Apple, Inc. v. Motorola, Inc. et al

Filing 92

Declaration of Christine Saunders Haskett filed by Plaintiffs Apple, Inc., NEXT SOFTWARE, INC. re: 90 Motion Requesting Claims Construction (Attachments: # 1 Ex. 1 Moto Infring. Cont. Ex. A, # 2 Ex. 2 '157 patent, # 3 Ex. 3 '179 patent, # 4 Ex. 4 '329 patent, # 5 Ex. 5 '230 file history, # 6 Ex. 6 Oxford dictionary definition, # 7 Ex. 7 '559 file history, # 8 Ex. 8 The OSI Model, # 9 Ex. 9 ISO Standard, # 10 Ex. 10 Japanese file history, # 11 Ex. 11 Japanese prosecution appeal, # 12 Ex. 13 Moto Infring. Cont. Ex. E, # 13 Ex. 14 IEEE Standard, # 14 Ex. 15 '333 patent, # 15 Ex. 16 '721 file history, # 16 Ex. 17 '193 file history, # 17 Ex. 18 Moto Infring. Cont. Ex. F, # 18 Ex. 19 Merriam Webster Dictionary, # 19 Ex. 20 Webster's Dictionary) (Haslam, Robert)

EXHIBIT 3

United States Patent

[11]

[19]

Fette

[45]

[54]

VARIABLE FRAME RATE, FIXED BIT RATE
VOCODING METHOD

[75]

Inventor:

Bruce A. Fette, Mesa, Ariz.

[73]

Assignee:

Motorola, Inc., Schaumburg, Ill.

[21]

Appl. No.: 104,698
Oct. 5, 1987

Filed:

[51]
[52]

Int. Cl.4 ................................................ GI0L 5/00
U.S. Cl......................................... 381/29; 381/51;
364/513.5
Field of Search .................................... 381/29-53;
364/513.5

[56]

References Cited
U.S. PATENT DOCUMENTS
4,701,955 10/1987 Taguchi ................................ 381/51

OTHER PUBLICATIONS
"The Application of a Functional Perceptual Model of
Speech to Variable-Rate LPC System", Viswanathan
et aI., IEEE, A.S.S.P. Conference, 1977, pp. 219-222.
"Vector Quantization in Speech Coding", Makhoul et
aI., Proceedings of the IEEE, vol. 73, No. 11, Nov.
1985, pp. 1551-1588.
"Vector Quantization: A Pattern-Matching Technique
for Speech Coding", Gevsho et aI., Proceedings of the
IEEE, Dec. 1983, pp. 87-105.

A method of operating a vocoder so that a variable
frame rate results while maintaining a constant bit rate is
disclosed. A base frame rate remains constant. However, spectral change is measured to determine an appropriate sub frame rate within the base frame. Only one
of a plurality of separate vector quantization processes
is selected in response to spectral change for performance of a vector quantization operation on LPC spectrum coefficients. Each of the plurality of vector quantization operations utilizes its own codebook that contains a different quantity of reference patterns from that
contained in the other codebooks. Each vector quantization operation produces a reference pattern descriptor code that contains a different number of bits from
that produced by the other vector quantization processes. Vector quantization operations producing
smaller, less spectrally accurate outputs are selected
when more subframes are included with a base frame,
and vector quantization operations producing larger,
more spectrally accurate outputs are selected when
fewer subframes are included within a base frame.
16 Claims, 3 Drawing Sheets

-=;..;;.;.~,

34

EX CIT ATION
QUANTIZA TION

ENERGY
QUANTIZA TION

56

54

ENERGY
EXPANSION

Jul. 25, 1989

ABSTRACT

1--....,...----IsP~m~~Es~~~GE
32

4,852,179

Primary Examiner-Patrick R. Salce
Assistant Examiner-Emanuel Todd Voeltz
Attorney, Agent, or Firm-Frank J. Bogacz
[57]

[22]

[58]

Patent Number:
Date of Patent:

ENERGY
EXPANSION

58

LPC SYNTHESIS
66

51

~

rn
~

t

~

~

i

t""'t--

ZfIG-. 1-

~

~

~
E-<

~

o

~

I-ol

~
TEMPORAL RESOLUTION--"

ZfIG-. :2
30 a

....

--

",18

I
28a/

t

16a. . .

J•

I/O

l

PROCESSOR

!

/22a

J

v

U"

()
)

J

I

I/O

-I

28b/

t

t

t

DSP

MEMORY

PROCESSOR

!

I

,

--

~

/22b

t

t

DSP

MEMORY

Ob

~

l

I-ol

~
VJ

'16b
....Ji;a.
QC

20a/

7
24a

'26a

20b/

24b/

'26b

~
....

~

~

u.s. Patent
32
38,

Jul. 25, 1989

180 SAMPLES
22.5 ms .J
36

40\

..

SPECTRAL CHANGE
MEASUREMENT

1

ENERGY
QUANTIZA TION

1

34'

1

/42

.,I
SWITCH

44
EXCITATION
QUANTIZA TION

1

LPC ANALYSIS:

.

\

4,852,179

Sheet 2 of3

I

--'

~ /46c

VQ

VQ

VQ

460

1

46b
WECTRUM
QUANTIZATION

i

I

FORMAT FOR APPLICATION
/ TO FIXED RATE CHANNEL
50

,

31

FIXED RATE CHANNEL
54 BITS
22.5 ms
54\

..

,

56,

18/

~

..

