Gain-shape vector quantization apparatus

2~198~4 GAIN-8XAPE VECTOR QU~NTIZATION APP~RATUS BACRGRO~ND OF T~E INVENTION 1. Field of the Invention The present invention relates to a gain-shape vector 5 quantization apparatus, and particularly to a gain-shape vector quantization apparatus for compressing the data of - speech signals.

In recent years, a vector quantization system for compressing the data of speech signals yet maintaining the 10 quality thereof has been employed for communication systems within companies and for digital mobile wireless systems or - the like.

As is well known, the vector quantization system consists of passing signal vectors of a code book through a synthesis filter to reproduce signals, and evaluating error - 15 electric powers between the reproduced signals and the input speech signals in order to determine the index of a signal vector having the smallest error. Among a variety of such vector quantization systems, attention has now been given particularly to a gain-shape (amplitude,~phase) vector 20 quantization system used as a high quality speech coding method.

-~ 2.

Description of the Related Art The prior art illustrates the construction of a conventional gain-shape vector quantization apparatus 25 according to which a shape vector (phase) is selected out of shape vectors obtained by normalizing electric power of a plurality of vectors that constitute a code book having a given size and a given dimension (or order), and is multiplied by the gain (amplitude) through a variable gain circuit to 30 separate an amplitude component from a phase component.

According to such a conventional gain-shape vector quantization system, however, there exists only one code book, ; and only one shape vector is selected from the code book by the evaluation unit resulting in an increase in the 35 quantization distortion (quantize error) and making it ~ di~icult to maintain quality of the reproduced speech.

-~ Such a quantization distortion can be decreased by increasing the number of code vectors ~shape vectors) included , in the code book and by increasing the dimension of the code .'.' .~. , .

.-............................... .

::: . :, . ....

. , .

.

.. , , : .

' ' :

- 2 - 2~1~804 . book. However, the amount of operation for finding an optimum shape vector increases, too, with the increase in the size of the code book, and increased memory is required for storing the shape vectors, posing a serious hindrance for realizing 5 the hardware and causing the amount of shape vector index transmission to increase.

When the size of the code book is decreased to quicken the speed of operation, on the other hand, the reproduced speech loses quality due to quantization error.

~3UNNARY OF THB INVENTION A feature of one aspect of the present invention therefore is to realize a gain-shape vector quantization apparatus which is capable of decreasing the quantization distortion at the time or reproducing the speech without 15 increasing the size of the code book.

In order to attain the above-mentioned object, the present invention deals with a gain-shape vector quantization apparatus for an input speech signal, the apparatus comprising:

a code book portion containing a plurality of shape vectors and selecting among the shape vectors upon receipt of an index signal;

a gain circuit portion including a plurality of variable gain circuits coupled in parallel to the code book 25 portion to impart predetermined gains to the shape vectors selected by the code book portion;

- a synthesis filter portion including a p~urality of synthesis filters coupled in parallel to the gain circuit portion to reproduce signals from the shape vectors as output 30 by the variable gain circuits in the gain circuit portion;

a first adder portion coupled to the synthesis ~ilter portion to add the signals reproduced by the synthesis ! ~ilters in the synthesis filter portion and to produce an output;

a second adder portion coupled to the first adder portion to add the output of the ~irst adder portion and the input speech signal to determine an error; and an evaluation unit coupled to the second adder and the code book portion to determine the index signal used in ._ , p~ ... .

.,~ . . ..

_ 3 _ 2~1~804 the selection of the shape vectors stored in the code book portion to minimize the error between the output of the first adder portion and the input speech signal produced by the second adder portion, and supplied to the code book portion.

BRIEF DESCRIPTION OF THE DRAWINGS The above object and features of the present invention will be more apparent from the following description of the preferred embodiments with reference to the accompanying drawings, wherein:

Fig. l is a block diagram illustrating the construction of a conventional gain-shape vector quantization apparatus;

Fig. 2 is a block diagram illustrating the principle and construction of a gain-shape vector quantization apparatus according to the present invention;

Fig. 3 is a diagram illustrating the principle of vector quantization using a synthesized vector of a plurality of shape vectors;

Figs. 4 and 5 are block diagrams illustrating gain-shape vector quantization apparatuses according to ernbodiments of the present invention; and Fig. 6 is a diagram illustrating the construction on the receiving side in the embodiment of Fig. 5.

