JPH0670754B2 - Polar zero analyzer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a pole-zero analyzer, and more particularly to a pole-zero analyzer that extracts a pole-zero parameter value that approximates the spectrum of a signal such as voice.

[Conventional technology]

In the fields of speech synthesis and speech recognition, it is important to extract the values of parameters that approximate the speech spectrum. Furthermore, it may be necessary to extract the values of parameters that approximate the spectrum of general signals.

A pole or pole zero is often used as a parameter for approximating the spectrum of a signal such as voice (hereinafter simply referred to as voice). This is because poles and zeros have characteristics such as clear physical meanings and are advantageous in application.

Conventionally, a high-order algebraic equation whose coefficient is a linear prediction coefficient extracted from speech or the like is used to extract the value of a pole parameter that approximates a spectrum of speech or the like, for example, Newton
A solution method using a successive approximation method such as Rapson method is known. This first conventional example is, for example, Jay. Dee. Markel and A. H. Gray (JDMarkel and AHG
Ray) 's book "Linear Prediction of Speech (Translated by Suzuki)" (Linear Pr
ediction Speech) chapter 7.

As a second conventional example of extracting the value of the pole parameter, an inverse filter link in the autocorrelation region is described in Fushikida's paper "Multi-stage formant estimation method using inverse filtering in the autocorrelation region" It is shown in the study group material S81-41). In the second conventional example, the error power is obtained by sequentially inverse filtering the autocorrelation value of the input signal using the coefficient corresponding to the pole parameter of the voice generation model. A table of pole parameter values is prepared in advance, and the coefficient of each stage is calculated from the table to calculate the error power. The minimum value of the error power is given to various candidates for the pole parameter value, and the value is output as the pole parameter value that gives the optimum approximation.

On the other hand, as a third conventional example for extracting the value of the pole-zero parameter, C.I. tea. Maris and Earl. Tei. "The Youth of Second Order Information in the Aproximation of Discrete" by Roberts (CTMullis and RARoberts), EE-Transaction (IEEE Transaction 24), No. 3, pages 226-238. Time Linear Systems (The Use
of Second-Order Information in the Approximation o
f Discrete-Time Linear Sys-tems) ”.

This speech spectral envelope H (ej ^omega), the transfer function is intended to approximate the pole-zero system represented by the following formula.

Where a ₀ = 1 where n is the degree of the denominator polynomial, that is, the degree of the polar circuit,
m is the order of the numerator polynomial, that is, the order of the zero circuit.

This method is based on the principle that ai and qk that minimize the value of the following equation give the optimal approximation.

In the time domain, this corresponds to solving a system of simultaneous equations in which ai is an unknown number, and a coefficient is a value determined from an autocorrelation value of an input signal such as a voice and its associated impulse response.

The value of the autocorrelation value and the value of the impulse response can be easily obtained by, for example, a linear prediction method of an order higher than n or m, that is, the method of the first conventional example. In that case, it is not necessary to find the root of the higher-order algebraic equation, and only the impulse response needs to be found.

In the third conventional example, coefficients ai and qk of the denominator polynomial of the transfer function and the numerator polynomial are obtained. In order to obtain the value of the pole-zero parameter from this, it is necessary to find the root of a high-order (nth-order or mth-order) algebraic equation as in the first conventional example.

[Problems to be solved by the invention]

Among the above-mentioned conventional examples, the first conventional example, that is, one that solves a high-order algebraic equation having a linear prediction coefficient as a coefficient requires many operations when solving a high-order algebraic equation, and There is a problem that it is difficult to stably obtain the frequency and the bandwidth.

Further, in the second conventional example, an optimum value is extracted in advance from candidates that are considered to be valid as a pole parameter representing a spectrum of voice or the like, so that stable calculation is possible, but there is a problem that many calculations are required. there were.