ENERGY
EXPANSION

52

GET
SUBFRAME
RATE

1

I

1.
SWITCH

60
ENERGY
EXPANSION

L

VQ

I

--'

~

1

VQ

VQ

620

62C

62b

~PECTRUM
EXPANSION
I
Z?IG. 3

W

J
I

I.

LPC SYNTHESIS
180 SAMPLES
22.5 ms

'58

.J

'66
68

51
-

r-

~

o~~
EXCITATION

I~

A

11 1 2 1

ENERGY

~

'

27

V

SUBFRAME
I

I
I

ENERGY

SPECTRUM

.I.

ENERGY

I
I

SPECTRUM

y

17

12

SUBFRAME
I

ENERGY

SPECTRUM

I

I

r - - - - - - - - ----------,r---- - - - . - ----

11

I

2

I
12
I G-. 4 C

5

zr

BASE ~ A
ENERGY ENERGY

EXCITATION

I

,&

T

11121

7

I

4

A
ENERGY

SPECTRUM
A

W

121

8

I
..

I

I
y

8

ENERGY

SPECTRUM

I

8

A
ENERGY
1--"--..

121

SUBFRAME
~- - - - -

I

gJ
1

•

SPECTRUM

~~

8

A
ENERGY
~

I

SPECTRUM

I

SUBFRAME
~

..

121

~

5

SPECTRUM

SUBFRAME
•

,&

T

ENERGY

~

I

------

I

5

SUBFRAME

.I.

----,

~

8

~

SUBFRAME

I

I

I-ol

12

5
_----'-___

T---

I

EXCITATION

~

SUBFRAME

...I...

r

~

y

5

G-. 4 B

~

,,-----

I

11 1 2 1

t""'t--

16

SUBFRAME

EXCITATION

~

A

8

G-. 4.A

~

SPECTRUM

A

r

zrI

~

-------

~ ~~

zrI

rn

SUBFRAME

I

W

8

SPECTRUM

I

,&,

,&

..

l
VJ

~
VJ

121

8

zrIG-. 4.IJ
EXCITATION

SPECTRUM

ENERGY
y

111219

1

zrIG-. 4E:

6

-..Ji;a.
------ -

36

"lI

QC

~
-..

~

~

1

4,852,179

VARIABLE FRAME RATE, FIXED BIT RATE
VOCODING METHOD
BACKGROUND OF THE INVENTION
The present invention relates generally to vocoders
and methods of operating vocoders. For purposes of the
present invention, vocoders receive digitized samples of
an analog voice signal and compress or encode the
samples so that a resulting code characterizes the analog
voice signal. The resulting code may then be applied to
a channel, such as a transmission channel or a storage
device. Such channels typically have a bandwidth
which accommodates the resulting code, but is too low
to accommodate the digitized samples. The resulting
code, characterizes the original analog voice signal so
that it may be decoded or expanded by a vocoder to
produce samples that reproduce the voice signal as
perceptually accurately as possible. The present invention relates to vocoders which seek to achieve optimal
voice quality in the reproduced voice signal for a given
bit rate. Specifically, the present invention relates to
vocoders which utilize a variable frame rate in the compression or encoding operations.
Voice represents a complicated analog signal which
is not easily compressed so that an accurate reproduction will result. For example, vowel sounds require a
relatively long analysis window so that a relatively high
degree of spectral accuracy can be achieved. The relatively high degree of spectral accuracy is required so
that a later synthesized vowel sound will appear to
accurately reproduce the original analog voice signal to
a listener. On the other hand, consonant sounds require
a relatively short analysis window so that a relatively
high degree of temporal resolution may be achieved.
The high degree of temporal resolution is required so
that a later synthesized consonant sound will appear as
an accurate reproduction of the original voice signal to
a listener.
FIG. 1 shows the relationship between spectral accuracy and temporal resolution. Generally speaking, at a
given bit rate a vocoder can achieve a high spectral
accuracy by sacrificing temporal resolution, or can
achieve a high degree of temporal resolution by sacrificing spectral accuracy.
Many conventional vocoders which apply coded
voice to a fixed rate channel do not vary frame rate.
Accordingly, designs of such systems attempt to trade
off temporal resolution, which is needed to achieve
accurate reproduction of consonants, with spectral accurate, which is needed to achieve accurate reproduction of vowels, and vice versa. Consequently, noticeably inaccurate reproductions for both vowels and consonants results. Reproduced consonants become
slightly slurred and vowels do not faithfully reproduce
nasal perceptions and voiced fricative perceptions.
A conventional solution to the problem of noticeably
inaccurate reproductions of vowel and consonant
sounds varies the analysis window, or frame, over
which samples are coded so that short frames are used
for analysis of consonants and long frames are used for
analysis of vowels. However, a cumbersome vocoder
architecture results from conventional implementations
which adapt such variable frame rate vocoding methods
for use with fixed rate channels. Such conventional
implementations typically require elaborate buffering
schemes with feedback systems to maintain a constant
bit rate in spite of the variable frame rate. In some con-

2

ventional systems, the buffering introduces an unacceptable delay.
SUMMARY OF THE INVENTION
5

