CA 02598683 2010-05-31
1
DESCRIPTION
A SPEECH ENCODER AND
METHOD OF SPEECH ENCODING
Technical Field
The present invention relates to a speech coder for
efficiently coding speech information and a speech decoder
for efficiently decoding the same. Background Art
A speech coding. technique for efficiently coding
and decoding speech information has been developed in
recent years. In Code Excited Linear Prediction: "High
Quality Speech at Low Bit Rate", M. R. Schroeder, Proc.
ICASSP'85, pp. 937-940, there is described a speech coder
of a CELP type, which is on the basis of such a speech
coding technique.
In this speech coder, a linear prediction for an
input speech is carried out in.every frame, which is
divided=at a fixed time. A prediction residual
(excitation signal) is obtained by the linear prediction
for each frame. Then, the prediction residual is coded
using an adaptive codebook in which a previous excitation
signal is stored and a random codebook in which a
plurality of random code vectors is stored.
FIG. 1 shows a functional block of a conventional
CELP type speech coder.
A speech signal 11 input to the CELP type speech
coder is subjected to a linear prediction analysis in a
linear prediction analyzing section 12. A linear
CA 02598683 2007-09-10
2
predictive coefficients can be obtained by the linear
prediction analysis. The linear predictive
coefficients are parameters indicating an spectrum.
envelop of the speech signal 11. The linear predictive
coefficients obtained in the linear prediction analyzing
section 12 are quantized by a linear predictive
coefficient coding section 13, and the quantized linear
predictive coefficients are sent to a linear predictive
coefficient decoding section 14. Note that an index
obtained by this quantization is output to a code
outputting section 24 as a linear predictive code. The
linear predictive coefficient decoding section 14
decodes the linear predictive coefficients quantized by
the linear predictive coefficient coding section 13 so
as to obtain coefficients of a synthetic filter. The
linear predictive coefficient decoding section 14
outputs these coefficients to a synthetic filter 15.
An adaptive codebook 17 is one, which outputs a
plurality of candidates of adaptive codevectors, and
which comprises a buffer for storing excitation signals
corresponding to previous several frames. The adaptive
codevectors are time series vectors, which express
periodic components in the input speech.
A random codebook 18 is one, which stores a
plurality of candidates of random codevectors. The
random code vectors are time series vectors, which
express non-periodic components in the input speech.
In an adaptive code gain weighting section 19 and
CA 02598683 2007-09-10
3
a random code gain weighting section 20, the candidate
vectors output from the adaptive codebook 17 and the
random codebook 18 are multiplied by an adaptive code
gain read from a weight codebook 21 and a random code
gain, respectively, and the resultants are output to an
adding section 22.
The weighting codebook stores a plurality of
adaptive codebook gains by which the adaptive codevector
is multiplied and a plurality of random codebook gains
by which the random codevectors are multiplied.
The adding section 22 adds the adaptive code vector
candidates and the random code vector candidates, which
are weighted in the adaptive code gain weighting section
19 and the random code gain weighting section 20,
respectively. Then, the adding section 22 generates
excitation vectors so as to be output to the synthetic
filter 15.
The synthetic filter 15 is an all-pole filter. The
coefficients of the synthetic filter are obtained by the
linear predictive coefficient decoding section 14. The
synthetic filter 15 has a function of synthesizing input
excitation vector in order to produce synthetic speech
and outputting that synthetic speech to a distortion
calculator 16.
A distortion calculator 16 calculates a distortion
between the synthetic speech, which is the output of the
synthetic filter 15, and the input speech 11, and outputs
the obtained distortion value to a code index specifying
CA 02598683 2007-09-10
4
section 23. The code index specifying section 23
specifies three kinds of codebook indicies (index of
adaptive codebook, index of random codebook, index of
weight codebook) so as to minimize the distortion
calculated by the distortion calculation section 16.
The three kinds of codebook indicies specified by the
code index specifying section 23 are output to a code
outputting section 24. The code outputting section 24
outputs the index of linear predictive codebook obtained
by the linear predictive coefficient coding section 13
and the index of adaptive codebook, the index of random
code, the index of weight codebook, which have been
specified by the code index specifying section 23, to
a transmission path at one time.
FIG. 2 shows a functional block of a CELP speech
decoder, which decodes the speech signal coded by the
aforementioned coder. In this speech decoder apparatus,
a code input section 31 receives codes sent from the
speech coder (FIG. 1). The received codes are
disassembled into the index of the linear predictive
codebook, the index of adaptive codebook, the index of
random codebook, and the index of weight codebook. Then,
the indicies obtained by the above disassemble are output
to a linear predictive coefficient decoding section 32,
an adaptive codebook 33, a random codebook 34, and a
weight codebook 35, respectively.
Next, the linear predictive coefficient decoding
section 32 decodes the linear predictive code number
CA 02598683 2007-09-10
obtained by the code input section 31 so as to obtain
coefficients of the synthetic filter, and outputs those
coefficients to a synthetic filter 39. Then, an adaptive
codevector corresponding to the index of adaptive
5 codebook is read from adaptive codebook, and a random
codevector corresponding to the index of random codebook
is read from the random codebook. Moreover, an adaptive
codebook gain and a random codebook gain corresponding
to the index of weight codebook are read from the weight
codebook. Then, in an adaptive codevector weighting
section 36, the adaptive codevector is multiplied by the
adaptive codebook gain, and the resultant is sent to an
adding section 38. Similarly, in a random codevector
weighting section 37, the random codevector is
multiplied by the random codebook gain, and the resultant
is sent to the adding section 38.
The adding section 38 adds the above two codevectors
and generates an excitation vector. Then, the generated
excitation vector is sent to the adaptive codebook 33
to update the buffer or the synthetic filter 39 to excite
the filter. The synthetic filter 39, composed with the
linear predictive coeffcients which are output from
linear predictive coefficient decoding section 32, is
excited by the excitation vector obtained by the adding
section 38, and reproduces a synthetic speech.
Note that, in the.distortion calculator 16 of the
CELP speech coder, distortion E is generally calculated
by the following expression (1):
CA 02598683 2007-09-10
6
E = I) v - (gaHP+gcHC) II2 = = = (1)
where v: an input speech signal (vector),
H: an impulse response convolution matrix
for a synthetic filter
h(0) 0 === === 0 0
h(1) h(0) 0 === 0 0
H- h(2) h(1) h(0) 0 0 0
0 0
h(0) 0
h(L-1) . . . . . . . . . h(1) h(0)
wherein h is an impulse response of a synthetic
filter, L is a frame length,
p: an adaptive codevector,
c: a random codevector,
ga: an adaptive codebook gain
gc: a random codebook gain
Here, in order to minimize distortion E of
expression (1), the distortion is calculated by a closed
loop with respective to all combinations of the adaptive
code number, the random code number, the weight code
number, it is necessary to specify each code number.
However, if the closed loop search is performed with
respect to expression (1), an amount of calculation
processing becomes too large. For this reason,
generally, first of all, the index of adaptive codebook
is specified by vector quantization using the adaptive
codebook. Next, the index of random coodbook is
specified by vector quantization using the random
codebook. Finally, the index of weight codebook is
CA 02598683 2007-09-10
7
specified by vector quantization using the weight
codebook. Here, the following will specifically
explain the vector quantization processing using the
random codebook.
In a case where the index of adaptive codebook or
the adaptive codebook gain are previously or temporarily
determined, the expression for evaluating distortion
shown in expression (1) is changed to the fdl lowing
expression (2):
EC a II X-gCHC 112 ... (2)
where vector x in expression (2) is random
excitation target vector for specifying a random code
number which is obtained by the following equation (3)
using the previously or temporarily specified adaptive
codevector and adaptive codebook gain.
x v-gaHP ... (3)
where ga: an adaptive codebook gain,
v: a speech signal (vector),
H: an impulse response convolution matrix for a
synthetic filter,
p: an adaptive codevector.
For specifying the random codebook gain gc after
specifying the index of random codebook, it can be
assumed that gc in the expression (2) can be set to an
arbitrary value. For this reason, it is known that a
quantization processing for specifying the index of the
random codebook minimizing the expression (2) can be
replaced with the determination of the index of the
CA 02598683 2007-09-10
8
random codebook vector maximizing the following
fractional expression (4):
(xtHc)2 (4)
IIHC112
In other words, in a case where the index of
adaptive codebook and the adaptive codebook gain are
previously or temporarily determined, vector
quantization processing for random excitation-becomes
processing for specifying the index of the random
codebook maximizing fractional expression (4)
calculated by the distortion calculator 16.
In the CELP coder/decoder in the early stages, one
that stores kinds of random sequences corresponding to
the number of bits allocated in the memory was used as
a random codebook. However, there was a problem in which
a massive amount of memory capacity was required and the
amount of calculation processing for calculating
distortion of expression (4) with respect to each random
codevector was greatly increased.
As one of methods for solving the above problem,
there is a CELP speech coder/decoder using an algebraic
excitation vector generator for generating an excitation
vector algebraically as described in "8KBIT/S ACELP
CODING OF SPEECH WITH 10 MS SPEECH-FRAME: A CANDIDATE
FOR CCITT STANDARDIZATION*: R. Salami, C. Laflamme, J-P.
