perm filename FMX[JC,MUS] blob sn#007333 filedate 1972-07-16 generic text, type T, neo UTF8
%BOTTOM,,,,1
		The Synthesis of Complex Audio Spectra
		   by Means of Frequency Modulation

			   John M. Chowning
		          Stanford University  

One of the monkeys sitting at his typewriter finally produced a  work
by Shakespeare, he was then asked to explain how.


Although frequency modulation is well understood as applied in  radio
transmission,  the  relevant  equations  have not been applied in any
significant way to the generation of spectra where both  the  carrier
and  modulating  frequencies  are in the audio band and the sidebands
are used  directly  to  form  the  spectrum.   Given  an  audio  band
oscillator  capable of producing the proper wave when the sign of its
frequency input becomes negative, it will be shown that  a  startling
variety  of spectra can be produced which are altogether predictable,
and with an extraordinarily simple system.


The   research  described  here  was  done  on  a  Digital  Equipment
Corporation PDP-10 computer  for  which  there  is  a  special  sound
synthesis  program  which  is  designed  to  make  optimum use of the
time-shared  capability  of  the  machine.  Implementation  of   this
research,  however,  will be described for mUSIC V, a sound synthesis
program which is both well documented and generally available (1).

MUSIC  V  is  a  program  which  generates  samples  or  a  numerical
representation  of  a  sound  pressure  wave  according to data which
specify the physical  characteristics  of  the  desired  sound.   The
samples are stored on a memory device such as a digital magnetic tape
or disk as they are computed.  On completion of the  computation  the
samples  are  passed  at  a  fixed rate, the sampling rate (typically
10000 to 30000 samples/sec), to a digital to analog converter,  which
generates   a   sequence  of  voltage  pulses  whose  amplitudes  are
proportional to the samples.  The pulses are smoothed by a filter and
the resulting signal is passed to an audio system.

The  program  is designed such that the computation of the samples is
done by program blocks,  called  unit  generators.   A  typical  unit
generator  is  the  oscillator which has two inputs, an output, and a
function.  The first input specifies the amplitude of the output, the
second  the  frequency of the output, and the function determines the
shape of the output.  The value of an input can be either a  constant
or  the  output  of  another  unit generator, thereby allowing higher
level operations on  waves.   A  collection  of  interconnected  unit
generators  is called an "instrument", which is supplied data through
a set of parameters (P1 to Pn) set  by  the  user.   P1  and  P3  are
reserved  by  the program for begin time and duration, P2 is reserved
for the  "instrument"  number  of  a  set  of  instruments,  and  the
remaining  parameters are assigned their function by the user.  Shown
in Figure 1, is an instrument diagram which consists  of  three  unit
generators,  two  oscillators  and  an  adder.   This  instrument  is
designed to produce a tone which  has  a  "vibrato",  or  a  periodic
variation  of a frequency around some average.  The function for each
oscillator in this case is a sinusoid.  In order to produce a tone of
440  Hz which has a vibrato width of 4 Hz above and below the average
440 at a rate of 5 Hz, the following assignments would be required.

	P4 = 1000 (arbitrary loudness scaling)
	P5 = 440
	P6 = 4
	P7 = 5

Oscillator  1  produces a sinusoidal output whose amplitude is scaled
by P6 to be + and - 4, at  a  frequency  of  5  periods/second.  This
output is then added to the constant 440 in the adder 2, whose output
is the frequency input of the other oscillator  3.   This  oscillator
produces an output wave whose amplitude is scaled by P4 to be + and -
1000 amplitude units and whose frequency varies 4 Hz + and -  440  Hz
at  the  rate  of  5  Hz.   Figure  2 is a graph of the instantaneous
frequency vs. time for this case.

In order to generate this sound, the "instrument"  would  be  encoded
and then the paramenters passed to it with the addition of P1,P2, and
P3 as defined above.   The program would then compute the numbers and
store them for conversion.

"Vibrato" is a special case of frequency modulation and will serve to
define three basic parameters.