10

15

20

25

30

Accordingly, it is an object of the present invention
to provide an improved vocoding method which may
be implemented using a wide variety of vocoder architectures and which permits variable frame rates at constant bit rates without the use of significant buffering or
feedback schemes.
Another object of the present invention concerns
providing an improved vocoding method that utilizes a
base frame that occurs at a constant rate but contains a
variable number of sub frames depending upon whether
a voice signal being analyzed resembles a vowel sound
or a consonant sound.
The above and other objects and advantages of the
present invention are carried out in one form by a
method of operating a vocoder to compress voice data
samples. The method first selects a particular one of a
plurality of possible sub frame rates. Subsequently, an
analyzing step operates on at least a portion of the voice
data samples to produce a predictive code which represents the analyzed samples. Additionally, a quantizing
step transforms the predictive code into a quantized
code wherein the bit length of the quantized code is
defined by the selected subframe rate and a bit rate
parameter of a fixed rate channel into which the quantized code is applied.
BRIEF DESCRIPTION OF THE DRAWINGS

A more complete understanding of the present invention may be derived by reference to the detailed de35 scription and claims when considered in connection
with the accompanying drawings, wherein like reference numbers indicate like features throughout the .
drawings, and wherein;
FIG. 1 shows a rate distortion boundary curve;
FIG. 2 shows a vocoder architecture utilized in im40
plementing the present invention;
FIG. 3 shows a flow diagram of the present invention; and
FIGS. 4A through 4F show exemplary definitions of
45 various fields within a base frame of data output by the
vocoder of FIG. 2.
DETAILED DESCRIPTION OF THE
PREFERRED EMBODIMENT
FIG. 1 shows a rate-distortion boundary curve for
applying encoded voice signals to a fixed rate channel.
It represents a monotonic, decreasing, convex curve.
The present invention causes a vocoder (discussed below) to operate at a plurality of points on this rate-dis55 tortion boundary rather than at only one point. However, during anyone base frame, the present invention
operates at only one point. Accordingly, the present
invention selects only one of a plurality of points on this
rate-distortion boundary at which to operate. This se60 lection occurs once during each base frame.
A point 10 shown in FIG. 1 represents a situation
where a voice signal may be described with relatively
high spectral accuracy, but relatively low temporal
resolution. This situation is appropriate for encoding
65 vowel sounds. On the other hand, a point 12 causes a
code generated by the present invention to exhibit a
high temporal resolution, but only a relatively low spectral accuracy. Such a point of operation is appropriate

50

3

4,852,179

for describing certain consonant sounds. Additionally, a
point 14 which resides between points 10 and 12 on the
rate-distortion boundary, describes the sound using a
moderate degree of temporal resolution and a moderate
degree of spectral accuracy.
FIG. 2 shows a structure which the present invention
utilizes in performing the methods described herein.
FIG. 2 shows an analysis vocoder 160 and a synthesis
vocoder 16b. As shown in FIG. 2, each of vocoders 16
may be identical in structure. In fact, vocoder 16b may
represent vocoder 160 at a later point in time if a fixed
rate channel 18 through which an encoded signal is sent
represents a memory channel. Conversely, vocoder 16b
may represent a different vocoder from vocoder 160
which, having knowledge of the coding scheme utilized
by vocoder 160, may decode and synthesize a signal
received from fixed rate channel 18.
V ocoders 160 and 16b each represent computerized
signal processing devices. Thus, vocoder 160 has a processor 20a. Processor 20a couples to a computer bus 22a
which in turn couples to a digital signal processor 24a,
a memory 260 and an input/output (I/O) circuit 28a.
I/O circuit 28a has a first I/O port which connects to a
terminal 30a. This first I/O port transfers analog signals
into and out from vocoder 160. I/O circuit 28a additionally has a second port which couples to fixed rate channel 18. This second port transfers digitally encoded
voice signals into and out from vocoder 160.
Vocoder 16b has a processor 20b which couples to a
computer bus 22b. Computer bus 22b couples to a digital signal processor 24b, a memory 26b and an I/O
circuit 28b. I/O circuit 28b has a first port which couples to a terminal 30b. This first I/O port transfers analog signals into and out from vocoder 16b. Additionally,
I/O circuit 18b has a second port which couples to fixed
rate channel 18. This second port transfers digitally
encoded data into and out from vocoder 16b.
Fixed rate channel 18 may represent a memory or
storage device for which compression of voice data
samples is needed because a quantity of voice data sampies occurring within a given period of time would be
too great for such memory device or storage. Alternatively, fixed rate channel 18 may represent a transmission channel, such as a telephone line, an RF tnlnsmission channel, or the like, which accepts data at a con~
stant rate. The application of data to a telephone line
represents one common use of a fixed rate channel.
In operation, analysis vocoder 160 inputs voice data
from terminal 30a and digitizes such voice data in I/O
circuit 28a. Such digitized samples are generally manipulated and compressed to a point where they may be
applied to fixed rate channel 18. Such manipUlation and
compression occurs primarily through the operation of
processor 20a acting upon the digitized samples with
the use of digital signal processor 24a and memory 260.
Synthesizer vocoder 16b performs an inverse operation. Compressed data samples are applied to I/O circuit 28b from fixed rate channel 18. Processor 20b,
utilizing digital signal processor 24b and memory 26b,
decompresses and expands this compressed data to a
point where I/O circuit 28b outputs a multiplicity of
voice data samples at terminal 30b. After some analog
signal conditioning, the voice data samples output from
synthesizer vocoder 16b represent a reproduction of the
input voice data.
In the preferred embodiment, analysis vocoder 160
takes 180 samples of analog voice data every 22.5 milliseconds (ms), and outputs data to fixed rate channel 18

