JPS6367200B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to an apparatus for extracting pole-zero parameter values of a speech output frequency characteristic pattern used for speech analysis and synthesis, speech recognition, and the like.

It is known that the frequency spectrum of a sound waveform has a frequency component with concentrated energy called a formant, which corresponds to the resonance frequency of the sound path. It is also known that this formant approximately corresponds to the polar parameters when the frequency spectrum of the speech waveform is approximated by an all-pole model. A typical example of a method for extracting polar parameters (formant parameters) from a speech waveform is to use a successive approximation method such as the Newton-Raphson method to calculate a high-order algebraic equation whose coefficients are linear prediction coefficients extracted from the speech waveform. There is a known method to solve the problem using However, this method has the disadvantage that it is difficult to stably determine the frequency and bandwidth of the formant, and in particular, it may not be possible to accurately correlate the formants extracted in adjacent analysis frames. .

In addition, based on the speech generation model, frequency spectra are synthesized for various formant patterns, and the so-called so-called synthesis method approaches the spectrum of natural speech.
The AbS method (synthetic analysis method) is also known. This method makes it easy to avoid errors in formant mapping,
Although it is possible to perform relatively accurate extraction,
It has the disadvantage of requiring a huge amount of calculation.

Also, for example, by preparing a large number of extreme parameter values and inputting the audio signal to an inverse filter that uses a linear prediction system derived from each extreme parameter value,
There is also a method of finding a pole parameter that minimizes the error power obtained by integrating the squares of the output values of the inverse filter. In other words, the transfer relationship T(z) when the frequency spectrum envelope of the speech waveform is approximated by an all-pole model
(z=jωT, T: sampling period) is T(z)= _M 〓 ^m=1 H _n (z) = _M 〓 ^m=1 1/1+α _1n Z ^-1 +α _2n Z ^-2 …(1) However, , α _1n = b ² _n α _2n = 2b _n cos2π _n TM; Number of poles _n ; Pole frequency b _n ; Pole band width H _n (z); Since it is expressed by the transfer function of the m-th pole, Energy (error power) of the output waveform after passing the actual audio signal through the inverse filter T ^-1 (z) of the filter expressed by equation (1)
This is a method of selecting the polar parameters so that . The same can be done for the zero parameter by changing the coefficients.

Inverse filter H ^-1 _n corresponding to one formant
(z) outputs an output signal _{e o} given by the following formula for the input signal S o when two linear prediction coefficients α _1n and α _2n are given.

e _o =S _o −α _1n S _o-1 −α _2n S _o-2 …(2) Therefore, the error power E is given by E= _oB 〓 ^n=nA e ² _o …(3). However, n _A and n _B are the sampling numbers at the beginning and end of the analysis window. It is known that the time width of the audio analysis window is approximately 30 m _sec . For example, if the audio waveform is sampled at 10 KHz, the cumulative interval length (n _B −
_nA ) will be around 300. Therefore, calculating the error power of equation (3) for linear prediction coefficients corresponding to various extreme parameter values requires a huge amount of calculation. Furthermore, it is difficult to combine prediction coefficients for, for example, four formants.

An object of the present invention is to provide a pole-zero parameter value extraction device that can calculate the error power given by the above equation (3) with a small amount of calculations and obtain an optimal pole-zero parameter value.

The parameter extraction circuit of the present invention extracts autocorrelation values V _i (i=0, 1, 2,...) within a time window from an audio signal.
an autocorrelation value calculation circuit that calculates an autocorrelation value calculation circuit; an address/polarity value conversion circuit that stores various pole and/or zero parameter values in advance and outputs the pole parameter value stored at the corresponding address in response to an address input; A linear prediction coefficient generation circuit that generates linear prediction coefficients α ₁ and α ₂ from the output parameter values of the conversion circuit, and a prediction given by the linear prediction coefficient generation circuit to which the output value V _i of the autocorrelation value calculation circuit is input. Calculated output value using the following formula using coefficients α ₁ and α ₂
r _i , i.e. r _i = (1+α ₁ ² + α ₂ ² )V _i −(α ₁ −α ₁ α ₂ )V _i+1 −(α ₁ −α ₁ α ₂ )V _i-1 −α ₂ V _{i -2} −α ₂ V _i+2 , an arithmetic circuit that calculates and outputs r _i ; an autocorrelation value temporary memory circuit that stores the output value of the arithmetic circuit; and r ₀ , that is, the error power stored in the memory circuit. a minimum value detection circuit that detects the minimum value and outputs the corresponding address of the address/pole parameter conversion circuit _;
It is characterized by extracting the pole-zero parameter that gives the minimum value of .

