KR0179248B1 - High-speed ALP-VISIELPI speech coding method - Google Patents

Abstract

A high-speed VSELP speech encoding method of the present invention includes: obtaining M Q-dimension mutually orthogonal regular orthogonal regular pulse magnitude vectors Bm (1? M? M) and storing them in a codebook; (Where D > = 4) 0's are inserted so that the first pulse is located at the first component (k = 0) of the M basis vectors, and at the same time the M mutually orthogonal regular pulse basis vectors Vk , m), and a step of backward-filtering the target vector y using the weighted synthesis filter H after the long-term prediction, Obtaining an inner product Ψk (m) of the backward-filtered filtered vector and a normal vector normal pulse basis vector (Vk, m); obtaining a sign of the inner product Ψk (m) (M), obtaining an absolute value of? K (m) and obtaining a sum Ck for 1? M? M, and determining whether the pulse position index k is a final pulse position of the basis vector 1 is added to the pulse position index, if k = K-1, and if k = k + 1, the backward filtered target vector To the left by a D / K sample interval and then to the next step of backward filtering the target vector y until the pulse position index k becomes a final pulse position index K-1; And selecting an optimal basis vector by searching a pulse position index k for maximizing the sum Ck and a component? K (m) of the codebook index at that time when the index k is K-1.

Description

High speed ALPI-VISELPHI (RP-VSELP) speech coding method

The present invention relates to a high-speed speech coding method using a new vector sum codebook, which is very effective in code search and memory. More particularly, the present invention relates to a vector sum codebook composed of mutually regular regular pulse size vectors and a reverse vector- (Fast Regular Pulse Vector Sum Excited Linear Prediction) speech coding which can greatly reduce the amount of calculation required for a codebook search in VSELP (Vector Sum Excited Linear Prediction) coder while maintaining the sound quality of the existing VSELP coder ≪ / RTI >

Conventional Code Excited Linear Prediction (CELP) has been recognized as the most promising method for producing a high quality sound at a low bit rate of 4800 bps or less. However, the CELP has a disadvantage that the calculation amount is very large. In order to overcome these drawbacks, a Ternary codebook, an Overlapping Sparse codebook (WB Kleijin, DJ Krasinsky and R. Ketchum, An Efficient Stochastically Excited Linear Predictive Coding Algorithm for High Quality Low Bit Rate Transmission M. Alperbratic codebook (JP Adoul, P. Mabilleau, M. Delprat and S. Morisette, eds., "Speech Communication, Vol. 7, No. 3, pp.305-316, Oct. 1988.) (See, for example, I Gerson, and MA Jasuik, "Fast CELP Coding Based on Algebraic Codes, IEEE Proc. Int. Conf. Acoust. Speech and Signal Proc., Pp.1957-1960, RA Sumami, Binary Pulse Exciatation: A Novel Approach to Low Complexity CELP Coding, Advances in Speech (in Japanese), Vol. Coding Kluwer Academic Publisher pp.145-156, 1991.). Among these methods, VSELP using a vector sum codebook has shown remarkable performance in all aspects of computational complexity, sound quality, and robustness against channel errors.

That is, as shown in the block diagram of the codebook search process of FIG. 1, the conventional VSELP speech coding method comprises the steps of (1) storing M N-dimensional basis vectors in a codebook so that a linear combination of random Gaussian basis vectors is formed (2) generating an excitation signal filtered through a weighted synthesis filter of the basis matrix of the M × 1 matrix, and (3) generating M × 1 binary vectors corresponding to the codebook index θ ④ multiplying the signals by the multiplication factors and multiplying the signals by the multiplication factors and multiplying them by the gain and multiplying the result by the gain to obtain an N-dimensional synthesized signal; ⑤ obtaining an error by the difference between the target data signal and the synthesized signal; The above process is repeated for the binary vector to obtain the binary vector? That minimizes the error.

According to the VSELP speech coding method as described above, the optimal codeword is selected by maximizing the scaling function expressed by Equation (1) such that the error is minimized.

In Equation (1), y represents a target data signal that is a psycho-weighted signal after past long section prediction, and H is an NxN Toepliz impulse response matrix of weighted synthesis filter 1 / A (2 /?).

The code word search in the VSELP speech coding method is performed by calculating Equation 1 through the above procedure. The search procedure is simplified using the characteristics of the vector sum codebook. In the codebook search process, By structuring the index and the single bit differently, we have made the codebook search using binary gray codes simple.

However, the above-described VSELP speech coding method has a greatly reduced computational complexity compared with the CELP method. However, as can be seen from Equation (1), the codeword retrieval process requires both denominator and numerator to be computed, .

Accordingly, an object of the present invention is to design a basis vector of a VSELP as a mutually regular orthogonal regular pulse size vector and perform a codebook search using a target vector after a backward-filtered long section prediction to maintain the sound quality of a VSLEP encoder, The present invention is to provide a high-speed RP-VSELP speech coding method using a regular basis regular basis vector with a considerably reduced calculation amount.

FIG. 1 is a diagram illustrating modeling of a speech encoding method by conventional vector sum-excitation linear prediction; FIG.