Adoul, ICASSP'94, pp.II-97-II-100, 1994.
However, in the above CELP speech coder/decoder
using an algebraic excitation vector generator, random
CA 02598683 2007-09-10
9
excitation (target vector for specifying an index of
random codebook) obtained by equation (3) is
approximately expressed by a few signed pulses. For this
reason, there is a limitation in improvement of speech
quality. This is obvious from an actual investigation
of an element for random excitation x of expression (3)
wherein there are few cases in which random excitations
are composed only of a few signed pulses.
Disclosure of Invention
An object of the present invention is to provide
an excitation vector
generator, which is capable of
generating an excitation vector whose shape has a
statistically high similarity to the shape of a random
excitation obtained by analyzing an input speech signal.
Also, an object of the present invention is to
provide a CELP speech coder/decoder, a speech signal
communication system, a speech 'signal recording system,
which use the above excitation vector generator as a
random codebook so as to obtain a synthetic speech having
a higher quality than that of the case in which an
algebraic excitation vector generator is used as a random
codebook.
A first aspect of the present invention is to provide
an excitation vector generator comprising a pulse vector
generating section having N channels (NZ 1) for generating pulse
vectors each having a signed unit pulse provided to one
element on a vector axis, a storing and selecting section having
a function of storing M (M? 1)kinds of dispersion patterns every
CA 02598683 2007-09-10
channel and a function of selecting a certain kind of dispersion
pattern from M kinds of dispersion patterns stored, a pulse vector
dispersion section having a function of convolving the dispersion
pattern selected from the dispersion pattern storing and selecting
5 section to the signed pulse vector output from the pulse vector
generator, so as to generator N dispersed vectors, and a dispersed
vector adding section having a function of adding N dispersed
vectors generated by the pulse vector dispersion section so as
to generate an excitation vector. The function for
10 algebraically generating (N?1) pulse vectors is provided to
the pulse vector generator, and the dispersion pattern storing
and selecting section stores the dispersion patterns obtained by
pre-training the shape (characteristic) of the actual vector,
whereby making it possible to generate the excitation vector, which
is well similar to the shape of the actual excitation vector as
compared with the conventional algebraic excitation generator.
Moreover, the second aspect of the present invention
is to provide a CELP speech coder/decoder using the above
excitation vector generator as the random codebook, which
is capable of generating the excitation vector being
closer to the actual shape than the case of the
conventional speech coder/decoder using the algebraic
excitation generator as the random codebook. Therefore,
there can be obtained the speech coder/decoder, speech
signal communication system, and speech signal recording
system, which can output the synthetic speech having a
higher quality.
CA 02598683 2007-09-10
11
Brief Description of Drawings
FIG. 1 is a functional block diagram of a
conventional CELP speech coder;
FIG. 2 is a functional block diagram of a
conventional CELP speech decoder;
FIG. 3 is a functional block diagram of an
excitation vector generator according to a first
embodiment of the present invention;
FIG. 4 is a functional block diagram of a CELP speech
coder according to a second embodiment of the present
invention;
FIG. 5 is a functional block diagram of a CELP speech
decoder according to the second embodiment of the present
invention;
FIG. 6 is a functional block diagram of a CELP speech
coder according to a third embodiment of the present
invention;
FIG. 7 is a functional block diagram of a CELP speech
coder according to a fourth embodiment of the present
invention;
FIG. 8 is a functional block diagram of a CELP speech
coder according to a fifth embodiment of the present
invention;
FIG. 9 is a functional block diagram of a vector
quantization function according to the fifth embodiment
of the present invention;
FIG. 10 is a view explaining an algorithm for a
target extraction according to the fifth embodiment of
CA 02598683 2007-09-10
12
the present invention;
FIG. 11 is a functional block diagram of a
predictive quantization according to the fifth
embodiment of the present invention;
FIG. 12 is a functional block diagram of a
predictive quantization according to a sixth embodiment
of the present invention;
FIG. 13 Is a functional block diagram of- a CELP
speech coder according to a seventh embodiment of the
present invention; and
FIG. 14 is a functional block diagram of a
distortion calculator according to the seventh
embodiment of the present invention.
Best Mode for Carrying Out the Invention
Embodiments will now be described with reference
to the accompanying drawings.
(First embodiment)
FIG. 3 is a'functional block diagram of an
excitation vector generator according to a first
embodiment of the present invention.
The excitation vector generator comprises a pulse
vector generator 101 having a plurality of channels, a
dispersion pattern storing and selecting section 102
having dispersion pattern storing sections and switches,
a pulse vector dispersion section 103 for dispersing the
pulse vectors, and a dispersed vector adding section 104
for adding the dispersed pulse vectors for the plurality
CA 02598683 2007-09-10
13
of channels.
The pulse vector generator 101 comprises N (a case
of N=3 will be explained in this embodiment) channels
for generating vectors (hereinafter referred to as pulse
vectors) each having a signed unit pulse with provided
to one element on a vector axis.
The dispersion pattern storing and selecting
section 102 comprises storing sections M1 to-M3 for
storing M (a case of M=2 will be explained in this
embodiment) kinds of dispersion patterns for each
channel and switches SW1 to SW2 for selecting one kind
of dispersion pattern from M kinds of dispersion patterns
stored in the respective storing sections Ml to M3.
The pulse vector dispersion section 103 performs
convolution of the pulse vectors output from the pulse
vector generator 101 and the dispersion patterns output
from the dispersion pattern storing and selecting
section 102 in every channel so as to generate N dispersed
vectors.
The dispersed vector adding section 104 adds up N
dispersed vectors generated by the pulse vector
dispersion section 103, thereby generating an excitation
vector 105.
Note that, in this embodiment, a case in which the
pulse vector generator 101 algebraically generates N
(N=3) pulse vectors in accordance with the rule described
in Table 1 set forth below will be explained.
CA 02598683 2007-09-10
14
TABLE 1
Channel Number Polarity Pulse Position Candidates
CH1 1 P'(0, 10, 20, 30, = . = , 60, 70)
CH2 t 1 PZ 2,12, 22,32,---, 62, 72
~6, 1626, 36, = = =, 66, 76
CH3 1 P3 4, 14, 24, 34,...,
8,18,28,38,===,68,78
An operation of the above-structured excitation
vector generator will be explained.
The dispersion pattern storing and selecting
section 102 selects a dispersion pattern by one kind by
one from dispersion patterns stored two kinds by two for
each channel, and outputs the dispersion pattern. In
this case, the number is allocated to each dispersion
pattern in accordance with the combinations of selected
dispersion patterns (total number of combinations:
MN=8 ) .
Next, the pulse vector generator 101 algebraically
generates the signed pulse vectors corresponding to the
number of channels (three in this embodiment) in
accordance with the rule described in Table 1.
The pulse vector dispersion section 103 generates
a dispersed vector for each channel by convolving the
dispersion patterns selected by the dispersion pattern
storing and selecting section 102 with the signed pulses
generated by the pulse vector generator 101 based on the
following expression (5):
CA 02598683 2007-09-10
L-1
ci(n) _ Xwij(n-k)di(k) ... (5)
where n: 0-L-1,
L: dispersion vector length,
is channel number,
5 j: dispersion pattern number (j=1-M),
ci: dispersed vector for channel 1,
wij: dispersed pattern for channel i,j wherein the
vector length of wij(m) is 2L-1 (m: -(L-1)-L-1), and it
is the element, Li j , that can specify the value and the
10 other elements are zero,
di: signed pulse vector for channel i,
di = 6(n-pi), n = 0-L-1, and
pi: pulse position candidate for channel I.
The dispersed vector adding section 104 adds up
15 three dispersed vectors generated by the pulse vector
dispersion section 103 by the following equation (6) so
as to generate the excitation vector 105.
N
c(n) _ ci(n) === (6)
where c: excitation vector,
ci: dispersed vector,
is channel number (i= 1-N), and
n: vector element number (n = 0-L-1: note thatL
is an excitation vector length).
The above-structured excitation vector generator
can generate various excitation vectors by adding
variations to the combinations of the dispersion
CA 02598683 2007-09-10
16
patterns, which the dispersion pattern storing and
selecting section 102 selects, and the pulse position
and polarity in the pulse vector, which the pulse vector
generator 101 generates.
Then, in the above-structured excitation vector
generator, it is possible to allocate bits to two kinds
of information having the combinations of dispersion
patterns selected by the dispersion pattern stofing and
selecting section 102 and the combinations of the shapes
(the pulse positions and polarities) generated by the
pulse vector generator 101. The indices of this
excitation vector generator are in a one-to-one
correspondence with two kinds of information. Also, a
training processing is executed based on actual
excitation information in advance and the dispersion
patterns obtainable as the training result can be stored
in the dispersion pattern storing and selecting section
102.
Moreover, the above excitation vector generator is
used as the excitation information generator of speech
coder/decoder to transmit two kinds of indices including
the combination index of dispersion patterns selected
by the dispersion pattern storing and selecting section
102 and the combination index of the configuration (the
pulse positions and polarities) generated by the pulse
vector generator 101, thereby making it possible to
transmit information on random excitation.
Also, the use of the above-structured excitation
CA 02598683 2007-09-10
17
vector generator allows the configuration
(characteristic) similar to actual excitation
information to be generated as compared with the use of
algebraic codebook.
