Note : Les descriptions sont présentées dans la langue officielle dans laquelle elles ont été soumises.
CA 02283203 1999-09-09
1
DESCRIPTION
METHOD AND APPARATUS FOR LAG PARAMETER ENCODING
AND METHOD FOR MAKING CODEBOOK
Technical Field
The present invention relates to a speech
processing apparatus applicable to a digital cellular
telephone and personal computer, etc. and relates in
particular to a method and apparatus for lag parameter
encoding, that is, encoding of lag parameters expressing
pitch period or related parameter which is one of
parameters indicating the features of speech signals,
and a method for making codebooks used for these.
Background Art
One of important parameters expressing the
features of speech signals is pitch period and lag
parameters. These parameters are used as encoded
parameters inspeechencoding processingforefficiently
encoding speech signals and synthesis parameters in
speech synthesis. When transmitting or storing a lag
parameter, it is necessary to encode the parameter value
to a code corresponding to that value according to a
specific rule.
The lag parameter encoding method for speech
encoding is described in the international organization
ITU-T Recommendation 6.729 (8 kbps CS-ACELP speech
CA 02283203 1999-09-09
2
encoding method).
The lag parameter encoded according to the
Recommendation is transmitted together with codes of
other encoded parameters. The lag parameter in this
conventional example is a value ( lag value ) indicating
which segment of a signal contained in a codebook called
"adaptive codebook" is used when making an excitation
signal used for synthesizing a decoded speech in the
CS-ACELP algorithm, a speech encoding algorithm of this
conventional example. This lag value T consists of
integer part T1 (T1=19 to 143) and fraction part frac/3
(frac=-1, 0, 1) .
This lag value T is encoded by a encoder as code
P ( P=0 to 255 ) in expression ( 1 ) below using T1 and frac
above.
3 X (T1-19)+frac-1, T1=19~-85,frac=- 1,0,1
P-
(T1- 85)+197, T1=86~-143,frac=0
... ( 1 )
On the other hand, decoded lag value T1 and frac
are decoded by a decoder based on code P according to
a rule opposite to expression (1).
The lag parameter is an amount of delay from time
tl of a speech signal to t0 preceding T1, at which a
waveform is similar to the waveform at tl. That is, the
lag parameter is typically a parameter indicating a pitch
CA 02283203 1999-09-09
3
period in a periodic waveform and is a pitch period of
speech itself . However, the lag parameter is a parameter
which has a wide concept in a sense that it includes an
amount of delay up to a position where a waveform is simply
similar in non-periodic signal such as the speech onset .
However, in the lag parameter code obtained by the
conventional lag parameter encoding method above, if a
bit error occurs in the process of transmission or
storage, the decoded lag value is by far different from
the correct lag value free of errors, which may cause
great deterioration of the decoded speech.
Generally, one of the methods for suppressing
deterioration of the quality due to bit errors in a code
is to provide a certain correlation between distortion
like an Euclidean distance between parameter values of
encoded parameters and a distance(Hamming distance)
between codes indicating those parameter values and
reduce an influences of a bit error.
If an Euclidean distance and differential value,
etc. between those lag values are used as a measure of
distortion between the parameter values of lag
parameters, they are valid as long as such a value is
small. However, if the value exceeds a certain value,
it is no longer possible to keep the correspondence with
perceptual distortion and the use of the general method
above is not so effective for encoding/decoding
processing of the lag parameters.
To handle such bit errors, a method of detecting
CA 02283203 1999-09-09
4
bit errors and preventing use of lag values containing
errors is also available, but error detection of this
method itself is complicated, and moreover adding a
redundant bit such as a check bit to a low bit rate
communication method such as speech communication is not
appropriate.
The present invention has been made in view of such
a situation and it is an objective of the present
invention to provide an excellent method and apparatus
for encoding lag parameters and a method for making
codebooks capable of suppressing, in the event of a bit
error with lag parameter codes, deterioration of the
perceptual speech quality caused thereby.
