JPS62180398A - Pole parameter value extractor - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention]

E. Field of Industrial Application] The present invention relates to a polar parameter value extraction device, and more particularly to a polar parameter value extraction device for extracting polar parameter values that approximate the spectrum of a signal such as voice. 1'Prior art] In the fields of speech synthesis and speech recognition, it is important to extract values for parameters that approximate the speech space.
Furthermore, even for general signals, it may be necessary to extract the values of parameters that approximate the spectrum of the signal. Specs of signals such as audio (hereinafter simply referred to as audio, etc.)
Polar parameters, that is, formant frequency and bandwidth, are often used as parameters to approximate the waveform. This is because it has characteristics such as a clear physical meaning and is advantageous in terms of application. Conventionally, to extract values of polar parameters that approximate the spectrum of speech, etc., high-order algebraic equations whose coefficients are linear prediction coefficients extracted from speech, etc., have been used, for example, by Newtonian equations.
A method of solving using a successive approximation method such as the Rapson method is known. This first conventional example is, for example, developed by J.D. Markel and A.H.G.
, A, 11. Gray)'s book [Linear Prediction of Speech (translated by Suzuki) J (Linear Prediction)
or 5pecl+), Chapter 7. As a second example of extracting the value of a polar parameter, inverse filtering in the autocorrelation region is used in the paper by Fushikida [Multi-stage formant estimation method using inverse filtering in the autocorrelation region] (Acoustical Society of Japan Speech Study Group). Document 581-41). In this second conventional example, the coefficients of the second-order grinding circuit in the first stage are set to 1 for α and 2 for α,
When the input autocorrelation value is rk-1, 1, the output autocorrelation value rk+1 is determined by the following equation. r,, +1 ” <1+ (t ”14.14 (!
214,2) ry-1,1-(α, I-α, l
°αh, 2) rk-1,141-tl h, 2
(rk-1,1-24rb-L 142> -(1)
Autocorrelation of speech, etc. is used for the input of the first stage. The autocorrelation value of the output is determined by the sequential formula (1), and the autocorrelation value rk of the output of the final stage (this is the first (stage)) is determined. This is a method called inverse filtering in the autocorrelation region. Note that rk and o are called prediction residual power or error power, and to calculate them, the following autocorrelation values of the input voice are required from order 0 to order 2. , prepare a table of polar parameter values, and then calculate the coefficients 1kah, I for each stage.
Hak. 2 and calculate the above rk, 0. The minimum value of rt<, o is given to various candidates for polar parameter values, and λ is outputted as a polar parameter value that provides the optimal approximation.2 [Problems to be Solved by the Invention] The above-mentioned conventional example Among these, the first example, that is, solving a high-order algebraic equation with linear prediction coefficients as coefficients, requires many operations when solving a high-order algebraic equation, and it is difficult to stabilize the frequency and bandwidth of the poles. There was a problem that it was difficult to determine. In addition, in the second example, the optimal value is extracted in advance from candidates that are considered to be appropriate as polar parameters representing the spectrum of speech, etc., so it can be stably determined, but there is a problem that it requires a large number of calculations. Ta. The purpose of the present invention is to improve the quality of signals such as audio! - An object of the present invention is to provide an extreme parameter value extracting device that can extract the extreme value of a meter that approximates R with a relatively small number of calculations IMt. Means for Solving the Problems ゛1 The polar parameter value extracting device of the present invention has a means for calculating and temporarily storing an autocorrelation value within a time window from an audio signal, and a means for storing the value of the polar parameter and calculating it for an input code. A polar parameter value table that outputs the value of a polar verameter, a means for calculating coefficients for performing inverse filtering in the autocorrelation region from the polar parameter values, and a means for automatically calculating the temporarily stored autocorrelation value using the coefficients. In a polar parameter value extracting device of a type having means for inverse filtering in the correlation region, linear prediction coefficients of the fourth order or less and prediction residuals are obtained from autocorrelation values of the fourth order or less among the outputs of the means for inverse filtering in the autocorrelation region. means for calculating the power of the prediction residual; means for calculating the polar parameters from the linear prediction coefficient; means for detecting the minimum value of the power of the prediction residual; It is comprised of means for controlling the output of the polar parameter value table and the output of the means for calculating polar parameters from the linear prediction coefficients to be output as extracted polar parameter values. [Operation] Inverse filtering in the autocorrelation region in the present invention is as follows:
In comparison to the second conventional example described above, inverse filtering is performed from the 12th stage to the 11th stage (this is referred to as the second stage), and then inverse filtering is performed up to the final stage. do not have. At this time, the autocorrelation value rL,l of the output of the second stage is determined for a value of order 2(K-L) (denoted as M) or less. That is, when K-L is 2, M is 4.1, then M is 2. Next, using this autocorrelation value, an M-th order linear prediction coefficient (this is al; i=1...,M) is determined. This can be obtained by solving the following simultaneous equations using the first conventional method. However, the error power p in this case is given by the following formula. In the inverse filtering from the first stage to the second stage,
The value of the polar parameter that calculates r) in equation (3) for various polar parameter candidates and gives its minimum value, and the X obtained as a solution to equation (2):