In frequency modulation the instantaneous frequency of a carrier wave
is  varied according to a modulating wave such that the rate at which
the carrier varies around its average is  at  the  frequency  of  the
modulating  wave  while  the  amount the carrier varies away from the
average is the deviation which is proportional to  the  amplitude  of
the  modulating  wave.  The  parameters  in  frequency modulation are
referred to as carrier frequency c (average), frequency  deviation  d
(peak   deviation  of  carrier  from  its  average),  and  modulation
frequency m (rate at which the deviation occurs).   The parameters of
Figure 1 would be defined to be then:


	P4 = A = amplitude of carrier wave
	P5 = c = carrier frequency
	P6 = d = frequency deviation
	P7 = m = modulating frequency

As  long  as  the  modulating  frequency is well below the audio band
(Figure 2), the ear has little difficulty tracking the  instantaneous
frequency  of  the  carrier  which  results from the modulation.  For
example, by changing the frequency deviation to 220 Hz the  deviation
will range between 220 and 660, as in Figure 3, which is perceived as
a periodic frequency sweep over an interval of a twelfth at the  rate
of 5 times per second.(2)

If  the  modulating  frequency  were  changed  to  440  Hz, where the
modulating frequency equals the carrier frequency,  the  ear  can  no
longer  track  the  instantaneous change in frequency as a sweep, but
rather perceives a complex  spectrum.   In  order  to  determine  the
spectral  components  it  is necessary to consider in some detail the
theory of frequency modulation.

When  a  sinusoid  is  frequency  modulated  by  a   sinusoid,   side
frequencies  are  produced  above  and below the carrier frequency at
intervals of the modulating frequency. The number of side frequencies
produced is related to the frequency deviation, in such a way that as
the deviation is increased  from  0,  energy  is  "stolen"  from  the
carrier  frequency and distributed between more and side frequencies.
The frequency band which contains  the  side  frequencies  above  the
carrier  is  called  the  upper sideband and that containing the side
frequencies below the carrier  is  called  the  lower  sideband.  The
amplitudes of the carrier and side frequencies in the upper and lower
sidebands are determined by Bessel functions (3) of  the  first  kind
and  nth  order,  Figure 4, the argument to which is the ratio of the
frequency deviation  to  the  modulating  frequency  and  called  the
modulation index.

	modulation index = frequency deviation / modulating frequency
		       I = d/m	


The 0th order Bessel function and some  modulation  index  I,  J0(I),
yields  an  amplitude  scaling coefficient for the carrier frequency,
the 1st order Bessel function, J1(I), the coefficient for  the  first
upper  and  lower  side  frequencies,  the 2nd order Bessel function,
J2(I),  the  coefficient  for  the  second  upper  and   lower   side
frequencies,  etc.  In general, there are frequencies of significance
on either side of the  carrier  frequency  within  a  band  which  is
slightly larger than the frequency deviation .(4)

The equation for a frequency modulated wave of peak amplitude A is

	e = A sin (αt + I sin βt)			Eq. 1

where

	e = the instantaneous amplitude of the modulated carrier
	α = 2π times the carrier frequency c
	β = 2π times the modulating frequency m
	I = the modulation index (d/m).

The trigonometric expansion of the above equation is in a form  which
allows  the  determination of all sideband frequencies and amplitudes
for a carrier and modulating wave which are both sinusoidal (5).

	e = A{J0(I) sin αt
	    + J1(I)[sin(α+β)t - sin(α-β)t]
	    + J2(I)[sin(α+2β)t + sin(α-2β)t]		Eq. 2
	    + J3(I)[sin(α+3β)t - sin(α-3β)t]
	    + J4(I)[sin(α+4β)t + sin(α-4β)t]
	    + . . . . . . . . . . . . . . . . }.

It  should  be  noted  that the lower side frequencies have alternate
signs and therefore inverting phase relationships.

The frequency potential is, then,

	  	   c		   amplitude = J0(I)A
	+(c+m)		-(c-m)	   amplitude = J1(I)A
	+(c+2m)		+(c-2m)	   amplitude = J2(I)A
	+(c+3m)		-(c-3m)	   amplitude = J3(I)A
	+(c+4m)		+(c-4m)	   amplitude = J4(I)A
	  .		  .			           Table 1
	  .		  .
	  .		  .
	+(c+nm)		+(c-nm)	   amplitude = Jn(I)A	
			-

	upper		lower
	sideband	sideband
		 carrier

for  all  cases  where  the  carrier  and  modulating frequencies are
invariant.   The  amplitudes  are   proportional   to   the   scaling
coefficient  of  the  nth order Bessel function at modulaion index i.
The  energy  is  distributed  between  the  carrier  and   the   side
frequencies according to the relation,

	Amp = 2A{J0(I) + J1(I) + .......Jn(I)} .

From  Table  1  and  Figure  4,  it  is now possible to determine the
spectrum  in  the  case  above  where  the  carrier  and   modulating
frequencies  are  both  440 Hz and the frequency deviation is 220 Hz.
The modulation index is

	I = d/m
	  =220/440
	  =.5   .

