JP3209247B2 - Excitation signal coding for speech - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-efficiency speech coding method for digitally coding a speech signal with a minimum amount of information, and more particularly to an excitation signal coding method for determining an excitation signal to be supplied to a linear prediction synthesis filter. It is.

[0002]

2. Description of the Related Art As a speech encoding method of about 8 kbit / s or less, CELP (Code Excited Lin) is used.
Ear Prediction (Code Excited Linear Prediction) coding is well known. As shown in FIG. 5, this encoding method uses a random codebook 11 and an adaptive codebook 12 in frame units.
, The excitation vector is selected from the
The gain is adjusted in steps 3 and 14 and input to the linear prediction synthesis filter 15 to synthesize speech. The difference between the synthesized speech and the input speech is obtained by a difference circuit 16, the difference output is passed through an audibility correction filter 17, and a distortion is calculated by a distortion calculator 18 from the filter output. Select and adjust the gain. The criterion for selecting the excitation vector is to minimize the audible error between the synthesized signal and the input signal. Since the excitation vector is determined by feeding back the finally synthesized waveform in this manner, high quality can be obtained.

On the other hand, on the other hand, there are disadvantages that it is necessary to store a large amount of noise excitation vectors and that the amount of calculation for searching the excitation vectors becomes enormous. For this reason, several methods have been considered to reduce the storage capacity and the amount of calculation. However, the basic principle of the search will be described below, and typical calculation amount reduction methods related to the present invention will be summarized. introduce.

In searching for a noise excitation vector in CELP coding, as described above, a linear prediction synthesis filter 15 is used.
Then, the difference between the output synthesized waveform and the input voice is obtained, and the vector having the minimum distortion after the difference output is passed through the audibility correction filter 17 is selected. Specifically, an n-dimensional vertical vector (a current component obtained by subtracting a response component from a previous frame and a synthetic component from an adaptive code vector from an input speech) in a certain frame (the frame length is n samples). X is the signal vector obtained by passing the input speech of the frame through the audibility correction filter, C is the ^j- dimensional vertical vector of m kinds of noise excitation vectors in the codebook (j = 0,..., M−1), The impulse response matrix (n × n) of the synthesis filter 15 is H, and the gain of the noise excitation vector is g.
Then, j that minimizes the following distortion d is searched.

[0005] Here, assuming that an optimum gain is given and that the gain is quantized after the determination of the noise excitation vector, g = X ^T HC ^j / ((HC ^j ) ^T (HC ^j )) (3) next (X ^T is the transpose matrix of X), the distortion in this case is ^{d = X T X- (X T} HC j) 2 / ((HC j) T (HC j)) (4). Eventually, it is sufficient to search for j that maximizes the second term, f = (X ^T HC ^j ) ² / ((HC ^j ) ^T (HC ^j )) (5). Where HC ^j
The synthesis operation requires an enormous amount of computation to synthesize C ^j for all j required sum of products about n ^2/2 times. Therefore, various proposals have been made.

[0006] First molecules of formula (5) calculates the X ^T H above once and method to reduce the amount of calculation without causing approximation error Taking C ^j and the inner product is known for its later I have. The inner product requires n operations. However, if an attempt is made to reduce the amount of calculation in the calculation of the denominator, an approximation error occurs. For example the denominator term as ^{(C j) T H T HC} j, approximates the H ^T H in the correlation matrix of the Toeplitz type, a method of calculating by the inner product of the autocorrelation function of the autocorrelation function and C ^j of H is known ing. As a result, the denominator term can be calculated by n times of product sum, but the approximation error is large, causing an increase in distortion and a deterioration in quality.
In order to store the correlation value for each noise excitation vector C ^j, additional storage capacity and comparable storage capacity for the codebook is required.

[0007] Apart from this, there is also known a method in which a low-pass filter is applied to C to obtain an excitation vector downsampled, and the term of the denominator is approximated by this vector. Although the amount of calculation can be reduced by the downsampled ratio, this method also has a problem that the approximation error and the storage capacity for the downsampled vector increase. There is also known a method in which the search is divided into two stages, and a plurality of ^Cj are preliminarily selected as candidates in an approximate calculation in the first stage, and only those candidates in the second stage are selected so as to maximize Expression (5). ing. As the first approximation method, a method using only the numerator term, a method using the above-described approximation method, and the like can be considered. Distortion can be reduced by reducing the approximation error in the first stage or increasing the number of candidates, but the effect of reducing the amount of computation is reduced.

