United States Patent [191
`Taniguchi et al.
`
`I lllll llllllll Ill lllll lllll lllll lllll lllll 111111111111111111111111111111111
`US005115469A
`5,115,469
`[11] Patent Number:
`[45] Date of Patent: May 19, 1992
`
`[75]
`
`[54] SPEECH ENCODING/DECODING
`APPARATUS HAVING SELECTED
`ENCODERS
`Inventors: Tomohiko Taniguchi, Yokohama;
`Kohei Iseda; Koji Okazaki, both of
`Kawasaki; Fumio Amano, Tokyo;
`Shigeyuki Unagami, Atsugi;
`Yoshinori Tanaka, Yokohama; Yasuji
`Ohta, Kawasaki, all of Japan
`[73] Assignee: Fujitsu Limited, Kawasaki, Japan
`[21] Appl. No.:
`460,099
`[22] PCT Filed:
`Jun. 7, 1989
`PCT I JP89/00580
`[86] PCT No.:
`§ 371 Date:
`Feb. 8, 1990
`§ 102(e) Date:
`Feb. 8, 1990
`[87] PCT Pub. No.: W089/12292
`PCT Pub. Date: Dec. 14, 1989
`Foreign Application Priority Data
`[30]
`Jun. 8, 1988 [JP]
`Japan ................................ 63-141343
`Mar. 14, 1989 [JP]
`Japan .................................... 1-61533
`[51]
`Int. Cl.s ................................................ GOlL 5/00
`[52] U.S. Cl. ......................................... 381/36; 381/34
`[58] Field of Search .................................... 381/29-40;
`364/513.5; 375/22-24, 34, 122
`References Cited
`U.S. PATENT DOCUMENTS
`3,067,291 12/1962 Lewinter ............................... 381/31
`
`[56]
`
`3,903,366 9/1975 Coulter ................................. 381/38
`4,005,274 1/1977 Vagliani et al. ...................... 381/32
`4,303,803 12/1981 Yatsuzuka ............................. 581/31
`4,546,342 10/1985 Weaver et al. .................... .-. 375/122
`4,622,680 11/1986 Zinser ................................... 381/31
`Primary Examiner-Emanuel S. Kemeny
`Assistant Examiner-Michelle Doerrler
`Attorney, Agent, or Firm-Staas & Halsey
`ABSTRACT
`[57]
`Several encoders perform a local decoding of a speech
`signal and extract excitation information and vocal tract
`information from a speech signal for an encoding opera(cid:173)
`tion. The transmission rate ratio between the excitation
`information and the vocal tract information are differ(cid:173)
`ent for each encoder. An evaluation/selection unit eval(cid:173)
`uates the quality of decoded signals subjected to a local
`decoding in each of the encoders, determines the most
`suitable encoders from among the several encoders
`based on the result of the evaluation, and selects the
`most suitable encoder, thereby outputting the selection
`result as selection information. The decoder decodes a
`speech signal based on selection information, vocal
`information. The
`tract information and excitation
`evaluation/selectiop unit selects the output from the
`encoder in which the quality of a locally decoded signal
`is the most preferable. When voc11I tract information
`changes little, the vocal tract information is not output,
`thereby allowing for increased quality of information.
`As much of the surplus of unused vocal tract informa(cid:173)
`tion as possible is assigned to a residual signal. Thus, the
`quality of a decoded speech signal is improved.
`
`12 Claims, 7 Drawing Sheets
`
`303
`INPUT
`
`ENCODER # 2
`
`301-m
`ENCODER #m
`307-m
`DECODED
`SIGNALS
`
`308
`
`309
`
`DECODER
`
`#n
`
`w
`z
`...J
`
`-
`
`0:::
`0
`I-
`(.)
`w
`...J
`w
`
`(/)
`
`.~
`
`306-m
`302-2
`
`307-1
`
`307
`-2
`
`302-1
`EVALUATION & DECISION
`OF OPTIMUM ENCODER
`
`310
`
`310 SELECT
`INFORMATION
`
`IPR2016-01710
`UNIFIED EX1012
`
`

`
`U.S. Patent
`
`May 19, 1992
`
`Sheet 1 of 7
`
`5,115,469
`
`01-
`lJJ :::>
`Oo..
`
`(!)
`0
`
`C\J
`0
`
`-Cl
`
`0:::
`f:?
`u
`lJJ a:::
`a..
`
`t-
`a:::
`-
`.
`<{
`(!) a::
`-o
`LL.~
`a..
`
`01-u :::> Zo
`lJJ
`r - - - - - - - --------,
`I
`x
`¢
`I
`0
`:::>
`I
`~
`I
`z en
`oa::
`_w
`I
`t-t-
`I
`uw
`I
`o:?