The carrier and sideband components are then,

		  440 Hz		J0(.5) = .95A
	+(+880) Hz	-(0) Hz		J1(.5) = .24A
	+(+1320) Hz	+(-440) Hz	J2(.5) = .02A
	+(+1760) Hz	-(-880) Hz	j3(.5) = .0A

As mentioned above, the sign for the odd order lower side frequencies
is negative, which means that these components will have an  inverted
phase.  It  can  be  seen in Figure 4 that any higher order side band
component, Jn(.5) where n > 3, will have a Bessel  coefficient  which
is  near  0  and  therefore is of decreasing significance.  Figure 5a
represents the spectrum for this case.

The two lower side frequencies need some explanation.  The  component
at  0  Hz  can  be  ignored  since it is a constant in the wave.  The
component at -440 Hz is reflected into the positive domain at +440 Hz
and since

	sin(-α) = -sin(α)

its  phase  will be inverted and subtract from the carrier component,
Figure 5b.  It is important, therefore, that one consider the sign of
the  Bessel  coefficient and the sign in the equation in plotting the
spectrum since the reflected side frequencies will in some cases  add
and other cases subtract.

By increasing the frequency deviation to 440 Hz, the modulation index
increases to 1, and from Figure 4 it can be seen that  the  amplitude
of  the carrier component will be reduced, while the amplitude of the
1st, 2nd, and 3rd side frequencies will be increased. should be noted
that  in  this  case  the instantaneous frequency of the carrier will
range between 0 Hz and 880  Hz.   If  the  frequency  deviation  were
increased  still  more,  yet  higher  order side frequency components
should appear. However, as can be seen in Figure 6, for a  modulation
index of 4, where

	d = I * m or
	d = 4 * 440
	d = 1760 ,

the  instantaneous  frequency  of the carrier will become less than 0
Hz! "Negative frequency" may be a conceptual  paradox,  but  it  does
have mathematical meaning in this case.

In  order  to  understand the application of "negative frequency", we
will first consider the case above where c = m =  d  =  440  Hz  and,
therefore,  I  =  1.   Figure  7  is  a  graph  of  the instantaneous
frequency.  For Eq. 1

	e = A sin (αt + I sin βt)				(Eq.
1)

we substitute

	e = A sin (αt - I cos βt + I)				Eq. 3

which  includes  the  phase  information and will allow us to see the
exact relationship between the instantaneous frequency, as in  Figure
7, and the modulated wave resulting from Eq. 3. We have for

	αt

the graph in Figure 8, where αt reaches the value 2π at the time t0 =
1/440.  We have for

	- I cos βt + I

the graph in Figure 9, having the period 1/440, since c =  m.  Figure
10 is a graph of the sum of the two or,

	αt - I cos βt + I

where  the  slope  of  the  curve  represents  the  rate of change in
frequency for a duration of 1/440 sec.  When the slope of  the  curve
is  0 at k, there is no change in angle and, therefore, the frequency
is 0.  Figure 11 is the sine of the curve in Figure 10, or

	A sin (αt - I cos βt + I) ,

and represents the instantaneous amplitude of the  modulated  carrier
wave.

If  the  modulation  index is increased at all, the deviation becomes
greater  than  the  carrier,  the  instantaneous  frequency   becomes
negative,  and  the slope in Figure 10 becomes negative.  Such a case
can be seen in the graphs, Figures 12 - 16, where I  =  1.5.  As  the
frequency  becomes  negative  between  k  and l, Figure 12, the slope
becomes negative between k and l, Figure 15, and the angle  DECREASES
with  time causing the mirror-like deformation of the carrier, Figure
16.  Figure 17 is 4 periods of the wave in Figure 16.

A negative slope will occur whenever the modulation index  times  the
modulating frequency is greater than the carrier frequency, or

	Im > c.


In order to realize this  condition  where  the  angle  has  negative
slope,  it  necessary  to make a small change to the algorithm of the
computer oscillator.

In MUSIC V, the oscillator is a block of instructions which  operates
on  a  function  stored  in an array of n elements. A variable, S, is
initialized such that it  indexes  the  first  element  of  a  stored
function,  F.   The  contents  of  F(S) is multiplied by an amplitude
scale factor, A, and the result is stored away.  The  index  is  then
added  to  an  increment,  I,  which is proportional to the frequency
input to the oscillator, where the higher the  frequency  the  larger
the increment.  Another value is then taken from the stored function,
then scaled and stored.  When the cumulative sum of increments stored
in the index reaches a value that is larger than the function length,
FL, a "wrap around" occurs by subtracting from the index the function
length.  The algorithm is, then

