JPH0731216B2 - Spectral analysis method and apparatus thereof - Google Patents

Description

The present invention relates to coded transmission of time series signals such as voice and sound, and
The present invention relates to a spectrum analysis method and apparatus for obtaining a frequency spectrum of a time-series signal in signal processing for synthesis or recognition.

(Prior Art) Conventionally, as a technology in such a field, Mikio Hino "Spectral Analysis" (1977-10-1) Asakura Shoten, P.210-223
There was something described in.

FIG. 2 is a functional block diagram showing a configuration example of a conventional spectrum analyzer.

This spectrum analyzer estimates a spectrum of random data by using a maximum entropy method (MEM) that determines a spectrum so as to maximize information entropy, and includes a covariance calculation unit 1 and a reflection coefficient calculation unit. 2, a prediction coefficient calculation unit 3, a prediction residual calculation unit 4, and a spectrum calculation unit 5.

Digitized time-series signal x (n) of voice, sound, etc.
Is input to the covariance calculation unit 1, the covariance calculation unit 1 outputs the covariance φ m ₋₁ (i, j; n) and the input power Po (n), and supplies it to the reflection coefficient calculation unit 2. . Covariance φ m _-1 (i, j; n)
Is an order m = 1,2, ..., M, an integer i, j = 0,1, ..., M, and a time is n
, The integer (i, j) lag (L
og; temporal deviation). The input signal power Po (n) is a prediction residual power value at the 0th order.

The reflection coefficient calculation unit 2 inputs the covariance φ m ₋₁ (i, j; n) and the prediction coefficient a (m−1, i; n) output from the prediction coefficient calculation unit 3, and the reflection coefficient γm (N) is generated and supplied to the prediction coefficient calculation unit 3 and the prediction residual calculation unit 4. The prediction coefficient a (m−1, i; n) is the i-th prediction coefficient value in the (m−1) th order, and the reflection coefficient γm (n) is the reflection coefficient value in the mth order. The prediction coefficient calculation unit 3 calculates the reflection coefficient γm.
Based on (n), the i-th prediction coefficient a (m, i;
n) is generated and given to the spectrum calculation unit 5. On the other hand, the prediction residual calculation unit 4 inputs the reflection coefficient γm (n) and the input signal power Po ₍ n _), and calculates the prediction residual Pm for the m-th order.
(N) is output and supplied to the spectrum calculation section 5.

Calculations in the reflection coefficient calculation unit 2, the prediction coefficient calculation unit 3, and the prediction residual calculation unit 4 recursively perform orders m = 1 to m = M.
(However, M is the optimum order). Prediction coefficient a (m, i; n) and prediction residual Pm obtained by the above processing
(N) is a feature quantity expressing a spectrum, and a spectrum is obtained from this. Therefore, the spectrum calculation unit 5
Then, based on the prediction coefficient a (m, i; n) and the prediction residual Pm (n), the m-th order spectrum Am (f, f of the time series signal x (n) is calculated.
n) and output it. Am (f, n) indicates the frequency f and the amplitude at time n when the m-th order spectrum analysis is performed.

(Problems to be Solved by the Invention) However, the spectrum analysis method and the apparatus thereof having the above configurations have the following problems.

(I) The covariance φ m _-1 (i, j, n) is the average of the time-series signal x (n) products within a frame, so the longer the frame length, the larger the amount of data and the accuracy of analysis. Becomes higher. Further, when the frame period (frame shift width) between the frames in the analysis is maximized, that is, when the frames are not overlapped, the maximum distortion occurs at the transition between the frames. Therefore, if each frame is overlapped and the frame cycle is shortened, the distortion decreases in inverse proportion to it, but since the amount of data processing increases and the analysis time increases, it is also possible to shorten the frame cycle. There is.

In this way, distortion corresponding to the frame cycle in the analysis occurs, and it is difficult to remove the distortion without shortening the frame cycle, that is, without increasing the data processing amount.

(Ii) The maximum entropy method (MEM) is based on the premise of stationary signal processing, and therefore, for example, when analyzing a signal with strong non-stationarity (a signal having a sharp amplitude with respect to time) such as "consonant" in speech. Causes an error in analysis.