`I
`LLJ <(
`a:: a::
`I
`Q.<(
`I
`Q.
`I
`0
`I
`t<)
`0
`0
`I
`~
`I
`I
`I
`I
`I
`I
`I
`I
`.....
`I
`Zz
`I
`I
`Cl: 0 :::>
`I
`<t t;cn
`I
`w -
`I
`I
`zocn
`:JW ~ 0
`I
`I o:::
`I w
`I
`Cl: <(
`18
`O.z
`I
`<t
`I~
`I
`~--- _.:_ _______ _J
`
`..J
`<( ..J
`::> <(
`Oz
`
`-Cf) l!>
`
`lJJ -
`0::: Cf)
`
`LO
`0
`
`.J
`t- <(
`=>z
`
`a,;-zcn -
`
`

`
`\0
`O'\
`
`... .s;;.
`UI
`.,...
`... .,...
`UI
`
`......
`Q ....
`N
`.....
`l'T>
`00 =(cid:173)
`l'T>
`
`~ ....
`'< ....
`~
`
`~
`
`N
`\C
`\C
`
`('!) =
`'"'d =
`•
`00
`~ •
`
`f""f'-
`
`f""f'-
`
`PRIOR ART
`
`FIG. 2
`
`CN
`
`212
`
`EVALUATION
`POWER
`ERROR
`
`211
`
`202
`
`FUNCTION
`
`I >t WEIGHT
`
`I
`
`QUANTIZER
`
`I
`
`204 I
`
`2067 207(Cp)
`
`203
`
`(z-m)l--f'>J 205(G)
`
`PREDICTOR
`
`PATTERN
`RESIDUAL
`
`208
`
`209
`
`210
`
`201
`
`ANALYSIS
`LP C
`
`SIGNAL
`
`INPUT
`
`

`
`°' \C>
`.t:;.
`...
`(II
`.....
`... .....
`
`(II
`
`0 ....
`w
`00 =(cid:173)~ a
`
`........
`
`~ ....
`'< ....
`~ =
`
`~
`\C
`\C
`
`('!) =
`"'C =
`•
`00
`~ •
`
`f"'I"
`
`f"'I"
`
`UTPUT
`
`0
`31
`
`INFORMATION
`
`310 SELECT
`
`~
`
`I 310
`
`OF OPTIMUM ENCODER
`EVALUATION & DECISION
`30211
`
`302~ 2
`
`-m
`
`300
`
`~ -2
`307---J t---30 7-I
`---
`~ ·
`
`SIGNALS
`
`DECODED
`
`307-m
`
`~ DECODER #n
`
`309}
`
`a::
`
`ENCODER #2 hi~ ~~·~
`
`~
`
`_J
`
`Cf)
`'W
`_J
`
`I
`I
`
`-301-211306-2:
`
`'
`
`I
`I
`
`301-m
`
`ENCOD~R nm I II~
`
`lf308
`
`INPUT • ~EN CODER # I
`
`.301-1
`
`303
`
`PARAMETERS
`304 VOCAL TRACT
`
`PARAMETERS
`
`305 EXCITATION
`
`FIG. 3
`
`

`
`U.S. Patent
`
`May 19, 1992
`
`Sheet 4 of 7
`
`5,115,469
`
`FIG. 4
`
`403-1
`
`RESIDUAL
`QUANTIZER 1-6_B~IT1-S ___ S_l_GN_A_L __ ~ 6 BITS
`/FRAME
`408
`
`406-1
`
`404-1
`
`PREDICTOR
`
`405-1
`INPUT SIGNAL
`
`401
`
`402
`
`LPC
`ANALYSIS
`
`LPC
`PARAMETERS
`407-1
`SN (A)
`-----------EVALUATION---
`
`2BITS
`/FRAME
`409
`
`405-2
`
`403-2
`
`RESIDUAL SIGNAL
`
`QUANTIZER BBlTS
`1 - - - - - - - -1___ . - " ' ( )
`
`BBITS
`/FRAME
`412
`
`....------~+ 406-2
`404-2
`
`~13
`
`PREDICTOR
`
`411
`
`MODE
`DETERMINA
`-TION
`
`FRAME
`DELAY
`
`A
`fa
`410
`......_ ________ __.,.EVALUATION
`......_ ____ SN (B)
`
`407-2
`
`

`
`ft> -UI
`
`00 =(cid:173)
`
`ft>
`
`0 .....
`
`N
`\Q
`\Q
`~
`
`~ =
`
`~\Q
`~
`'-<
`
`~ a (D = «-1'-
`
`•
`00
`~ •
`
`503
`
`DMUX
`
`505
`
`503
`
`REPRODUCED .