	Oi = Ai * F(Si mod FL)
	Si+1 = Si +Ii

where

	Oi = the ith output sample
	Ai = the ith amplitude input
	Ii = the ith increment input
	F = a stored function
	Si = the ith sum of increments
	FL = the length of the function in samples

Normally,  the  frequency  is positive and therefore the increment is
positive, but in the case of interest where the  frequency  deviation
is  greater  than  the  carrier  as  a  result of extreme modulation,
"negative frequencies" and therefore negative increments will  occur.
The  oscillator must be altered such that when the index is less than
0 the array length will be added to  the  index  thereby  allowing  a
"wrap around" in either direction. The algorithm is

	IF Si ≥ 0
		Oi = Ai * F(Si mod FL)
	ELSE
		Oi = Ai * F((Si mod FL) + FL) .(6)

With this alteration we have a rather exceptional oscillator, for  it
is  now  capable  of  reversing  the sampling direction of its stored
function when the frequency  becomes  "negative".   The  reversal  of
direction  is equivalent to DECREASING the angle, thereby meeting the
requirements of the equation. With this form of the oscillator, then,
the higher order sideband components can be used in forming spectra.

The  spectrum which results from an instantaneous change in frequency
as in Figure 6, where the modulation index is 4, is shown  in  Figure
18a.    The   adjusted  spectrum,  with  the  "negative"  frequencies
reflected in the positive domain, is shown in Figure 18b.  It can  be
seen  that  the  1st order side frequencies are very near 0 amplitude
since J1(3.8) is a zero crossing of the function (Figure 4).

Normally, in radio applications, one thinks of the  side  frequencies
as  being symmetrical about the carrier, however, in this case, where
the frequency deviation is greater than the carrier, the  lower  side
frequencies   add   algebraically  to  the  upper  side  frequencies,
eliminating the symmetry and causing all of the components to fall at
or above the carrier except for the component at 0 Hz.




		HARMONIC SPECTRA

The significance of the case above, where the ratio of the carrier to
the  modulating frequency is 1/1, is that it is a member of the class
of ratios of integers, all of whose sideband components are a  subset
of  the  harmonic  series;  for  example,  a  ratio of c/m = 1/2 will
produce sideband components which are all odd numbered partials.  The
algorithm  for  determining  the  partial  numbers  of  the  sideband
components is,

		partial	k = |c + nm| ,			Eq. 3
			       -
		where
			n = 0,1,2,3......j
		and the ratio c/m is a rational number.

Tables 2  and  3  give  the  partial  of  the  carrier  and  sideband
components  in  the  harmonic  series  through four orders for simple
ratios of c/m.  For orders 1 through 4, the left number of each  pair
is  the  partial  in the harmonic series of the upper side frequency,
while the right number is the partial in the series of the lower side
frequency. A "0" means that the component is at 0 Hz.  In cases where
the carrier is  greater  than  the  modulating  frequency  the  lower
sideband  may  span  both the positive and negative frequency domain.
As an example, in Table 3, column 2, where c/m = 3/1, the  first  two
low  order  side  frequencies are not reflected (3 - 1 and 3 -2, from
Eq. 3), the third is at 0, and the fourth is reflected (3 - 4).


			c/m   where c ≤ m

          1/1   1/2   1/3   1/4   1/5   2/3   2/5   3/4   3/5
	_____________________________________________________
 nth  0 |  1  |  1  |  1  |  1  |  1  |  2  |  2  |  3  |  3
order 1 |2   0|3   1|4   2|5   3|6   4|5   1|7   3|7   1|8   2
side  2 |3   1|5   3|7   5|9   7|11  9|8   4|12  8|11  5|13  7    Table 2
freq. 3 |4   2|7   5|10  8|13 11|16 14|11  7|17 13|15  9|18 12
      4 |5   3|9   7|13 11|17 15|21 19|14 10|22 18|19 13|23 17

			c/m   where c > m

          2/1   3/1   4/1   5/1   3/2   4/3   5/2   5/3   5/4
	_____________________________________________________
 nth  0 |  2  |  3  |  4  |  5  |  3  |  4  |  5  |  5  |  5
order 1 |3   1|4   2|5   3|6   4|5   1|7   1|7   3|8   2|9   1
side  2 |4   0|5   1|6   2|7   3|7   1|10  2|9   1|11  1|13  3    Table 3
freq. 3 |5   1|6   0|7   1|8   2|9   3|13  5|11  1|14  4|17  7
      4 |6   2|7   1|8   0|9   1|11  5|16  8|13  3|17  7|21 11