Note that the number of calculations can be significantly reduced by estimating the polar parameter values roughly in the first stage and increasing the precision successively in the second stage.

Next, the present invention will be explained in detail with reference to the drawings.

First, the principle of the present invention will be explained. Said (2)
Considering the autocorrelation value r _i of the inverse filter output value (approximation error) e _o given by the formula, r _i = _oB 〓 ^n=nA e _o e _oi …(4) However, e _oi is e This is the output value i time slots before _o . Therefore, r ₀ corresponds to the error power E. Substituting equation (2) into equation (4) and expanding it, r _i = _oB 〓 〓 ^n=nA (S _o −α ₁ S _o-1 −α ₂ S _o-2 )×(S _oi −α _{1 Soi-1} −
α ₂ S _oi-2 ) = _oB 〓 〓 ^n=nA (S _o S _oi −α ₁ S _oi S _oi −α ₂ S _o-2 S _oi −α ₁ S _o S _oi-
1 +α ₁ ² S _o-1 S _oi-1 +α ₁ α ₂ S _o-2 S _oi-1 −α ₂ S _o S _oi-2 +α ₁ α ₂ S _o-1 S
_oi-2 + α ₂ ² S _o-2 S _oi-2 )…(5). Considering that the analysis window (for example, Hamming window) becomes 0 outside a certain time interval and S _o also becomes 0, for example, S _o S _oi can be considered to be equal to S _o-1 S _oi-1 . Therefore, equation (5) can be expressed as equation (6) below.

r _i = (1+α ₁ ² + α ₂ ² )V _i −(α ₁ −α ₁ α ₂ )V _i+1 −
(α ₁ −α ₁ α ₂ )V _i-1 −α ₂ V _i-2 −α ₂ V _i+2 …(6) However, V _i = _oB 〓 ^n=nA S _o S _oi (i: time slot (7) That is, V _i is the autocorrelation value of the input signal S _o . Furthermore, due to the nature of the analysis window, V _i =V _-i . Therefore, from the autocorrelation value V _i of the input signal S _o and the linearity prediction coefficients α ₁ and α ₂ , the error power _oB 〓 ^n=nA e _o ² (i.e. r ₀ )
can be found. When there are multiple poles (6)
By substituting r _i obtained by the equation into V _i on the right side of equation (6) to obtain a new r _i , and repeating this process, the final r ₀ can be obtained. For example, when there are four formants, i=0 _, 1,
. . 8 is determined, and from this _, r _{0 ,} . . . _{r 6} for the first formant is determined. Substitute r ₆ into _Vi on the right side of equation (6) to find r ₀ ,...r ₄ for the second formant, and then similarly calculate r ₀ , that is, the error power Σe ² taking into account both the third and fourth formants. _o can be found. It is possible to obtain the error power Σe ² _o for various polar parameters by the above-mentioned calculation, and extract the polar parameter value that gives the minimum value. In the calculation using equation (6) above, (n _B −n _A ) is 300
Since the value is of the order of magnitude, both the number of multiplications and the number of additions can be significantly reduced compared to calculating the error power using equations (2) and (3) described above.