SUMMARY OF THE INVENTION The present invention provides a spectrum analysis method and an apparatus therefor, which solve the problems of the above-mentioned prior art with respect to distortion depending on the frame period and analysis error due to non-stationarity of the input signal.

(Means for Solving the Problem) In order to solve the problem, the invention of claim 1 calculates a covariance value or an autocorrelation value from a time-series signal, and calculates the frequency spectrum of the time-series signal from the calculated value. In the spectrum analysis method to be obtained, the signal product of the time-series signal in a frame is obtained, high-frequency components of the signal product are removed, and the signal products after removal of the high-frequency components in the frame are added to obtain the covariance value or The autocorrelation value is calculated.

According to a second aspect of the present invention, a first calculation means for calculating a covariance value or an autocorrelation value from the time series signal, and a frequency spectrum of the time series signal from the covariance value or the autocorrelation value are calculated. In the spectrum analyzer including the second calculation means, the first calculation means includes a signal product calculation unit that calculates a signal product of the time-series signals in a frame, and a distortion that removes a high-frequency component of the signal product. A removal filter and an addition unit that inputs the output of the distortion removal filter, calculates the sum within the frame, and calculates the covariance value or the autocorrelation value.

(Operation) According to the inventions of claims 1 and 2, since the spectrum analysis method and the apparatus thereof are configured as described above, the signal product calculation unit obtains the signal product of the time-series signal and calculates the high-frequency component of the signal product. Is removed by the distortion removal filter, so that the distortion depending on the frame period and the analysis error due to the non-stationarity of the input signal are removed. After that, the output of the distortion removal filter is calculated by the addition unit, and the covariance value or the autocorrelation value is output. Therefore, if a spectrum is obtained from these output values, highly accurate spectrum analysis becomes possible. Therefore, the above problem can be solved.

(Embodiment) FIG. 1 shows a first embodiment of the present invention and is a functional block diagram of a spectrum analyzer for performing spectrum analysis by the maximum entropy method using covariance.

This spectrum analyzer is based on a digitized time series signal x (n) of voice, sound, etc., and is an integer (i,
j) and a distortion removal covariance calculation unit 10 which is a first calculation means for obtaining the 0th-order input signal power Po (n) and the covariance ψm (i, j; n) The second computing means is connected to the output side of the distortion removal covariance calculation unit 10.

The distortion removal covariance calculation unit 10 is configured by a digital signal processor, an individual arithmetic circuit, or the like, and includes a signal product calculation unit 11, a distortion removal filter 12, and an addition unit 13. The signal product calculator 11 calculates the time-series signal x
It has a function of calculating the signal product v (τ, n) of (n). In x (n), v (τ, n), n is time and τ = j−
i = 0,1, ..., M, i, j are integers. Distortion removal filter 12
Input the signal product v (τ, n) and perform filtering, and remove the high-frequency component, that is, the distortion component,
n) is output, for example, "Applications of Digital Signal Processing", 3rd edition (Sho 58-7-10) The Institute of Electronics and Communication Engineers, P.14-
It is composed of a non-recursive filter (FIR) and a recursive filter (IIR) described in 59. The adder 13 has a function of inputting the signal product (τ, n), outputting the covariance ψm (i, j; n) and the input signal power Po (n), and giving them to the second arithmetic means. is doing. In ψm (i, j; n), the order is m = 1,2, ..., M and the integers i, j = 0,1, ... M.

The second calculation means is a covariance ψm (i, j;
n) and the input signal power Po (n), the time series signal x
(N) The amplitude Am at the m-th order in the waveform component is set to the frequency f.
It has a function of calculating the spectrum Am ₋₁ (f, n) expressed as a function of. The second calculation means is composed of a digital signal processor, an individual calculation circuit, or the like, and includes a reflection coefficient calculation unit 20, a prediction coefficient calculation unit 21, a prediction residual calculation unit 22, and a spectrum calculation unit. It consists of 23. The reflection coefficient calculation unit 20 determines the covariance ψm
_-1 (i, j; n), and output from the prediction coefficient calculation unit 21 (m
-1) Input the i-th prediction coefficient a (m−1, i; n) at the next time, output the m-th order reflection coefficient γm (n), and output it to the prediction coefficient calculation unit 21 and the prediction residual. It has a function of supplying it to the difference calculation unit 22. The prediction coefficient calculation unit 21 calculates the reflection coefficient γm (n)
Based on, the i-th prediction coefficient a (m, i; n) for the m-th order is output and supplied to the spectrum calculation unit 23. The prediction residual calculation unit 22 inputs the reflection coefficient γm (n) and the input signal power Po (n), and calculates the m-th order prediction residual Pm.
It has a function of outputting (n) and giving it to the spectrum calculation section 23. The spectrum calculation unit 23 uses the prediction coefficient a.
From (m, i; n) and the prediction residual Pm (n), the spectrum An (f,
n) is output.