`
`SIGNAL
`
`(/)
`w
`(.) w _.
`
`1-
`0
`0:::
`
`~ °' \0
`...
`U1
`~
`~
`...
`U1
`
`--.l
`
`INFOMATION i
`
`NUMBER
`ENCODER
`
`I
`_____ '---____ _J
`I
`I
`
`FIG. 5
`
`502
`
`---------------
`
`OMPUTING CEPSTRUM DISTANC
`CD
`
`LPC
`
`507-2
`
`INFORMATION
`
`SIGNAL
`RESIDUAL
`
`---iTso~-2
`
`I
`I
`BOOK
`I
`I
`CODE
`NOISE :
`WHITE I
`
`MUX
`
`--------;
`
`t-+----' I PARAMETERS
`
`IPREDlCTION
`
`soa-2..----
`
`-TlON
`
`EVALUA
`ERROR
`512-2
`
`,..,,,.~~
`
`515
`
`sis--
`L_ ________ S /NA
`
`J I
`
`I
`
`51~ B
`L
`1
`
`L
`I
`I
`I
`
`:
`I
`:
`I
`
`I
`:
`
`( ~
`517
`
`~gg~ I INFOR~1J~
`NOISE I SIGNAL
`WHITE I RESIDUAL
`
`I
`I
`
`~F'
`
`:
`
`..----
`
`507-1
`
`!
`
`508 -I
`
`_
`T~ou
`LONG-'-
`509-1
`
`P
`~
`m
`512-f 5i
`
`-TTON
`EVALUA
`ERROR
`
`---REPROOUCW-----------------
`
`_}501-1
`
`w
`a:
`
`ANALYSIS
`PREDICTION
`LINEAR
`
`I
`I
`r---
`I
`I
`I
`I
`I
`I
`
`510-1
`
`511-1
`SIGNAL A
`
`-
`
`(SPEECH,
`SIGNAL I
`I
`INPUT
`r
`
`

`
`U.S. Patent
`
`May 19, 1992
`
`Sheet 6 of 7
`
`5,115,469
`
`START
`
`YES
`
`NO
`
`YES
`
`S2
`
`NO
`
`54
`
`A-MODE
`
`8-MODE
`
`FIG. 6
`
`

`
`U.S. Patent
`
`May 19, 1992
`
`Sheet 7 of 7
`
`5,115,469
`
`CODE NUMBER
`
`1600 bps
`
`GAIN
`
`PITCH FREQUENCY
`
`PITCH COEFFI ClENT
`
`LPC PARAMETERS
`TOTAL
`
`1000
`
`600
`
`600
`
`1000
`
`"
`•
`
`"
`
`•
`
`4800 bps
`
`FIG. 7A
`PRIOR ART
`
`A-MODE
`
`8-MODE
`
`CODE NUMBER
`
`1600_ bps
`
`GAIN
`
`PITCH FREQUENCY
`
`PITCH COEFFICIENT
`LPC PARAMETERS
`MODE SIGNAL
`
`1000
`
`600
`
`600
`950
`50
`
`H
`
`"
`
`"
`,,
`
`,,
`
`2200 bps
`,,
`
`1350
`
`600
`
`600
`
`"
`
`.,
`
`50
`
`,,
`
`TOTAL
`
`4800bps
`
`4800 bps
`
`FIG. 78
`
`

`
`1
`
`5, 115,469
`
`SPEECH ENCODING/DECODING APPARATUS
`HAVING SELECTED ENCODERS
`
`JO
`
`15
`
`BACKGROUND OF THE INVENTION
`1. Field of the Invention
`The present invention relates to a speech encoding
`and decoding apparatus for transmitting a speech signal
`after information compression processing has been ap-
`plied.
`Recently, a speech encoding and decoding apparatus
`for compressing speech information to data of about 4
`to 16 kbps at a high efficiency has been demanded for
`in-house communication systems, digital mobile radio
`systems and speech storing systems.
`2. Description of Related Art
`As the first prior art structure of a speech prediction
`encoding apparatus, there is provided an adaptive pre(cid:173)
`diction encoding apparatus for multiplexing the predic(cid:173)
`tion parameters (vocal tract information) of a predictor 20
`and residual signal (excitation information) for transmis(cid:173)
`sion to the receiving station.
`FIG. 1 is a block diagram of an encoder used in the
`speech encoding apparatus of the first prior art struc·
`ture. Encoder 100, comprises linear prediction analysis 25
`unit 101, predictor 102, quantizer 103, multiplexing unit
`104 and adders 105 and 106.