Disclosure of Invention
To solve the above problem, the present invention
performs encoding of lag parameters using a codebook
defined as shown below. The codebook is generated in
such a way that the rate of lag parameter code, which
has a decoded lag parameter value with a bit error close
to n times ( including one time ) or 1/n times ( n is integer )
of the decoded value without bit error, increases.
The codebook is generated in such a way that a total
sum of distortion of decoded values between codes with
Hamming distance within a prescribed number of bits in
the codebook is minimized or nearly minimized, and by
using such a distortion measure that distortion is
measured smaller between one decoded lag parameter value
CA 02283203 1999-09-09
and another value with n times or 1/n times (n is integer)
of that value.
As a result, the codebook is generated in such a
way that the rate of lag parameter code, which has a
5 decoded lag parameter value with a bit error close to
n times (including one time) or 1/n times(n is integer)
of the decoded value without bit error, increases,
allowing the speech signal to be encoded/decoded with
less deterioration of the perceptual speech quality.
Brief Description of Drawings
FIG.1 is a schematic block diagram of a radio
transmitter to which a method and apparatus for lag
parameter encoding of Embodiment 1 of the present
invention is applied;
FIG.2 is a schematic block diagram of a speech
encoding section of the radio communication apparatus
of Embodiment 1;
FIG.3 is a block diagram of the main part of the
speech encoding section of the radio communication
apparatus of Embodiment 1;
FIG.4 is a block diagram of the main part of a speech
decoding section of the radio communication apparatus
of Embodiment 1; and
FIG.5 is a drawing showing the procedure for making
a codebook applicable to the radio communication
apparatus of Embodiment 1.
CA 02283203 1999-09-09
6
Best Mode for Carrying out the Invention
(Embodiment 1)
With reference to FIG.1 to FIG.5, Embodiment 1 of
the present invention is explained below.
FIG.1 is a schematic block diagram of a radio
transmitter to which the present invention is applied.
Speech transmission processing is carried out as
follows : A speech signal input from microphone 101 is
converted from analog to digital by A/D converter 102 ,
output to speech encoding section 103 and encoded
according to a CELP algorithm, for example. The encoded
output is modulated by modulator/demodulator 104
according to a CDMA system, etc., and transmitted via
radio transmission section 105 and antenna 106.
Speech reception processing is carried out as
follows : A modulated signal received via antenna 107 and
radio reception section 108 is demodulated by
modulator/demodulator 104, then decoded by speech
decoding section 109, converted from digital to analog
by D/A converter 110 and output from speaker 111 as a
speech.
The present invention is applied to part of adaptive
codebook search processing used in speech encoding
section 103 and speech decoding section 109 of the radio
communication apparatus above.
FIG.2 is a schematic block diagram of speech
encoding section 103 of the radio communication
apparatus and shows a general configuration of a CELP
CA 02283203 1999-09-09
7
type speech encoder/decoder. The A/D converted speech
signal is input from terminal 201 and output to LPC
analyzing section 202. LPC analyzing section 202
carries out a linear predictive analysis based on the
input speech signal and outputs linear predictive
coefficients. LPC parameter quantization section 203
quantizes the linear predictive coefficients(L) and
outputs the quantization result to synthesis filter 204
and multiplexer 205.
Synthesis filter 204 constitutes a filter with a
given characteristic according to the linear predictive
coefficients above, filters an excitation signal input
from adder 206 and outputs the result to adder 207. This
adder 207 calculates an error between the input speech
signal from terminal 201 and the output from synthesis
filter 204 and outputs the error signal to perceptual
weighting section 208. Perceptual weighting section
208 carries out weighting processing corresponding to
the perceptual sense of the error signal and outputs the
result to error minimization section 209.
Error minimization section 209 selects code
vectors for adaptive codebook 210 and fixed excitation
codebook 211 so that the error signal output from
perceptual weighting section 208 may be minimized and
selects a gain for gain codebook 212 as well.
Adaptive codebook 210 is an excitation signal table
which stores past excitation vectors and selectively
outputs a specific code vector selected by error
CA 02283203 1999-09-09
8
minimization section 209. Multiplier 213 multiplies
the output by a gain select by gain codebook 212 and
outputs the result to adder 206.