5

10

15

20

25

30

35

40

45

50

55

60

65

4

at a constant base frame rate of 54 bits every 22.5 milliseconds, or a 2400 bits per second constant rate. Those
skilled in the art will recognize that the vocoder architecture described in FIG. 2 and other similar architectures represent general purpose vocoders which are
extremely flexible and may be operated to perform
many ·different methods. Furthermore, the methods
performed by such vocoders are established to a large
degree by the programming which controls processors
20.
FIG. 3 describes the operation of vocoders 16 in
compressing and expanding voice data. FIG. 3 is divided into an analysis process 31 which is performed
prior to application of data to fixed rate channel 18 by
vocoder 160. (see FIG. 2) and a synthesis process 51
which is applied to data obtained from fixed rate channel18 and is practiced by vocoder 16b. Analysis process
31 starts with the receipt of a multiplicity of voice data
samples. In the preferred embodiment, analysis process
31 receives 180 voice data samples every 22.5 milliseconds, as shown in beginning state 32.
The present invention compresses the voice data
samples into a base frame of data that exhibits a constant
base frame rate. Furthermore, the base frame includes
one or more subframes. The base frame will include
only one subframe when process 31 analyzes a vowel
sound. This situation describes point 10, shown in FIG.
1. In this situation, the vowel sound will be described
with a high degree of spectral accuracy, but a low degree of temporal resolution.
Several subframes, a maximum three or four in the
preferred embodiment, are utilized to describe consonant sounds. For a consonant sound, the sounds described by each of the subframes within a single base
frame exhibit relatively low spectral accuracy. However, due to the quantity of subframes a relatively high
temporal resolution results. Thus, such subframes may
be characterized as being operations at point 12 in FIG.
1.
From beginning state 32, a spectral change measurement task 34 is performed. Vowel sounds tend to exhibit
a relatively constant spectrum over the duration of a
base frame. Consonants tend to exhibit a relatively variable spectrum over the duration of a base frame. Accordingly, vowel sounds may be distinguished from
consonant sounds by measuring spectral changes. Thus,
spectral change measurement process 34 performs a
linear predictive coding (LPC) analysis to obtain coefficients useful in making a measurement of spectral
change.
However, in order to save computer processing time,
the preferred embodiment does not perform a complete
LPC analysis to measure spectral change. Since the
present invention contemplates only a relatively small
number of possible subframe rate choices, only a few
spectrum coefficients which tend to influence the spectrum representation most significantly are generated.
The preferred embodiment generates only four spectrum coefficients in spectral change measurement process 34. Those skilled in the art will recognize that such
coefficients may represent predictor coefficients, or
may represent predictor coefficients transformed into
reflector coefficients or other coefficient representations such as log area ratios or uniform spectral sensitivities.
Furthermore, spectral change measurement process
34 obtains sets of these few LPC spectrum coefficients
for each sub frame occurring within a base frame at the

5

4,852,179

maximum subframe rate. For example, if the maximum
subframe rate is four subframes per base frame, and the
base frame contemplates 180 voice data samples every
22.5 ms, then a first set of spectrum coefficients is obtained for the first 45 of the 180 samples, or the first 5
5.625 ms of the 22.5 ms base frame. A second set of
spectrum coefficients is obtained for the second 45 of
the 180 samples, or for the period occurring between
5.625 and 11.25 ms after the start of the base frame.
Likewise, a third set of spectrum coefficients is obtained 10
for the third 45 of the 180 samples. This corresponds to
the period of time occurring between 11.25 and 16.875
ms after the start of the base frame. Finally, a fourth set
of spectrum coefficients is obtained for the fourth 45 of
the 180 samples, or for the time period between 16.875 15
and 22.5 ms after the start of the base frame. If each set
of spectrum coefficients consists of 4 coefficients, then a
total of 16 coefficients are obtained.
Measurement process 34 next detects the amount of
change occurring between each element of the obtained 20
set of coefficients and corresponding elements of a set of
coefficients obtained from the immediately previous
subframe. Such detection of change may occur through
the use of subtraction operations between corresponding elements of successive sub frames. These change 25
values are then combined into a single number that may
be compared against predetermined thresholds to indicate appropriate subframe rates. Such combining may
be accomplished by adding coefficient changes together.
30
As an example, a resulting combined spectral change
number which is greater than a greatest predetermined
threshold would suggest that a consonant sound is described by the voice data samples contained in the base
frame and that a maximum subframe rate should be 35
utilized throughout the current base frame. Alternatively, if the combined number is less than a least predetermined threshold, then a sound resembling a vowel
sound is described by the voice data samples in the base
frame, and a minimum sub frame rate is appropriate. The 40
minimum subframe rate would typically be one subframe per base frame. If the combined number exhibited
a value between the greatest and the least predetermined thresholds, then a medium subframe rate, such as
two sub frames per base frame, would be appropriate. 45
Alternatively, the spectral change between each subframe may be compared with the threshold, and the
number of changes greater than the threshold used as an
indication of location on the rate distortion board.
The number of subframes per frame represents the 50
subframe rate and is the output from spectral change
measurement process 34. This output increases in value
in response to an increase in spectral change. The output from spectral change measurement process 34 is
used by several other processes within analysis process 55
31.
An LPC analysis process utilizes the output from
spectral change measurement process 34 in performing
a conventional LPC analysis of the voice data samples
obtained from state 32. This subframe rate information 60
tells LPC analysis process 36 the number or portion of
samples from the base frame to analyze and the number
of times such analyses are required for the base frame of
180 data voice samples and the approximate temporal
boundary between similar regions. If spectral change 65
measurement process 34 outputs a value of only one
subframe per base frame, then only one analysis is required and this analysis utilizes all 180 of the voice data