Furthermore, since only the pole parameter values are obtained in the first and second conventional examples, the parameter value that gives the optimum approximation cannot be obtained at a relatively low order depending on the shape of the spectrum of speech or the like. On the contrary, in such a case, there has been a problem that a higher order parameter has to be obtained and more calculations are required.

The third conventional example, that by Maris and Roberts, obtains the values of the pole parameter and the zero parameter, so the parameter value gives an optimum approximation at a relatively low order,
As in the first example, the high-order algebraic equations had to be solved, and there were the same problems as in the first conventional example. In the case of speech, this is not a problem because the number of zeros is small, but a pole order of at least about 10 is required, which is a problem. Furthermore, in this conventional example, there is a problem that many calculations are required because the estimated value of the impulse response must be obtained.

An object of the present invention is to provide a pole-zero analyzer which can accurately extract the value of the pole-zero parameter approximating the spectrum of a signal such as voice with a relatively small amount of calculation.

[Means for solving problems]

The pole-zero analysis apparatus of the present invention is a pole-zero analysis apparatus that estimates pole-zero parameter values based on a generation model of a pole-zero type speech using a value of a cepstrum extracted from a signal such as a voice, from an input signal. It has a means for calculating and temporarily storing the cepstrum within the time window, a secondary pole circuit and a primary zero circuit, and its output is the first output, and the value obtained by multiplying the output by a coefficient is the second output. And a circuit in which a value obtained by multiplying the output of the first-order delay element by a coefficient is the third output is a basic filter,
A first parallel filter in which the first outputs of at least one basic filter are connected in parallel, and having a second parallel filter of the same configuration, the output of the first parallel filter and the second parallel filter An output of a value obtained by multiplying the output difference by a coefficient, and the difference between the cepstrum extracted from the input signal and the impulse response of the analysis filter is calculated, and then the estimation error of the pole-zero parameter is calculated. Means, means for calculating correction information of the estimated value of the pole-zero parameter from the difference between the second and third outputs of each of the basic filters and the cepstrum, and a candidate value of the pole-zero parameter of the voice generation model is selected. , Calculate the estimation error at that time,
The candidate value of the pole-zero parameter is output as the value of the extracted pole-zero parameter when the estimation error of the pole-zero parameter is minimized while modifying the candidate value of the pole-zero parameter based on the correction information of each parameter. And means for controlling so as to do so.

[Action]

The principle of the present invention is to find the optimum pole-zero parameter by minimizing the square error r of the logarithmic spectrum given by the following equation (3).

This is equivalent to minimizing the squared error between the cepstrum extracted from the speech and the cepstrum of the speech production model in the time domain. Since the cepstrum that represents the spectral envelope of speech is a component of several tens or less, only a component of that level is sufficient when evaluating a squared error.

The value of the speech cepstrum within a predetermined time window can be easily calculated using a fast Fourier transformer or the like, as is well known in the art. On the other hand, if the cepstrum of the speech production model is to be calculated using a fast Fourier transformer as well, an estimated value of the pole-zero parameter is set, synthetic speech is generated, and its cepstrum is generated by the fast Fourier transformer. Since it has to be obtained, a huge amount of calculation is required.

However, in the present invention, when the speech production model is a pole-zero model, it is utilized that the value obtained by multiplying the impulse response of the component analysis filter by the reciprocal of time is equivalent to the cepstrum of the speech production model. The amount of calculation is reduced. At this time, the coefficient of the basic filter is given by the following equation.

a = 2 · p · cos (θ) b = −p ² (4) c = −p · cos (θ) where p is the absolute value of the pole or zero of the generative model, and θ is its argument. The coefficients a, b and c respectively correspond to the coefficients of the multipliers 203, 204 and 205 of the basic filter shown in FIG.

Furthermore, the fact that the partial differential coefficient of the square error of the cepstrum with respect to the pole-zero parameter is equivalent to the second and third outputs of the basic filter is used to determine the correction of the candidate value of the pole-zero parameter. The coefficient for calculating the partial differential coefficient at this time is given by the following equation.

d = −1 / p e = −2 · p · sin (θ) (5) Here, d and e correspond to the coefficients of the multipliers 206 and 207 of the basic filter shown in FIG.