2~ DESCRIPTION OF THE PREFERRED EMBODIMENTS In order to realize the above-mentioned object, the - basic technical structure of the gain-shape vector quantization apparatus of the present invention consists of compressing the amount of data, and using a reduced ' 30 number of indexes or, preferably, using a single index ~ data to decrease the number of bits, yet maintaining .'~ vector data in an amount sufficient for maintaining the same quality as that of the conventional apparatus ; during the operation.

Concretely speaking, the opera. 35 tion is carried out while permitting a plurality of Qhape vector data to be produced in response to a single index data.

.~ .

... . ..

:,........... 1~ . . .

; :.

' . :- ' ., , _ 4 _ 2~1~8~4 .~ Referring initially to the prior art shown in Figure 1 of the drawings, Figure 1 illustrates the construction of a conventional gain-shape vector quantization apparatus according to which a shape vector (phase) is selected out of 5 shape vectors obtained by normalizing electric power of a plurality of vectors that constitute a code book 11 having a size S and a dimension (or order) L, and is multiplied by the gain (amplitude) through a variable gain circuit 12 to separate an amplitude component from a phase component. Then, 10 a signal X is reproduced through a synthesis filter 13 that includes a pitch synthesis filter 13a ad an LPC (linear predictive coding) synthesis filter 13b and that has a transfer function H(Z), and an error E between the reproduced signai X and tHe input signal A is evaluated by an 15 evaluation unit 14, whereby a shape vector Ck in the code book 11 is newly selected such that the error E becomes the smallest according to the following equations (1) and (2), ¦Ek¦ = (A - Xk)2 ....................... (1) Xk = H-(g Ck) .......................... (2) (k = 1, 2, ..., S) Furthermore, the gain g for a shape vector is determined based on the error electric power of E, the gain being calculated for each shape vector.

Then, the index of the shape vector of the code book 11, gain of the variable gain circuit 12, pitch ; delay of pitch prediction coefficients the pitch synthesis filter 13a, and LPC coefficient of when the input voice is analyzed through an LPC analyzer 15, 30 which are obtained as described above, are transmitted to compress the speech data.

As explained hereinabove, such a conventional gainshape vector qyantization system, however, there exists only one code book, and only one shape vector is selected from the 35 code book by the evaluation unit resulting in an increase in the quantization distortion ~quantize error) and making it - di~icult to maintain quality of the reproduced speech.

.,.

" __ ,~ ,:

- 4A - 2 ~ 1 9 ~ ~ 4 Such a quantization distortion can be decreased by increasing the number of code vectors (shape vectors) included in the code book and by increasing the dimension of the code book. However, the amount of operation for finding an optimum shape vector increases, too, with the increase in the size of the code book, and increased memory is required for storing the shape vectors, posing a serious hindrance for realizing the hardware and causing the amount of shape vector index transmission to increase.

When the size of the code book is decreased to ~-; quicken the speed of operation, on the other hand, the reproduced speech loses quali'y due to quantization error.

The basic technical idea of the present invention will now be described with reference to Fig. 2 starting, first, with the principle thereof. Referring to Fig. 2 which shows the constitution in principle, the apparatus comprises a code book portion 1 which is constituted by a plurality of shapé vectors lli (i = 1 - N) and produces a plurality of selected shape vectors, a , , ~,, , ~ .

., '; .

:,.i :' , . ..

,: t.~ :. . : _ 5 _ 2~1~8~4 plurality of variable gain circuits 2 that impart gains to the shape vectors produced from the code book portion 1, a plurality of synthesis filters 3 that reproduce signals from the outputs of the variable gain circuits 2, a first adder 4 that adds up the signals reproduced by the synthesis filters 3, a second adder 12 that adds up the output of the first adder 4 and the input speech, and an evaluation unit 5 that selects a plurality of shape vectors in the code book portion 1 in response to an index such that the error between the output of the first adder 4 and the input speech becomes the smallest.

According to the present invention in this case, the code book portion 1 may be constituted by code books C of a number M, each code book C being constituted by shape vectors of a numbe N (N ~ 1).

In the gain-shape vector quantization apparatus of the present invention, the code book portion 1 includes at least one code book C which may be a conventionally known one.