`Linear prediction analysis unit 101 analyzes input
`speech signals and outputs prediction parameters, and
`predictor 102 predicts input signals using an output 30
`from adder 106 (described below) and prediction pa(cid:173)
`rameters from linear prediction analysis unit 101. Adder
`105 outputs error data by computing the difference
`between an input speech signal and the predicted signal,
`quantizer 103 obtains a residual signal by quantizing the 35
`error data, and adder 106 adds the output from predic(cid:173)
`tor 102 to that of quantizer 103, thereby enabling the
`output to be fed back to predictor 102. Multiplexing unit
`104 multiplexes prediction parameters from linear pre(cid:173)
`diction analysis unit 101 and a residual signal from quan- 40
`tizer 103 for transmission to a receiving station.
`With such a structure, linear prediction analysis unit
`101 performs a linear prediction analysis of an input
`signal at every predetermined frame period, thereby
`extracting prediction parameters as vocal tract infon.na· 45
`tion to which appropriate bits are assigned by an en(cid:173)
`coder (not shown). The prediction parameters are thus
`encoded and output to predictor 102 and multiplexing
`unit 104. Predictor 102 predicts an input signal based on
`the prediction parameters and an output from adder 50
`106. Adder 105 computes the error data (the difference
`between the predicted information and the input signal),
`and quantizer 103 quantizes the error data, thereby
`assigning appropriate bits to the error data to provide a
`residual signal. This residual signal is output to multi- 55
`plexing unit 104 as excitation information.
`After that, the encoded prediction parameter and
`residual signal are multiplexed by multiplexing unit 104
`and transmitted to a receiving station.
`Adder 106 adds an input signal predicted by predic- 60
`tor 102 and a residual signal quantized by quantizer 103.
`An addition output is again input to predictor 102 and is
`used to predict the input signal together with the pre(cid:173)
`diction parameters.
`In this case, the number of bits assigned to prediction 65
`parameters for each frame is fixed at a-bits per frame
`and the number of bits assigned to the residual signal is
`fixed at /3-bits per frame. Therefore, the (a+/3) bits for
`
`2
`each frame are transmitted to the receiving station. In
`this case, the transmission rate is, for example, 8 kbps.
`FIG. 2 is a block diagram showing a second prior art
`structure of the speech encoding apparatus. This prior
`5 art structure is a Code Excited Linear Prediction
`(CELP) encoder which is known as a low bit rate
`speech encoder.
`Principally, a CELP encoder, like the first prior art
`structure shown in FIG. 1, is an apparatus for encoding
`and transmitting linear prediction . code parameters
`(LPC or prediction parameters) obtained from an LPC
`analysis and a residual signal. However, this CELP
`encoder represents a residual signal by using one of the
`residual patterns within a code book, thereby obtaining
`high efficiency encoding.
`Details of CELP are disclosed in Atal, B. S., and
`Schroeder, M. R. "Stochastic Coding of Speech at
`Very Low bit Rate" Proc.ICASSP 84-1610 to 1613,
`1984, and a summary of the CELP encoder will be
`explained as follows by referring to FIG. 2.
`LPC analysis unit 201 performs a LPC analysis of an
`input signal, and quantizer 202 quantizes the analyzed
`LPC parameters to be supplied to predictor 203. Pitch
`period m, pitch coefficient Cp and gain G, which are
`not shown, are extracted from the input signal.
`A residual waveform pattern (code vector) is sequen(cid:173)
`tially read out from the code book 204 and its respective
`pattern is, at first, input to multiplier 205 and multiplied
`by gain G. Then, the output is input to a feed-back loop,
`namely, a long-term predictor comprising delay circuit
`206, multiplier 207 and adder 208, to synthesize a resid·
`ual signal. The delay value of delay circuit 206 is set at
`the same value as the pitch period. Multiplier 207 multi-
`plies the output from delay circuit 206 by pitch coeffici(cid:173)
`ent Cp.
`A synthesized residual signal output from adder 208 is
`input to a feed-back loop, namely, a short term predic(cid:173)
`tion unit comprising predictor 203 and adder 209, and
`the predicted input signal is synthesized. The prediction
`parameters are LPC parameters from quantizer unit
`202. The predicted input signal is subtracted from an
`input signal at subtracter 210 to provide an error signal.
`Weight function unit 211 applies weight to the error
`signal, taking into consideration the acoustic character(cid:173)
`istics of humans. This is a correcting process to make
`the error to a human ear uniform as the influence of the
`error on the human ear is different dependfog on the
`frequency band.
`The output of weight function unit 211 is input to
`error power evaluation unit 212 and an error power is
`evaluated in respective frames.