By the way, this adaptive codebook 210 includes a
buffer storing a history for a certain period of the
excitation vector finally determined as the output of
adder 206 and outputs a lag value indicating which
segment of the signal sequence stored in said buffer
should be extracted to lag parameter encoding section
215 according to the code vector selected by error
minimization section 209. This lag parameter encoding
section 215 includes lag parameter codebook 215a
generated beforehand according to a prescribed rule and
lag parameter encoding section 215b and encodes the lag
value of adaptive codebook 210 according to a certain
rule and outputs it to multiplexer 205. This lag
parameter encoding section 215 will be described in
detail later.
Fixed excitation codebook 211 selectively outputs
a specific fixed excitation code vector selected by error
minimization section 209. Multiplier 216 multiplies
the output by a gain set by gain codebook 212 and outputs
the result to adder 206.
Adder 206 finds a sum of the outputs of multiplier
213 and multiplier 216 and outputs it to synthesis filter
204 as an excitation vector. At the same time, the output
is fed back to adaptive codebook 210 and accumulated.
In this way, error minimization section 209
CA 02283203 1999-09-09
9
measures error signals for all excitation vectors stored
in adaptive codebook 210 and outputs the output ( P ) of
lag parameter encoding section 215b , the output ( S ) of
fixed excitation codebook 211 and the output (G) of gain
codebook 212 corresponding to a minimum value of the
error signal from perceptual weighting section 208 to
multiplexes 205. Multiplexes 205 multiplexes quantized
linear predictive coef f icient ( L ) and outputs ( P ) , ( S )
and (G) above and outputs the result to modulator 104
in FIG.1.
Speech decoding section 110 of the radio
communication apparatus (FIG.1) also consists of a
general CELP type speech decoder, but its explanation
is omitted here.
Then, lag parameter encoding section 215 to which
the present invention is applied is explained in detail
below.
FIG.3 and FIG.4 show a configuration of the main
part of lag parameter encoding section 215 to which the
present invention is applied and FIG. 3 shows a
functional block on the lag parameter encoding section
side and FIG.4 shows a functional block on the lag
parameter decoding section side. Such a lag parameter
encoding section is not limited to cellular telephones
but is applicable to all apparatuses carrying out speech
encoding/decoding.
As shown in FIG . 3 , lag parameter encoding section
215 comprises lag parameter codebook 215a and lag
CA 02283203 1999-09-09
parameter encoding section 215b that encodes lag values
with reference to this lag parameter codebook 215a. Lag
parameter codebook 215a is a table that stores input lag
values with corresponding output codes and is generated
5 beforehand according to a certain rule.
In the same way, as shown in FIG . 4 , the lag parameter
decoding section of the speech decoder comprises same
lag parameter codebook 215a as that in the lag parameter
encoding section above and lag parameter decoding
10 section 401 that decodes lag parameters corresponding
to the encode received/input with reference to this lag
parameter codebook 215a.
Lag parameter encoding section 215 configured as
shown above is further explained in detail.
Lag parameter codebook 215a is a table that shows
a relationship between lag parameter value T and
corresponding code P. For example, if the codebook size
is N, lag value T corresponding to code P (=0 to N-1)
is stored. Furthermore, it is also possible to find
intermediate code PO (0 to N-1) from a calculation
expression such as expression (1) for lag parameter
encoding in the ITU-T Recommendation 6.729 (8 kbps
CS-ACELP) mentioned in the conventional technology and
store a table of correspondence with final code P (=0
to N-1) corresponding to P0.
Lag parameter codebook 215a of the present
invention is characterized by a configuration such that
the codebook is generated in such a way that the rate
CA 02283203 1999-09-09
11
of lag parameter code , which has a decoded lag parameter
value with a bit error close to n times ( including one
time) or 1/n times(n is integer) of the decoded value
without bit error, increases. The method of its
generation will be explained later.
