral density of the audio
Downloaded 12 Jul 2013 to 128.237.233.75. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
260
260
low i Hz. The peaks between i and 10 Hz are due to the rhythmic structure of the music.
I0-
5
b)
as announcements and commercials. Figures 4(b) through(d) showthe spectral densities of the audiopower fluctuations for three radio stations characterized by
different motifs. Figure 4(b) showsS.(f) for a classical station. The spectral densityexhibitsa smooth1/f dependence. Figure 4(c) showsSw.(f)for a rock station. The spectral densityis 1/f-like above2x 10 '3 Hz,
and flattens for lower frequencies, indicating that the correlation of the audio power fluctuations does not extend over times longer than a single selection, roughly
c) 8
I
I I I I I
d)
100 s. Figure 4(d) showsSw.(f)for a news and talk station, and is representative of Sw.(f)for speech. Once againthe spectral densityis 1/f-like. In Fig. 4(b) andFig. 4(d), Sv.(f) remains 1/f-like downto the low-
-5
-8
-I
LOGo(f/1Hz)
FIG. 5. Spectral density of frequencyfluctuations, Sz(f) vs f for four radio stations (a) classical; (b) jazz andblues; (c) rock; and (d) news and talk.
est frequencymeasured, 5x 10'4 Hz, implying correlations over times of a least 5 min.
nal of low frequency will have few zero crossings per second and a small Z, while a high-frequency signal
will have a high Z. For the case of music, Z(t) roughly follows the melody. Correlations in Z(t) represent correlations in the frequencies of successive notes. Figure 5 shows the spectral density of the rate of zero
crossings, Sz(f), for four radio stations averagedover approximately 12 h. Z(t) was also smoothedby a 20-Hz
low-pass filter before the spectral density was mea-
The spectral densityvaries closely as 1/f above4x 10 '4 Hz. Figures 5(b) and 5(c) showSz(f) for a jazz and
blues station and a rock station.
(d)
a news and talk station, exhibits a quite different spectral density. The spectral density is that of a quantity characterized by two correlation times: The average length of an individual speech sound, roughly 0.1 s, and the average length of time for which a given announcer talks, about 100 s. For most musical selections the
Downloaded 12 Jul 2013 to 128.237.233.75. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
261
261
the other pieces show decreased correlations for times longer than about 10 s.
(b)
=E
(c)
(e)
interest to investigate the music and spee6h of other cultures, such as Chinese, in which pitch plays an important role in communication.
-3 -2_ -I 0
LOGlo(f/1Hz)
(o)
(b)
(c)
(e)
,
-3 -2
I
-I
I
o
Davidovsky [Fig. 6(a)] andStockhausen [Fig. 6(e)] show the correlationscharacteristicof the 1/f spectral density whilethoseby Babbit[Fig. 6(b)], Jolas[Fig. 6(c)], andCarter [Fig. 6(d)] showdecreasing correlationsat
times longer than several seconds. For the frequency
LOG1o(f/1 Hz)
fluctuations (Fig. 7), Davidovsky's Synchronism [Fig. 7(a)] remainsclosestto the 1/f spectraldensity,while
J. Acoust.Soc.Am., Vol. 63, No. 1, January1978
Downloaded 12 Jul 2013 to 128.237.233.75. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
262
262
III.
COMPOSITION
number generator (white noise source) which is uncorrelated in time. In the simplest case the white noise source can be used to determine the frequency and dura-
'
_:_,-
,
't--
........
MUSIC
dee:mned by 1/ nose
processes (in which the probability of a given note dependson its immediate predecessors) were able to impose some local structure but lacked long time correlations. Attempts at increasing the number of preceding notes on which the given note depended gave increasingly repetitious r,esults rather than interesting long term
a standard musical scale (pentatonic, major, or 12 tone chromatic) over a two-octave range: A high number specified a high frequency and vice versa. This process was then repeated with another noise source to produce an independent series of stored numbers whose values corresponded to the durations of successive notes.
The PDP-11 was then used to "perform" the stochastic composition by controlling a single amplitude modulated voltage controlled oscillator. We used sinusoidal, square, and triangle waveforms and a variety of attack and decay rates. The computer was also used to put the stochastic compositions in more conventional form.
The 1/f noise sourceitself has the sametime correlations as we have measured in various types of music. To illustrate this process at an elementary level, we
frequencyandduration. In Fig. 9 a 1/f noisesource was used, while in Fig. 10 a 1/f ' noisesourcewasused.
Although Figs. 8-10 are not intended as complete formal compositions, they are representative of the correlations between successive notes that can be achieved
when the three types of noise sources of Fig. 1 are used to control various musical parameters. In each case the noise sources were "Gaussian" implying that values
near the mean were more likely than extreme values.
Over a period of about two years we have played samples of our music to several hundred people at nine universities and research laboratories. The listeners
ranged from those with little technical knowledge of music to professional musicians and composers. We
WHITE MUSIC
playedselections of white, l/f, and1/f musicvarying in lengthfrom oneto ten minutes. Our 1/f musicwas
judged by most listeners to be far more interesting than
FIG. 8.
source.
Downloaded 12 Jul 2013 to 128.237.233.75. Redistribution subject to ASA license or copyright; see http://asadl.org/terms
263
263
ACKNOWLEDGMENTS
We are pleased to acknowledge helpful conversations with and encouragement from Dr. M. Mathews and A.
Moorer.
Dugger for his advice, comments, and a critical reading of the manuscript. Finally, we are grateful to Dr. Robert Wo Young for his helpful comments during the revision of the manuscript. 1/f 2 MUSIC
1j. B. Johnson, Phys. Rev. 26, 71-85 (1925). 2C.J. Christenson andG. L. Pearson,Bell Syst.Tech. J. 15,
197-223 (1936).
scalelike 1/f ' music (which was "too correlated"). Indeedthe surprising sophistication of the 1/f music
(which was close to being "just right") suggeststhat the
3For a review, seeA. van der Ziel, Noise: Sources,Characterization, Measurement (Prentice-Hall, NJ., 1970).
386--392 (1969).
Englewood Cliffs,
IV.
DISCUSSION
9j. ClarkeandG. Hawkins, Phys.Rev. B14, 2826-2831 (1976). 1A. A. Verveen andH. E. Derkson,Proc. IEEE 56, 906-916 (1968).
composition, the variation of only two parameters (frequency and duration of the notes of a single voice) can,
at best, produce only a very simple form of music. More structure is needed, not all of which can be pro-
12The spectral density of voltage fluctuations hasbeen measured down to 5x10-?Hz in semiconductors [M. A. Caloyannides,
J. Appl. Phys. 45, 307-316 (1974)]; in the case of the river Nile, the spectral density of the annual flood levels extends downto 3x10 'l Hz.
videdby 1/f noise sources. We improvedon these elementary compositions by using two voices that were
ments) coupledwith external constraints (for example, rhythm or the rejection rules of Hiller) as offering
promising possibilities for stochastic composition. A
Downloaded 12 Jul 2013 to 128.237.233.75. Redistribution subject to ASA license or copyright; see http://asadl.org/terms