JPH04351018A - Gain shape vector quantization system - Google Patents

Abstract

PURPOSE:To suppress deterioration in the reproduced voice quality when the quantization. system is subjected to variable rate processing by referencing the combination of shape vectors and quantized gains stored in a table so as to send an index corresponding to a set where error power is minimized. CONSTITUTION:A retrieval section 5 for an optimum set of a shape and a gain references a table 1 for a set of a shape vector and a gain in response to a coded bit number commanded externally by a variable rate processing request in a coder. Then the set corresponding to the command is processed by a synthesis filter 2 and a multiplier 3 and the error power between the obtained reproduction voice signal and input voice signal is calculated by an error power calculation section 4. An optimum set is decided by an index signal from the retrieval section 5 based on the result of calculation for the minimum error power. Through the constitution above, the deterioration in the quality of the reproduction voice signal in the case of making variable rate for the quantization system is minimized.

Description

[Detailed description of the invention]

0001

[Field of Industrial Application] The present invention relates to a gain shape vector quantization method, and in particular to CELP (Code Excited LP), which is known as a highly efficient audio encoding method.
This relates to the gain/shape vector quantization method used in the C) method and the like.

[0002] In recent years, vector quantization methods have been used to compress information while preserving the quality of voice signals in in-house communication systems and digital mobile radio systems. The CELP method, which performs pitch search and fixed codebook search using two codebooks to find the optimal driving excitation signal, is attracting attention as an even more efficient speech encoding method. A more efficient gain/shape vector quantization method that has the advantage of being able to search for an optimal pair of shape vectors and gains among the shape vectors constituting a codebook and requiring a small amount of calculation is attracting attention.

0003

2. Description of the Related Art FIG. 6 shows a general configuration of a gain-shape vector quantization method, in which 11 is a codebook consisting of a plurality of shape vectors 121 to 12n;
31 to 13n are each shape vector 121 to 12
Converters 141 to 14n are converters 13 that provide linear predictive synthesis filters or perceptual weighting transformation matrices that pass n.
A gain multiplier 15 gives a gain to be multiplied by each output of 1 to 13n to generate a gain/shape vector of the reproduced audio, and 15 is one of the gain/shape vectors output from the gain multipliers 141 to 14n. A selector selects one of the vectors, and 17 is an error power calculation unit that calculates error power from the error between the target vector based on the input audio signal and the output of the selector 15, and sequentially controls selection of the selector 15.

Next, in the operation of this conventional example, converters 131 to 12n are used for each shape vector 121 to 2n.
Gain multiplier 141 for the vector passed through 13n
n(n
>1) The selector 15 selects the error powers of the calculation unit 16 between the gain shape vector and the target vector one by one, and selects the optimal gain shape vector that is the minimum among these error powers. is determined and only its index (shape vector number and gain) is sent to the receiving side, thereby realizing transmission of the compressed audio signal.

[0005]

[Problem to be Solved by the Invention] In such a conventional gain/shape vector quantization method, when audio encoding is performed at a variable rate, it is possible to reduce the number of quantization bits of the gain and shape vector. In this case,
We don't know which gain bits to remove and which shape vectors to remove. In general, the importance of the pair of gain and shape vector (hereinafter sometimes referred to as gain and shape) is different for each shape vector, and the number of gains combined with this is also different for each shape vector, so it has been deleted. There was a problem in that if the wrong target was selected, the quality of audio playback would deteriorate significantly.

Therefore, it is an object of the present invention to minimize the deterioration in audio reproduction quality when changing the gain shape vector quantization method to a variable rate.

[0007]

[Means for Solving the Problems] In order to achieve the above object, the gain shape vector quantization method according to the present invention is based on 2N (N is a positive integer) as shown in principle in Fig. 1(a). ) shape vectors and 2M (M is a positive integer) quantized gains.
(2K = 2N × 2M) Table 1 in which the optimal combination of shape vector and gain that can be obtained for each number of encoded bits is determined and stored in advance, and the shape vector output from Table 1 are combined. A filter 2, a multiplier 3 that generates a reproduced audio signal by multiplying the output signal of the synthesis filter 2 by the gain output from the table 1, and an error between the reproduced audio signal from the multiplier 3 and the input audio signal. An error power calculation unit 4 that calculates the power, and when encoding with a desired number of bits, sends each shape vector to the synthesis filter 2 and sequentially multiplies the gain corresponding to the number of encoded bits in each shape vector. The search unit 5 searches for an optimal pair of shape vector and gain from the error power and transmits the index to the unit 3.

