JP2553745B2 - Speech analysis method and speech analysis device - Google Patents

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a new voice analysis method and voice analysis device for obtaining a linear prediction cepstrum coefficient from a linear prediction coefficient obtained from an arbitrary voice waveform in the field of digital voice signal processing. Is.

2. Description of the Related Art A linear prediction cepstrum coefficient is an efficient representation of an analysis spectrum obtained by a linear prediction analysis method with a relatively small number of coefficients.In recent years, a voice analysis method or apparatus for obtaining a linear prediction cepstrum coefficient from a linear prediction coefficient It plays an important role as a pretreatment method or device for recognition.

An example of the conventional voice analysis method described above will be described below with reference to mathematical expressions. The description of the conventional voice analysis device is omitted because it is easily inferred from an example of the voice analysis method described below. For simplicity, the linear prediction coefficient will be referred to as an a coefficient, and the linear prediction cepstrum coefficient will be referred to as a c coefficient hereinafter.

According to JD Markel and AH Gray, the relationship between the a coefficient and the C coefficient is given by the following fourth equation. ("Voice Linear Prediction", Corona Publishing, translated by Kuki Suzuki,
(Excerpt from p.284).

Generally, it is often expressed by the following fifth formula obtained by transforming this formula by variable conversion of n−k = m.

DISCLOSURE OF THE INVENTION Problems to be Solved by the Invention By the way, each term inside the summation symbol of the above-mentioned formula 5 has a total of three kinds of variables of a natural number m, an m-th order c coefficient _cm, and an nmth order a coefficient a _nm. It is the product of When this is obtained by a two-variable input type multiplier that is widely used in the field of voice signal processing, the product of the two variables out of the above three variables is once taken out from the output of the multiplier and the result is retained again. The final product must be obtained by inputting to the multiplier at the same time as one variable of. The above operation is necessary for each of the product terms in the summation symbol. When carrying out such a procedure, it is possible to repeatedly use a single multiplier or to separately use a plurality of multipliers for each product term. However, in the former case, there is a problem that it takes time to complete all the multiplications, and in the latter case, there is a problem that the device scale becomes large due to the large number of multipliers.

In addition, since the range of the values of the a coefficient and the c coefficient is not limited in principle, when calculating using the adder and the multiplier of finite word length and fixed decimal point expression, the coefficient itself is used. It is necessary to pay attention to the prevention of overflow during or during calculation. Conventionally, in consideration of the statistical frequency of appearance of the values of the a coefficient and the c coefficient, each coefficient was reduced to about 1/4 and expressed in 2's complement, and the calculation was performed by the following formula 6. However, the multiplication by the constant 4 is added to the second term on the right side of this equation, and there is a problem in that the operation is more complicated than in the case of the above fifth equation.

In view of the above problems, the present invention requires a small number of multiplications,
The present invention also provides a new speech analysis method capable of obtaining linear prediction cepstrum coefficients by a simple procedure.

Means for Solving the Problems Based on the following seventh equation, the nth-order linear prediction coefficient a _n to the nth
The procedure for obtaining the next surrogate variable A _n , the procedure for obtaining the nth-order surrogate variable F _n based on the recurrence formula shown in the following eighth formula, and the following ninth
And a procedure for obtaining the nth-order linear prediction cepstrum coefficient c _n from the n-th order surrogate variable F _n based on the equation.

However, a _n : n-th order linear prediction coefficient c _n : n-th order linear prediction cepstrum coefficient In any formula, 1 ≦ n ≦ p p: maximum order of linear prediction analysis.

Action The present invention newly introduces the proxy variables A _n and F _n
By introducing and, it is possible to eliminate the reduction operation of the a coefficient and the c coefficient as in the sixth equation and the complicated multiplication operation by the three variables and the one constant in the summation symbol.

Example Hereinafter, a voice analysis procedure as an example of the voice analysis method according to claim 1 will be described with reference to the drawings and mathematical formulas.

FIG. 1 shows a general-purpose digital signal processor whose data representation format and data operation format are finite word length and fixed point type (for simplicity, DSP
Is a procedural diagram showing a calculation procedure when it is carried out.

In the figure, 1 is a procedure for obtaining the nth-order proxy variable A _n from the nth-order a coefficient a _n based on the 7th equation, and 2 is the nth-order proxy based on the recurrence equation shown in the 8th equation. Procedure for obtaining variable F _n , 3
Is a procedure for obtaining the nth-order c coefficient c _n from the nth-order proxy variable F _n based on the above-mentioned equation 9. Both procedures have order n
Flows from 1 to the maximum order p of the linear predictive analysis.

The principle and operation of the operation procedure configured as described above will be described below using mathematical expressions. The formula below is
It is the conventional c coefficient calculation formula already described.

As already mentioned, this formula is complicated, so it is desirable to add some innovation so that the calculation procedure on the DSP is simplified by some modification of the formula. Therefore, once multiply each side of this equation by n / 8 to obtain the following equation, Introducing surrogate variables A _n and F _n Substituting Becomes the accumulation of product terms with only two variables in the sum signal,
It becomes a formula that can be calculated quickly and simply by DSP. This equation is the eighth one already mentioned in the section of means for solving the problem.
It is understood that this is the same as the formula and that the formula allows the proxy variable F _n to be calculated sequentially. The division by 8 in the equation can be easily realized by shifting the dividend by 3 bits in the LSB direction, so that there is no problem in implementation. Finally, the c coefficient is calculated from the surrogate variable F _n by the ninth formula. In this case, the division by n is relatively short by substituting right bit shift when n is a power of 2 and obtaining 1 / n in advance otherwise and multiplying by this. Can be calculated in time.