6

samples. However, if spectral change measurement
process 34 outputs a value of 4 subframes per base
frame, then 4 separate independent analyses are needed.
The first of the 4 LPC analyses utilizes the first 45 of the
180 samples, the second analysis utilizes the second 45
of the 180 voice data samples, the third analysis utilizes
the third 45 of the 180 voice data samples and the fourth
utilizes the fourth 45 of the 180 voice data samples.
The output from LPC analysis process 36 is a predictive code which includes a set of numbers describing
excitation, energy, and spectrum for each analysis performed on the base frame of voice data samples. Those
skilled in the art will recognize that a conventional LPC
analysis may compress voice data samples into a smaller
amount of data than are used to describe the voice data
samples. However, additional compression may occur
before such information is applied to fixed rate channel
18.
Thus, an excitation quantization process 38 inputs the
excitation output from LPC analysis process 36. Excitation quantization process 38 provides additional compression of the excitation information. The present invention contemplates the application of conventional
quantization techniques to excitation information 38.
Such techniques may introduce a wide variation in the
amount of compression achieved. For lower rate channels, excitation information may advantageously be
quantized as a single pitch value. Additionally, a predetermined unique code may be established to indicate
whether the excitation is voiced or unvoiced. For
higher channel rates, the excitation may describe a complex waveform having many different frequency components and phase relationships. The precise quantization process utilized depends upon the number of available bits in an excitation field of a frame of data applied
to channel 18 (discussed below). In general, a greater
number of bits available within the excitation portion of
the frame causes in a more accurate reproduction of an
analog voice signal by synthesis process 51. The preferred embodiment of the present invention contemplates utilizing only one excitation for an entire base
frame regardless of the number of subframes contained
therein. Thus, this excitation information is assumed to
remain relatively constant throughout the base frame.
Similarly, an energy quantization process 40 receives
energy information from LPC analysis process 36 for
quantization. Those skilled in the art will recognize that
the energy may also be called gain or amplitude. The
present invention contemplates a conventional energy
quantization process. In the preferred embodiment of
the present invention, a unique energy value is obtained
for each sub frame within the base frame. Thus, energy
quantization process 40 generates a quantity of energy
values which depends upon the number of subframes
contained within the base frame. The quantity of bits
used to described energy values depends upon the size
of fields allocated for energy information in a frame of
data which is to be applied to channel 18.
A spectrum quantization process 42 utilized by the
present invention contemplates the use of vector quantization (VQ) of spectrum information into a quantized
code. VQ represents a method of compressing data
which takes advantage of linear and non-linear redundancies. An "off-line" process (not shown) that occurs
during the design of vocoder 16 (see FIG. 2) populates
various codebooks. This populating process derives a
set of spectrum coefficients, which may be referred to
as reference patterns, code vectors, or templates. The

7

4,852,179

entire set of references patterns within a codebook represents all possible speech spectral sounds. The set is
then arranged in a memory device, such as memory 26
(see FIG. 2), to form a codebook, or library, so that an
efficient search may be performed in real time to detect
which one of all the reference patterns contained in the
codebook represents a nearest matching pattern to a
particular set of spectrum coefficients which may be
produced by the LPC analysis process 36. A pointer, or
reference pattern descriptor code, that uniquely identifies the nearest matching reference pattern is then selected as the quantized value and serves as the output
from spectrum quantization process 42. Thus, the spectrum coefficients are transformed into reference pattern
descriptor codes.
The use of a greater quantity of reference patterns in
a codebook permits closer matches between analyzed
spectrum coefficients from LPC analysis process 36 and
the reference patterns contained within the codebook.
Closer matches are desirable because they permit a
more accurate spectral representation of the spectral
coefficients in the output from spectrum quantization
process 42. However, a greater number of reference
patterns in a given codebook requires a greater number
of bits for description of reference pattern descriptor
codes that uniquely define nearest matching reference
patterns.
An N bit reference pattern descriptor code can
uniquely identify a reference pattern obtained from a
codebook having less than 2N reference patterns. For
example, if a codebook contains 256 reference patterns,
then all possible sounds are characterized as being a set
of only 256 sounds. A listener can detect many more
than 256 sounds. Thus, much spectral inaccuracy is
introduced by using a codebook having only 256 reference patterns. However, the reference pattern descriptor code for such a codebook requires no more than
8-bits to uniquely describe anyone of the 256 reference
patterns.
On the other hand, if a codebook contains 65,536
reference patterns, then a greater likelihood exists that a
given set of spectrum coefficients from LPC analysis
process 36 will find a very close match within the reference patterns. As a result, spectral accuracy is greatly
improved over the use of a codebook containing only
256 samples. On the other hand, a corresponding reference pattern descriptor code for a codebook containing
approximately 65,000 samples requires at least 16 bits
for a unique identification of anyone of the 65,536 reference patterns.
In the present invention, a switching step 44 chooses
only one of a plurality of quantization tasks 46 to operate upon spectrum coefficients generated by LPC analysis process 36. The choice performed by switching
step 46 occurs in response to the output from spectral
change measurement process 34. A VQ task 46a is selected when the output from spectral change measurement process 34 indicates a maximum rate for subframes
within the base frame. VQ task 46a makes a smallest
codebook formed in memory 26 (see FIG. 1) available
for the VQ operation performed thereby. Consequently, a resulting reference pattern descriptor code
output from VQ task 46a may not demonstrate a high
degree of spectral accuracy, but contains only a few
bits. For example, the codebook utilized by VQ task 46a
may contain only 256 reference patterns and VQ task
46a may output an 8-bit reference pattern descriptor
code. As discussed above, when the maximum subframe

