KR19980031885A - An Adaptive Codebook Search Method Based on a Correlation Function in Code-Excited Linear Predictive Coding - Google Patents

Abstract

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an adaptive codebook search method in code-excited linear prediction, and more particularly, to a method and apparatus for estimating a gain vector from a past excitation signal, To an adaptive codebook search method that can be expressed as a scalar.

The adaptive codebook search method according to the present invention can model a multi-gain with a single gain using the correlation of excitation signals. Therefore, through the adaptive codebook search method according to the present invention, the encoding performance of the Multi-tap Multi-gain Adaptive Codebook (MMAC) scheme in the CELP structure can be reduced and a saving effect of 2 bits per subframe can be obtained.

Description

An Adaptive Codebook Search Method Based on a Correlation Function in Code-Excited Linear Predictive Coding

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an adaptive codebook search method in code-excited linear prediction, and more particularly, to a method and apparatus for estimating a gain vector from a past excitation signal, To an adaptive codebook search method that can be expressed as a scalar.

The code excitation linear prediction (hereinafter referred to as CELP) coder consists mainly of a pitch filter and a noise codebook. The pitch filter is used to model the periodicity of speech, and it is common to use a structure called an adaptive codebook as a method of implementing the pitch periodicity.

The adaptive codebook search is to obtain the most suitable period and gain value for the speech to be analyzed with the past excitation signal. When the pitch of the speech is represented by the sample interval, the adaptive codebook shows excellent performance for the analysis speech in which the pitch interval is an integer value exactly. On the other hand, if the pitch interval is not an integral multiple of the sample interval, the adaptive codebook exhibits a sharp fall performance. In this case, a fractional pitch method and a multi-tap adaptive codebook method have been proposed to maintain the performance.

Fractional pitch (hereinafter referred to as FP) assumes that the pitch of the analysis speech is a prime rather than an integer. Generally, considering a transmission capacity of an encoder, only a fraction of 0.25 units is assumed as a pitch. First, the speech is oversampled to obtain a resolution of 0.25 units from the analyzed speech. Also, the oversampling of the past excitation signal is performed four times, and the period and the gain value are obtained through the adaptive codebook search as described above.

This FP method can maintain the performance of the adaptive codebook even if the pitch change is not an integer, while it can maintain the performance of the adaptive codebook for the oversampling and the impulse response filtering for comparison with the analyzed speech. Calculation is required. In addition, an additional bit for transmitting FP is required. For example, for FP method of 0.25 unit, 2 bits should be added.

The integer type adaptive codebook can be called a single-tap single-gain adaptive codebook (hereinafter referred to as SSAC) since the delay and the gain are each configured to be one. The adaptive codebook can be generalized to a Multi-tap Multi-gain Adaptive Codebook (hereinafter referred to as MMAC) structure having an arbitrary number of delays and gains. Because of the limit of the transmission capacity, the tap-delay exists near the reference value (pitch delay) in the multi-tap structure. In other words, when the pitch delay is T and the tap number is (2M + 1), the tap delay is (TM, T-M + 1 ,,,, T-1, T, T + 1, . At this time, .

The FP scheme requires a bit increase for FP, whereas the MMAC scheme requires an additional bit in the gain vector compared to the SSAC scheme.

It is an object of the present invention to provide a new adaptive codebook search method that simultaneously takes advantage of the low transmission rate of the SSAC scheme and the performance improvement of the MMAC scheme.

1 is a flowchart illustrating a conventional CELP encoding process.

FIG. 2 is a flowchart illustrating a conventional adaptive codebook search process.

3 is a flowchart illustrating a conventional noise codebook search process.

4 is a graph showing the range of gain values according to the change of the decimal part of the pitch in the MMAC system.

5 is a flowchart illustrating an adaptive codebook search method according to the present invention.

According to another aspect of the present invention, there is provided an adaptive codebook search method,

Past excitation signal Modeling with a secondary smoothing filter as follows:

remind A coefficient that minimizes the sum of the square of the error signal of , By the following equation;

The coefficient , To obtain an excitation signal represented by the following equation;

remind A pitch delay T satisfying the following expression, and a gain value?

here, Wow As the signal after passing through the adaptive codebook excitation signal and the filter respectively

being.

The method comprising the steps of:

The adaptive codebook search method according to the present invention can model a multi-gain with a single gain using the correlation of excitation signals. Therefore, through the adaptive codebook search method according to the present invention, the coding performance of the MMAC scheme in the CELP structure can be reduced and a saving effect of 2 bits per subframe can be obtained.

BRIEF DESCRIPTION OF THE DRAWINGS FIG.

FIG. 1 is a flowchart illustrating a coding process of a conventional CELP encoder.

In the process shown in FIG. 1, in step 101, a certain period (frame) of a voice to be analyzed is sampled. Generally, one frame corresponds to 20 to 30 ms, that is, 160 to 240 sample values for 8 KHz sampling.

In operation 102, high-pass filtering is performed on the sampled audio data of one frame to remove direct current components.

In operation 103, the feature parameters of the speech . These feature parameters (hereinafter referred to as LPC coefficients) correspond to the coefficients of a polynomial obtained when a speech signal weighted by a window function is approximated to a linear polynomial of a p-th order as shown in Equation (1).

[Equation 1]

here, And corresponds to a coefficient that minimizes the value of Equation (2).

&Quot; (2) "

here, to be.

The LPC coefficients thus obtained are converted into Line Spectrum Pairs (LSP) coefficients having better subframe interpolation characteristics in step 104 before being quantized and transmitted.

The LSP coefficients are quantized in operation 105.

In operation 106, the LSP coefficient is inversely quantized to synchronize the encoder and the decoder.