The above embodiment explained the case in which
the dispersion pattern storing and selecting section 102
stored two kinds of dispersion patterns per one channel.
However, the similar function and effect can be obtained
in a case in which the dispersion patterns other than
two kinds are allocated to each channel.
Also, the above embodiment explained the case in
which the pulse vector generator 101 was based on the
three-channel structure and the pulse generation rule
described in Table 1. However, the similar function and
effect can be obtained in a case in which the number of
channels is different and a case in which the pulse
generation rule other than Table 1 is used as a pulse
generation rule.
A speech signal communication system or a speech
signal recording system having the above excitation
vector generator or the speech coder/decoder is
structured, thereby obtaining the functions and effects
which the above excitation vector generator has.
(Second embodiment)
FIG. 4 shows a functional block of a CELP speech
coder according to the second embodiment, and FIG. 5
shows a functional block of a CELP speech decoder.
The CELP speech coder according to this embodiment
CA 02598683 2007-09-10
18
applies the excitation vector generator explained in the
first embodiment to the random codebook of the CELP
speech coder of FIG. 1. Also, the CELP speech decoder
according to this embodiment applies the excitation
vector generator explained in the first embodiment to
the random codebook of the CELP speech decoder of FIG.
2. Therefore, processing other than vector
quantization processing for random excitation is the
same as that of the apparatuses of FIGS. 1 and 2. This
embodiment will explain the speech coder and the speech
decoder with particular emphasis on vector quantization
processing for random excitation. Also, similar to the
first embodiment, the generation of pulse vectors are
based on Table 1 wherein the number of channels N = 3
and the number of dispersion patterns for one channel
M = 2.
The vector quantization processing for random
excitation in the speech coder illustrated in FIG. 4 is
one that specifies two kinds of indices (combination
index for dispersion patterns and combination index for
pulse positions and pulse polarities) so as to maximize
reference values in expression (4).
In a case where the excitation vector generator
illustrated in FIG. 3 is used as a random codebook,
combination index for dispersion patterns (eight kinds)
and combination index for pulse vectors (case
considering the polarity: 16384 kinds) are searched by
a closed loop.
CA 02598683 2007-09-10
19
For this reason, a dispersion pattern storing and
selecting section 215 selects either of two kinds of
dispersion patterns stored in the dispersion pattern
storing and selecting section itself, and outputs the
selected dispersion pattern to a pulse vector dispersion
section 217. Thereafter, a pulse vector generator 216
algebraically generates pulse vectors corresponding to
the number of -channels (three in this embodiment) in
accordance with the rule described in Table 1, and
outputs the generated pulse vectors to the pulse vector
dispersion section 217.
The pulse vector dispersion section 217 generates
a dispersed vector for each channel by a convolution
calculation. The convolution calculation is performed
on the basis of the expression (5) using the dispersion
patterns selected by the dispersion pattern storing and
selecting section 215 and the signed pulses generated
by the pulse vector generator 216.
A dispersion vector adding section 218 adds up the
dispersed vectors obtained by the pulse vector
dispersion section 217, thereby generating excitation
vectors (candidates for random codevectors).
Then, a distortion calculator 206 calculates
evaluation values according to the expression (4) using
the random code vector candidate obtained by the
dispersed vector adding section 218. The calculation on
the basis of the expression (4) is carried out with
respect to all combinations of the pulse vectors
CA 02598683 2007-09-10
generated based on the rule of Table 1. Then, among the
calculated values, the combination index for dispersion
patterns and the combination index for pulse vectors
(combination of the pulse positions and the polarities),
5 which are obtained when the evaluation value by the
expression (4) becomes maximum and the maximum value are
output to a code number specifying section 213.
Next, the-dispersion pattern storing and selecting
section 215. selects the combination for dispersion
10 patterns which is different from the previously selected
combination for the dispersion patterns. Regarding the
combination for dispersion patterns newly selected, the
calculation of the value of expression (4) is carried.
out with respect to all combinations of the pulse vectors
15 generated by the pulse vector generator 216 based on the
rule of Table 1. Then, among the calculated values, the
combination index for dispersion patterns and the
combination index for pulse vectors, which are obtained
when the value of expression (4) becomes maximum and the
20 maximum value are output to the code indices specifying
section 213 again.
The above processing is repeated with respect to
all combinations (total number of combinations is eight
in this embodiment) selectable from the dispersion
patterns stored in the dispersion pattern storing and
selecting section 215.
The code indices specifying section 213 compares
eight maximum values in total calculated by the
CA 02598683 2007-09-10
21
distortion calculator 206, and selects the highest value
of all. Then, the code indices specifying section 213
specifies two kinds of combination indices (combination
index for dispersion patterns, combination index for
pulse vectors), which are obtained when the highest value
is generated, and outputs the specified combination
indices to a code outputting section 214 as an index of
random codebook.
On the other hand, in the speech decoder of FIG.
5, a code inputting section 301 receives codes
transmitted from the speech coder (FIG. 4), decomposes
the received codes into the corresponding index of LPC
codebook, the index of adaptive codebook, the index of
random codebook (composed of two kinds of the combination
index for dispersion patterns and combination index for
pulse vectors) and the index of weight codebook. Then,
the code inputting section 301 outputs the decomposed
indicies to a linear prediction coefficient decoder 302,
an adaptive codebook, a random codebook 304, and a weight
codebook 305. Note that, in the random code number, that
the combination index for dispersion patterns is output
to a dispersion pattern storing and storing section 311
and the combination index for pulse vectors is output
to a pulse vector generator 312.
Then, the linear prediction coefficient decoder
302 decodes the linear predictive code number, obtains
the coefficients for a synthetic filter 309, and outputs
the obtained coefficients to the synthetic filter 309.
CA 02598683 2007-09-10
22
In the adaptive codebook 303, an adaptive codevector
corresponding to the index of adaptive codebook is read
from.
In the random codebook 304, the dispersion pattern
storing and selecting section 311 reads the dispersion
patterns corresponding to the combination index for
dispersion pulses in every channel, and outputs the
resultant to a pulse vector dispersion section 313. The
pulse vector generator 312 generates the pulse vectors
corresponding to the combination index for pulse vectors
and corresponding to the number of channels, and outputs
the resultant to the pulse vector dispersion section 313.
The pulse vector dispersion section 313 generates a
dispersed vector for each channel by convolving the
dispersion patterns received from the dispersion pattern
storing and selecting section 311 on the singed pulses
received from the pulse vector generator 312. Then, the
generated dispersed vectors are output to a dispersion
vector adding section 314. The dispersion vector adding
section 314 adds up the dispersed vectors of the
respective channels generated by the pulse vector
dispersion section 313, thereby generating a random
codevector.
Then, an adaptive codebook gain and a random
codebook gain corresponding to the index of weight
codebook are read from the weight codebook 305. Then,
in an adaptive code vector weighting section 306, the
adaptive codevector _is multiplied by the adaptive
CA 02598683 2007-09-10
23
codebook gain. Similarly in a random code vector
weighting section 307, the random codevector is
multiplied by the random codebook gain. Then, these
resultants are output to an adding section 308.
The adding section 308 adds up the above two code
vectors multiplied by the gains so as to generate an
excitation vector. Then, the adding section 308 outputs
the generated excitation vector to the adaptive codebook
303 to update a buffer or to the synthetic filter 309
to excite the synthetic filter.
The synthetic filter 309 is excited by the
excitation vector obtained by the adding section 308,
and reproduces a synthetic speech 310. Also, the
adaptive codebook 303 updates the buffer by the
excitation vector received from the adding section 308.
In this case, suppose that the dispersion patterns
obtained by pre-training are stored for each channel in
the dispersion pattern storing and selecting section of
FIGS. 4 and 5 such that a value of cost function becomes
smaller wherein the cost function is a distortion
evaluation expression (7) in which the excitation vector
described in expression (6) is substituted into c of
expression (2).
N 2
Ec = - gcH ~ci
k~(x(n) - gcH ,iAci(n))2
_ o(x(n) - gcHy "1 , k.owij(n-k)di(k))2 ... (7)
where x: target vector for specifying index of
CA 02598683 2007-09-10
24
random codebook,
gc: random codebook gain,
H: impulse response convolution matrix for synthetic
filter,
c: random codevector,
is channel number (ii = 1-N),
j: dispersion pattern number (j = 1-M)
ci: dispersion vector for channel i,
wij: dispersion patterns for channels i-th, j-th
kinds,
di: pulse vector for channel i, and
L: excitation vector length (n = O-L-1).
The above embodiment explained the case in which
the dispersion patterns obtained by pre-training were
stored M by M for each channel in the dispersion pattern
storing and selecting section such that the value of cost
function expression (7) becomes smaller.
However, in actual, all M dispersion patterns do not have
to be obtained by training. If at least one kind of
dispersion pattern obtained by training is stored, it
is possible to obtain the functions and effects to
improve the quality of the synthesized speech.
Also, the above embodiment explained that case in
which from all combinations of dispersion patterns
stored in the dispersion pattern storing and selecting
section stores and all combinations of pulse vector
position candidates generated by the pulse vector
generator, the combination index that maximized the
CA 02598683 2010-05-31
reference value of expression (4) was specified by the
closed loop. However, the similar functions and effects
can be obtained by carrying out a pre-selection based
on other parameters (ideal gain for adaptive codevector,
5 etc.) obtained before specifying the index of the random
codebook or by a open loop search.