A lag parameter is a parameter related to a pitch
period included in a speech signal. There are cases
where a correct lag value is not obtained due to a bit
error, etc. However, the inventor of the present
invention has found that if a wrong decoded lag value
is close to n times ( including one time ) or 1 /n times ( n
is integer) of the correct lag value, perceptual
deterioration can be relatively small. This is because
the spectrum of the speech signal decoded or synthesized
using that wrong lag value includes the frequency
component of the correct pitch period as its part as long
as the above condition is satisfied.
As shown above , Embodiment 1 is capable of reducing
deterioration of the perceptual speech quality when a
bit error occurs with a code taking advantage of the
characteristic that perceptual deterioration is small
in the case of a lag parameter with a bit error having
a value close to n times ( including one time ) or 1/n times
(n is integer) of the correct decoded value.
Then, the method of making the lag parameter
codebook above used in the present invention is explained.
This lag parameter codebook is generated in such a way
that the rate of lag parameter code, which has a decoded
CA 02283203 1999-09-09
12
lag parameter value with a bit error close to n times
(including one time) or 1/n times(n is integer) of the
decoded value without bit error, increases.
FIG. 5 shows the processing procedure for making the
lag parameter codebook above.
First, initial codebook Table(i) (i=0 to N-1; N:
codebook size) is set in step 501. Here, Table(i)
denotes a decoded value ( either a scalar value or vector
value). If this codebook is a lag parameter codebook,
Table ( i ) may be set to indicate intermediate code PO for
code i as explained in lag parameter codebook 215a of
Embodiment 1. Furthermore, the correspondence between
a code and decoded value in the initial codebook can be
determined arbitrarily.
Then, in step 502, for all combinations of codes
whose Hamming distance dH is within a specific number
of bits (suppose it is MB) (dH S MB) in Table(i),
distortion of the decoded value between those combined
codes is calculated one by one and total sum DO is
obtained.
Here, distortion of the decoded value between codes
depends on the parameter indicated by the code , but an
Euclidean distance between decoded values or the like
is used. The method of expressing a distortion measure
of lag parameters is one of the features of the present
invention. It will be described further in Embodiment
2.
Then, in step 503, a code pair i_a and i_b whose
CA 02283203 1999-09-09
13
Hamming distance dH exceeds said specific number of bits
MB ( dH>MB ) is randomly selected from codebook Table ( i ) .
In step 504, after the decoded values are mutually
exchange between said code pair, total sum D of
distortion of decoded values between codes with a Hamming
distance within said specific number of bits is
calculated.
Then in step 505, it is judged whether total sum
D of distortion in step 504 is smaller than total sum
Do of distortion calculated before. If it is smaller,
the decoded values are exchange between said code pair
and the total sum of distortion is updated in step 506.
In step 507, the convergence of said total sum DO
of distortion is judged and the operations from said
steps 503 to 507 are repeated until said total sum of
distortion converges.
Making the lag parameter codebook using the
processing above can reduce the total sum of distortion
measure of decoded values between codes within a specific
Hamming distance, making the decoded values when a bit
error occurs with a code closer to the correct decoded
value, suppressing deterioration of the perceptual
speech quality.
Especially, by limiting the minimization of the
total sum of distortion to between codes with a Hamming
distance within a specific number of bits, it is possible
to suppress deterioration more effectively when a bit
error occurs with fewer bits occurs. By randomly
CA 02283203 1999-09-09
14
selecting code pair i_a and i_b whose Hamming distance
exceeds a specific number of bits, it is possible to
achieve higher efficiency and reduce the total sum of
distortion. Thus, even if a bit error occurs it is
possible to suppress deterioration of the perceptual
speech quality.
In step 503 above, code pair i_a and i_b selected
randomly from codebook Table (i) is limited to those
whose Hamming distance exceeds a specific number of bits,
but the present invention is not limited to this.
(Embodiment 2)
Embodiment 2 is implemented on the same hardware
and software as those in Embodiment 1. The difference
from the method for making the lag parameter codebook
applied in Embodiment 1 is a change in the distortion
measure.