Each code book C in the code book portion 1 ; contains a plurality (N) of shape vectors lli (i = 1 N), each of the shape vectors lli having a vector structure expressed by any random numbers and being - stored therein in the form of digital values.

Further, though there is no particular limitation on the length of the shape vectors, i.e., in the dimension, a preferred example is a forty-dimensional one.

The code book portion 1 used in the present invention receive~ an index data from the evaluation unit 5 that will be described later and produces a plurality of shape vectors 13i (i - 1 - M) simultaneously.

That is, : according to the present invention, a lot of data are obtained from small amounts of data.

As described above, the shape vectors simultaneously produced from the code book portion 1 in response to an index data, pass through pairs of the variable gain circuit~ 2i (i = 1 - M) in the gain .

2~1~8~4 ~ - 6 - circuit portion 2 arranged for each of the shape vectors 13i (i = 1 - M) and the synthesis filter circuits 3i (i = 1 - M) in the prediction filter portion 3 so as to be reproduced into the original signals, and are then input to the first adder 4 which forms a synthesized vector of reproduced speech. Next, the reproduced synthetic vector is added to a vector of practical speech input from an external unit at the second adder 12, thereby to find an error between the two, and the result is input to the evaluation unit 5 which may be constituted as a unitary structure together with the second adder.

The evaluation unit 5 executes the operation for all shape vectors lli (i = 1 - N) stored in the code book C in the code book portion 1 such that the error lS calculated by the second adder 12 will become the smallest, selects plural combinations of optimum shape vectors, and sends them as index data to the code book portion 1. In this case, though the number of index ' data is only one, the instruction is given so that a plurality of shape vectors are produced simultaneously from the code book portion 1. For this purpose, a table is prepared in advance to impart index data for plural combinations of all shape vectors, operation is carried out b~ the evaluation unit 5 for all of these combinations to find a combination that makes the error smallest, and an index of the combination of shape vectors that makes the error the smallest is produced.

The same table is also possessed in the code book portion 1, and a required combination of a plurality of ,; shape vectors is produced based upon the index data from ` the evaluation unit 5.

In the present Lnvention, there is no particular limitation in the number of shape vectors that conv stitute combinations of shape vectors.

Next, the evaluation unit 5 is so constituted as to supply gain ad~usting signals to change the gains of the plurality of variable gain cLrcuits 2i ~i = 1 - M) that .

:,, , 2019804 -- 7 constitute the gain circuit portion 2.

The gain adjusting signals are to so change the gains of the variable gain circuits that the error in the second adder 12 will become the smallest.

That is, in the present invention, the error of the synthesized speech signal from the input speech signal is minimized by selectively adjusting the shape vectors 11 which are the data related to phase, so that a synthesized vector having a small phase difference is obtained, adjusting the gains that are related to the amplitude. This is to reproduce synthetic speech which is as close to the original speech as possible.

That is, the present invention reproduces the vector which is synthesized from a plurality of shape vectors unlike the conventional art in which a single shape vector in a single code book is used.

Fig. 3 shows the principle of vector quantization according to the present invention for two-dimensional case.

; 20 First, in the case of the conventional vector quantization using a single vector shown in Fig. 3(a), a shape vector C4 is selected out of finite shape vectors Cl to CN to minimize the quantization distortion with respect to the speech input A that is to be quantized when the shape vectors are multiplied by a scalar g related to a variable gain.

In this case, the number of shape vectors stored in the code book is finite.

To reduce the quantization ~ error, it is required to increase the code book size, - 30 which results in increasing computational complexity.

On the other hand, Fig. 3(b) shows an example of quantization when the apparatus according to the present invention is used, wherein two shape vectors are used to quantize the two-dimensional vector A.

That i8, as shown in Fig. 3(b), the two independent two-dimensional shape vectors Cl , C2 are multiplied by gains g1 and g2 to weight them, and the input speech A 2~1~8G4 - 8 is quantized by the synthesized vector.

In this case, the gains gl t g2 are so adjusted as to correctly realize any vectors.

That is, the error can be brought to zero.

Here, what can be quantized without distortion is - the case where the number of independent shape vectors used for the synthesis is greater than the number of dimensions of the shape vectors.

Even when this condition is not satisfied, however, the quantization distortion can be more decreased with the vector quanti zation based on a plurality of shape vectors than withthe vector quantization based on a single shape vector.