Moreover, a speech signal communication system or
a speech signa-1 recording system having the above the
speech coder/decoder is structured, thereby obtaining
10" the functions and effects which the excitation vector
generator described in the first embodiment has.
(Third embodiment)
FIG. 6 is a functional block of a CELP speech coder
according to the third embodiment. According to this
15 embodiment, in the CELP speech coder using the excitation
vector generator of the first embodiment in the random
codebook, a pre-selection for dispersion patterns stored
in the dispersion pattern storing and selecting section
is carried out using the value of an ideal adaptive
20 codebook gain obtained before searching the index of
random codebook. The other portions of the random
codebook peripherals are the same as those of the CELP
speech coder of FIG. 4. Therefore, this embodiment will
explain the vector quantization processing for random
25 excitation in the CELP speech coder of FIG. 6.
This CELP speech coder comprises an adaptive
codebook 407, an adaptive codebook gain weighting
section 409, a random codebook 408 constituted by the
CA 02598683 2007-09-10
26
excitation vector generator explained in the first
embodiment, a random codebook gain weighting section 410,
a synthetic filter 405, a distortion calculator 406, an
indices specifying section 413, a dispersion pattern
storing and selecting section 415, a pulse vector
generator 416, a pulse vector dispersion section 417,
a dispersed vector adding section 418, and a distortion
power juding section 419.
In this case, according to the above embodiment,
suppose that at least one of M (M = 1'2) kinds of dispersion
patterns stored in the dispersion pattern storing and
selecting section 415 is the dispersion pattern that is
obtained from the result by performing a pre-training
to reduce quantization distortion generated in vector
quantization processing for random excitation
In this embodiment, for simplifying the
explanation, it is assumed that the number N of channels
of the pulse vector generator is 3, and the number M of
kinds of dispersion patterns for each channel stored in
the dispersion pattern storing and selecting section is
2. Also, suppose that one of M (M = 2) kinds of dispersion
patterns is dispersion pattern obtained by the
above-mentioned training, and other is random vector
sequence (hereinafter referred to as random pattern)
which is generated by a random vector generator.
Additionally, it is known that the dispersion pattern
obtained by the above training has a relatively short
length and a pulse-like-shape as in wll of FIG. 3.
CA 02598683 2007-09-10
27
In the CELP speech coder of FIG. 6, processing for
specifying the index of the adaptive codebook before
vector quantization of random excitation is carried out.
Therefore, at the time when vector quantization
processing of random excitation is carried out, it is
possible to refer to the index of the adaptive codebook
and the ideal adaptive codebook gain (temporarily
decided). In this embodiment, the pre-selection for
dispersion patterns is carried out using the value of
the ideal adaptive codebook gain.
More specifically, first, the ideal value of the
adaptive codebook gain stored in the code indices
specifying section 413 just after the search for the
index of adaptive codebook is output to the distortion
calculator 406. The distortion calculator 406 outputs
the adaptive codebook gain received from the code indices
specifying section 413 to the adaptive codebook gain
judging section 419.
The adaptive gain judging section 419 performs a
comparison between the value of the ideal adaptive
codebook gain received from the distortion calculator 409
and a preset threshold value. Next, the adaptive
codebook gain judging section 419 sends a control signal
for a pre-selection to the dispersion pattern storing and
selecting section 415 based on the result of the
comparison. The contents of the control signal will be
explained as follows.
More specifically, when the adaptive codebook gain
CA 02598683 2007-09-10
28
is larger than the threshold value as a result of the
comparison, the control signal provides an instruction
to select the dispersion pattern obtained by the pre
training to reduce the quantization distortion in vector
quantization processing for random excitations. Also,
when the adaptive code gain is not larger than the
threshold value as a result of the comparison, the control
signal provides an instruction to carry out the pre-
selection for the dispersion pattern different from the
dispersion pattern obtained from the result of the
pre-training.
As a consequence, in the dispersion pattern storing
and selecting selection 415, the dispersion pattern of
M (M = 2) kinds, which the respective channels store, can
be pre-selected in accordance with the value of the ideal
adaptive codebook gain, so that the number of combinations
of dispersion patterns can be largely reduced. This
eliminates the need of the distortion calculation for all
the combinations of the dispersion patterns, and makes
it possible to efficiently perform the vector
quantization processing for random excitation with a
small amount of calculations.
Moreover, the random codevector is pulse-like
shaped when the value of the adaptive gain is large (this
segment is determined as voiced) and is randomly shaped
when the value of the adaptive gain is small (this segment
is determined as unvoiced). Therefore, since the random
code vector having a suitable shape for each of the voice
CA 02598683 2007-09-10
29
segment the speech signal and the non-voice segment can
be used, the quality of the synthtic speech can be
improved.
Due to the simplification of the explanation, this
embodiment explained limitedly the case in which the
number N of channels of the pulse vector generator was
3 and the number M of kinds of the dispersion patterns
was 2 per channel stored in the dispersion pattern storing
and selecting section. However, similar effects and
functions can be obtained in a, case in which the number
of channels of the pulse vector generator and the number
of kinds of the dispersion patterns per channel stored
in the dispersion pattern storing and selecting section
are different from the aforementioned case.
Also, due to the simplification of the explanation,
the above embodiment explained the case in which one of
M kinds (M = 2) of dispersion patterns stored in each
channel was dispersion patterns obtained by the above
training and the other was random patterns. However, if
at least one kind of dispersion pattern obtained blip' the
training is stored for each channel, the similar effects
and functions can be expected instead of the above-
explained case.
Moreover, this embodiment explained the case in
which large and small information of the adaptive codebook
gain was used in means for performing pre-selection of
the dispersion patterns. However, if other parameters
showing a short-time character of the input speech are
CA 02598683 2007-09-10
used in addition to large and small information of the
adaptive codebook gain, the similar effects and functions
can be further expected.
Further, a speech signal communication system or a
5 speech signal recording system having the above the speech
coder/decoder is structured, thereby obtaining the
functions and effects which the excitation vector
generator described in the first embodiment his.
In the explanation of the above embodiment, there
10 was explained the method in which the pre-selection of
the dispersion pattern was carried out using the ideal
adaptive codebook gain of the current frame at the time
when vector quantization processing of random excitation
was performed. However, the similar structure can be
15 employed even in a case in which a decoded adaptive
codebook gain obtained in the previous frame is used
instead of the ideal adaptive codebook gain in the
current frame. In this case, the similar effects can be
also obtained.
20 (Fourth embodiment)
FIG. 7 is a functional block diagram of a CELP speech
coder according to the fourth embodiment. In this
embodiment, in the CELP speech coder using the excitation
vector generator of the first embodiment in the random
25 codebook, a pre-selection for a plurality of dispersion
patterns stored in the dispersion pattern storing and
selecting section is carried out using available
information at the_ time of vector quantization
CA 02598683 2007-09-10
31
processing for random excitations. It is characterized
that a value of a coding distortion (expressed by an S/N
ratio) , that is generated in specifying the index of the
adaptive codebook, is used as a reference of the
pre-selection.
Note that the other portions of the random codebook
peripherals are the same as those of the CELP speech coder
of FIG. 4. Therefore, this embodiment will specifically
explain the vector quantization processing for random
excitation.
As shown in FIG. 7, this CELP speech coder comprises
an adaptive codebook 507, an adaptive codebook gain
weighting section 509, a random codebook 508 constituted
by the excitation vector generator explained in the first
embodiment, a random codebook gain weighting section 510,
a synthetic filter 505, a distortion calculator 506, a
code indices specifying section 513, a dispersion
pattern storing and selecting section 515, a pulse vector
generator 516, a pulse vector dispersion section 517,
a dispersed vector adding section 518, and a coding
distortion judging section 519.
In this case, according to the above embodiment,
suppose that at least one of M (M = Z2) kinds of dispersion
patterns stored in the dispersion pattern storing and
selecting section 515 is the random pattern.
In the above embodiment= for simplifying the
explanation, the number N of channels of the pulse vector
generator is 3 and the number M of kinds of the dispersion
CA 02598683 2007-09-10
32
patterns is 2 per channel stored in the dispersion
pattern storing and selecting section. Moreover, one of
M (M = 2) kinds of dispersion patterns is the random
pattern, and the other is the dispersion pattern that
is obtained as the result of pre-training to reduce
quantization distortion generated in vector
quantization processing for random excitations.
In the CELP speech coder of FIG. 7, processing for
specifying the index of the adaptive codebook is
performed before vector quantization processing for
random excitation. Therefore, at the time when vector
quantization processing of random excitation is carried
out, it is possible to refer to the index of the adaptive
codebook, the ideal adaptive codebook gain (temporarily
decided), and the target vector for searching the
adaptive codebook. In this embodiment, the pre-
selection for dispersion patterns is carried out using
the coding distortion (expressed by S/N ratio) of the
adaptive codebook which can be calculated from the above
three information.