As described above, according to the present embodiment, based on the procedure for obtaining the nth-order surrogate variable A _n from the nth-order linear prediction coefficient a _n shown in the seventh equation, and the recurrence equation shown in the eighth equation. Nth surrogate variable
A step of obtaining the F _n, by providing a procedure for obtaining the ninth n-th linear prediction cepstrum coefficients c _n on the basis of the expression from the n-th proxy F _n, required when conventional linear prediction cepstrum coefficient calculation Therefore, it is possible to realize a speech analysis method based on the accumulation of product terms of only two variables which can be calculated quickly and simply by DSP without the need of product terms consisting of three or more variables in the summation symbol.

Hereinafter, an embodiment of the voice analysis device according to claim 2 will be described with reference to the drawings.

FIG. 2 shows a general-purpose digital signal processor whose data representation format and data operation format are finite word length and fixed point type (for simplicity, DSP from now on).
Is a block diagram when it is realized by ().

21 in the figure is a set of a coefficients {a _n | n = 1, 2, ...
p} is the first storage means, 22 is the second storage means for storing the set {A _n | n = 1, 2, ..., P} of the proxy variables A _n , and 23 is the proxy variable F _n. Set {F _n | n = 1, 2, ..., P}
Is a third storage means for storing, and 24 is a set of c coefficients {C _n | n =
Fourth storage means for storing 1, 2, ..., P}, 25 is an nth-order surrogate variable from the nth-order a coefficient a _{n according} to the above seventh formula
First computing means for obtaining the A _n, second arithmetic means for obtaining a proxy variable F _n in accordance with the eighth formula 26, n-th proxy F from _n n-th c coefficients c in accordance with the eighth formula 27 Third to find _n
Is a calculation means. In any means, the order n is 1
To the maximum order p of the linear predictive analysis.

The principle and operation of the speech analysis apparatus configured as described above are the same as those described above with respect to the speech analysis procedure that is one embodiment of the speech analysis method according to claim 1. Will not be mentioned again. Further, according to the present embodiment, the product term consisting of three or more variables in the summation symbol, which is required in the conventional linear prediction cepstrum coefficient calculation, is unnecessary, and the accumulation of the cumulative terms of only two variables that can be calculated quickly and simply by DSP is performed. It is equally clear that a good speech-based analyzer can be realized.

EFFECTS OF THE INVENTION According to the present invention, a product term consisting of three or more variables in a summation symbol, which is required in the conventional linear prediction cepstrum coefficient calculation, is unnecessary, and a product term of only two variables that can be calculated quickly and simply by DSP. It is possible to realize a voice analysis method mainly by accumulating.

[Brief description of drawings]

FIG. 1 is a calculation procedure diagram of a voice analysis method in one embodiment of the present invention, and FIG. 2 is a block diagram of a voice analysis device in one embodiment of the present invention. 11 ...... steps from the n-th a coefficient a _n seek n th proxy A _n, the procedure for obtaining the 12 ...... n th proxy F _n, 13 ...... first n
A procedure for obtaining a next c coefficient c _n from the n th surrogate variable F _n , 21
The first storage means for storing a set of a coefficients {a _n | n = 1, 2, ..., P}, 22 ... A set of proxy variables A _n {A _n | n = 1,
Second storage means for storing 2, ..., P}, 23 ... Third storage means for storing a set of proxy variables F _n {F _n | n = 1, 2, ..., P} , 24 ... C coefficient set {c _n | n = 1, 2, ...
.., p} fourth storage means, 25 ... nth order a
First computing means for obtaining the n th proxy A _n from the coefficient a _n according to the above seventh equation, 26 ...... eighth proxy variable according formula
Second calculating means for calculating a F _n, third arithmetic means for obtaining the n-th c coefficients c _n from the n-th proxy F _n according 27 ...... eighth equation.

Claims (2)

(57) [Claims] 1. A method of calculating a set of linear prediction cepstrum coefficients from a set of linear prediction coefficients obtained from an arbitrary speech waveform using an adder and a multiplier of finite word length, in the calculation process. The procedure of obtaining the nth-order surrogate variable A _n from the nth-order linear prediction coefficient a _n based on the following a-1 expression, and the nth-order surrogate variable F based on the recurrence expression shown in the following a-2 expression procedures and, speech analysis method characterized in that it comprises a procedure for determining the n-th linear prediction cepstrum coefficients c _n from the n-th proxy F _n based on the first a-3 formula below to obtain the _n. However, a _n : n-th order linear prediction coefficient c _n : n-th order linear prediction cepstrum coefficient In any formula, 1 ≦ n ≦ p p: maximum order of linear prediction analysis. 2. An apparatus for calculating a set of linear prediction cepstrum coefficients from a set of linear prediction coefficients obtained from an arbitrary speech waveform by using a finite word length adder and a multiplier, in the calculation process. Means for obtaining the _n- th order proxy variable A _n from the n-th order linear prediction coefficient a _n based on the following c-1 expression, and the nth order proxy variable F based on the recurrence formula shown in the following c-2 expression. means for determining _n, sound analysis apparatus characterized by having a means for obtaining the n-th linear prediction cepstrum coefficients c _n from the n-th proxy F _n based on the c-3 formula below. However, a _n : n-th order linear prediction coefficient c _n : n-th order linear prediction cepstrum coefficient In any formula, 1 ≦ n ≦ p p: maximum order of linear prediction analysis.