In this way, the squared error of the cepstrum is calculated for various candidates of pole-zero parameter values, and the minimum value of the squared error is corrected while correcting the candidate values based on the second and third outputs of each basic filter. The candidate for the pole-zero parameter value when given is output as the optimum pole-zero parameter value.

〔Example〕

 Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram of an embodiment of the present invention.

101 is a control circuit, 102 is a cepstrum calculation circuit, 103 is a cepstrum memory, 104 is an impulse generator, 105 is a first parallel filter, 106 is a second parallel filter, 107 is a first coefficient calculation circuit, and 108 is a second coefficient calculation. Circuit, 109 is a coefficient multiplier,
110 is an error calculation circuit, 111 is a pole correction information calculator, and 112 is a zero correction information calculator.

Further, FIG. 2 is a block diagram of a basic filter which is a constituent element of two parallel filters. As shown in FIG. 2 (a), the basic filter includes two delay elements 201 and 202, coefficient multipliers 203 and 204 of the pole circuit, coefficient multiplier 205 of the zero circuit, and correction information of the absolute value of the pole from its output. And a coefficient multiplier 207 for multiplying a first-order delay element to generate deflection angle correction information. One input signal line 211 and three output signal lines 212, 213 and 214 are provided. have.

On the other hand, FIG. 2B is a block diagram of the parallel filter, in which the signal input from the signal line 221 is distributed and input to each input signal line of the plurality of basic filters, and the first output of each basic filter is input. The signal line, that is, the signal line 212 in FIG.
Are output from the signal line 222, and the other signal lines, that is, the signal lines 213 and 214 in FIG. 2A, are output to the signal line bundle 223 as they are.

When a signal is input from the signal line 121 in FIG. 1, the cepstrum calculation circuit 102 calculates the cepstrum of the input signal within the time window in accordance with the control signal sent from the control circuit 101 via the signal line 123. It is sent to the cepstrum memory 103 via 133 and temporarily stored.

On the other hand, the first and second coefficient calculation circuits 107 and 108 are
From the polar parameter candidate value and the zero parameter candidate value of the voice generation model sent from the respective signal lines 127 and 128 respectively, the respective coefficients of the respective basic filters are calculated, and the first and second parallel lines are respectively calculated via the signal lines 136 and 137. Send to filter.

The first and second parallel filters 105 and 106 have coefficient values sent from the first and second coefficient calculation circuits 107 and 108 after clearing internal delay elements according to control signals sent from the control circuit 101 via signal lines 125 and 126, respectively. Is set to the multiplier in each basic filter, the response to the impulse sent from the impulse generator 104 via the signal line 135 is sequentially calculated, and the results are sent to the signal lines 138 and 139, respectively, and the difference between the data passes through the signal line 140. And is sent to the multiplier 109. Multiplier 10
The reciprocal of the time from the impulse generation is sent to the control circuit 101 from the control circuit 101 via the signal line 131, and the output of the multiplier 109 is sent to the signal line 141 as the cepstrum data of the voice generation model.
Sent to.

The cepstrum memory 103 sends the stored voice cepstrum data to the signal line 134 in synchronization with the timing signal sent from the control circuit 101 via the signal line 124, and outputs the difference between the data and the cepstrum of the voice generation model. Is sent to the error calculation circuit 110, the pole correction information calculator 111, and the zero correction information calculator 112 via the signal line 142.

The error calculation circuit 110 calculates the sum of squares of the cepstrum difference data sent via the signal line 142 in synchronism with the timing signal sent from the control circuit 101 via the signal line 132, and via the signal line 143. And sends it to the control circuit 101.

From the first and second parallel filters 105 and 106, the bundle of signal lines 1
Information obtained from the second and third outputs of the respective basic filters via 44 and 145 are sent to the pole correction information calculator 111 and the zero correction information calculator 112, respectively.