More specifically, the index of adaptive codebook
and the value of the adaptive codebook gain (ideal gain)
stored in the code indices specifying section 513 just
after the search for the adaptive codebook is output to
the distortion calculator 506. The distortion
calculator 506 calculates the coding distortion (S/N
ratio) generated by specifying the index of the adaptive
codebook using the index of adaptive codebook received
CA 02598683 2007-09-10
33
from the code indices specifying section 513, the
adaptive codebook gain, and the target vector for
searching the adaptive codebook. Then, the distortion
calculator 506 outputs the calculated S/N value to the
coding distortion juding section 519.
The coding distortion juding section 519 performs
a comparison between the S/N value received from the
distortion calculator 506 and a preset threshold value.
Next, the coding distortion juding section 519 sends a
control signal for a pre-selection to the dispersion
pattern storing and selecting section 515 based on the
result of the comparison. The contents of the control
signal will be explained as follows.
More specifically, when the S/N value is larger than
the threshold value as a result of the comparison, the
control signal provides an instruction to select the
dispersion pattern obtained by the pre-training to reduce
the quantization distortion generated by coding the
target vector for searching the random codebook. Also,
when the S/N value is smaller than the threshold value
as a result of the comparison, the control signal provides
an instruction to select the non-pulse-like random
patterns.
As a consequence, in the dispersion pattern storing
and selecting selection 515, only one kind is pre-selected
from M (M = 2) kinds of dispersion patterns, which the
respective channels store, so that the number of
combinations of dispersion patterns can be largely
CA 02598683 2007-09-10
34
reduced. This eliminates the need of the distortion
calculation for all the combinations of the dispersion
patterns, and makes it possible to efficiently specify
the index of the random codebook with a small amount of
calculations.
Moreover, the random codevector is pulse-like
shaped when the SIN value is large, and is non-pulse-
like shaped when the S/N value is small. Therefore, since
the shape of the random codevector can be changed in
accordance with the short-time characteristic of the
speech signal, the quality of the synthetic speech can
be improved.
Due to the simplification of the explanation, this
embodiment explained limitedly the case in which the
number N of channels of the pulse vector generator was
3 and the number M of kinds of the dispersion patterns
was 2 per channel stored in the dispersion pattern storing
and selecting section. However, similar effects and
functions can be obtained in a case in which the number
of channels of the pulse vector generator and the number
of kinds of the dispersion patterns per channel stored
in the dispersion pattern storing and selecting section
are different from the aforementioned case.
Also, due to the simplification of the explanation,
the above embodiment explained the case in which one of
M kinds (M = 2) of dispersion patterns stored in each
channel was dispersion patterns obtained by the above
pre-training and the other was random patterns. However,
CA 02598683 2007-09-10
if at least one kind of random dispersion pattern is stored
for each channel, the similar effects and functions can
be expected instead of the above-explained case.
Moreover, this embodiment explained the case in
5 which only large and small information of coding
distortion (expressed by S/N value) generated. by
specifying the index of the adaptive codebook was used
in means for pre-selecting the dispersion pattern.
However, if other information, which correctly shows the
10 short-time characteristic of ,the speech. signal, is
employed in addition thereto,,the similar effects and
functions can be further expected.
Further, a speech signal communication system or a
speech signal recording system having the above the speech
15 coder/decoder is structured, thereby obtaining the
functions and effects which the excitation vector
generator described in the first embodiment has.
(Fifth embodiment)
FIG. 8 shows a functional block of a CELP speech
20 coder according to the fifth embodiment of the present
invention. According to this CELP speech coder, in an
LPC analyzing section 600 performs a self-correlation
analysis and an LPC analysis of input speech data 601,
thereby obtaining LPC coefficients. Also, the obtained
25 LPC coefficients are quantized so as to obtain the index
of LDC codebook, and the obtained index is decoded so as
to obtain decoded LPC coefficients.
Next, an excitation generator 602 takes out
CA 02598683 2007-09-10
36
excitation samples stored in an adaptive codebook 603 and
a random codebook 604 (an adaptive codevector (or adaptive
excitation) and random codevector (or a random
excitation)) and sends them to an LPC synthesizing section
605.
The LPC synthesizing section 605 filters two
excitations obtained by the excitation generator 602 by
the decoded LPC coefficient obtained by the LPC analyzing
section 600, thereby obtaining two synthesized
excitations.
In a comparator 606, the relationship between two
synthesized excitations obtained by the LPC synthesizing
section 605 and the input speech 601 is analyzed so as
to obtain an optimum value (optimum gain) of two
synthesized excitations. Then, the respective
synthesized excitations, which are power controlled by
the optimum value, are added so as to obtain an integrated
synthesized speech, and a distance calculation between
the integrated synthesized speech and the input speech
is carried out.
The distance calculation between each of many
integrated synthesized speeches, which are obtained by
exciting the excitation generator 602 and the LPC
synthesizing section 605, and the input speech 601 is
carried out with respect to all excitation samples of the
adaptive codebook 603 and the random codebook 604. Then,
an index of the excitation sample, which is obtained when
the value is the smallest-in the distances obtainable from
CA 02598683 2007-09-10
37
the result, is determined.
Also, the obtained optimum gain, the index of the
excitation sample, and two excitations responding to the
index are sent to a parameter coding section 607. In the
parameter coding section 607, the optimum gain is coded
so as to obtain a gain code, and the index of LPC codebook
and the index of the excitation sample are sent to a
transmission path 608 at one time.
Moreover, an actual excitation signal is generated
from two excitations responding to the gain code and the
.index, and the generated excitation signal is stored in
the adaptive codebook 603 and the old excitation sample
is abandoned at the same time.
Note that, in the LPC synthesizing section 605, a
perceptual weighting filter using the linear predictive
coefficients, a high-frequency enhancement filter, a
long-term predictive filter, (obtained by carrying out
a long-term prediction analysis of input speech) are
generally employed. Also, the excitation search for the
adaptive codebook and the random codebook is generally
carried out in segments (referred to as subframes) into
which an analysis segment is further divided.
The following will explain the vector quantization
for LPC coefficients in the LPC analyzing section 600
according to this embodiment.
FIG. 9 shows a functional block for realizing a
vector quantization algorithm to be executed in the LPC
analyzing section 600._ The vector quantization block
CA 02598683 2007-09-10
38
shown in FIG. 9 comprises a target extracting section 702,
a quantizing section 703, a distortion calculator 704,
a comparator 705, a decoding vector storing section 707,
and a vector smoothing section 708.
In the target extracting section 702, a quantization
target is calculated based on an input vector 701. Here,
a target extracting method will be specifically
explained.
In this embodiment, the "input vector" comprises two
kinds of vectors in all wherein one is a parameter vector
obtained by analyzing a current frame and the other is
a parameter vector obtained from a future frame in a like
manner. The target extracting section 702 calculates a
quantization target using the above input vector and a
decoded vector of the previous frame stored in the decoded
vector storing section 707. An example of the
calculation method will be shown by the following
expression (8).
X(i) - {St(1)+p(d(i)+S,+:(i)/2}l(1+p) ... (8)
where X(i) : target vector,
is vector element number,
St (i) , St+1(i) : input vector,
t: time (frame number),
p: weighting coefficient (fixed), and
d(i): decoded vector of previous frame.
The following will show a concept of the above. target
extraction method. In a typical vector quantization,
parameter vector SL(i)_is used as target X(i) and a
CA 02598683 2007-09-10
39
matching is performed by the following expression (9):
En = ~(X (i)-Cn(i))2 ... (9)
where En: distance from n-th code vector,
X(i): target vector,
Cn(i): code vector,
n: code vector number,
is order of vector, and
I: length of vector.
Therefore, in the conventional vector quantization,
the coding distortion directly leads to degradation in
speech quality. This was a big problem in the ultra-
low bit rate coding in which the coding distortion cannot
be avoided to some extent even if measurements such as
prediction vector quantization is taken.
For this reason, according to this embodiment,
attention should be paid to a middle point of the decoded
vector as a direction where the user does not perceptually
f eel an error easily, and the decoded vector is induced
to the middle point so as to realize perceptual
improvement. In the above case, there is used a
characteristic in which time continuity is not easily
heard as a perceptual degradation.
The following will explain the above state with
reference to FIG. 10 showing a vector space.
First of all, it is assumed that the decoded vector
of one previous frame is d(i) and a future parameter vector
CA 02598683 2007-09-10
is SL+1(i) (although a future coded vector is actually
desirable, the future parameter vector is used for the
future coded vector since the coding cannot be carried
out in the current frame. In this case, although the code
5 vector Cn (i) : (1) is closer to the parameter vector St(j)
than the code vector Cn (i) : (2). the code vector Cn (i) :
(2) is actually close onto a line connecting d(i) and
St+1 (i) . For this reason, degradation is not easily heard
as compared with (1). Therefore, by use of the above
10 characteristic, if the target X(i) is set as a vector
placed at the position where the target X(i) approaches
to the middle point between d(i) and St+l(i) from St(i)
to some degree, the decoded vector is induced to a
direction where the amount of distortion is perceptually
15 slight.
Then, according to this embodiment, the movement of
the target can be realized by introducing the following
evaluation expression (10)
X(i) - {Sr (i) + p (d (i) + S,+1(i) / 2} /(1 + p) ... (10)
20 where X(i): target vector,
is vector element number,
St (i) , St+i (i) : input vector,
t: time (frame number),
p: weighting coefficient (fixed), and
25 d(l): decoded vector of previous frame.