FIG. 3 is a block diagram of the distortion removal filter 12 in FIG.
The distortion removal filter 12 is provided with each signal product v (0, n), ..., v
Filter that outputs the signal products (0, n), ..., (τ, n), ..., (M, n) from which the distortion components of (τ, n), ..., v (M, n) are removed respectively 12- It is composed of 0, ..., 12−τ, ..., 12-M. Here, τ = 0, 1, ..., M.

The operation of the spectrum analyzer configured as above will be described.

When the time series signal x (n) digitized by an analog / digital converter or the like is input to the signal product calculation unit 11, the signal product calculation unit 11 calculates the product v of the time series signal x (n) from the following equation.
Find (τ, n).

v (τ, n) = x (n) · x (n−τ) (1) where τ = 0,1, ..., M When the distortion removal filter 12 is composed of, for example, a non-recursive filter, the filter Is the signal product (τ, n) with the distortion component removed from the signal product v (τ, n) by performing the convolution operation of
Is output.

However, the number of filter taps η = 0.1, ..., H H (η); filter tap coefficient The addition unit 13 inputs the signal product (τ, n), compares and judges the magnitude relationship between integers i and j, and then The m-th order covariance ψm (i, j; n) is obtained from the equation (3-1) and the 0th-order input signal power Po (n) is obtained from the following equation (3-2).

When j ≧ i However, when n, k; time m; order n; frame length τ = j−i i <j ψm (i, j; n) = ψm ₋₁ (i, j; n) (3-2) Next, the reflection coefficient calculation unit 20 determines the covariance ψ m ₋₁ (i,
j; n) and the prediction coefficient a (m-1, i; n) are input, and the reflection coefficient γm (n) is obtained from the following equation.

However, a (m, 0; n) = 1 The prediction coefficient calculation unit 21 inputs the reflection coefficient γm (n) and obtains the prediction coefficient a (m, i; n) from the following equation.

a (m, i; n) = a (m−1, i; n) + γm (n) · a (m−1, m−i; n)
(6) The prediction residual calculation unit 22 inputs the reflection coefficient γm (n) and the input signal power Po (n) and calculates the prediction residual Pm (n) from the following equation.

Pm (n) = Pm ₋₁ (n) {1- (γm (n)) ² } ...
(7) The above calculations of the equations (5), (6) and (7) are recursively repeated from the order m = 1 to m = M. Also, (1),
While the calculation of the expressions (2), (3-1), (3-2), and (4) is performed every time n, (5), (6),
The calculation of the equation (7) does not have to be performed every time n, but may be performed once for each frame period (interval) that actually requires a spectrum.

Prediction coefficient a (m, i; n) and prediction residual P obtained by the above processing
Since m (n) is a feature quantity expressing a spectrum, it is possible to directly use these for voice coding transmission and signal processing such as synthesis or recognition.

Next, if necessary, the spectrum calculation unit 23 inputs the prediction coefficient a (m, i; n) and the prediction residual Pm (n) to obtain the spectrum Am (f, n) from the following equation.

However, Am; amplitude f; frequency Δt; sampling interval As described above, in the first embodiment, the time series signal x (n)
The signal product v (τ, n) is calculated from
Since the covariance ψm (i, j; n) and the input signal power Po (n) are calculated after obtaining the signal product (τ, n) from which the distortion component has been removed, the frame cycle is not shortened. That is, the analysis error corresponding to the frame period and the non-stationarity of the time-series signal x (n) can be accurately removed without increasing the data processing amount, and thereby the accuracy under high noise can be improved. Spectrum analysis becomes possible.