      	   1     2     3     4     5     6     7     8     9
			        Column

Reading  down  a  column  of  the  tables coresponds to an increasing
modulation index which, therefore, indicates the order  of  entry  of
the  side  frequencies.  For example, if the modulation index is in a
range where the 4th order side frequencies  become  significant,  the
frequency  potential  includes  all  lower  order side frequencies as
well.  The term "potential" has special significance.  The amplitudes
of  the  side  frequencies  are  determined  by  a  specific index in
reference to the Bessel functions.  Since  the  functions  are  quasi
periodic  around  0,  there is a possibility within a range of values
for the index, where a side frequency is of "significance", that  the
coefficient will be at or near 0.  As an example, an index that is in
a range between 4 and 6 will produce  components  including  the  2nd
order  side frequencies; howver, a specific index of 5.2 will yield a
coefficient near 0 for that order.

Where a frequency component is redundant within a column,  it  should
be  noted that the redundancy is always between different orders. The
components will not cancel  then,  since  the  coefficients  for  all
orders  are different at a given index. There is one case where there
is cancellation of side frequencies.  Figure 19 is the  spectrum  for
the case where,

	c = 0
	m = 1
	I = 3 .

Because  the side frequencies are symmetrical about 0, each order has
the  same  coefficient  and  magnitude,  and  as  a  result  of   the
alternating  sign  in  the lower sideband, the odd orders add and the
even orders cancel.  The spectrum is composed  of  the  odd  partials
where m is the fundamental.

Other  spectra  composed of the odd partials occur when the ratio c/m
is such that c is odd and m is even ( 1/2, 1/4, 3/4, 3/2, 5/2,  5/4).
Spectra  composed  of  all  the partials occur when the ratio is such
that m = 1 ( 1/1, 2/1, 3/1, 4/1, etc.).  See Tables 2 and 3.



		INHARMONIC SPECTRA

Inharmonic spectra will result from ratios of large integers or  real
numbers.  Although  the  spectrum may be composed of components which
are technically a subset of the harmonic series,  the  spread  and/or
location in the series causes the spectrum to be heard as inharmonic.
Ratios in a form where the c is unity and m is a real number  express
the  frequency  relations  more  directly.  Table 4 gives examples of
large integer and real number ratios.


			c/m

          2/11    11/2    9/11    11/9    1/1.4   1/.7    1/2.1
	_____________________________________________________
 nth  0 |   2   |  11   |   9   |  11   |   1   |   1   |   1   
order 1 |13    9|13    9|20    2|20    2|2.4  .4|1.7  .3|3.1 1.1
side  2 |24   20|15    7|31   13|29    7|3.8 1.8|2.4  .4|5.2 3.2  Table 4
freq. 3 |35   31|17    5|42   24|38   16|5.2 3.2|3.1 1.1|7.3 5.3
      4 |46   42|19    3|53   35|47   25|6.6 4.6|3.8 1.8|9.4 7.4

            1       2       3       4       5       6       7
				 Column

If  m  >  2c  as  in  column  7,  Table  4, all of the side frequency
components will be greater than the carrier frequency.

Figure 20 is the spectrum for the ratio of 1/1.4 of  the  carrier  to
the  modulating  frequency  and  a  modulation index of 5.  The lower
sideband  components  are  interleaved  with  the  upper   components
producing an inharmonic spectrum.

In  summary,  the  ratio  c/m determines the distribution of the side
frequencies in the spectrum and the degree to which they are harmonic
or  inharmonic,  while  the  ratio  d/m,  or  the  modulation  index,
determines the amount of energy of the side frequencies as  an  index
to   the  Bessel  coefficients.   It  should  be  understood  that  a
particular spectrum can be tranposed in  frequency  as  long  as  the
above  ratios  are  unchanged. For example, the shape of the spectrum
defined by the relations

	c = 220
	m = 440
	d = 1540 where
	c/m = .5
	d/m = 3.5 = I ,

is identical to the shape of the spectrum defined by

	c = 440
	m = 880
	d = 3080 where

	c/m = .5
	d/m = 3.5 = I

since  the  ratios  are  the  same.   The  deviation  d, then, is the
dependent variable and should be computed according to

	d = Im



		DYNAMIC SPECTRA

A characteristic of most natural sounds is that the amplitudes of the
frequency  components  of the spectrum are time-dependent or dynamic.
The energy proportions of the components often evolve in  complicated
ways,  in  particular  during  the  attack  and decay portions of the
sound(7). The evolution of the  spectrum  is  in  some  cases  easily
followed  as  with bells.  In other cases the evolution is not easily
followed because it occurs in a very short time period,  but,  it  is
nevertheless  an  important  cue  in  our  perception of timbre. Many
natural sounds  have  characteristic  spectral  evolutions  which  in
addition  to  providing  cues to their timbral uniqueness, are largly
responsible for what we judge to  be  their  "lifelike"  quality.  In
contrast,  it  is  largly  due to a fixed proportion spectrum in some
synthesized sounds that  so  readily  imparts  to  the  listener  the
"electronic cue" and "lifeless" quality.