The first half of expression (10) is a general
evaluation expression, and the second half is a perceptual
component. In order to carry out the quantization by the
CA 02598683 2007-09-10
41
above evaluation expression, the evaluation expression
is differentiated with respect to each X(i) and the
differentiated result is set to 0, so that expression (8)
can be obtained.
Note that the weighting coefficient p is a positive
constant. Specifically, when the weighting coefficient
p is zero, the result is similar to the general
quantization when the weighting coefficient p is infinite,
the target is placed at the completely middle point. If
the weighting coefficient p is' too large, the target is
largely separated from the parameter St (i) of the current
frame so that articulation is perceptually reduced. The
test listening of decoded speech confirms that a good
performance with 0.5<p<1.0 can be obtained.
Next, in the quantizing section 703, the
quantization target obtained by the target extracting
section 702 is quantized so as to obtain a vector code
and a decoded vector, and the obtained vector index and
decoded vector are sent to the distortion calculator 704.
Note that a predictive vector quantization is used
as a quantization method in this embodiment. The
following will explain the predictive vector
quantization.
FIG. 11 shows a functional block of the predictive
vector quantization. The predictive vector quantization
is an algorithm in which the prediction is carried out
using the vector (-synthesized vector) obtained by coding
and decoding in the past _and the predictive error vector
CA 02598683 2007-09-10
42
is quantized.
A vector codebook 800. which stores a plurality of
main samples (codevectors) of the prediction error
vectors, is prepared in advance. This is prepared by an
LBG algorithm (IEEE TRANSACTIONS ON COMMUNICATIONS, VOL.
COM-28, NO. 1, PP84-95, JANUARY 1980) based on a large
number of vectors obtained by analyzing a large amount
of speech data.
A vector 801 for quantization target is predicted
by a prediction section 802. The prediction is carried
out by the post-decoded vectors stored in a state storing
section 803, and the obtained predictive error vector is
sent to a distance calculator 804. Here, as a form of
prediction, a first prediction order and a fixed
coefficient are used. Then, an expression for calculating the
predictive error vector in the case of using the above
prediction is shown by the following expression (11).
Y(i) = X(i) - /3D(i) ... (11)
where Y(i): predictive error vector,
X(i): target vectpr,
/3: prediction coefficient (scalar)
D(i): decoded vector of one previous
frame, and
is vector order.
In the above expression, it is-general that the
prediction coefficient (3 is a value of 0 < Q < 1.
Next, the distance calculator 804 calculates the
distance between the predictive error vector obtained by
CA 02598683 2007-09-10
43
the prediction section 802 and the codevector stored in
codebook 800. An expression for obtaining the above
distance is shown by the following expression (12):
1 z
En - J (T (i) - Cn(i)) = = = (12)
where En: distance from n-th code vector,
Y(i): predictive error vector,
Cn(i): codevector,
n: codervector number,
I: vector order,land
I: vector length.
Next, in a searching section 805, the distances for
respective codevectors are compared, and the index of
codevector which gives the shortest distance is output
as a vector code 806.
In other words, the vector codebook 800 and the
distance calculator 804 are controlled so as to obtain
the index of codevector which gives the shortest distance
from all codevectors stored in the vector codebook 800,
and the obtained index is used as vector code 806.
Moreover, the vector is coded using the code vector
obtained from the vector codebook 800 and the past-decoded
vector stored in the state storing section 803 based on
the final coding, and the content of the state storing
section 803 is updated using the obtained synthesized
vector. Therefore, the decoded vector here is used in
the prediction when a next quantization is performed.
The decoding of the example (first prediction order, -
CA 02598683 2007-09-10
44
fixed coefficient) in the above-mentioned prediction
form is performed by the following expression (13):
Z(i) = CN(i) + (3D(i) ... (13)
where Z(i): decoded vector (used as D(i) at a next coding
time,
N: code for vector,
CN(i): code vector,
Q: prediction coefficient (scalar),
D(i): decoded vector of one previous
frame, and
is vector order.
On the other hand, in a decoder, the code vector is
obtained based on the code of the transmitted vector so
as to be decoded. In the decoder, the same vector
codebook and state storing section as those of the coder
are prepared in advance. Then, the decoding is carried
out by the same algorithm as the decoding function of the
searching section in the aforementioned coding algorithm.
The above is the vector quantization, which is executed
in the quantizing section 703.
Next, the distortion calculator 704 calculates a
perceptual weighted coding distortion from the decoded
vector obtained by the quantizing section 703, the input
vector 701, and the decoded vector of the previous frame
stored in the decoded vector storing section 707. An
expression for calculation is shown by the following
expression (14):
CA 02598683 2007-09-10
Ew a ( V (i) - S, (i) )Z + P{V (i) - (d (i) + Sr+1(i) / 2}2 ... (14)
where Ew: weighted coding distortion,
St (i) , St+1(i) : input vector,
t: time (frame number)
5 is vector element number,
V(i): decoded vector,
p: weighting coefficient (fixed), and
d(i): decoded vector of previous frame.
In expression (14) , the weighting efficient p is the
10 same as the coefficient of the' expression of the target
used in the target extracting, section 702. Then, the
value of the weighted coding distortion, the encoded
vector and the code of the vector are sent to the
comparator 705.
15 The comparator 705 sends the code of the vector sent
from the distortion calculator 704 to the transmission
path 608, and further updates the content of the decoded
vector storing section 707 using the vector sent from the
distortion calculator 704.
20 According to the above-mentioned embodiment, in the
target extracting section 702, the target vector is
corrected from St (i) to the vector placed at the position
approaching to the middle point between D(i) and St+1(i )
to same extent. This makes it possible to perform the
25 weighted search so as not to arise perceptual degradation.
The above explained the case in which the present
invention was applied to the low bit rate speech coding
technique used in such-as- a cellular phone. However, the
CA 02598683 2007-09-10
46
present invention can be employed in not only the speech
coding but also the vector quantization for a parameter
having a relatively good interpolation in a music coder
and an image coder.
In general, the LPC coding executed by the LPC
analyzing section in the above-mentioned algorithm,
conversion to parameters vector such as LPS (Line Spectram
Pairs), which are easily coded, is commonly performed,
and vector quantization (VQ) is carried out by Euclidean
distance or weighted Euclidean distance.
Also, according to the above embodiment, the target
extracting section 702 sends the input vector 701 to the
vector smoothing section 708 after being subjected to the
control of the comparator 705. Then, the target
extracting section 702 receives the input vector changed
by the vector smoothing section 708, thereby re-
extracting the target.
In this case, the comparator 705 compares the value
of weighted coding distortion sent from the distortion
calculator 704 with a reference value prepared in the
comparator. Processing is divided into two, depending
on the comparison result.
If the comparison result is under the reference
value, the comparator 705 sends the index of the
codevector sent from the distortion calculator to the
transmission path 608, and updates the content of the
decoded vector storing section 707 using the coded vector
sent from the distortion- calculator 704. This update is -
CA 02598683 2007-09-10
47
carried out by rewriting the content of the decoded vector
storing section 707 using the obtained coded vector.
Then, processing moves to one for a next frame parameter
coding.
While, if the comparison result is more than the
reference value, the comparator 705 controls the vector
smoothing section 708 and adds a change to the input vector
so that the target extracting section 702, the quantizing
section 703 and distortion calculator 704 are functioned
again to perform coding again.
In the comparator 705, coding processing is repeated
until the comparison result reaches the value under
reference value. However, there is a case in which the
comparison result can not reach the value under the
reference value even if coding processing is repeated many
times. In case, the comparator 705 provides a counter
in its interior, and the counter counts the number of times
wherein the comparison result is determined as being more
than the reference value. When the number of times is
more than a fixed number of times, the comparator 705 stops
the repetition of coding and clears the comparison result
and counter state, then adopts initial index.
The vector smoothing section 708 is subjected to the
control of the comparator 705 and changes parameter vector
St (i) of the current frame, which is one of input vectors,
from the input vector obtained by the target extracting
section 702 and the decoded vector of the previous frame
obtained decoded vector storing section 707 by the
CA 02598683 2007-09-10
48
following expression (15), and sends the changed input
vector to the target extracting section 702.
St(i) E- (1-q) - St(i) + q(d(i) + St+1(i))/2 ... (15)
In the above expression, q is a smoothing
coefficient, which shows the degree of which the parameter
vector of the current frame is updated close to a middle
point between the decoded vector of the previous frame
and the parameter vector of the future frame. The coding
experiment shows that good performance can be obtained
when the upper limitation of the number of repetition
executed by the interior of the comparator 705 is 5 to
8 under the condition of 0. 2 < q <0.4.
Although the above embodiment uses the predictive
vector quantization in the quantizing section 703, there
is a high possibility that the weighted coding distortion
obtained by the distortion calculator 704 will become
small. This is because the quantized target is updated
closer to the decoded vector of the previous frame by
smoothing. Therefore, by the repetition of decoding the
previous frame due to the control of the comparator 705,
the possibility that the comparison result will become
under the reference value is increased in the distortion
comparison of the comparator 705.