The procedure for making a codebook is the same as
that in FIG.5 shown in Embodiment 1. What is different
from Embodiment 1 is the use of a measure shown in
expression ( 2 ) as distortion of decoded values between
codes used in steps 502 and 504.
CA 02283203 1999-09-09
d{fa,fb)=min(wl X d0(fb,fa), w2 x d0(fb,2 X fa), w3 X d0(fb,3 X fa))
where fa=Fs/Ta(Hz) ~~~ { 2 )
fb=Fs/Tb (Hz)
fb z fa
d0(fx,fy)= I fx-fy I /(fx x fy)' ~ z
Where, Ta and Tb are decoded lag values (unit:
sample) of target codes i_a and i_b; fa and fb, frequency
values ( Hz ) for Ta and Tb; Fs , sampling frequency ( Hz ) ;
5 and d(fa, fb) , distortion of decoded values between code
pairs.
Expression ( 2 ) does not simply express distortion
of lag parameter values by anything similar to an
Euclidean distance. Expression (2) is an example of
10 definition taking account of a difference between one
lag value and n times(n is integer) of another lag
value(wl, w2 and w3 are weighting constants
corresponding to distortion from n times(n is integer)
of that value) and another definition implementing a
15 similar concept can also be used.
Using such a distortion measure, one decoded value
of codes with a Hamming distance within a specific number
of bits becomes a value close to n times(n is integer)
of the other decoded value. As already explained, the
lag parameter is a parameter related to a pitch period
contained ~in the speech signal . If the decoded lag value
is close to n times ( including one time ) or 1/n times ( n
CA 02283203 1999-09-09
16
is integer) of the correct lag value due to a bit error,
etc., the spectrum of the speech signal decoded or
synthesized using that value contains the frequency
component of the correct pitch period as a part, and
therefore perceptual deterioration can be relatively
small.
It is possible to further reduce the total sum of
distortion by defining a value close to such n times(n
is integer) as small distortion and making a codebook
through minimization of distortion by limiting to
between codes with a Hamming distance within a specific
number of bits. Therefore, if a lag parameter codebook
is generated by the method described above, it is
possible to suppress deterioration of the perceptual
quality more effectively in the event of a bit error even
for parameters like lag parameters susceptible to
dislocation of decoded values due to the error.
As described above, the invention of the lag
parameter encoding method of the present invention is
a method for encoding lag parameters which are parameters
for encoding speech signals , and is intended to encode
lag parameters using a lag parameters codebook generated
in such a way that the rate of lag parameter code , which
has a decoded lag parameter value with a bit error close
to n times (including one time) or 1/n times(n is integer)
of the decoded value without bit error, increases.
In the invention of the lag parameter decoding
method of the present invention, the lag parameter coded
CA 02283203 1999-09-09
17
by the encoding side using the encoding method described
above is decoded using the same lag parameter codebook
as that on the encoding side.
As shown above, it is possible to suppress
deterioration of the perceptual speech quality when a
bit error occurs with a code by taking advantage of the
characteristic that perceptual deterioration is small
in the case of a lag parameter with a bit error having
a value close to n times ( including one time ) or 1/n times
(n is integer) of the correct decoded value by using a
codebook generated in such a way that the rate of lag
parameter code, which has a decoded lag parameter value
with a bit error close to n times ( including one time )
or 1 /n times ( n is integer ) of the decoded value without
bit error, increases.
Furthermore, the invention of the codebook
generation method of the present invention is the method
for making a codebook in such a way that a total sum of
distortion of decoded values between codes with Hamming
distance within a prescribed number of bits in the
codebook is minimized or nearly minimized. Setting the
decoded value when a bit error occurs with a code to a
value close to the correct decoded value can suppress
deterioration of the perceptual speech quality and
limiting the target of minimization of the total sum of
distortion to between codes with a Hamming distance
within a specific number of bits can more effectively
suppress deterioration of the speech quality when a bit
CA 02283203 1999-09-09
18
error occurs with fewer bits.