[0008]

[Operation] As shown in FIG. 1(a), Table 1 used in the present invention has 2N shape vectors and 2M quantized gains. For example, if the number of encoded bits is "1" bit,
In the shape vector V(0), G
(p1), G(p2) are selected, and for the shape vector V(1), G(s1), G(s2) are selected, and for the last V(m), G(v1 )
, G(v2), but in Fig. 1(a)
These gains shown in are shown in a generalized form, and in reality, when the number of encoding bits is "1", two of the above G(p1) to G(v2) are optimal. The combinations are prepared in advance in this table 1.

Similarly, when the specified number of encoding bits is "2" bits, the shape vector V(0) has G(q1), G(q2), G(q3) (not shown). , G(q4) (not shown) is selected, and in the case of shape vector V(1), G(t1),
G(t2), G(t3) (not shown), G(t4)
(not shown) is selected, and for the last V(m), G(w1), G(w2), G(w3) (not shown), G(w4) (not shown). The four optimal shape vector and gain combinations selected as shown in FIG.
Regarding the bits, G(r1), G(r2), ...
(not shown) is selected, and the shape vector V(1
), then G(u1), G(u2), ...(
(not shown) is selected, and in the case of the last V(m), G(x1), G(x2), ... (not shown)
2L optimal combinations selected as follows are prepared in Table 1.

[0010] When actually encoding the table 1 of optimal combinations of gain shapes that have been selected and stored in advance in this way, the search unit 5 performs A synthesis filter 2 synthesizes each shape vector, and a multiplier 3 combines and multiplies the gains corresponding to the number of encoded bits in the shape vector one by one to reproduce an audio signal and combine it with the input audio signal. By calculating this in the error power calculation unit 4, a combination of shape vector and gain with the smallest error can be searched for, and the index of this combination is transmitted to the receiving side via the transmission path.

[0011] Then, on the receiving side, as shown in FIG.
As shown in FIG. 6, when the gain/shape vector inverse quantization method receives an index indicating the above-mentioned optimum shape vector and gain pair from the transmitting side, the shape vector and gain are taken out from the table 6 and the synthesis filter 7 A reproduced audio signal can be obtained by synthesizing the signals.

[0012] In this way, in the present invention, the optimum gain shape combinations are determined in advance for each number of encoded bits, and from among them, the combination with the least error in relation to the actual input audio signal is selected. Therefore, deterioration in audio quality due to variable rate can be minimized.

[0013]

[Example] Fig. 2 shows an example of the gain shape vector quantization method according to the present invention shown in Fig. 1(a). This figure shows the optimal combination of shape vector and gain when N=3, M=1, and the maximum number of encoding bits L=4 bits.

An explanatory diagram of an embodiment of Table 1 shown in FIG. 2 is shown in FIG. 3. To explain Table 1 using FIG. 3, the combination of shape vector and gain is 4 bits in this example, In other words, all can be shown with 16 pieces,
Table 1 shows the possible gains of G(0) (=2/3) and G(1) (=1) when each shape vector has 4 bits. In Figure 3, this is the set S(
4) It is shown as.

[0015] When encoding with ``3'' bits, which is 1 bit less than the maximum ``4'' bits in such an example due to the request for variable rate, 1 in the case of 4 bits described above
8 out of 6 shape vector and gain combinations
A combination of these items will be selected. Therefore, these 16
There are 8 combinations (16C8) of these, and the set obtained as the optimal combination is S(3). Accordingly, in Table 1, the shape vector V(0) and Only V(1), V(2), V(3), V(4), and V(7) are shape vectors with optimal gains.

Similarly, when 2 bits are reduced and encoded using "2" bits, the result is a combination of four shape vectors and gains, of which the optimal one is shown in FIG. The set S(2) is the set S(2), and accordingly Table 1 contains V(0), V(1) and V(
Only V(2) and V(3) have the optimal gain, and if further encoding is done with "1" bit, the combination of set S(1) is obtained as shown in Figure 3, and Table 1 shows that V(3) has the optimal gain.
(1) and V(2) are the only two that have the optimal gain.

FIG. 4 is a flowchart showing how Table 1 described above and shown in FIGS. 2 and 3 is determined. This is executed in advance prior to encoding.

First, a speech sample for performance testing is prepared (step S1). This is obtained by collecting representative audio samples and, for example, averaging them.

Next, the following process is executed for the combinations in which 2k (1≦k≦4) sets are selected from 24 = 16 combinations of gain and shape vector (step S
2).