The  special application of frequency modulation described above, has
an inherent and  desirable  characteristic;  the  complexity  of  the
spectrum  is  directly related to the modulation index. If, then, the
modulation index were made to be a function of time, the evolution of
the  spectrum  could  be  generally  described  by  the  shape of the
function.

In order to specify the modulation index as a function  of  time  and
control  the  attack and decay of the output wave, it is necessary to
alter the instrument, Figure 1, by adding three more unit generators.
In Figure 21, oscillator 1 is used to impose an amplitude envelope on
the modulated wave, oscillator 2 and adder 3 together form a  dynamic
control  of the modulation index, and oscillators 4 and 6 and adder 5
have the same functions as 1,2, and 3 in Figure 1. The parameters for
this instrument with which the user is concerned are:

		(User Parameters)

	P1 = Begin time of instrument
	P2 = Instrument number of a set of instruments
	P3 = Duration of the note
	P4 = Amplitude of output wave
	P5 = Carrier frequency
	P6 = modulating frequency
	P7 = modulation index 1
	P8 = modulation index 2

A special routine is  used  to  convert  the  parametric  information
specified  by  the user into suitable data for the computation of the
samples.  In MUSIC V, this routine is usually supplied by  the  user,
and  in  this case it would perform the following functons, where V =
functionlength / sampling rate to convert frequency into increments:

	P5 = P5 * V
	P6 = P6 * V
	P7 = P7 * P6 , d = Im for deviation (P6 is already scaled  by
	     V)
	P8 = (P8 * P6) - P7 , for deviation, (the relationship of  P7
	     and P8 will be explained below)
	P9 = 1 /P3 * V , or V/P3 , the frequency input to oscillators
1  and  2,  where  the  relation  1  /  note duration, scaled by V to
increments, causes the functions associated with these oscillators to
be  sampled  at  a  rate  such  that  one  period is completed in the
duration P3.

Oscillator 2 and adder 3 are related in such a way  that  P7  becomes
the effective value of the function F2 at x,y = 0,0, while P8 becomes
the effective value of the function at x,y = 0,1.   For  example,  if
Figure 22 represents the stored function F2, and

	P7 = 2
	P8 = 8
	P6 = 100 Hz
	P3 = .6 sec

then the output of the adder 3 would be a deviation  increasing  from
200  to  800 Hz in the first 1/6 sec, 800 to 450 in the next 1/6 sec,
etc. On the other hand, if

	P7 = 8
	P8 = 2
	P6 = 100 Hz

the output of the adder would be a deviation decreasing from  800  to
200  Hz  in the first 1/6 sec, etc.  Having the capability of scaling
the deviation, in direct  or  inverse  proportion  to  the  function,
between  any two values as a function of index 1 and index 2, will be
shown to be very useful in generating a variety of dynamic spectra.

In the follwing section, techniques for simulating three  classes  of
timbres  will  be  defined,  where  the  system will be that shown in
Figure 21 controlled by the user parameters listed above.

In visualizing the effect of sweeping between modulation  indexes,  a
careful   study   of   Figure  23  will  be  helpful.(8)  This  is  a
representation of the orders J(0) to J(15) for indexes 0 to  20,  and
is  a  sufficient  range of orders and indexes for most audio spectra
discussed below. Contour lines A, B, and C are for constant values of
the  Bessel function Jn(I) = .01, .001, and .0001 respectively.  Line
A, then, indicates  which  order  side  frequency  is  just  becoming
significant  for  a  given index.  Line D represents the order of the
function equal to the argument or Jn(I) where I = n.  This  relation,
easily  remembered, indicates that any order side frequencies greater
than the value of the index I decrease rapidly in significance.  Line
E represents the absolute maximum amplitude value for each order, for
0 ≤ I ≤ ∞ .  Lines  F,G,H,I,J,  and  K  show  the  zero  crossings,(0
amplitude),  of  the  functions.  Because the ear is not sensitive to
small  changes  in  amplitude,  this  representation  of  the  Bessel
functions  has  proven  to provide sufficient information for most of
the synthesis presented below.