Also, in the decoder, there is prepared a decoding
section corresponding to the quantizing section of the
coder in advance such that decoding is carried out based
on the index of the codevector transmitted through the
transmission path.
CA 02598683 2007-09-10
49
Also, the embodiment of the present invention was
applied to quantization (quantizing section is
prediction VQ) of LSP parameter appearing CELP speech
coder, and speech coding and decoding experiment was
performed. As a result, it was confirmed that not only
the subjective quality but also the objective value. (S/N
value) could be improved. This is because there is an
effect in which the coding distortion of predictive VQ
can be suppressed by coding repetition processing having
vector smoothing even when the spectrum drastically
changes. Since the future prediction VQ was predicted
from the past-decoded vectors, there was a disadvantage
in which the spectral distortion of the portion where the
spectrum drastically changes such as a speech onset
contrarily increased. However, in the application of the
embodiment of the present invention, since smoothing is
carried out until the distortion lessens in the case where
the distortion is large, the coding distortion becomes
small though the target is more or less separated from
the actual parameter vector. Whereby, there can be
obtained an effect in which degradation caused when
decoding the speech is totally reduced. Therefore,
according to the embodiment of the present invention, not
only the subjuctive quality but also' the objective value
can be improved. _
In the above-mentioned embodiment of the present
invention, by the characteristics of the comparator and
the vector smoothing section, control can be provided to
CA 02598683 2007-09-10
the direction where the operator does not perceptually
feel the direction of degradation in the case where the
vector quantizing distortion is large. Also, in the case
where predictive vector quantization is used in the
5 quantizing section, smoothing and coding are repeated
until the coding distortion lessens, thereby the
objective value can be also improved.
The above explained the case in which the present
invention was applied to the low bit rate speech coding
10 technique used in such as a cellular phone. However, the
present invention can be employed in not only the speech
coding but also the vector quantization for a parameter
having a relatively good interpolation in a music coder
and an image coder.
15 (Sixth embodiment)
Next, the following will explain the CELP speech
coder according to the sixth embodiment. The
configuration of this embodiment is the same as that of
the fifth embodiment excepting quantization algorithm of
20 the quantizing section using a multi-stage predictive
vector quantization as a quantizing method. In other
words, the excitation vector generator of the first
embodiment is used as a random codebook. Here, the
quantization algorithm of the quantizing section will be
25 specifically explained.
FIG . 12 shows the functional, block of the quantizing
section. In the multi-stage predictive vector
quantization, the vector quantization of the target is
CA 02598683 2007-09-10
51
carried out, thereafter the vector is decoded using a
codebook with the index of the quantized target, a
difference between the coded vector. Then, the original
target (hereinafter referred to as coded distortion
vector) is obtained, and the obtained coded distortion
vector is further vector-quantized.
A vector codebook 899 in which a plurality of
dominant samples (codevectors) of the predictive error
vector are stored and a codebook 900 are generated in
advance. These codevectorsare generated by applying the
same algorithm as that of the codevector generating method
of the typical "multi-vector quantization'. In other
words, these codevectors are generally generated by an
LBG algorithm (IEEE TRANSACTIONS ON COMMUNICATIONS, VOL.
COM-28, NO. 1, PP84-95, JANUARY 1980) based on a large
number of vectors obtained by analyzing many speech data.
Note that, a training date for designing codevectors 899
is a set of many target vectors, while a training date
for designing codebook 900 is a set of coded distortion
vectors obtained when the above-quantized targets are
coded by the vector codebook 899.
First, a vector 901 of the target vector is predicted
by a predicting section 902. The prediction is carried
out by the past-decoded vectors stored in a state storing
section 903, and the obtained predictive error vector is
sent to distance calculators 904 and 905.
According to the above embodiment, as a form of
prediction, a fixed coefficient is used for a first order
CA 02598683 2007-09-10
52
prediction. Then, an expression for calculating the
predictive error vector in the case of using the above
prediction is shown by the following expression (16).
Y(i) = X(i) - (3D(i) . (16)
where Y(i): predictive error vector,
X(i): target vector,
Q: predictive coefficient (scalar),
D(i'): decoded vector of one previous
frame, and
is vector order.
In the above expression, it is general that the
predictive coefficient 0 is a value of 0 < (3 < 1.
Next, the distance calculator 904 calculates the
distance between the predictive error vector obtained by
the prediction section 902 and code vector A stored in
the vector codebook 899. An expression for obtaining the
above distance is shown by the following expression (17) :
Z
En=~(X(i) - Cln(i)) === (17)
where En: distance from n-th code vector A
Y(i): predictive error vector,
Cin(i): codevector A,
n: index of codervector A,
I: vector order, and
I: vector length.
Then, in a searching section 906, the respective
distances from the codevector A are compared, and the
index of the code vector A having the shortest distance
CA 02598683 2007-09-10
53
is used as a code for code vector A. In other words, the
vector codebook 899 and the distance calculator 904 are
controlled so as to obtain the code of codevector A having
the shortest distance from all codevectors stored in the
codebook 899. Then, the obtained code of codevector A
is used as the index of codebook 899. After this, the
code for codevector A and decoded vector A obtained from
the codebook 89-9 with reference to the code for coddevector
A are sent to the distance calculator 905. Also, the code
for codevector A is sent to a searching section 906 through
the transmission path.
The distance calculator 905 obtains a coded
distortion vector from the predictive error vector and
the decoded vector A obtained from the searching section
906. Also, the distance calculator 905 obtains amplitude
from an amplifier storing section 908 with reference to
the code for codevector A obtained from the searching
section 906. Then, the distance calculator 905
calculates a distance by multiplying the above coded
distortion vector and codevector B stored in the vector
codebook 900 by the above amplitude, and sends the
obtained distance to the searching section 907. An
expression for the above distance is shown as follows:
Z(i) = Y(i) C1N(i)
2
Em - ;(Z(i)-aNC2m(i)) === (18)
where Z(i) : decoded vector,
Y(i): predictive error vector,
CA 02598683 2007-09-10
54
C1N(i): decoded vector A,
Em : distance from m-th code vector B,
aN : amplitude corresponding to the code for
codevector A,
C2m(i): codevector B,
m : index of codevector B,
i : vector order, and
I : vector length
Then, in a searching section 907, the respective
distances from the codevector B are compared, and the
index of the codevector B having the shortest distance
is used as a code for codevector B. In other words, the
codebook 900 and the distance calculator 905 are
controlled so as to obtain the code of codevector B having
the shortest distance from all codevectors stored in the
vector codebook 900. Then, the obtained code of
codevector B is used as the index of codebook 900. After
this, codevector A and codevector B are added and used
as a vector code 909.
Moreover, the searching section 907 carries out the
decoding of the vector using decoded vectors A, B obtained
from the vector codebooks 899 and 900 based on the codes
for codevector A and codevector B, amplitude obtained from
an amplifier storing section 908 and past decoded vectors
stored in the state storing section 903. The content of
the state storing section 903 is updated using the
obtained decoded vector. (Therefore, the vector as
decoded above is used in the prediction at a next coding
CA 02598683 2007-09-10
time). The decoding in the prediction (a first
prediction order and a fixed coefficient) in this embodiment
is performed by the following expression (19):
Z(i) = C1N(i) + aN=C2M(i) + (3D(i) _ (19)
5 where Z (i) : decoded vector (used as D (i) at the next
coding time),
N: code for codevector A,
M: code for codevector B,
C1N(i): decoded codevector A,
10 C2M(i): decoded codevector B,
aN : amplitude corresponding to the code for
codevector A,
Q: predictive coefficient (scalar),
D(i): decoded vector of one previous
15 frame, and
is vector order.
Also, although amplitude stored in the amplifier
storing section 908 is preset, the setting method is set
forth below. The amplitude is set by coding much speech
20 data is coded, obtaining the sum of the coded distortions
of the following expression (20), and performing the
training such that the obtained sum is minimized.
f 2
EN = ;(Y, (i)-C1N(i)-aNC2m, (i)) = = = (20)
where EN: coded distortion when the code for
25 codevector A is N,
N: code for codevector A,
t: time when the code for codevector A is
CA 02598683 2007-09-10
56
N,
Yt(I) : predictive error vector at time t,
C1N(i) decoded codevector A,
aN: amplitude corresponding to the code for
codevector A,
C2mt(i): codevector B,
is vector order, and
I: vector length.
In other words, after coding, amplitude is reset
such that the value, which has been obtained by
differentiating the distortion of the above expression
(20) with respect to each amplitude, becomes zero, thereby
performing the training of amplitude. Then, by the
repetition of coding and training, the suitable value of
each amplitude is obtained.
On the other hand, the decoder performs the decoding
by obtaining the codevector based on the code of the vector
transmitted. The decoder comprises the same vector
codebooks (corresponding to codebooks A, B) as those of
the coder, the amplifier storing section, and the state
storing section. Then, the decoder carries out the
decoding by the same algorithm as the decoding function
of the searching section (corresponding to the codevector
B) in the aforementioned coding algorithm.
Therefore, according to the above-mentioned
embodiment, by the characteristics of the amplifier
storing section and the distance calculator, the code
vector of the second stage is applied to that of the first
CA 02598683 2007-09-10
57
stage with a relatively small amount of calculations,
thereby the coded distortion can be reduced.