The pole-correction-information calculator 111 and the zero-correction-information calculator 112 divide the data sent via the signal line bundles 144 and 145, respectively, and the signal line 142 in accordance with the control signals sent from the control circuit 101 via the signal lines 129 and 130, respectively. The partial differential coefficient of the squared error with respect to the change of each pole parameter and zero parameter is calculated from the data of the cepstrum difference sent via the signal and sent to the control circuit 101 via the signal lines 146 and 147, respectively.

The control circuit 101 sends the candidate values of the pole-zero parameter of the voice generation model to the first and second coefficient calculation circuits 107 and 108, receives the estimation error at that time from the error calculation circuit 110, and also determines each pole parameter and zero parameter. Of the correction information of the pole correction information calculator 111, the zero correction information calculator 112, while correcting the candidate value of the pole-zero parameter, the candidate value when the estimation error is the minimum, as the value of the optimum pole-zero parameter. Output from the signal line 122.

The above processing is repeated for each analysis frame.

〔The invention's effect〕

As described above, according to the present invention, the pole-zero parameter value that can approximate the spectrum more accurately than the first and second conventional examples can be obtained. Moreover, since it is not necessary to solve a high-order algebraic equation, there is an effect that it can be obtained more stably than the third conventional example. At this time, unlike the third conventional example, it is not necessary to obtain the impulse response of the voice generation model.
The calculation for that amount is unnecessary.

Further, according to the present invention, there is an advantage that, in pole-zero analysis or the like for speech synthesis, by limiting restrictions on the pole-zero parameter value candidates, ones that match the speech synthesis model can be extracted.

Furthermore, when extracting the cepstrum from speech, it is expected that the accuracy of analysis can be improved by converting the frequency axis to mel scale or by weighting the extracted cepstrum data. .

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of the present invention, and FIG. 2 is a block diagram of two parallel filters and a basic filter which is a component thereof. 101 ... Control circuit, 102 ... Cepstrum calculation circuit, 103
...... Cepstral memory, 104 …… Impulse generator, 1
05 …… First parallel filter, 106 …… Second parallel filter, 1
07 …… First coefficient calculation circuit, 108 …… Second coefficient calculation circuit, 1
09 …… coefficient multiplier, 110… error calculation circuit, 111 …… pole correction information calculator, 112 …… zero correction information calculator, 201,202 …… delay element, 203,204 …… pole circuit coefficient multiplier, 205 …… Coefficient multiplier of zero circuit, 206 ... Coefficient multiplier for generating correction information of absolute value of pole from output of basic filter, 207 ...
... A coefficient multiplier for multiplying the first-order delay element to generate the correction information of the argument.

Claims (1)

[Claims] 1. A pole-zero analyzer for estimating pole-zero parameter values based on a pole-zero voice generation model using a cepstrum value extracted from a signal such as a voice signal. It has a means for calculating and temporarily storing the cepstrum, a secondary pole circuit and a primary zero circuit, and its output is the first output, and a value obtained by multiplying the output by a coefficient is the second output. A circuit in which the value obtained by multiplying the output of the delay element by a coefficient is the third output is the basic filter, and the first parallel filter in which the first outputs of at least one basic filter are connected in parallel, An analysis filter having two parallel filters, the output of which is a value obtained by multiplying the difference between the output of the first parallel filter and the output of the second parallel filter, and a cepstrum extracted from the input signal. Of the analysis filter And a correction means for calculating the estimation error of the pole-zero parameter from the difference, and the correction information of the estimation value of the pole-zero parameter from the difference between the second and third outputs of the basic filters and the cepstrum. Means for selecting the pole-zero parameter candidate value of the speech generation model, calculating the estimation error at that time, and correcting the pole-zero parameter candidate value based on the correction information of each parameter while controlling the pole-zero parameter. And a means for controlling so that the candidate value of the pole-zero parameter when the estimation error of the parameter becomes the minimum is output as the value of the extracted pole-zero parameter.