When the two-dimensional vectors are to be quantized as will be obvious from the foregoing description, it is desired that the two shape vectors that are selected are at right angles to each other. That is, inthe present invention, when an index is given to the code book portion l from the evaluation unit 5 based on an error between the reproduced speech and the input speech, a plurality of shape vectors lll , ll2 are -~ selected and produced from the code book portion l.

The shape vectors are then weighted through the corresponding variable gain circuits 2l , 22 and, then, -~ signals are reproduced from the prediction filters 31 ' 32 to effect quantization with little distortion.

Fig. 4 illustrates a concrete embodiment of the ~J, code book portion l which produces a plurality of shape . vectors based on an index data according to the present ~ invention.

-~ 30 In the concrete embodiment of Fig. 4, the code book portion l is divided into M code books (Cl , C2 , CM), each code book being constituted by N (N > l) shape vector~.

The sys~tem shown in Fig. 4 will be described in further detail.

Reference is made independently to M code books Cl , C2 , ..., CM each consisting of N shape vectors, in response to an index produced from the .. .. .

2~1~8~4 . g evaluation portion 5, and whereby M code shape vectors (Ck l ~ Ck 2 ~ ~ Ck M) are produced from the code book portion l.

That is, when the evaluation unit 5 produces an index k that so instructs that a k-th shape vector is selected from each of the code books C, there are produced k-th shape vectors (Ck l ~ Ck 2 ~ ' Ck M) that are stored in M code books Cl , C2 , ..., CM respectively. In this case, the shape vectors stored in the code books are different from one another, and hence, M shape vectors are selected to form vectors that are all different from one another. In this embodiment, the number of code books C is determined depending upon the reproduced speech quality that are to be required.

The plurality of shape vectors thus selected are multiplied by a weighting coefficient gk I are processed ; through the synthesis filter 3, and are added together through the first adder 4 to obtain a synthesized shape vector ~k ' k gl H Ck,l + g2 H Ck,2 + -- + gM'H-Ck M , ... (3) An error is found between this synthesized vector Yk and the input speech A to be quantized, and an index that : minumizes the square error in the evaluation unit 5 and a gain vector G at that moment, G = (gl ' g2 ' M are transmitted.

Then, even when the code books have a small size, the vector Yk expressed by a linear combination of shape rs Ck,l Ck,2 ' --~ Ck,M representS any vector in ~! partial space including shape vectors Ck l ~ Ck 2 ~ ~ Ck M as shown in Fig. 3(b), which is substantially equal to the code book of an increased size.

Therefore, when the quantization distortion equal to that of the conventional large codebook system is acceptable, then the code books may have a size smaller than that of the conventional system.

2~198~4 -- 10 - Though Fig. 4 shows a circuit on the transmitting side only in the communication system, the circuit, in practice, is also constituted in the same manner on the receiving side. Namely, when a gain data and an index data determined by the evaluation unit 5 are transmitted, the receiving side selectively adjusts the code book portion 1 and the gain circuit portion 2 that includes variable gain circuits in response to these data, and reproduces a synthesized voice.

In the aforementioned embodiment, the code book portion is equipped with a plurality of code books (C to CM) to generate a plurality of shape vectors in response to an index data.

Figs. 5A and 5B shows a gain-shape vector quanti15 zation apparatus based on a system different from the above-mentioned one.

That is, Figs. 5A and 5B show the gain-shape vector quantization apparatus according to another embodiment of the present invention, wherein Fig. 5A shows a coding device and Fig. 5B shows a 20 decoding device.

In the embodiment of Fig. SA, the code book portion 1 is comprised of a code book which consists of S : shape vectors, and a mapping table T which ~: simultaneously selects M (M > 1) different shape vectors v 25 in response to an index produced from the evaluation unit 5, In Fig. S, M = 2.

~, As will be obvious from Figs. 5A and 5B, on either the coding device side or the decoding device side, the code book portion 1 is comprised of a single code : 30 book 11 having an L-dimension and a size S, and a mapping table T which generates an index j of a code vector corresponding to a code vector of an index k (k 1, 2, ..., S) (S' ~ S), the code vector of the index being different by one from the code vector of the index k.