FIGS. 4 (a) and 4 (b) are spectrum analysis measurement diagrams of the conventional example and the first example. In this figure, the result of spectrum analysis is shown when a voice signal is used as the test data and "pau" is uttered. The horizontal axis is frequency f (H
z), the vertical axis is the time n, and the amplitude Am is three-dimensionally expressed in the vertical axis direction. As is apparent from FIGS. 4 (a) and 4 (b), in the first embodiment shown in FIG. 4 (b), the change in the amplitude Am in the vertical direction is gentle and corresponds to the frame period. It can be seen that the distortion has been removed.

FIG. 5 shows a second embodiment of the present invention, and is a functional block diagram of a spectrum analyzer for performing spectrum analysis by a covariance method using covariance. Elements common to those shown in FIG. Are given the same reference numerals.

In this spectrum analyzer, the distortion removing covariance calculating unit 10 as the first calculating unit has the same configuration as that of the first embodiment, but the second calculating unit connected to the output side thereof has the matrix calculating unit. It is different from the first embodiment in that it is composed of 30 and a spectrum calculation unit 23.

The matrix calculation unit 30 outputs the covariance ψm output from the addition unit 13.
(I, j; n) is input and the prediction coefficient a (m, i; n) is calculated from the following equation.

The spectrum calculation unit 23 inputs the prediction coefficient a (m, j; n), and in the equation (7), the prediction residual Pm (n) = 1 of the numerator.
Then, the spectrum Am (f, n) is obtained. This allows
The same effect as that of the first embodiment can be obtained.

The present invention is not limited to the illustrated embodiment, and various modifications can be made. The following are examples of such modifications.

In the adder 13 shown in FIGS. 1 and 5, (3-1)
Expression If the calculation is performed, the autocorrelation value can be obtained instead of the covariance value. Therefore, using the autocorrelation value, the spectrum Am (f,
If n) is calculated, the same operation and effect as those of the first and second embodiments can be obtained.

(Effects of the Invention) As described in detail above, according to the spectrum analysis method and the apparatus thereof of the present invention, the signal product is calculated from a time-series signal, the distortion component is removed from the signal product, and then the covariance is obtained. Since the value or autocorrelation value is calculated, the distortion corresponding to the frame period can be accurately removed without shortening the frame period, that is, without increasing the data processing amount, and when the non-stationarity is strong. It is possible to reduce the analysis error of the serial signal.

[Brief description of drawings]

FIG. 1 is a functional block diagram of a spectrum analyzer showing the first embodiment of the present invention, FIG. 2 is a block diagram of a conventional spectrum analyzer, and FIG. 3 is a configuration of a distortion elimination filter in FIG. FIGS. 4 (a) and 4 (b) are spectrum analysis measurement diagrams, and FIG. 5 is a functional block diagram of the spectrum analyzer showing the second embodiment of the present invention. 10 …… Distortion removal covariance calculator, 11 …… Signal product calculator, 12…
... Distortion removal filter, 13 ... Addition unit, 20 ... Reflection coefficient calculation unit, 21 ... Prediction coefficient calculation unit, 22 ... Prediction residual calculation unit, 23
...... Spectrum calculator, 30 …… Matrix calculator, x (n) ...
… Time series signal, Am (f, n) …… Spectrum.

Claims (2)

[Claims] 1. A spectrum analysis method for calculating a covariance value or an autocorrelation value from a time-series signal and obtaining a frequency spectrum of the time-series signal from the value, for obtaining a signal product of the time-series signal in a frame, A spectrum analysis method, wherein a high frequency component of the signal product is removed, and the signal products after the removal of the high frequency component in the frame are added to calculate the covariance value or the autocorrelation value. 2. A first arithmetic means for calculating a covariance value or an autocorrelation value from the time series signal, and a second arithmetic means for calculating a frequency spectrum of the time series signal from the covariance value or the autocorrelation value. In the spectrum analyzer including the above, the first calculation means includes a signal product calculation unit that calculates a signal product of the time-series signals in a frame, a distortion removal filter that removes a high-frequency component of the signal product, A spectrum analysis device comprising: an addition unit that receives the output of the distortion removal filter and calculates the sum within the frame to calculate the covariance value or the autocorrelation value.