The following equation (,
The root of the algebraic equation in 1) is output as the value of the optimal polar parameter. Since the degree M of the root of equation (4) is less than or equal to 4, it is known that it can be stably found with a finite number of operations.
When solving, it is possible to directly calculate from the autocorrelation value without solving as in the conventional example. For example M = 2
In the case of , it can be found by the following formula. al”-<rL, 1) rL, 1-rL, lrL,
z)/(re, Q-rt, 1) a2'-(+・L,
Ill'L, 2-rl↑, + )/(r-, n-
rt", l)...(5) 1st Embodiment] Next, the present invention will be explained with reference to the drawings. FIG. 1 is a process diagram of the first embodiment of the present invention. . Signal lines 123, 124, 126, 127,
128, 129 and 130, respectively, to the autocorrelation calculation circuit 102. Buffer 103, autocorrelation inverse filter 106. Linear prediction circuit 107. Extreme value calculation circuit 108. Control information instructing initialization is sent to the buffer 109 and comparison circuit 110, and each circuit, buffer, or autocorrelation inverse filter 106 is initialized. Autocorrelation calculation port i? In 8102, ilJ R circuit 10
1, the autocorrelation coefficient of the signal input from the signal line 121 is calculated and sent to the barnofa 103 via the signal line 131. After the interior of the bass sofa 103 is cleared according to the initialization instruction from the control circuit 101, the autocorrelation coefficient sent from the autocorrelation calculation circuit 102 is stored, and the autocorrelation coefficient is transmitted via the signal line 132. It is sent to inverse filter 106. The autocorrelation coefficients stored here are thereafter repeatedly sent to the autocorrelation inverse filter 106 each time there is an instruction to send them from the control circuit 101. The power control circuit 101 sends a code to the extreme value table 104 through a signal line 125 to indicate a value candidate for the extreme parameter. circuit 10
Sent to 5. The coefficient calculating circuit 105 calculates the coefficients necessary for inverse filtering in the autocorrelation region, that is, each coefficient of the above-mentioned formula <1) from the values, and sends them to the autocorrelation inverse filter 106 via the signal line 134. This process is repeated each time the control circuit 101 sends a code of a candidate for a polar parameter. In accordance with instructions from the control circuit 101, the autocorrelation inverse filter 106 performs inverse filtering in the autocorrelation region using the coefficients sent from the coefficient calculation circuit 105 for the autocorrelation coefficient data sent from the bath sofa 1()3. Processing, that is, the above-mentioned formula (1) is repeatedly applied to calculate the value of the output autocorrelation coefficient.
The value of the ``11 coefficient'' is calculated and sent to the linear output circuit 107 via the f line 135. This process is also performed by the control circuit 1.
Repeated every time an instruction is sent from 01. Line Lula Yo J! 1 circuit 107 uses the autocorrelation coefficients up to the fourth order sent from the autocorrelation inverse filter 106 with tjf=in response to the instruction from the control circuit 101, and calculates the optimal linear prediction up to the first order J] for that value. The jlJ coefficient and the error power at that time are calculated, and the linear pre-four law coefficient is connected to the signal line 13.
6 is sent to the pole 1 direct calculation circuit 108, and the error power is sent to the comparison circuit 11() via the signal line 138. This process is repeated every time an instruction from the control circuit 101 is sent. The extreme value calculation circuit 108 responds to instructions from the control circuit 101 by
(=The root of the algebraic equation whose coefficients are the linear predictions up to the fourth order sent from the linear prediction circuit 107 and !!11 coefficients is determined and sent to the buffer 109 via the signal line 137. In the comparison circuit 110 When an initialization instruction is sent from the control circuit 101, the initial value for comparison when detecting the minimum value of error power is sent to the memory 111 via the signal line 140.Next, in the comparison circuit 110, the control circuit According to the indication t from 101, the error power sent from the linear pre-1ijj1 circuit 107 and the memory 111 to 13 line 13
The values read through 9 are compared, and the comparison result is sent to signal lines 1 and 11 fFl. In addition, the linear prediction circuit 1
When the error power sent from 07 is smaller, its value is sent to the memory 1]1 via the signal line 140, and its contents are rewritten every time there is an instruction from the control circuit 01. repeated. When the control circuit 101 receives a comparison result from the comparison circuit 110 indicating that the error power calculated in the linear prediction circuit 1[]7 is smaller, the control circuit 101 updates the extreme value table 104.
The value of the polar parameter at that time for the signal line 142
It also instructs the extreme value calculation circuit 108 to convert linear prediction coefficients to polar parameters, that is, the east root of an algebraic equation, and sends the result to the buffer 109. In the buffer 109, the control circuit 1
Every time there is an instruction from 01, it is rewritten and stored with the value of the extreme parameter sent from the extreme value table 104 and the extreme value calculation circuit 108. In this way, at the time when inverse filtering has been performed on various candidates for the pole parameters, the buffer 109 stores the optimal value of the pole parameters in the sense that the error power is the minimum. Buffer 109 from the control circuit lot to the processing of each analysis frame
An output instruction is sent to , and the stored optimal value of the polar parameter is output from the signal line 122 . After that, this process is repeated for each analysis frame. In this embodiment, the linear prediction Jj is obtained from the autocorrelation inverse filter 106.
Although the order of the autocorrelation coefficient sent to one circuit 107 is up to 4th order, it may be up to 2nd order as described in Moe's article. FIG. 2 is a schematic diagram of a second embodiment of the present invention. In this embodiment, when extracting polar parameters for speech synthesis, etc., it is possible to limit the values of the extracted polar parameters so that analysis can be performed in accordance with the speech synthesis model. For this purpose, a comparison circuit 212 and a memory 213 are added to the first embodiment. Autocorrelation calculation circuit 202 in this embodiment. buffer 2
03. Extreme value table 204. Coefficient calculation circuit 205. Autocorrelation inverse filter 206. Linear prediction circuit 207. Extreme value calculation circuit 208. The buffer 209, the comparison circuit 210, and the memory 211 are the autocorrelation calculation circuit 102, the buffer 103. extreme value daple 104
i-series calculation circuit 105. autocorrelation inverse filter 106,
Linear prediction circuit 107, extreme value calculation circuit 108. Buffer 109. It corresponds to comparison circuit 110 and memory 111 and operates in the same way. With the control performed by the control circuit 201, the first. The points that are different from the embodiment will be explained below. First, in the case of W periodization, the limit value to be changed to the value of the polar parameter is sent to the comparison circuit 212 via the signal line 229, and the ratio 1 bit circuit 212 further transmits the value to the signal line 243. The data is sent to the memory 213. Each time a code of an extreme parameter candidate that meets the model limit is sent to the extreme value table 204, the extreme value calculation circuit 2
08, the conversion of the polar parameter I\ from the linear prediction coefficient is instructed each time, and the value is sent to the comparator circuit 212. The comparator circuit 212 reads the value from the memory 213 via the signal line 244. It is determined whether the issued limit value is met or not, and the result is sent to the control circuit 20 via the signal line 245. In the control circuit 201, the judgment word usage from the comparison circuit 212 is
Only when the limit is met, the error power comparison result from the comparator circuit 210 is referred to and the optimal extreme parameter value is extracted. In this embodiment, since only the values of the extreme parameters that fall within the limit can be extracted, it has the advantage that more stable formant extraction can be performed in speech formant analysis. 1. Effects of the Invention As explained above, the present invention has the advantage that equivalent polar parameter values can be obtained with a smaller amount of calculation than the conventional method. Furthermore, since the present invention does not use a method of solving high-order algebraic equations, the polar parameter values can be stably determined.