`A white noise code book 204 has a plurality of sam(cid:173)
`ples of residual waveform patterns (code vectors), and
`the above series of processes is repeated with regard to
`all the samples. A residual waveform pattern whose
`error power within a frame is minimum is selected as a
`residual waveform pattern of the frame.
`As described above, the index of the residual wave(cid:173)
`form pattern obtained for every frame as well as LPC
`parameters from quantizer 202, pitch period m, pitch
`coefficient Cp and gain G are transmitted to a receiving
`station (not shown). The receiving station forms a long(cid:173)
`term predictor with transmitted pitch period m and
`pitch coefficient Cp as is similar to the above case, and
`the residual waveform pattern corresponding to a trans-
`mitted index is input to the long-term predictor, thereby
`reproducing a residual signal. Further, the transmitted
`
`

`
`3
`LPC parameters form a short-term predictor as is simi(cid:173)
`lar to the above case, and the reproduced residual signal
`is input to the short-term predictor, thereby reproduc(cid:173)
`ing an input signal.
`Respective dynamic characteristics of an excitation 5
`unit and a vocal tract unit in a sound producing struc(cid:173)
`ture of a human are different, and the respective data
`quantity to be transmitted at arbitrary points by the
`excitation unit and vocal tract unit are different.
`However, with a conventional speech encoding ap- 10
`paratus as shown in FIGS. 1 or 2, excitation information
`and vocal tract information are transmitted at a fixed
`ratio of data quantity. The above speech characteristics
`are not utilized. Therefore, when the transmission rate
`is low, quantization becomes coarse, thereby increasing 15
`noise and making it difficult to maintain satisfactory
`speech quality.
`The above problem is explained as follows with re(cid:173)
`gard to the conventional examples shown in FIGS. 1 or
`1
`
`In a speech signal there exists a period in which char(cid:173)
`acteristics change abruptly and a period in which the
`state is constant, and the latter value of the prediction
`parameters do not change too much. Namely, there are
`cases where co-relationship between the prediction 25
`parameters (LPC parameters) in continuous frames is
`strong, and cases where they are not strong. Conven(cid:173)
`tionally, prediction parameters (LPC parameters) are
`transmitted at a constant rate with regard to each frame.
`Consequently, the characteristics of the speech signals 30
`are not fully utilized. Therefore, the transmission data
`causes redundancies and the quality of the reproduced
`speech in the receiving station is not sufficient for the
`amount of transmission data.
`
`SUMMARY OF THE INVENTION
`An object of the present invention is to provide a
`mode-switching-type speech encoding/decoding appa(cid:173)
`ratus for providing a plurality of modes which depend
`on the transmission ratio between excitation informa- 40
`tion and vocal tract information, and, upon encoding,
`switching to the mode in which the best reproduction of
`speech quality can be obtained.
`Another object of the present invention is to suppress
`redundancy of transmission information, which pre- 45
`vents relatively stable vocal tract information from
`being transmitted, and instead assigning a lot of bits to
`excitation information, which is useful for an improve(cid:173)
`ment of quality, thereby increasing the quality of the
`reproduced speech. In order to achieve the above ob- 50
`ject, the present invention has adopted the following
`structure.
`The present invention relates to a speech encoding
`apparatus for encoding a speech signal by separating the
`characteristics of said speech signal into articulation SS
`information (generally called vocal tract information)
`representing articulation characteristics of said speech
`signal, and excitation information representing excita(cid:173)
`tion characteristics of said speech signal. Articulation
`characteristics are frequency characteristics of a voice 60
`formed by the human vocal tract and nasal cavity, and
`sometimes refer to only vocal tract characteristics.
`Vocal tract information representing vocal tract char(cid:173)
`acteristics comprise LPC parameters obtained by form(cid:173)
`ing a linear prediction analysis of a speech signal. Exci- 6S
`tation information comprises, for example, a residual
`signal. The present invention is also based on a speech
`decoding apparatus. The present invention based on
`
`5,115,469
`
`w
`
`4
`above speech encoding/decoding apparatus has the
`structure shown in FIG. 3.
`A plurality of encoding units (or "ENCODERS #1
`to #m")301-1to301-m locally decode speech signal (or
`"INPUT") 303 by extracting vocal tract information
`(or "VOCAL TRACT PARAMETERS") 304 and
`excitation information (or "EXCITATION PARAME(cid:173)
`TERS") 305 from the speech signal 303, by performing
`a local decoding on it. The vocal tract information and
`excitation information are generally in the form of pa(cid:173)
`rameters. The transmission ratios of respective encoded
`information are different, as shown by the reference
`numbers 306-1to306-m in FIG. 3. The above encoding
`units comprise a first encoding unit for encoding a
`speech signal by locally decoding it, and extracting
`LPC parameters and a residual signal from it at every
`frame, and a second encoding unit for encoding a
`speech signal by performing a local decoding on it and
`extracting a residual signal from it using the LPC pa(cid:173)
`rameters from the frame several frames before the cur(cid:173)
`rent one, the LPC parameters being obtained by the
`first encoding units.