FIG. 2 is a diagram illustrating modeling of a high-speed speech coding method by regular pulse vector sum-excitation linear prediction of the present invention;

FIG. 3 schematically shows a process of the high-speed speech encoding method of the present invention. FIG.

FIG. 4 is a flowchart showing a high-speed speech coding method according to the present invention.

According to another aspect of the present invention, there is provided a high-speed speech encoding method comprising the steps of: obtaining M Q-dimension normal regularity regular pulse size vectors Bm (1? M? M) (Where D > = 4) 0s are inserted between the M components of the basis vectors, and the first pulse is located at the first component (k = 0) of each of the M basis vectors and at the same time, (Vk, m), generating a normalized regular pulse basis vector (Vk, m), and backward-filtering the target vector y by using a weighted synthesis filter (H) (M) of the backward filtered filtered object vector and the mutually orthogonal regular pulse basis vector (Vk, m); obtaining a sign of the inner product Ψk (m) as a component of the codebook indexer (m), obtaining an absolute value of? k (m) and obtaining a sum Ck for 1? m? M, and calculating a sum Ck as a value of the pulse position index k K-1, and if k = K-1, adding 1 to the pulse position index to set k = k + 1 and outputting the backward- Repeatedly until the pulse position index k becomes K-1, which is the last pulse position index, after the backward filtering step of y of the target vector after shifting to the left by the D / K sample interval, And selecting an optimal basis vector by finding a pulse position index k that maximizes the sum Ck and a component θk (m) of the codebook index at that time when the position index k is K-1. .

BRIEF DESCRIPTION OF THE DRAWINGS Fig.

FIG. 2 illustrates modeling of a high-speed speech coding method of Fast Regular Pulse-Vector-Sum Excited Linear Prediction (RP-VSELP) of the present invention. The high speed RP-VSELP speech coding In the method, the basis vectors are composed of M Q-dimension reciprocal regular periodic pulse magnitude vectors.

That is, in the present invention, a regular vector matrix B is defined as a regular pulse size matrix that is not a zero-order constant Q × M matrix, and a basis vector designed in a regular form is regularized using a known Gram-Schmidt (GS) method.

Let D and K denote the pulse interval and the maximum possible number of the first position, and Q is the number of pulses, and Pk = [Pij] is NxQ indicating the regular pulse position at a given index k, Matrix.

Therefore, for a given pulse position index using equations (1) and (2), the basis matrix Vk is modeled as follows.

In this model, the basis vector consists of statistical pulses at predetermined pulse positions, and the basis vector of length N consists of Q pulses separated by D-1 zeros. Where Q = [N / D] and [x] is the largest integer less than or equal to x. A pulse position index and a codebook index are sequentially determined using a regular pulse excitation (RPE) approach to find an optimal code modeled by Equation (3). Equation (3) is substituted into Equation (1) and Equation (1) becomes Equation (4) if Hk = HPk.

Where T denotes transpose matrix. In the present invention, the autocorrelation function of the impulse response of the weighted synthesis filter is defined as Equation (5).

Then, it can be observed through experiments that γ (i) «γ (0) for i≥4, and the product of the matrix Can be approximated to < RTI ID = 0.0 > y (0) < / RTI >

Where γ (0) is the energy of the impulse response and I is the unit matrix. Therefore, Equation (4) is simplified as shown in Equation (6).

In Equation (6), B is a mutually orthogonal matrix ego, , Equation (6) becomes more simplified as follows.

The optimal codevector of the present invention is shown in Equation 7 for a given regular pulse position index k Is selected to maximize the cost function < RTI ID = 0.0 > e, < / RTI >

here The target signal y is back-filtered before searching the codebook.

Since the denominator of epsilon is constant regardless of the binary vector [theta] k, the scaling function [epsilon] Lt; RTI ID = 0.0 > Pk. ≪ / RTI >

The optimal pulse position k can be selected by maximizing? K in equation (8). Since the signs of θk (m) and Ψk (m) are the same, the optimum pulse position is obtained by maximizing the positive value Ck.

Ck can be obtained as shown in Equation (11).

Where Z is a delay matrix, to be.

In Equation (11), the object signal y is obtained by backward-filtering Is a vector irrespective of k, and a mutually orthogonal regular pulse basis vector end (Right-shifted) by the D / K sample and can be converted into data as follows.

The process of obtaining? K from Equations (11) and (12) can be formulated as shown in Equation (13).

In other words, Equation (13) represents a backward-filtered object vector And the normalized regular pulse basis vector, the position against the zero value should be excluded in the inner product calculation.

The embodiments of the present invention will be described in detail with reference to the accompanying drawings on the basis of the constitution principle of the present invention.

FIG. 3 is a diagram schematically illustrating a high-speed RP-VSELP speech encoding method of the present invention, and FIG. 4 is a flowchart of a high-speed RP-VSELP speech encoding method of the present invention.

First, M Q dimension (Q pulses of different size) mutually orthogonal regular pulse size vectors are made Bm (1? M? M) and stored in the codebook.