[0008]

SUMMARY OF THE INVENTION An object of the present invention is to improve the quality of a coded speech while encoding the speech with a small amount of information while minimizing the amount of computation as much as possible. It is an object of the present invention to provide a speech excitation signal encoding method capable of significantly reducing the amount of computation in spite of small distortion degradation in combination with a two-stage search.

[0009]

According to the present invention, when a noise excitation vector is selected (or preselected), equation (5) is used.
Although the term of the denominator of, that is, the term of the energy of the vector is also taken into consideration, the term of the energy is characterized in that it is obtained by reading an estimated value stored in advance. The estimated value of the energy is stored in advance for each noise excitation vector and each classification of the synthesis filter. FIG. 1 shows a comparison between the method of the present invention and the conventional method when a two-stage search is performed. Conventionally, the energy term is completely ignored by ignoring the energy term as shown in FIG. 1B, or the energy term is obtained by performing an approximation operation as shown in FIG. It has been known that the equation (5) is approximated and preliminarily selected using this, but in the present invention, the energy term is estimated by referring to a table based on the classification of the impulse response as shown in FIG. 1A. Using this, equation (5) is approximated and preliminarily selected.

[0010]

FIG. 2 shows the structure of a main part of the most basic embodiment of the present invention. In addition to the codebook 11 composed of m normal n-dimensional noise excitation vectors, in this embodiment, a storage unit 21 storing a plurality of types of impulse response patterns as a type of a synthesis filter, and the storage unit 21
A selector 22 for selecting a code corresponding to an impulse response sequence from the above, an impulse response pattern as a vertical axis, a noise excitation vector as a horizontal axis, and an energy estimated value after synthesis or its reciprocal is pre-calculated and stored in each element. Energy term table 23 is provided. A search procedure when a two-step search is performed is as follows.

1. When the characteristics of the synthesis filter and the audibility correction filter are given in a certain frame, the impulse response of the synthesis filter is determined. This impulse response and the storage unit 21
The selector 22 compares the response pattern with the response pattern stored in advance in the selector 22 to determine the pattern of the impulse response of the current frame. 2. Next, the term of the numerator of the equation (5) is calculated for all the noise excitation vectors.

3. The energy estimation value corresponding to the determined impulse response from the selector 22 is obtained with reference to Table 23, and for each noise excitation vector, (numerator term / energy estimation value) is calculated. Leave a set number of candidates from the largest. 4. Formula (5) is calculated only for the preselected candidates, and the sign of the optimal noise excitation vector is determined.

In the above process, the impulse response is classified.
To determine the pattern of the impulse response
The impulse response, that is, the distance of the FIR filter coefficient
Separation scale method, all poles approximating impulse response
There is a method that uses a measure of the type spectral model. Fret
H (where h
^T= ｛H₀, ..., h_n-1｝), M types of representative patterns and
The impulse response vector to be storedⁱ(I =
0,..., M−1), in the former case, minimize e.
You will select a representative pattern.

E = {h−h ′ ⁱ } ² (6) As a modification, weighting of a triangular window W in which a higher-order coefficient decreases as in the following equation, and distance calculation only in a lower order are also performed. Conceivable. e = {W (h−h ′ ⁱ )} ² (7) This takes into account that since the impulse response matrix is a lower triangular, lower-order coefficients contribute more to energy calculation.

On the other hand, when an all-pole spectral model is used, a representative pattern is selected using a maximum likelihood distance scale that minimizes the following e. e = Σα _i Φ _i (8) where Σ is from i = 0 to p, p is the order of the model, α
_i is the i-th linear prediction coefficient of the representative model, Φ _i is the i-th autocorrelation function of the impulse response, and this measure is a measure frequently used in speech recognition. In order to create a small number of representative patterns prepared in advance, the representative pattern design algorithm used in normal vector quantization may be used as it is.

In the above embodiment, the basic CE shown in FIG.
Although the description has been made on the assumption that the LP coding is used, the present invention can be applied to a method of orthogonalizing all the noise excitation vectors with the pitch component before searching for the noise excitation vector, or to a case where the noise excitation vector has a plurality of channels. . That is, a case where the noise excitation vector is orthogonalized to the pitch component (output from the adaptive codebook) n-dimensional vector P will be described as a second embodiment. In this case, the expression (5) that finally maximizes is modified as follows.