5

10

15

20

25

30

35

40

45

50

55

60

65

8

rate is output from spectral change measurement process 34, the base frame of 180 voice data samples is
partitioned into subframes and LPC analyses are performed on each of the subframes independently of the
other subframes. Likewise, VQ task 46a performs a
vector quantization operation for each of the subframes
defined by the indicated subframe rate. Consequently,
VQ task 46a outputs separate reference pattern descriptor codes for each of the subframes occurring within a
base frame at the maximum subfame rate. Each of the
reference pattern descriptor codes describes the spectrum for only one of the subframes.
When switching step 44 selects task 46c, and task 46c
represents a vector quantization task, a largest codebook contained within memory 26 (see FIG. 2) is made
available for a VQ search. A resulting reference pattern
descriptor code output from VQ task 46c demonstrates
a high degree of spectral accuracy, but also requires a
large number of bits. For example, if the codebook
utilized by VQ task 46c contains 65,536 reference patterns, then a resulting reference pattern descriptor code
would contain at least 16 bits.
When switching step 44 selects VQ task 46b, a medium size codebook formed in memory 26 is used for
the VQ operation. Spectral accuracy and bit requirements fall between those outlined above for VQ task
46a and 46c. For example, a codebook containing 4,096
reference patterns may be used, and a reSUlting reference pattern descriptor code might contain 12 bits.
A task 50 formats or establishes a base frame of data
from quantized code output from excitation quantization process 38, energy quantization process 40, and
spectrum quantization process 42 for application to
fixed rate channel 18. FIGS. 4A-4E show examples of
possible predetermined formats that task 50 might utilize in forming the base frame of data. Each of FIGS.
4A-4E describe a base frame containing 54 bits which,
in the preferred embodiment, would be applied to channel 18 over 22.5 milliseconds to achieve a 2400 bits per
second constant bit rate.
FIG. 4A shows an example of a format for a base
frame of data in a single subframe per base frame situation. In this base frame of data, one bit is allocated for a
sink bit, and two bits are allocated to describe the subframe rate. The subframe rate bits apply to all subframes
in the base frame, which is only one sub frame in this
example, but permit the use of up to 4 different subframe rates. In the FIG. 4A example, 27 bits of the base
frame are allocated to describing the excitation information output from excitation quantization process 38,
eight bits are allocated to describing the energy information obtained from energy quantization process 40,
and 16 bits are allocated to describing the spectrum
information obtained from spectrum quantization process 42.
FIG. 4B shows base frame of data which contains
two subframes. In FIG. 4B, one bit has been allocated
for synchronization and two bits have been allocated
for defining subframe rate as described above in connection with FIG. 4A. Seventeen bits are allocated to describing the excitation. This single excitation information field is utilized in connection with each of the two
subframes within the base frame of data. Each of the
two subframes allocate 5 bits to describing energy for
that particular subframe and 12 bits for describing the
spectrum for that particular subframe.
FIG. 4C shows an example of a base frame of data
that contains three subframes. The FIG. 4C base frame

9

4,852,179

allocates 1 bit for synchronization and 2 bits for subframe rate definition as described above in connection
with FIG. 4A. Additionally, the FIG. 4C base frame
allocates 12 bits for excitation, and the 12 excitation bits
apply to all three subframes in this base frame. Each of 5
the 3 subframes allocates 5 bits for an energy information field and 8 bits for a spectrum information field.
FIG. 40 shows a base frame of data that contains
four subframes. FIG. 40 allocates 1 bit to synchronization and 2 bits to subframe rate, as discussed above in 10
connection with FIG. 4A. FIG. 40 allocates 7 bits to
excitation. The excitation field applies to each of the
four subframes contained within this base frame of data.
Additionally, a 4-bit field is allocated as a base energy
field. The base energy field also applies to each of the 15
four subframes within this base frame of data. Each of
the four sub frames contains a 2-bit energy difference
field and an 8-bit spectrum field. The 2-bit energy difference field is intended to describe a change in energy
from the base energy value contained in the base energy 20
field. Alternatively, energy difference fields may describe the change in energy from the energy description
of a previous subframe within the base frame of data.
FIG. 4E represents an alternative embodiment to the
format shown in FIG. 4A. As discussed above, FIG. 4A 25
allocates only 16 bits to a spectrum field. Using VQ, this
l6-bit spectrum field suggests that a l6-bit reference
pattern descriptor code and a codebook having less than
216 reference patterns are to be employed. The conventional LPC analysis process may generate 10 spectrum 30
coefficients for each subframe. If a codebook contains
65,536 reference patterns and each reference pattern
contains ten values, an undesirably large amount of
memory may be required for the codebook. Furthermore, an undesirably large processing capability may be 35
required in order to search through such a codebook for
a nearest matching pattern in real time. Accordingly,
the present invention contemplates the use of an alternative quantization process to VQ process 46c when a
slowest subframe rate is selected. Consequently, the 40
format described by FIG. 4E describes 1 bit allocated to
synchronization and 2 bits allocated to subframe rate as
discussed above in FIG. 4A. However, FIG. 4E allocates only 9 bits to an excitation field and 6 bits to an
energy field. In FIG. 4E, the spectrum field requires 36 45
bits.
Referring back to FIG. 3, process 50 formats data
into various base frame formats, as described in FIGS.
4. The base frame rate remains constant and the bit rate
remains constant. However, such formats contemplate a 50
variation in subframe rates through the selection of base
frame formats. Process 50 applies formatted base frames
of data to fixed rate channel 18 at the completion of
analysis process 31.
Synthesis process 51 receives base frames of data 55
formatted by process 50 and decompresses this data to
synthesize voice data samples. Synthesis process 51 is
performed in vocoder 16b (see FIG. 2). It represents the
inverse of the above-described analysis process.
In a "get subframe rate" process 52, the subframe rate 60
field, described above in connection with FIGS. 4, from
the base frame of data is obtained and output to other
processes within synthesis process 51.
An excitation expansion process 54 represents a conventional process which reverses excitation quantiza- 65
tion process 38 from analysis process 31. This process
operates on data obtained from the excitation field of
the base frame. In the preferred embodiment, one exci-