			Brass-like Sounds

Risset demonstrated in his revealing  analysis  of  brass  sounds(9),
that  there is a fundamental characteristic in this class of timbres;
the  amount  of  energy  in  the  spectrum  is  distributed  over  an
increasing  band,  in  proportion  to  the  increase of intensity.  A
simulation of this  class  of  timbres  would  be  developed  on  the
premises:
	1. The frequencies  in  the  spectrum  are  in  the  harmonic
	   series,
	2. Both odd even numbered harmonics are at some time present,
	3. The higher harmonics should increase in significance with
	   intensity,
	4.  The  rise-time  of  the  amplitude is rapid for a typical
attack	and  "overshoots"  the  steady  state.  Oscillators  1 and 2,
Figure 21, controlling  amplitude  and  modulation  index  (deviation
indirectly),  will  both  use  the  function  shown in Figure 22. The
parameter values for a brass-like sound are: (see user parameters)

	P3 = .6
	P4 = 1000 (arbitraary amplitude scaling)
	P5 = 440 Hz
	P6 = 440 Hz (ratio of c/m = 1/1)
	P7 = 0
	P8 = 5

The   modulation   index  (therefore  deviation)  changes  in  direct
proportion to the amplitude of the  carrier  wave,  with  the  result
being an increase or decrease in significance of the side frequencies
in direct proportion to the amplitude function.  The ratio  c/m  =  1
produces  components  that  are in the harmonic series (see Table 2).
By changing  the  values  and  function  shapes  a  large  number  of
variations can be achieved.  One particularly useful variation is the
addition of a small constant to the  modulating  frequency.   If  the
value  .5  Hz  were  added,  for  example,  the  reflected lower side
frequencies would not fall directly on the  upper  side  frequencies,
producing  a  beat  frequency  or  tremulant of 1 cps. The brass-like
quality is preserved in octave transpositions as long as  the  ratios
are maintained.

			Woodwind-like Sounds

It  is  sometimes  the  case  with woodwinds and organ pipes that the
first frequencies to become prominant with the attack are the  higher
harmonics  followed  by  the  lower harmonics as the amplitude of the
wave settles. This type of spectral  evolution  can  be  achieved  in
several ways. By making the carrier frequency an integral multiple of
the modulating frequency, any partial can be  given  emphasis  during
the  attack.  For example, a ratio of c/m = 3/1 will give emphasis to
the 3rd partial (see Table 2).  The  parameter  values  might  be  as
follows:

	P5 = 900 Hz
	P6 = 300 Hz
	p7 = 0
	P8 = 1 ,

and the amplitude and index function as in Figure 24.  The  perceived
fundamental  frequency  is  the modulating frequency, in this case. A
ratio of c/m = 5/1 will produce bassoon-like  quality  in  the  lower
octaves.  The functions remain as above.

	P5 = 500 Hz
	P6 = 100 Hz
	P7 = 0
	P8 = 1 .

The fundamental is 100 Hz.  Another reed quality can be  produced  by
choosing  a ratio of c/m which is a subset of the odd harmonics.  For
example, the parameters

	P5 = 900 Hz
	P6 = 600 Hz
	P7 = 4
	P8 = 2 ,

will  produce a clarinet-like quality where 300 Hz is the fundamental
(see Table 3, c/m = 3/2), and the index is inversely proportional  to
amplitude function.

In   all   of  the  above  examples,  the  realism  can  be  improved
considerably by making the function controlling the index,  the  same
as  the  amplitude  function only through the attack and steady state
portions.  For the decay portion the index function remains constant.
If  Figure 24 is the shape of the amplitude function, then, Figure 25
would be the shape of the  index  function.   The  evolution  of  the
spectrum  during  the  attack  is  apparently not reversed during the
decay.

			Percussive Sounds

A general characteristic of percussive  sounds,  is  that  the  decay
shape  of  the  amplitude envelope is roughly exponential as shown in
Figure 26.  A simulation of this class of timbres would be  developed
around the following premises:
	1. The spectral components are not nescessarily in the
	   harmonic series,
	2. The evolution of the spectrum is from the complex
	to the simple. Bell-like sounds can be produced by making the
change in the index directly proportional to the amplitude  function.
Figure  26  is  the function for amplitude and index.  The parameters
are set to the following:

	P3 = 15
	P4 = 1000
	P5 = 200 Hz
	P6 = 280 Hz
	P7 = 0
	P8 = 10 .