Furthermore, it is not necessary to perform the above calculations for all polar parameter values. The number of calculations until the final extraction can be further reduced significantly by first finding the minimum error power for the coarsely quantized polar parameter values, determining coarse polar parameter values, and increasing the precision of the polar parameter values sequentially. I can do it small.
Now, for example, if the number of formants is M, and the polar parameter value is selected from F prepared in advance for each formant, then when calculating all combinations using the brute force method,
F _M calculations are required. As a typical example, M
= 4 and F = 32, a huge number of calculations, ³²⁴ times, are required. In the present invention, it is possible to first roughly estimate the polar parameters and then use the results to gradually and finely estimate them. That is, in the first stage, an optimum value is determined for a decimal number of coarsely quantized parameter values from among the F pole parameter values by combining each pole. and,
In the latter stage, the estimation of the pole parameter values is limited to a small number of poles in the vicinity of the pole parameter values determined in the former stage. In other words, the F parameters may be expressed by quantization codes, the optimum values for the upper digits may be obtained in the first stage, and the results may be used to sequentially obtain the optimum values for the lower digits. For example, if the number of division stages is L and the quantization level of each pole in each stage is I, then the pole parameter values can be estimated with sufficient accuracy by setting the number of division stages L to approximately L = log _I F. . Therefore, the polar parameter value can be determined by calculating the error power about I ^M ×L times. As an example, M=5, F=32, I=2
Then, the number of division stages L is about 5 (since I = 2, the accuracy can be doubled for each stage), and it is possible to estimate the polar parameter value by calculating the error power about 160 times. be.

Furthermore, when estimating polar parameter values, constraints such as limits on the estimation range can be easily set, so estimation results of past analysis frames can be referenced to estimate polar parameter values in the current analysis frame. Limiting the range also has the advantage of preventing large discontinuities in extreme parameter values from occurring.

In the above explanation, for the sake of simplicity, the case of estimating the polar parameter value has been described, but the estimation of the parameter value at the zero point can be similarly performed using two linear prediction coefficients representing the zero point.

In addition, in the above, only linear prediction coefficients α ₁ and α ₂ were considered, but even when using third-order or higher prediction coefficients, the polar parameter values can be obtained in the same way, and the same effect can be achieved. it is obvious.

A specific embodiment of the present invention based on the above principle is shown in the drawings. First, an audio waveform is input to the analysis window circuit 2 via the audio waveform input terminal 1. The analysis window circuit 2 applies an analysis window to the input audio waveform in accordance with the frame synchronization signal generated by the frame synchronization signal generation circuit 9, and inputs the resulting output waveform to the autocorrelation value calculation circuit 3. The autocorrelation value calculation circuit 3 calculates the autocorrelation value V _i (for example, i=0 to 8) of the input waveform according to the frame synchronization signal,
It is stored in the autocorrelation value buffer 4. The autocorrelation value V _i is given to the inverse filtering circuit (arithmetic circuit) 5 by a control signal given from the control circuit 11 . On the other hand, various pole parameter values are stored in the address pole parameter value conversion circuit 8, and any parameter of any pole (for example, the first parameter of the first formant) is sent to the linear prediction coefficient generation circuit 10 via the control circuit 11. is given, the linear prediction coefficient generation circuit 10 calculates the linear prediction coefficients α ₁ and α ₂ from the pole parameter values, and the inverse filtering circuit 5
give to Inverse filtering circuit (arithmetic circuit) 5
calculates the above equation (6) using the given linear prediction coefficients α ₁ and α ₂ and the autocorrelation value _Vi , and stores the result in the autocorrelation value temporary storage circuit 6. For example, when the autocorrelation values V ₀ to V ₈ are given,
r ₀ to r ₆ are obtained from the calculation result of equation (6). The autocorrelation value temporary storage circuit 6 provides the above calculation results r ₀ to _{r 6} to the inverse filtering circuit 5 according to control data from the control circuit 11 and outputs an error power value. Next, new autocorrelation values r ₀ to r ₄ are calculated from the above calculation results r ₀ to r ₆ using, for example, prediction coefficients α ₁ and α ₂ corresponding to the first parameter value of the second formant. Thereafter, in the same manner, _r0 , that is, the error power for up to the fourth formant is calculated and stored in the autocorrelation value temporary storage circuit 6. Similarly, when the error power is calculated and accumulated for each combination of polar parameter values of each formant, the minimum value detection circuit 7 detects the minimum error power among the above-mentioned error powers according to the control data given from the control circuit 11. The address data storing the pole parameter value of each pole when the minimum error power is obtained is outputted and given to the address pole parameter value conversion circuit 8. The address pole parameter value conversion circuit 8 outputs the pole parameter value stored at the designated address. This makes it possible to obtain the pole parameters that give the minimum error power.