10

tation field applies to all subframes. Thus, the output
from excitation expansion process 54 applies to all subframes contained in a base frame.
An energy expansion process 56 receives the base
frame of data, obtains the energy fields, and performs a
conventional expansion process to obtain a unique energy value associated with each of the subframes contained within the base frame of data. Accordingly, an
independent energy predictor code is provided at the
output of energy expansion process 56 for each of the
subframes within the base frame of data.
A spectrum expansion process 58 receives the base
frame of data and transforms the code contained within
the spectrum field into a set of LPC coefficients. Since
the present invention contemplates vector quantization,
spectrum expansion process 58 contemplates an inverse
vector quantization operation for expansion of the
quantization caused by spectrum quantization process
42 in analysis process 31. Accordingly, a switching step
60 selects only one of a plurality of expansion tasks 62 to
operate upon the spectrum data obtained from the base
frame.
Switching step 60 selects VQ task 62a when subframe
rate data obtained from the base frame indicates a maximum subframe rate for the base frame of data. VQ task
52 performs an inverse vector quantization operation to
obtain an independent set of predictor code spectrum
coefficients for each of the subframes within the base
frame of data. The inverse vector quantization operation represents a simpler operation than the vector
quantization operation. This task resembles a table
lookup operation utilizing the same codebook as was
used by VQ task 46a in analysis process 31.
Likewise, switch 60 selects VQ task 62b when the
subframe rate data from the base frame of data suggests
a medium subframe rate within the base frame. Additionally, switch 60 selects task 62c for expansion of a
base frame of data that contains only a single subframe.
Although the vector quantization operations utilized
in analysis process 31 and the inverse quantization processes utilized by synthesis process 51 may require a
large amount of memory for the codebooks utilized
therewith, the same memory may be used for both analysis and snythesis processes. Accordingly, when the
same or identical vocoders are utilized to transfer data
through fixed rate channel 18, codebooks need not be
duplicated for synthesis and analysis operations.
An LPC synthesis process 66 obtains predictor coefficient inputs from each of excitation expansion process
54, energy expansion process 56, and spectrum expansion process 58. LPC synthesis process 66 performs a
conventional LPC synthesis operation for each of the
sub frames indicated by process 52. Thus, as discussed
above in connection with LPC analysis process 36, LPC
synthesis process 66 performs one or more synthesis
operations over the base frame of data. The output from
LPC synthesis process 66 represents a multiplicity of
voice data samples. In the preferred embodiment, LPC
synthesis 66 produces 180 samples every 22.5 milliseconds as shown at completion state 68 in FIG. 3.
Accordingly, the present invention utilizes a general
purpose vocoder architecture to implement a variable
frame rate, constant bit rate vocoder without the use of
buffering or feedback schemes to maintain the constant
bit rate. Rather, the constant bit rate is established
through the use of a constant base frame rate. A subframe rate is modulated by spectral change of an analog
voice signal being analyzed. The spectral change identi-

4,852,179
11

12

fies whether a given sound resembles a vowel or a consonant.
The present invention is described above with reference to a particular preferred embodiment which facilitates teaching the present invention. Those skilled in the
art will recognize that many alternative embodiments
also fall within the scope of the present invention. For
example, a wide variation in the number of samples per
frame, length of base frames, and particular formats of
base frames are contemplated within the scope of the
present invention. Likewise, a wide variation may
occur in the particular vocoder structures utilized to
implement the method of the present invention. These
and other modifications to the preferred embodiment
which are obvious to those skilled in the art are intended to be included within the scope of the present
invention.
I claim:
1. A method of operating a vocoder to compress a
multiplicity of voice data samples for application to a
constant rate channel, said method comprising the steps
of:
selecting one of a plurality of subframe rates;
analyzing at lest a portion of the samples to produce
a predictive code representative of the analyzed
samples;
quantizing the predictive code into a quantized code
having a bit length defined by the selected subframe rate and a rate parameter of the constant rate
channel; and
said steps of selecting, analyzing and quantizing being
performed at a bit rate of said constant rate channel.
2. A method as claimed in claim 1 additionally comprising the step ,a base frame of data to be applied to the
constant rate channel at a fixed base frame rate and to
include information which remains constant throughout
the base frame in addition to a quantized code for at
least one subframe.
3. A method as claimed in claim 2 additionally comprising the step of repeating said analyzing and quantizing steps when the selected subframe rate permits more
than one subframe to occur during a base frame.
4. A method as claimed in claim 1 wherein said selecting step comprises the steps of:
measuring spectral change occurring between the
multiplicity of voice data samples and a prior multiplicity of voice data samples; and
modulating sub frame rate in response to the spectral
change.
5. A method as claimed in claim 4 wherein said modulating step comprises the step of increasing subframe
rate in response to increasing spectral change.
6. A method as claimed in claim 1 wherein said quantizing step comprises the step of performing a vector
quantization of the predictive code.
7. A method as claimed in claim 6 wherein said quantizing step comprises the steps of:
making first and second vector quantization codebooks available for searching;
choosing only one of the first and second vector
quantization codebooks in response to said selecting step;
searching within the chosen one from said choosing
step of the first and second vector quantization
codebooks for a nearest matching reference pattern
to the predective code; and