When making the codebook above, the initial
codebook can be designed to comprise the steps of
calculating a total sum of distortion of decoded values
between codes with a Hamming distance within a prescribed
number of bits , randomly selecting a code pair from the
codebook, calculating a total sum of distortion of
decoded values between codes with a Hamming distance
within a prescribed number of bits after exchanging the
decoded values between said code pair, exchanging said
decoded values and updating the total sum of distortion
if said total sum of distortion of decoded values is
smaller than said total sum of distortion calculated
before and judging the convergence of said total sum of
distortion, and repeat said steps of randomly selecting
a code pair, exchanging decoded values and updating a
total sum of distortion and judging the convergence of
a total sum of distortion until said total sum of
distortion converges.
It is also preferable to use such a distortion
measure that distortion is measured smaller between one
decoded lag parameter value and another value with n
times or 1/n times (n is integer) of that value, and using
such a distortion measure makes it possible to suppress
deterioration of the perceptual speech quality when a
bit error occurs with a code using the characteristic
that perceptual deterioration is small in the case of
a lag parameter with a bit error having a value close
CA 02283203 1999-09-09
19
to n times(including one time) or 1/n times (n is integer)
of the correct decoded value.
Furthermore, it is possible to realize a lag
parameterencoding/decodingmethodforencoding/decoded
lag parameters using the encoding and decoding methods
above or a codebook generated by one of the codebook
generation methods above.
It is also possible to implement the present
invention as a speech encoder comprising a codebook
indicating the correspondence between the parameter
values of lag parameters which are encoded parameters
of speech signals and codes and a lag parameter encoder
that encodes lag parameters using said codebook. It is
further possible to implement the present invention as
a speech decoder comprising a lag parameter decoder that
decodes lag parameters codes encoded by the above encoder
using the same codebook as that on the encoding side.
Furthermore, it is possible to implement the lag
parameter encoder/decoder with a single apparatus.
It is also possible to implement the encoding method
above by computer software. More specifically, it is
possible to configure a system that comprises a
computer-readable medium and a programinstruction means
for instructing a computer processor to encode lag
parameters using a lag parameter codebook generated in
such a way that the rate of lag parameter code, which
has a decoded lag parameter value with a bit error close
to n times ( including one time ) or 1/n times ( n is integer )
CA 02283203 1999-09-09
of the decoded value without bit error, increases,
wherein said program instruction means is stored in said
medium in an executable format and loaded to a computer
memory when executed by said processor to operate the
5 computer.
Of course, the decoding method above can also be
implemented by computer software likewise.
It is also possible to use the encoding software
above by storing it in various types of storage medium.
10 It is a mechanically readable storage medium storing a
program that instructs the computer to encode lag
parameters using a lag parameter codebook generated in
such a way that the rate of lag parameter code, which
has a decoded lag parameter value with a bit error close
15 to n times ( including one time ) or 1/n times ( n is integer )
of the decoded value without bit error, increases. It
is then downloaded to the computer to operate the
computer, thus implementing the encoding method above.
Of course, the decoding software above can also be
20 used by storing it in various types of storage medium
likewise.
The present invention can also be implemented as
a codebook generation apparatus that comprises a
computer-readable medium and a programinstruction means
that instructs a computer processor to generate a
codebook in such a way that a total sum of distortion
of decoded values between codes with a Hamming distance
within a prescribed number of bits in the codebook is
CA 02283203 1999-09-09
21
set to a minimum value or a value close to a minimum value,
wherein said program instruction means is stored in said
storage medium in an executable format and loaded to a
computer memory when executed by said processor to
operate the computer.
The present invention is applicable when no error
detection is performed, and of course can be used with
error detection as well. It is further applicable to all
speech encoding/decoding methods carrying out encoding
of lag parameters.
This application is based on the Japanese Patent
Application No.HEI 10-29332 filed on January 27, 1998,
entire content of which is expressly incorporated by
reference herein.
Industrial Applicability
The encoder, decoder and encoding and decoding
methods of the present invention are applicable to a wide
range of equipment equipped with a speech encoder and
speech decoder. The use of the present invention for a
radio communication apparatus such as digital cellular
telephone is particularly preferable because it can
efficiently suppress deterioration of the perceptual
speech quality.