The speech periodicity is removed from the analyzed speech parameters, and the speech interval is divided into four subframes in detail to model it as the adaptive codebook. That is, the lengths of the voice segments of each subframe are .

The LSP coefficient for the i < th > Can be obtained by the following equation (3).

&Quot; (3) "

here, Wow Represents the i-th LSP coefficient of the immediately preceding frame and the current frame, respectively.

In operation 108, the LSP coefficients are converted into LPC coefficients. In steps 109, 110, and 111, a speech synthesis filter and an error weighting filter are constructed from the sub-frame LPC coefficients.

Speech synthesis filter And error weighting filter Is expressed by the following equations (4) and (5).

&Quot; (4) "

&Quot; (5) "

here, LSP coefficient Lt; / RTI >

Step 109 removes the influence of the synthesis filter of the previous subframe. Zero-Input-Response (ZIR), Can be obtained by the following Equation (6).

&Quot; (6) "

here, Indicates a synthesized signal in the immediately preceding frame.

The result of this ZIR is subtracted from the original speech signal S (n). This .

Step 113 and Step 114 are respectively the adaptive codebook and the noise codebook Which is the closest to the codebook search.

2 is a flowchart illustrating a conventional search process of an adaptive codebook. The error weight filter corresponding to Equation (5) The signal And a speech synthesis filter, respectively. A signal to which an error weighting filter is applied . First, the case of the SSAC scheme will be described.

Using the adaptive codebook, the value produced with a delay of L The signal filtered in operation 202 is And minimizes the difference between the two signals Wow Is found by the following Equation (7) to Equation (9) in Step 203.

&Quot; (7) "

&Quot; (8) "

&Quot; (9) "

The Wow The error signal Quot; This value is expressed by Equation (10).

&Quot; (10) "

Therefore, the search result of the SSAC scheme is expressed by Equations 8 and 9 .

The MMAC scheme is similar to the SSAC scheme. Here, description will be made only when the number of multi-tap in the MMAC scheme is 3, that is, when M = 1. The excitation signal generated with a delay of L The excitation signal for MMAC is , , And Are used.

The signals after passing through the filter of step 202 are , , And . Therefore, the difference from the target signal is expressed by Equation (11).

&Quot; (11) "

Minimizing the sum of the squares of Equation (11) , Is given by the following equations (12) and (13).

&Quot; (12) "

&Quot; (14) "

Error signal Is expressed by the following equation (14).

&Quot; (13) "

Since the MMAC scheme performs better than the SSAC scheme, but the gain is given as a vector, it must be quantized by a vector quantizer that requires more than 2 bits over the scalar quantizer.

3 is a flowchart illustrating a conventional noise codebook search process. The i-th code word of the M codebook The signal filtered in operation 301 is . The optimal code word and codebook gain are expressed by the following equations (15) to (17).

&Quot; (15) "

&Quot; (16) "

&Quot; (17) "

The finally obtained excitation signal of the voice filter is expressed by Equation (18) for the SSAC scheme,

&Quot; (18) "

In the case of the MMAC scheme, Equation 19 is used.

&Quot; (19) "

The result of Equation 19 is used to update the adaptive codebook for analysis of the next subframe.

4 is a graph showing the distribution of gain values according to the integer pitch T and the decimal pitch P in the MMAC system. First, it is assumed that the spectral characteristics of speech are constant regardless of the change in the pitch interval.

For example, if T = 30 and P = 30.25 Has a value between 0.0 and 1.0. Also, 0.0 > Conversely, when T = 30 and P = 29.75 0.0 < / RTI > Has a value between 0.0 and 1.0. If T = P Lt; / RTI & The FP method, the MMAC method, and the SSAC method are the same.

In the present invention, a gain vector is estimated from past excitation signals. From this estimated value, the third order gain vector can be represented by scalar. As a result, an adaptive codebook can be designed with a scalar gain while having a multi-tap structure.

The following describes the operation of the MSAC scheme that is devised through the present invention.

The excitation signal of the adaptive codebook can be modified as shown in FIG. 4 and the following equation (20).

&Quot; (20) "

That is, the gain values of the MMAC scheme satisfy the condition shown in the following Equation (21).

&Quot; (21) "

The relationship between the three gain values is shown in Equation (22).

&Quot; (22) "

5 is a flowchart illustrating a search process of the MSAC according to the present invention.

In the present invention, , , Instead of , Is estimated as an excitation signal in the past and only the entire gain? Is calculated in the adaptive codebook search process.

here, , Past excitation signal Can be modeled as a coefficient of a secondary smoothing filter as shown in the following Equation (23).

&Quot; (23) "

A coefficient that minimizes the sum of squares of the error signal of Equation (23) , Can be obtained by the following equation (24).

&Quot; (24) "

The coefficient obtained from equation (24) , The pitch delay T, and the gain value? As shown in Equation (25) as in Equation (20).

&Quot; (25) "

here, Wow Corresponds to the adaptive codebook excitation signal of Equation (20) and the signal after passing through the filter of Step 202 and is expressed as Equation (26).

&Quot; (26) "

The final error signal Is calculated by the following equation (26), and obtained in step 403.

&Quot; (27) "

As described above, the adaptive codebook search method according to the present invention has a multi-tap structure by estimating a gain vector from a past excitation signal and expressing a third-order gain vector from the estimated value in scalar it is possible to design an adaptive codebook with a scalar gain.

Claims (1)

1. An adaptive codebook search method of code-excited linear prediction coding, Past excitation signal Modeling with a secondary smoothing filter as follows: remind A coefficient that minimizes the sum of the square of the error signal of , By the following equation; The coefficient , To obtain an excitation signal represented by the following equation; And remind A pitch delay T satisfying the following expression, and a process of obtaining a gain value? (here, Wow As the signal after passing through the adaptive codebook excitation signal and the filter respectively being) / RTI >