retrieving a reference pattern descriptor code which
uniquely describes the nearest matching reference
pattern.
S. A method as claimed in claim 7 wherein said making step comprises the step of populating the first vector
quantization codebook with a quantity of reference
patterns that differs from the quantity of reference patterns that populate the second vector quantization codebook so that a first reference pattern descriptor code
retrievable from the first codebook for the predective
code is described utilizing a different number of bits
than describes a second reference pattern descriptor
code retrievable from the second codebook for the
predective code.
9. A method as claimed in claim 7 wherein said analyzing step performs a linear predictive coding of the
multiplicity of voice data samples to generate excitation, energy, and spectrum data, and said searching step
comprises the step of transforming the spectrum data
into an N-bit code wherein the chosen codebook contains no more than 2N unique reference patterns.
10. A method of operating a vocoder to decompress
a base frame of data which represents a multiplicity of
voice data samples and which is received from a constant rate channel, said method comprising the steps of:
obtaining information describing one of a plurality of
subframe rates from the constant rate base frame of
data;
separating the base frame of data into a least one
subframe of data in response to the subframe rate
from said obtaining step;
expanding each subframe of data from said separating
step into an independent predictive code;
synthesizing the mUltiplicity of voice data samples
from the predictive code for each subframe; and
said steps of obtaining, separating, expanding and
synthesizing being performed at a bit rate of said
constant rate channel.
11. A method as claimed in claim 10 wherein said
expanding step comprises the step of performing an
inverse vector quantization of each subframe of data.
12. A method as claimed in claim 11 wherein said
expanding step comprises the steps of:
making first and second vector quantization codebooks available for the inverse vector quantization
operations;
choosing only one of the first and second vector
quantization codebooks in response to the subframe
rate from said obtaining step;
operating with the chosen one from said choosing
step of the first and second vector quantization
codebooks to generate reference patterns that
uniquely match the predictive code for each subframe of data.
13. A method as claimed in claim 12 wherein said
making step comprises the step of populating the first
vector quantization codebook with a quantity of reference patterns that differs from the quantity of reference
patterns that populate the second vector quantization
codebook so that a first reference pattern descriptor
code utilized with the first vector quantization codebook is described utilizing a different number of bits
than describes a second reference pattern descriptor
code utilized with the second codebook.
14. A method as claimed in claim 12 wherein the
chosen codebook from said choosing step contains no
more than 2N reference patterns, wherein the predictive
code describes excitation, energy, and spectrum data,

5

10

15

20

25

30

35

40

45

50

55

60

65

13

4,852,179

wherein said synthesizing step performs a linear predictive coding of the predictive code to generate the multiplicity of voice data samples, and said operating step
comprises the step of transforming only an N-bit portion of the subframe of data into the spectrum data
portion of the predictive code.
15. A method of operating a vocoder to compress a
mUltiplicity of voice data samples for application to a
fixed rate channel, said method comprising the steps of:
measuring spectral change occurring between the
multiplicity of voice data samples and a prior multiplicity of voice data samples;
modulating a subframe rate in response to the spectral
change so that increasing spectral change causes
increasing subframe rate;
performing a linear predictive coding of at least a
portion of the multiplicity of voice data samples to
generate excitation, energy and spectrum data;

5

10

15

20

14

making first and second vector quantization codebooks available for searching;
choosing one of the first and second vector quantization codebooks in response to the subframe rate of
said modulating step;
searching within only the one codebook chosen in
said choosing step for a nearest matching reference
pattern to the spectrum data of said performing
step; and
retrieving a reference pattern descriptor code which
uniquely describes the nearest matching reference
pattern, wherein the reference pattern descriptor
code contains N bits and the one codebook chosen
in said choosing step contains less than 2N unique
reference patterns.
16. A method as claimed in claim 15 additionally
comprising the step of repeating said performing,
searching, and retrieving steps when the subframe rate
of said modulating step permits more than one sub frame
to occur during a base frame.

* * ,. ,. *

25

30

35

40

45

50

55

60

65

UNITED STATES PATENT AND TRADEMARK OFFICE

CERTIFICATE OF CORRECTION
PATENT NO.

:

DATED
INVENTOR(S) :

4,852,179
July 25, 1989
Bruce A. Fette

It is certified that error appears in the above-identified patent and that said Letters Patent is hereby
corrected as shown below:

In column 11, claim 2, line 35, after the word "step"
insert --of establishing--.

Signed and Sealed this
Twenty-ninth Day of May, 1990

Attest:
HARRY F. MANBECK, JR.

Attesting Officer

Commissioner of Patents and Trademarks