That is, the combinations just selected are set S' (k
), encode the test speech sample with k bits, and calculate the encoding error power (step S3). and,
It is determined whether the above-calculated encoding error power is the minimum among the combinations selected so far (step S4), and when it is determined that it is the minimum, the set S(k) so far is selected. The set S'(k) is replaced (step S5).

After this step S5, it is determined whether or not the encoding error power has been executed for all combinations in addition to the case where it is determined that the encoding error power is not the minimum in the above step S4 (
In step S6), if all the combinations have not yet been executed, the next combination is selected and executed in the same way (step S7).

[0022] In this way, a set of combinations closest to the performance test audio sample is determined from among the maximum 16 sets of gain shapes shown in FIG. The Rukoto.

FIG. 5 shows that Table 1 generated in this way is used to further calculate the combination closest to the current input audio signal from among the set of combinations of shape vectors and gains corresponding to each number of encoded bits. A flowchart diagram for searching shape vectors and gains is shown.

In this processing procedure, first, as an initial setting, the optimal error power is set to infinity, the optimal vector is set as a temporary number 0, the optimal gain is set as a temporary number 0, and the optimal gain/shape vector is set as Let a be the optimal code word (index) according to (step S1
1).

[0025] Then, as the initial setting of the current code word, a
= 0 (step S12), and the initial setting of the current shape vector number is set to m = 1 (step S13).

[0026] Then, shape vector V(m)
are synthesized by the synthesis filter 2 (step S14), and n
The current gain number is initialized as =1 (step S15).

After that, the shape vector synthesized by the synthesis filter 2 is multiplied by the nth value of the gain that can be obtained when encoding with k bits in the multiplier 3 (step S16).
, the error power is calculated by the calculation unit 4 (step S17).

As a result, it is determined whether or not the obtained error power is smaller than the optimal error power (step S18), and if it is smaller than the optimal error power, the current error power is set as the optimal error power, and a is set as the optimal code word (step S19).

After that, when the error power is larger than the optimal error power, a is incremented by "1" (step S20), and it is determined whether or not execution has been performed for all gains (step S21), and the execution is performed. If it has not been completed, n is incremented by "1" (step S22) and the process returns to step S16.

Furthermore, when the execution has been completed for all the gains, it is determined whether or not the execution has been executed for all the shape vectors (step S23), and if the execution has not been completed, m is incremented by "1" (step S23).
24), return to step S15.

[0031] In this way, it is possible to search for a more optimal combination of gain shapes from among the optimal set shown in FIG. Become.

[0032]

[Effects of the Invention] As explained above, according to the present invention,
Among all combinations of shape vectors and quantized gains, the optimal combinations that can be taken for each number of encoded bits are determined in advance and stored in a table, and such optimal combinations of shape vectors and gains are determined in advance. When encoding with a desired number of bits from a set of Since the index is transmitted using the variable rate, it is possible to minimize the deterioration in quality due to the variable rate.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the principle configuration of a gain shape vector quantization method according to the present invention.

FIG. 2 is a diagram showing an embodiment of a gain shape vector quantization method according to the present invention.

FIG. 3 is an explanatory diagram for obtaining the table shown in the embodiment of FIG. 2;

FIG. 4 is a flowchart showing a procedure for pre-determining a table used in the present invention.

FIG. 5 is a flowchart showing a procedure for searching for an optimal pair from a set of gain-shape vector pairs corresponding to each encoded bit.

FIG. 6 is a block diagram showing a conventional example.

[Explanation of symbols]

1 Table 2 Synthesis filter 3 Multiplication unit 4 Error power calculation unit 5 Optimal gain-shape vector pair search unit In the diagrams, the same reference numerals indicate the same or equivalent parts.

Claims (1)

[Claims] Claim 1: All combinations 2K (2K = 2N) of 2N (N is a positive integer) shape vectors and 2M (M is a positive integer) quantized gains.
x 2M) for each number of encoded bits and is stored in advance and stored, and the shape vector output from the table (1). Filter (2)
and the table (
1) a multiplier (3) that generates a reproduced audio signal by multiplying the gain output from the multiplier (3); and an error power calculation unit (3) that calculates the error power between the reproduced audio signal from the multiplier (3) and the input audio signal. 4) When encoding with a desired number of bits, each shape vector is sent to the synthesis filter (2), and the gain corresponding to the number of encoded bits in each shape vector is sequentially sent to the multiplier (3). A gain/shape vector quantization method comprising: a search unit (5) that searches for an optimal pair of shape vector and gain from the error power and transmits the index.