The  ratio  c/m  = 1/1.4, Table 4.  With the large initial index, the
spectrum is dense and as the amplitude decreases the spectrum becomes
ever  more simple.  At the time the amplitude function reaches 0, the
spectrum is composed of a  sinusoid  at  200  Hz.   By  changing  the
amplitude function to that shown in Figure 27, and with the following
parameter values, a drum-like sound can be produced.

	P3 = .2
	P4 = 1000
	P5 = 200 Hz
	P6 = 280 Hz
	P7 = 0
	P8  = 2 A wood drum-like sound can be produced by maintaining
the previous amplitude function, but modulating the  index  according
to the function shown in Figure 28.  The parameters are:

	P3 = .2
	P4 = 1000
	P5 = 80 Hz
	P6 = 56 Hz
	P7 = 0
	P8 = 25 .

These relations produce a burst of energy through  a  wide  frequency
band  for  a  duration  of  20ms followed by a sinusoid which has the
perceptual effect of  a  resonance.   It  should  be  noted  that  an
additional amplitude modulation occurs when there is a rapid sweep of
the index over a  wide  range.   Because  the  Bessel  functions  are
quasi-periodic  around  0, each of the frequency components undergoes
an asynchronous amplitude modulation as  the  index,  in  this  case,
decreases.


The above examples are intended to give some feeling for the power of
this  technique  of  synthesis, although they by no means exhaust the
potential of this instrument.  With the addition of  five  more  unit
generators  as  shown  in Figure 29, additional control can be gained
over the spectrum.  Oscillator 10 provides an additional carrier wave
whose  modulation  frequency  is  the  same as the other carrier, but
whose index function can be scaled up or down by  the  multiplier  8.
Since the new carrier frequency is independent, it can be set to be a
multiple of the first carrier frequency, thus providing components in
another  region of the spectrum.  The proportion of the two modulated
carriers  is  determined  by  the  multiplier  7,  which  scales  the
amplitude  function  before it is applied to the second carrier.  The
outputs are "mixed" by the adder 11.  With these parameter values:

	P4 = 1000
	P5 = 300
	P6 = 300
	P7 = 1
	P8 = 3
	P10 = .2
	P11 = .5
	P12 = 2100 ,

the  second  carrier  will  add  components  centered  around the 7th
partial ( c/m = 7/1), where the index will range between .5 and  1.5,
and  at  an  amplitude  ratio  of  1/5.   The  effect  is  that of an
additional resonant region added to the spectrum.


			Conclusion

This  means of synthesis provides a very simple temporal control over
a variety of spectra whose component frequencies can have  a  variety
of  relationships.  Because "nature" is doing most of the "work", the
system is far simpler than  additive  or  subtractive  systems  which
produce  similar spectra.  Perhaps the most surprising aspect of this
technique, is that the seemingly limited control  over  the  spectral
components,  a  function  of the Bessel coefficients, proves to be no
limitation at all in most cases.  This  suggests  that,  the  precise
amplitude  curve  for each frequency component in a complex spectrum,
is not nearly as important, perceptually, as the general character of
evolution of the components as a group.

		Foot Notes and References

1. Mathews, M.V., The Technology of Computer Music, MIT Press, 1968.
	(A complete description of computer synthesis and the MUSIC V
program.)

2. At this point  it  would  seem  that  any  useful  application  of
frequency  modulation should account for the logrithmic perception of
frequency since in Figure 3. it can be  seen  that  for  1/2  of  the
modulating  period the instantaneous frequency encompasses the octave
below the carrier while in the other 1/2 it encompasses only a  fifth
above  the carrier; however, for the application described below, the
frequency deviation must be linear.

3. Tables and equations for Bessel functions are available in most of
the standard books of tables.

4. Corrington, Murlan S.,Variation of Bandwidth with Modulation Index
in Frequency Modulation, Selected  Papers  on  Frequency  Modulation,
edited by Klapper, Dover Publications, 1970.

5.  Terman,  Frederick  E.,  Radio Engineering, McGraw Hill, 1947, pp
483-489.

6. The change in the code for the oscillator in MUSIC V follows.
	for 
		290  IF(SUM-XNFUN)288,287,287
		287  SUM=SUM-XNFUN
	substitute
		290  IF(SUM.GE.XNFUN)GO TO 287
		     IF(SUM.LT.0.0)GO TO 289
and
	for
		     GO TO 293
		292  J6=L1+J3-1
	substitute
		     GO TO 293
		287  SUM=SUM-XNFUN
		     GO TO 288
		289  SUM=SUM+XNFUN
		     GO TO 288
		292  J6=L1+J3-1
7. Risset J.C., ?????Study of Trumpet Tones????Bell Labs

8. Corrington, op. cit.

9. Risset, op. cit.