`Next, evaluation/selection units (or "EVALUA(cid:173)
`TION AND DECISION OF OPTIMUM EN(cid:173)
`CODER") 302-1/302-2 evaluate the quality of respec(cid:173)
`tive decoded signals 07-1 to 307-m subjected to local
`decoding by respective encoding units 301-1 to 301-m,
`thereby providing the evaluation result. Then they de(cid:173)
`cide and select the most appropriate encoding units
`from among the encoding units 301-1 to 301-m, based
`on the evaluation result, and output a result of the selec-
`tion (or "SELECT") as selection information 310. The
`evaluation/selection units each comprise evaluation
`decision unit 302-1 and selection unit 302-2, respectively
`35 as shown in FIG. 3.
`The speech encoding apparatus of the above struc(cid:173)
`ture outputs vocal tract information 304 and excitation
`information 305 encoded by the encoding units selected
`by evaluation/selection units 302-302-2, and outputs
`selection information 310 from evaluation/selection
`unit 302-1/302-2, to, for example, line 308.
`Decoding unit (or "DECODER #") 309 decodes
`speech signal 311 from selection information 310, vocal
`tract information 304 and excitation information 305,
`which are transmitted from the speech encoding appa(cid:173)
`ratus.
`With such a structure, evaluation/selection unit
`302-1/302-2 selects encoding output 304 and 305 of the
`encoding unit, which is evaluated to be of good quality
`by decoding signals 307-1 to 307-m subjected to local
`decoding.
`In the portions of the speech signal in which vocal
`tract information does not change, the LPC parameter
`is not output, thereby causing a surplus of information.
`As much of the surplus as possible is assigned to a resid(cid:173)
`ual signal, thereby improving the quality of decoded
`signal (or "OUTPUT") 311 obtained in a speech decod(cid:173)
`ing apparatus.
`· In the block diagram shown in FIG. 3, the speech
`encoding apparatus is combined with the speech decod(cid:173)
`ing apparatus through a line 308, but it is clear that only
`the speech encoding apparatus or only the speech de(cid:173)
`coding apparatus may be used at one time. Thus, the
`output from the speech encoding apparatus is stored in
`a memory, and the input to the speech decoding appara(cid:173)
`tus is obtained from the memory.
`Vocal tract information is not limited to LPC param(cid:173)
`eters based on linear prediction analysis, but may be
`
`

`
`5,115,469
`
`5
`cepstrum parameters based, for example, on cepstrum
`analysis. A method of encoding the residual signal by
`dividing it into pitch information and noise information
`by a CELP encoding method or a RELP (Residual
`Excited Linear Prediction) method, for example, may 5
`be employed.
`
`30
`
`6
`The A-mode encoder produces current frame predic(cid:173)
`tion parameters (LPC parameters) as vocal tract infor(cid:173)
`mation from output terminal 409, and a residual signal
`as excitation information through output terminal 408.
`In this case, the transmission rate of the LPC parame(cid:173)
`ters is f3 bits/frame and that of a residual signal is a
`bits/frame. The B-mode encoder outputs a residual
`BRIEF DESCRIPTION OF THE DRAWINGS
`signal from output terminal 412 by using LPC parame-
`ters of the previous frame or a frame which is several
`FIG. 1 shows a block diagram of a first prior art
`10 frames before the current frame. In this case, the trans-
`structure,
`mission rate of the residual signal is (a+/3)bits/frame,
`FI G. 2 shows a block diagram of a second prior art
`so the number of bits for the residual signal can be in-
`structure,
`creased by the number of bits that are not being used for
`FIG. 3 depicts a block diagram for explaining the
`the LPC parameters, as the LPC parameters vary little.
`principle of the present invention,
`FIG. 4 shows a block diagram of the first embodi- 15 Input signals to predictors 404-1 and 404-2 are locally
`ment of the present invention,
`decoded outputs from adders 406-1 and 406-2. They are
`FIG. 5 represents a block diagram of the second
`equal to signals that are decoded in the receiving sta-
`embodiment of the present invention,
`·
`1
`tton. Eva uation units 407-1 and 407-2 compare these
`FIG. 6 depicts an operation flow chart of the second
`locally decoded signals with their input signals from
`20 input terminal 401 to evaluate the quality of the de(cid:173)
`embodiment,
`FIG. 7A shows a table of an ass1·gnment of bt"ts to be
`coded speech. Signal to quantization noise ratio SNR
`transmitted in the second prior art, and
`within a frame, for example, is used for this evaluation,
`
`
`·
`t f b"t t b t
`
`f t bl FIG 7B · ts a a e o an ass1gnmen o 1 s o e rans-
`.