Then, a pulse having a size of D-1 (D? 4) 0s is inserted between each element of the vector Bm, so that the first pulse is located at the first component (k = 0) of each basis vector and D (M rows) of mutually orthogonal regular pulse basis vectors Vm (1? M? M) having Q pulses of different sizes, which are located at the sample intervals of?

After the long interval prediction, the target vector y is backward filtered to obtain the backward filtered target vector y .

The backward filtered target vector And the inner product Ψ ₀ (m) of the mutual orthogonal basis pulse V ₀ , m. At this time, the position corresponding to the 0 value between the pulses is excluded in the internal calculation so as to reduce the calculation amount Depending on the conditions of the dot product calculated Ψ ₀ (m).

Following is the sign of Ψ ₀ (m) to the value of the codebook index component θ ₀ (m). And the sum of the absolute values of? ₀ (m) for (1? M? M) .

Then pulse position index k = 1 dot product sum of the absolute values C ₁ and then the code book and then add 1 more to the position index k so as to obtain the composition θ ₁ (m) of the index object backward filtered vector of the Is shifted to the left by a sample interval of D / K (for example, D / K = 1 in the case of D = K = 4) to obtain a backward filtered target vector As described above, when calculating the inner product Ψ ₁ (m) of the regular regular basis pulse basis vector V ₀ , m, the position corresponding to the zero value is excluded in the inner product calculation.

The sign of the inner product Ψ ₁ (m) is taken as the value of the component θ ₁ (m) of the codebook index. The absolute value of? ₁ (m) is obtained, and the sum of all the values of ₁ ? M? M .

Next, k = 2 is added to the pulse position index k as in the case of k = 1, and then the backward filtered target vector k Is shifted to the left by the D / K sample interval and the backward filtered target vector for k = 2 To obtain a regular in the same manner described above in the case of the above-mentioned k = 1 mutually orthonormal basis vector V _0, inner Ψ ₂ (m) and the corresponding components of the codebook index in the m θ ₂ (m) and Ψ ₂ (m) Sum of absolute values of .

and k = K-1, the target vector (M) and the corresponding codebook index component θk (m) of the regular regular basis interval vector V ₀ , m, taking the absolute value of the inner product Ψk (m) Thereby obtaining the sum Ck.

When the final pulse position index k = K-1, the optimum basis vector is selected by finding the pulse position index k that maximizes Ck and the component? K (m) of the mode book index at that time.

In this embodiment, whenever the pulse position index k is incremented by 1, the previous reverse-filtered Is shifted to the left by the D / K sample interval and a new reverse-filtered And the inner product is obtained, but the inner product may be obtained by adopting a scheme of moving the basis vector Vk to the right by the D / K sample interval.

A high-speed RP-VSELP speech coding method using the backward filtering of the object vector of the present invention and a regular pulse vector sum-excitation linear prediction (RP-VSELP) method using the conventional VSELP scheme and forward filtering of basis vectors Table 1 compares the computation amount of the codebook search process and the memory required.

VSELP, RP-VSELP, and the codebook operation procedure of the fast RP-VSELP speech encoding method of the present invention when N = 60, M = 11, D = 4, Q = 15 and K = The number of calculations was 34660, 8088, and 2578, respectively. In the present invention, the amount of calculation was reduced to about 13 times and about 3 times that of the VSELP method and the RP-VSELP method, respectively. 165 and 165, the present invention is shown to be four times less than the conventional VSELP method.

As described above, according to the present invention, the basis vectors of the VSELP are designed as mutually regular regular basis pulse basis vectors, and the inner product with the normal regular basis pulse basis vectors is reverse filtered by the target vector, By selecting the optimum basis vector by searching the pulse position index that maximizes the sum of the absolute values and the components of the codebook index at that time, the amount of calculation and the memory in the codebook search are equivalent to those of the conventional VSELP method, There is an excellent effect.

Claims (1)

(1 ≤ m ≤ M) of M Q-dimension mutually regular orthogonal regular pulse size vectors and storing them in a codebook; (V ₀ , m) positioned at a sampling interval of D at the same time that the first pulse is located at the first component (k = 0) of each basis vector, and after the long interval prediction, The backward-filtered backward-filtered target vector Obtaining a backward filtered target vector, (M) of the inner product Ψk (m) with the mutual orthogonal regular pulse basis vector (V ₀ , m), taking the sign of the inner product Ψk (m) Obtaining a value Ck obtained by adding absolute values of 1? M? M to absolute values of? K (m), and checking whether the pulse position index k is a final pulse position index K-1 of the basis vector If the pulse position index k is not K-1, 1 is further added to the pulse position index, the backward filtered target vector u at the time k-1 is shifted to the left by D / K sample interval, And repeating the routine until k is the last pulse position index k = K-1; and if k = K-1, calculating a sum Ck of the inner absolute values to a maximum And the codebook index < RTI ID = 0.0 > k < / RTI & Fast RP-VSELP speech coding method which is characterized by comprising the step of: Find component θk (m) of the switch selecting an optimal basis vectors.