[Brief explanation of drawings]

FIGS. 1 and 2 are block diagrams showing first and second embodiments of the arrogant parameter value extracting apparatus of the present invention, respectively. 101.102...control circuit, 102,202...
Autocorrelation calculation circuit, 103.203...Ha sofa,
104.204... Extreme value table, 105,205...
... Coefficient calculation circuit, 106,206 ... Autocorrelation inverse filter, 107.207 - ... Linear prediction circuit, 108°208 ... Extreme value calculation circuit, 109.209 ... Bar.
Sofa, 110,210.212... Comparison circuit, 11
1゜211.213...Memory.

Claims (1)

[Claims] A means for calculating and temporarily storing an autocorrelation value within a time window from an audio signal, an extreme parameter value table for storing extreme parameter values and outputting the extreme parameter value for an input code, and an autocorrelation area from the extreme parameter value. A polar parameter value extracting device of the type, comprising means for calculating coefficients for performing inverse filtering in the autocorrelation region, and means for inverse filtering the temporarily stored autocorrelation value in the autocorrelation region using the coefficients. Means for calculating the power of a fourth-order or lower linear prediction coefficient and prediction residual from the fourth-order or lower autocorrelation value of the output of the means for inverse filtering in the region, and means for calculating a polar parameter from the linear prediction coefficient; a means for detecting the minimum value of the power of the prediction residual; and an output of the means for calculating the polar parameter from the polar parameter value table and the linear prediction coefficient when the minimum value of the power of the prediction residual is given. 1. A polar parameter value extracting device comprising means for controlling to output the extracted polar parameter value as the extracted polar parameter value.