~ = T(k) (5) 2~19804 That is, in response to the index k from the evaluation unit 5, the mapping table T generates a new index T(k) with k as a function, and a plurality of different shape vectors CT(k) are selected in response to the index T(k).

In this embodiment, k and T(k) maintain a suitable functional relationship. For instance, a shape vector which is two in advance may be selected at all times in response to the value k. In this case, T(k) is k+2 with respect to k.

In this embodiment, furthermore, the mapping table T need not be limited to the one that produces a new index T(k) for the index k but may be one that produces a plurality of indexes T(Kll), T(K12) .... In the embodiment of Fig. 5, reference is made to the mapping table T to find an index j = T(k) for an index k that is transmitted from the evaluaticn unit 5 on the coding device side or that is transmitted from the coding device side on the decoding device side, and two shape vectors Ck , CT(k) are obtained that correspond to the index k.

The shape vectors Ck and CT(k) are multiplied by the gains gk ' gT(k) I and are input to the synthesis filter 3 to regenerate signal vectors Xk , XT(k) of .~ 25 L-dimension.

At this moment, the gains gk and gT(k) corresponding to the two shape vectors Ck , CT(k) are - evaluated by the evaluation unit S on the coding device side shown in Fig. SA so as to minimize the distortion scale Dk of the synthesized vector Yk ~ Yk = gk.H-Ck ~ gT(k) H T(k) ... (6) and the vector A to be quantized.

That is, the evaluation unit 5 uses a Euclidean distance as a distortion scale to find a distortion Dk i.e., Dk = IA = Ykl (7) The distortion Dk is then partially differentiated with 2~1~8~4 - 12 gk and gT(k) , and is set to zero, in order to find gain vectors gk and gT(k) that minimize the distortion Dk.

Thus, the distortion Dk is found for each of the indexes k (k = 1 to S), and indexes k corresponding to 5 combinations of shape vectors Ck , CT(k) that give the smallest distortion are transmitted together with the gain vectOrS (gk ~ gT(k))On the decoding device side shown in Fig. 5B, on the other hand, the mapping table T which is the same as 10 that of the coding device side is referred to by the index k to find an index j = T(k).

Then, the code - book 11 is referred to by the indexes k and j to obtain ; shape vectors Ck and CT(k).

Then, the shape vectors are multiplied by the gains 15 gk and gT(k) , pass through the synthesis filter 3, and are added up through the adder 4 to regenerate a voice vector A', i.e., A~ = gk.H~Ck + gT(k) H cT(k) ... (8) In the embodiment shown in Fig. 5A, the index k , 20 produced from the evaluation unit 5 is directly input to S the code book portion 11 and is further input to the mapping table T to generate a new index T(k).

As described above, however, the index k may be input to a plurality of the mapping tables T1 , T2 1 --, TM only 25 to generate a plurality of indexes T(k), T(k'), T(k") from each mapping table.

The embodiment of Fig. 5A has dealt with the case where two indexes were formed from one index.

It needs not be pointed out, however, that two or more indexes can be formed by providing a plurality of mapping tables T to realize quantization with less distortion.

In the aforementioned embodiments, the synthesis filters 31 to 3M in the synthesis filter portion 3 may be constituted by a pitch prediction filter 13a to a long-term correlation and an LPC filter 13b to add a short term correlation as shown in Fig. 1.

The synthesis filter portion 3 was provided after - 13 _ 2~98~4 the gain circui.t portion 2 in the aforementioned two embodiments. It is, however, also allowable to provide the gain circuit portion 2 on the side of the first adder 4 and to provide the synthesis filter portion 3 on the side of the code book portion 1.

As shown in Fig. 6, furthermore, the outputs of the gain circuit portion 2 may be passed through the synthesis filter portion 3 which consists of a synthesis filter 3~ after they have been added together through the adder 4' on the receiving side, i.e., on the side of the decoding device in the embodiment of Fig. 5B.

According to the gain-shape vector quantization apparatus of the present invention as described above, a plurality of shape vectors are defined for one index and vectors are quantized using weighted and synthesized vectors, making it possible to decrease the quantization distortion.

Furthermore, by using this multiple vector synthesis technique, it becomes possibie to decrease the size of the code book while suppressing the quantization distortion. It is therefore allowed to reduce the memory quantity and the quantity of operation, presenting great advantage for realizing the hardware.

,