The above explained the case in which the present
invention was applied to the low bit rate speed coding
technique used in such as a cellular phone. However, the
present invention can be employed in not only the speech
coding but also the vector quantization for a parameter
having a relatively good interpolation in a music coder
and an image coder.
(Seventh embodiment)
Next, the following will explain the CELP speech
coder according to the sixth embodiment. This embodiment
shows an example of a coder; which is capable of reducing
the number of calculation steps for vector quantization
processing for ACELP type random codebook.
FIG. 13 shows the functional block of the CELP speech
coder according to this embodiment. In this CELP speech
coder, a filter coefficient analysis section 1002
provides the linear predictive analysis to input speech
signal 1001 so as to obtain coefficients of the synthesis
filter, and outputs the obtained coefficients of the
synthesis filter to a filter coefficient quantization
section 1003. The filter coefficient quantization
section 1003 quantizes the input coefficients of the
synthesis filter and outputs the quantized coefficients
to a synthesis filter .1004.
The synthesis filter 1004 is constituted by the
filter coefficients supplied from the filter coefficient
CA 02598683 2007-09-10
58
quantization section 1003. The synthesis filter 1004 is
excited by an excitation signal 1011. The excitation
signal 1011 is obtained by adding a signal, which is
obtained by multiplying an adaptive codevector 1006, 1. e. ,
an output from an adaptive codebook 1005, by an adaptive
codebook gain 1007, and a signal, which is obtained by
multiplying a random codevector 1009, i . e . , an output from
a random codebook 1008, by a random codebook gain 1010.
Here, the adaptive codebook 1005 is one that stores
a plurality of adaptive codevectors, which extracts the
past excitation signal for exciting the synthesis filter
every pitch cycle. The random codebook 1007 is one that
stores a plurality of random codevectors. The random
codebook 1007 can use the excitation vector generator of
the aforementioned first embodiment.
A distortion calculator 1013 calculates a
distortion between a synthetic speech signal 1012, i.e.,
the output of the synthesis filter 1004 excited by the
excitation signal 1011, and the input speech signal 1001
so as to carry out code search processing. The code
search processing is one that specifies the index of the
adaptive codevector 1006 for minimizing the distortion
calculated by the distortion calculator 1013 and that of
the random gain 1009. At the same time, the code search
processing is one that calculates optimum values of the
adaptive codebook gain 1007 and the random codebook gain
1010 by which the respective output vectors are
multiplied.
CA 02598683 2007-09-10
59
A code output section 1014 outputs the quantized
value of the filter coefficients obtainable from the
filter coefficient quantization section 1003, the index
of the adaptive codevector 1006 selected by the distortion
calculator 1013 and that of the random codevector 1009,
and the quantized values of adaptive codebook gain 1007
and random codebook gain 1009 by which the respective
output vectors-are multiplied. The outputs from-the code
output section 1014 are transmitted or stored.
In the code search processing in the distortion
calculator 1013, an adaptive codebook component of the
excitation signal is first searched, and a codebook
component of the excitation signal is next searched.
The above search of the random codebook component
uses an orthogonal search set forth below.
The orthogonal search specifies a random vector c,
which maximizes a search reference value Eort
Nort/Dort) of expression (21).
Eort s Nor, a [{(PrH'Hc)x-(x'Hp)Hp}HcJ ... (21)
( Dort ) (ctHtHc)(ptHrHp)-(ptHtHc)2
.20 where Nort: numerator term for Eort,
Dort: denominator term for Eort,
p: adaptive codevector already specified,
H: synthesis filter coefficient matrix,
Ht: transposed matrix for H,
X: target signal (one that is obtained by
differentiating a zero input response
of the synthesis filter from the
CA 02598683 2007-09-10
input speech signal), and
c: random codevector.
The orthogonal search is a search method for
orthogonalizing random codevectors serving as candidates
5 with respect to the adaptive vector specified in advance
so as to specify index that minimizes the distortion from
the plurality of orthogonalized random codevectors. The
orthogonal search has the characteristics in-which a
accuracy for the random codebook search can be improved
10 as compared with a non-orthogonal search and the quality
of the synthetic speech can be improved.
In the ACELP type speech coder, the random
codevector is constituted by a few signed pulses. By use
of the above characteristic, the numerator term (Nort)
15 of the search reference value shown in expression (21)
is deformed to the following expression (22) so as to
reduce the number of calculation steps on the numerator
term. f~ t
Nort {aot/ (~o)+ai (l1)+...+a~_I~V(1~-i)IZ ..: (22)
20 where a1: sign of i-th pulse (+1/-1),
li: position of i-th pulse,
N: number of pulses, and
4 : ((ptHtHp) x- (xtHp) Hp )H .
If the value of 4) of expression (22) is calculated
25 in advance as a pre-processing and expanded to an array,
(N-1) elements out of array 4) are, added or substituted,
and the resultant is squared, whereby the numerator term
of expression (21) can be calculated.
CA 02598683 2007-09-10
61
Next, the following will specifically explain the
distortion calculator 1013, which is capable of reducing
the number of calculation steps on the denominator term.
FIG. 14 shows the functional block of the distortion
calculator 1013. The speech coder of this embodiment has
the configuration in which the adaptive codevector 1006
and the random codevector 1009 in the configuration of
FIG. 13 are input to the distortion calculator 1013.
In FIG. 14, the following three processing is
carried out as pre-processing at the time of calculating
the distortion for each random codevector.
(1) Calculation of first matrix (N): power of
synthesized adaptive codevector (ptHtHp) and auto-
correlation matrix of synthesis filter's coefficients
(HtH) are computed, and each element of the auto-
correlation matrix are multiplied by the above power so
as to calculate matrix N (= (ptHtHp) HtH) .
(2) Calculate second matrix (M)-: time reverse
synthesis is performed to the synthesized adaptive
codevector for producing (ptHtH) , and outer products of
the above resultant signal (ptHtH) is calculated for
producing matrix M.
(3) Generate third matrix (L) : matrix M calculated
in item (2) is subtracted from matrix N calculated in item
(1) so as to generate matrix L.
Also, the denominator term (Dort) of expression (21)
can be expanded as in the following expressions (23).
Dort = (ctHtHc) (ptHtHp) - (ptH`Hc) a ... (23)
CA 02598683 2007-09-10
62
= ctNc - (rtc) 2
= ctNc-(rtc)t(rtc)
= ctNc- (ctr.rtc )
= ctNc- (ctMc )
= ct(N-M)c
= ctLc
where N : (ptHtHp) HtH the above pre-processing (1) ,
r: ptHtH the above pre-processing (2),
M: rrt the above pre-processing (2),
L: N-M the above pre-processing (3),
c: random codevector
Thereby, the calculation of the denominator term
(Dort) at the time of the calculation of the search
reference value (Eort) of expression (21) is replaced with
expression (23), thereby making it possible to specify
the random codebook component with the smaller amount of
calculation.
The calculation of the denominator term is carried
out using the matrix L obtained in the above pre-
processing and the random codevector 1009.
Here, for simplifying the explanation, the
calculation method of the denominator term will be
explained on the basis of expression (23) in a case where
a sampling frequency of the input speech signal is 8000
Hz, the random codebook has Algebraic structure and its
codevectors are constructed by five signed unit pulses
per 10 ms frame.
The five signed unit pulses constituting the random
CA 02598683 2007-09-10
63
vector have pulses each selected from the candidate
positions defined for each of zero to fourth groups shown
in Table 2, then random vector c can be described by the
following expression (24).
C=a,S (k-10) +a1S (k-11) +== +a4S (k-14) === (24)
(k=0, 1, ...79)
where a1: sign (+1/-1) of pulse belonging to
group i, and
li: position of pulse belonging to group i.
Table
Group Code Pulse Candidate Position
Number
0 1 0, 10, 20, 30, 60, 70
1 1 2, 12, 22, 32, 62, 72
2 1 2, 16, 26, 36, 66, 76
3 1 4, 14, 24, 34, ===, 64, 74
4 1 8, 18, 28, 38, 68, 78
At this time, the denominator term (Dort) shown by
expression (23) can be obtained by the following
expression (25):
4 4
Dort = I ;a,a;L(I1,l;) ... (25)
where a1: sign (+1/-1) of pulse belonging to
group i,
li: position of pulse belonging to group i, and
L (li , l): element (li row and 1j column) of matrix L.
As explained above, in the case where the ACELP type random
codebook is used, the numerator term (Nort) of the code search
CA 02598683 2007-09-10
04
reference value of expression (21) can be calculated by
expression (22), while the denominator term (Dort) can
be calculated by expression (25). Therefore, in the use
of the ACELP type random codebook, the numerator term is
calculated by expression (22) and the denominator term
is calculated by expression (25), respectively, instead
of directly calculating of the reference value of expression (21).
This makes it possible to greatly reduce the number of calculation
steps for vector quantization processing of random excitations.
The aforementioned embodiments explained the random code
search with no
pre-selection. However, the same effect as
mentioned above can be obtained if the present invention is applied
to a case in which pre-selection based on the values of expression
(22) is employed, the values of expression (21) are calculated
for only pre-selected random codevectors with expression (22) and
expression (25), then finally selecting one random codevector,
which maximize the above search reference value.