`mitted in the second embodiment of the present inven-
`enabling evaluation units 407-1 and 407-2 to output
`tion.
`25 SN(A) and SN(B). The mode determination unit 413
`compares these signals, and if SN(A)>SN(B), a signal
`designating A-mode is output, and if SN(A)<SN(B), a
`signal designating B-mode is output.
`A signal designating A-mode or B-mode is transmit(cid:173)
`ted from mode determination unit 413 to a selector (not
`shown). Signals from output terminals 408, 409, and 412
`are input to the selector. When the selector designates
`A-mode, the encoded residual signal and LPC parame(cid:173)
`ters from output terminals 408 and 409 are selected and
`output to the opposite station. When the selector desig(cid:173)
`nates B-mode, the encoded residual signal from output
`terminal 412 is selected and output to the opposite sta(cid:173)
`tion.
`Selection of A- or B-modes is conducted in every
`frame. The transmission rate is (a+f3) bits per frame as
`described above and is not changed in any mode. The
`data of(a+/3) bits per frame is transmitted to a receiv(cid:173)
`ing station after a bit per frame representing an A/B
`signal designating whether the data is in A-mode or
`B-mode is added to the data of (a+f3) bits per frame.
`The data obtained in B-mode is transmitted if B-mode
`provides better quality. Therefore, the quality of repro(cid:173)
`duced speech in the present invention is better than in
`the prior art shown in FIG. 1, and the quality of the
`reproduced speech in the present invention can never
`be worse than in the prior art.
`FIG. 5 is a structural view of the second embodiment
`of this invention. This embodiment corresponds to the
`second prior art structure shown in FIG. 2. In FIG. 5,
`501-1 and 501-2 depict encoders. These encoders are
`both CELP encoders, as shown in FIG. 2. One of them,
`501-1, performs linear prediction analysis on every
`frame by slicing speech into 10 to 30 ms portions, and
`outputs prediction parameters, residual waveform pat(cid:173)
`tern, pitch frequency, pitch coefficient, and gain. The
`other encoder, 501-2, does not perform linear prediction
`analysis, but outputs only a residual waveform pattern.
`Therefore, as described later, encoder 501-2 can assign
`more quantization bits to a residual waveform pattern
`than encoder 501-1 can.
`The operation mode using encoder 501-1 is called
`A-mode and the operation mode using encoder 501-2 is
`called B-mode.
`
`DESCRIPTION OF THE PREFERRED
`EMBODIMENTS
`The embodiment of the present invention will be
`explained by referring to the drawings.
`FIG. 4 shows a structural view of the first embodi(cid:173)
`ment of the present invention, and this embodiment
`corresponds to the first prior art structure shown in
`FIG. I.
`The first quantizer 403-1, predictor 404-1, adders 35
`405-1 and 406-1, and LPC analysis unit 402 correspond
`to the portions designated by 103, 102, 105, 106, and
`101, respectively, in FIG. 1, thereby providing an
`adaptive prediction speech encoder. In this embodi(cid:173)
`ment, a second quantizer 403-2, a second predictor 40
`404-2, and additional adders 405-2 and 406-2 are further
`provided. The LPC parameters applied to predictor
`404-2 are provided by delaying the output from LPC
`analysis unit 402 in frame delay circuit 411 through
`terminal A of switch 410. The portions in the upper 45
`stage of FIG. 4, which correspond to those in FIG. 1,
`cause output terminals 408 and 409 to transmit LPC
`parameters and a residual signal, respectively. This is
`defined as A-mode. The signal transmitted from output
`terminal 412 in the lower stage of FIG. 4 is only the 50
`residual signal, which is defined as B-mode. Evaluation
`units 407~1 and 407-2 evaluate the SIN of the encoder of
`the A- or B-mode. Mode determining (or "MODE
`DETERMINATION") portion 413 produces a signal
`A/B for determining which mode should be used (A- 55
`mode or B-mode) to transmit the output to an opposite
`station (i.e. receiving station) (not shown), based on the
`evaluation. Switch (SW) unit 410 selects the A side
`when the A-mode is selected in the previous frame.
`Then, as LPC parameters of the B-mode for the current 60
`frame, the values of the A-mode of the previous frame
`are used. When the B-mode is selected in the previous
`frame, the B side is selected and the values of the B(cid:173)
`mode in the previous frame, namely, the values of the
`A-mode in the frame which is several frames before the 65
`current frame, are used.