For example, when the above selection is made for pole parameter estimation only for the upper digits of the pole parameter value in the previous stage, the address pole parameter value conversion circuit 8 in the next stage estimates up to the lower digits in the vicinity of the above result. The included polar parameter values are sent to the linear prediction coefficient generation circuit 10 via the control circuit 11, and the arithmetic circuit 5, autocorrelation value temporary storage circuit 6, minimum value detection circuit 7, etc. perform the same operation as described above to generate an error again. Find the minimum power value and output the corresponding address. The above-mentioned operation is repeated to successively improve the estimation accuracy, and when the estimation up to the least significant digit is completed, the polar parameter value is outputted from the output terminal 17 as the final stage estimation value. That is, by roughly estimating the polar parameter values in the first stage, it is possible to reduce the number of combinations and improve the accuracy of successive estimation. Compared to the case where error power is determined for all combinations from the beginning, the number of combinations can be significantly reduced, and therefore the number of calculations can be reduced.

As described above, in the present invention, the error power is calculated from the autocorrelation value and the linear prediction coefficient of the input audio signal, and the pole-zero parameter corresponding to the minimum error power value is used as the estimated value. Therefore, it is possible to obtain pole-zero parameters with a relatively small number of calculations.

Note that if the estimation of polar parameters, etc. is roughly estimated in the first stage and the results are used to increase the accuracy and estimate, the number of combinations can be significantly reduced, and the number of calculations can be further reduced. This has the effect that it can be reduced to

Additionally, it is easy to place constraints on the estimation range based on the extreme parameter values obtained from the previous analysis frame, which has the advantage of avoiding large discontinuities in extreme parameter values, etc. between analysis frames. .

[Brief explanation of drawings]

The figure is a block diagram showing one embodiment of the present invention. In the figure, 1...audio waveform input terminal, 2...analysis window circuit, 3...autocorrelation value calculation circuit, 4...autocorrelation value buffer, 5...arithmetic circuit, 6...autocorrelation value temporary storage circuit, 7...minimum value detection Circuit, 8...Address/
Polar parameter value conversion circuit, 9... Frame synchronization signal generation circuit, 10... Linear prediction coefficient generation circuit, 11...
Control circuit, 12 to 16... Control data transmission line, 17
...Polar parameter value output terminal.

Claims (1)

[Claims] 1. An autocorrelation value calculation circuit that calculates an autocorrelation value V _i (i is zero or a natural number) within a time window from an audio signal, and a circuit that stores various pole or zero parameter values in advance and inputs them to an address. an address/polarity parameter value conversion circuit that outputs polar parameter values stored in corresponding addresses; a linear prediction coefficient generation circuit that generates linear prediction coefficients α ₁ and α ₂ from the output parameter values of the conversion circuit; The output value V _i of the autocorrelation value calculation circuit is input and the prediction coefficients α ₁ and α ₂ given from the linear prediction coefficient generation circuit are used to calculate the output value r _i =(1+α ₁ ² +α ₂ ² )V _i −( an arithmetic circuit that calculates α ₁ −α ₁ α ₂ )V _i+1 −(α ₁ −α ₁ α ₂ )V _i-1 −α ₂ V _i-2 −α ₂ V _i+2 ; and the arithmetic circuit. an autocorrelation value temporary storage circuit that stores the output value of the storage circuit, and detects the minimum error power r ₀ among the output values stored in the storage circuit, and determines the corresponding address of the address/polarity parameter conversion circuit. 1. A pole-zero parameter value extraction device, comprising: a minimum value detection circuit that outputs a minimum value.