F = (X ^T HC ^j −ρHP) ² / ((HC ^j ) ^T (HC ^j ) −ρ (HC ^j ) ^T (HP)) (9) where ρ = (HC ^j ) ^T (HP ) / ((HP) ^T (HP)) (10). In equation (9), the term of the numerator can be calculated accurately and with relatively few operations as in the case of equation (5). Denominator ρ
The same applies to the following second term and the calculation of the sum of products. Therefore, if the first term of the denominator is read from the table in the same procedure as in the first embodiment, the amount of calculation can be reduced. Further, even if the calculation of the second term of the denominator is omitted altogether and the evaluation is performed only with the first term, the distortion hardly increases.

As a third embodiment, a case in which noise excitation vectors of two channels are provided will be described. Here, if a two-stage search is performed, the processing shown in FIG. 4 is performed. Assuming that the excitation vector of the first channel is C ^j and the excitation vector of the second channel is C ′ ^k , the final evaluation scale is as follows.

[0019]

(Equation 1) However, in the preliminary selection for each channel,
The same processing as in the first embodiment is performed. That is, the approximate value of the energy term of the denominator is read from the table, and a plurality of candidates that maximize the equation (5) are selected. Then, a combination of j and k that maximizes Expression (11) is searched for in a combination of candidates from each channel.

In the above description, in the energy term table 23, the reciprocal of the estimated energy value after the synthesis of the noise excitation vector may be stored for each excitation vector and each type of synthesis filter. In some cases, the preliminary selection may be the final selection without correctly calculating equation (5).

[0021]

According to the present invention, the input voice and the SNR of the coding distortion and the amount of calculation (unit MO)
PS Mega Operation Per Sec
FIG. 4 shows a comparison between the conventional method and ond: the number of operations performed in units of one million operations per second required for real-time processing. In this case, the noise excitation vector includes two channels, and 7 bits (excluding polarity) are allocated to each channel. The dimension number n of the noise excitation vector is 4
At 0, orthogonal processing with the pitch component is used together. The impulse response was classified into four types. Here, the conventional method is a method in which the preliminary selection is performed using only the numerator of the equation (5). The numbers in the figure are the number of candidates to be left out of 128 for each channel. From this figure, according to the present invention, there is almost no increase in the amount of computation as compared with the conventional preliminary selection.
It can be seen that the SNR has been improved. If the number of candidates to be left in the present invention is the same, the subdivision of the increasing amount of computation is as follows.
The product of the reciprocal of the term and the denominator term is once for each noise excitation vector in the preliminary selection, which is a very small amount as compared with the calculation amount of the distance calculation for the search. The storage capacities increased in the present invention are the storage capacities of Table 21 and Table 23 for the classification of the impulse response. For example, when the impulse responses are classified into four types as in the case of FIG. 3 and a 40-dimensional noise excitation vector is used, the rate at which the storage capacity increases is about 10%.

[Brief description of the drawings]

FIG. 1 is a diagram showing processing of the present invention in a two-stage search in comparison with a conventional method.

FIG. 2 is a block diagram showing a main part of the embodiment of the present invention.

FIG. 3 is a block diagram showing a main part of an embodiment in which the present invention is applied to the case of a two-channel noise excitation vector.

FIG. 4 is a diagram comparing the relationship between the SNR and the amount of calculation according to the present invention with the conventional method.

FIG. 5 is a block diagram showing the basic principle of the CELP encoding method.

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁷ , DB name) G10L 19/12

Claims (1)

(57) [Claims] 1. A speech synthesis model having a pitch period component vector and a noise excitation vector as excitation sources of a linear predictive synthesis filter, and a noise excitation vector in a codebook so as to minimize an error between synthesized speech and input speech. In the speech excitation signal encoding method selected from the following, the inner product of the synthesized waveform and the input speech waveform is obtained for each noise excitation vector, and the type of the synthesis filter is determined. From the table of energy estimation values after synthesis of the excitation vectors created for each type or the table of the reciprocals of the estimation values, the estimated values are read for each noise excitation vector in the determined type, and the square of the inner product value is calculated. A method for encoding a speech excitation signal, comprising selecting or preselecting a noise excitation vector based on a value obtained by dividing the estimated energy value.