`In this circuit structure, the encoders of the A-and B
`modes operate in parallel with regard to every frame.
`
`

`
`5,115,469
`
`8
`7
`waveform distortion and spectral distortion of repro-
`In encoder 501-1, linear prediction analysis unit 506
`performs the same function as both LPC analysis unit
`duced speech signals A and B to evaluate the quality of
`speech reproduced by encoders 501-1 or 501-2. In other
`201 and quantizing unit 202. White noise code book
`words, unit 502 uses segmental S/N and LPC cepstrum
`507-1, gain controller 508-1, and error computing unit
`511-1, respectively, correspond to those features desig- 5 distance (CD) of respective frames in parallel to evalu-
`ate the quality of reproduced speech.
`nated by the reference numbers 204, 205, and 210 in
`FIG. 2. Long-term prediction (or "LONG-TERM
`Therefore, quality evaluation/encoder selection unit
`PREDICTOR") unit 509-1 corresponds to those fea-
`502 is provided with cepstrum distance (or "CD") com-
`tures designated by the reference numbers 206 to 208 in
`puting unit 515, operation mode judgement unit 516,
`FIG. 2. It performs an excitation operation by receiving 10 and switch 514.
`pitch data as described in conjunction with the second,
`Cepstrum distance computing unit 515 obtains the
`prior art structure. Short-term prediction (or "SHORT- .
`first LPC cepstrum coefficients from the LPC parame-
`TERM PREDICTOR"), unit 510-1 corresponds to
`ters that correspond to the present frame, and that have
`those features represented by the reference numbers 203
`been obtained from linear prediction analysis unit 516.
`and 209 in FIG. 2, and functions as a vocal tract by 15 Cepstrum distance computing unit 515 also obtains the
`receiving prediction parameters as described in the
`second LPC cepstrum coefficients from the LPC pa-
`second prior art. In addition, error evaluation unit 512-1
`rameters that are obtained from coefficient memory 513
`corresponds to those features designated by the refer-
`and are currently used in the B-mode. Then it computes
`ence numbers 211 and 212 in FIG. 2, and performs an
`LPC cepstrum distance CD in the current frame from
`evaluation of error power as described in conjunction 20 the first and second LPC cepstrum coefficients. It is
`with the second prior art structure. In this case, error
`generally accepted that the LPC cepstrum distance thus
`evaluation unit 512-1 sequentially designates addresses
`obtained clearly expresses the difference between the
`(phases} in white noise code book 507-1, and performs
`above two sets of vocal tract spectral characteristics
`evaluations of error power of all the code vectors (re-
`determined by preparing LPC parameters (spectral
`sidual patterns) as described in the second prior art 25 distortion).
`Operation mode judgement unit 516 receives segmen-
`structure. Then it selects the code vector that has the
`lowest error power, thereby producing, as the residual
`ta! SINA and SINB from encoders 501-1 and 501~2, and
`signal information, the number of the selected code
`receives the LPC cepstrum distance (CD) from cep-
`vector in white noise code book 507-1.
`strum distance computing unit 515 to perform the pro-
`Error evaluation unit 512-1 also outputs a segmental 30 cess shown in the operation flow chart of FIG. 6.
`SIN (SINA) that has waveform distortion data within a
`This process will be described later.
`frame.
`Where operation mode judgement unit 516 selects the
`Encoder 501-1, described in reference to FIG. 2,
`A-mode (encoder 501-1), switch 514 is switched to the
`produces encoded prediction (or "PREDICTION")
`A-mode terminal side. Where operation mode judge-
`parameters (LPC parameters) from linear prediction 35 ment unit 516 selects B-mode (encoder 501-2), switch
`514 is switched to the B-mode terminal side. Every time
`analysis unit 506. It also produces encoded pitch period,
`pitch coefficient and gain (not shown).
`A-mode is produced (output from encoder 501-1 is
`In encoder 501-2, the portions designated by the ref-
`selected) by a switching operation of switch 514, coeffi-
`erence numbers 507-2 to 512-2 are the same as respec-
`cient memory 513 is renewed. When the B-mode is
`tive portions designated by reference numbers 507-1 to 40 produced (so that the output from encoder 501-2 is
`512-1 in encoder 501-1. Encoder 501-2 does not have
`selected) coefficient memory 513 is not renewed and
`linear prediction analysis unit 506; instead, it has coeffi-
`maintains the current values. Multiplexing (or "MUX")
`cient memory 513. Coefficient memory 513 holds pre-
`unit 504 multiplexes residua