JPH08328593A - Spectrum analysis method - Google Patents

Abstract

PURPOSE: To improve precision in spectrum analysis and to increase a processing speed. CONSTITUTION: Respective first to Nb-th band signal extraction parts 101 -10Nb perform band division to inputted time sequential signals X(n). That is, they extract the time sequential signals at every band of Nb pieces to respectively impart to distortion removal covariance calculation parts 201 -20Nb . The distortion removal covariance calculation parts 201 -20Nb obtain signal products in frame in the time sequential signals at every band, and remove respectively their high frequency components by using a distortion removal filter. Then, the signal products removing the high frequency components are added, and the covariance at every band are obtained. Spectrum calculation parts 301 -30Nb respectively connected to respective distortion removal covariance calculation parts 201 -20Nb obtain spectra at every band, and a spectrum synthesis part 40 synthesizes these spectra at every band.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to time-series signals temporally in coded transmission of time-series signals such as voice or sound, and signal processing for synthesis or recognition of voice or sound. The present invention relates to a spectrum analysis method for obtaining a fluctuating frequency spectrum.

[0002]

2. Description of the Related Art Conventionally, techniques in such a field include:
For example, there was one shown in the following document. References: IEICE Transactions A, J73-A [6]
(1990-6), Yumi Takizawa, "Spectrum Estimation Method for Nonstationary Processes Based on Instantaneous Maximum Entropy Method" p.1083―
1093 is a functional block diagram showing a configuration example of a conventional spectrum analyzer. This device comprises a signal product calculator 1. The distortion remover 2 is connected to the output side of the signal product calculator 1, and the output side of the distortion remover 2 is connected to the adder 3. The reflection coefficient calculator 4 is connected to the output side of the adder 3. A prediction coefficient calculator 5 is connected to the reflection coefficient calculator 4, and the output side of the prediction coefficient calculator 5 is connected to the prediction coefficient-frequency characteristic converter 6. The output sides of the adder 3 and the reflection coefficient calculator 4 are also connected to the prediction residual calculator 7, and the output side of the prediction residual calculator 7 is also connected to the prediction coefficient-frequency characteristic converter 6. There is. The spectral analysis method in this device is a maximum entropy method (M) in which a spectrum is determined so as to maximize information entropy.
EM) is used to estimate the spectrum of random data. The signal product calculator 1 inputs the digitized time-series signal X (n) such as voice or sound, and calculates the signal product of the time-series signal X (n) in the frame. The distortion remover 2 filters the signal product output from the signal product calculator 1 to remove high frequency components of the signal product. As a result, the distortion depending on the frame period is removed. When this signal product after distortion removal is input, the adder 3 adds the signal products after distortion removal, and the distortion removal covariance φ _m−1 with the distortion removed.
(I, j; n) and the human power signal power P _O (n) are generated and given to the reflection coefficient calculator 4 and the prediction residual calculator 6. The distortion elimination covariance φ _m-1 (i, j; n) has the order m = 1, 2,
, M, integer i = 0, 1, ..., M. Integer j = 0,1,
, M, where n is the time, the covariance value has an integer (i, j) lag (lag; time lag) in the (m−1) th order. On the other hand, the input signal power P _O (n) is the prediction residual power in the 0th order.

The reflection coefficient calculator 4 calculates the distortion elimination covariance φ _m-1.
(I, j; n) and the prediction coefficient a (m-1, i; n) output from the prediction coefficient calculator 5 are input, and the reflection coefficient γ
_m (n) is generated and supplied to the prediction coefficient calculator 5 and the prediction residual calculator 6. Prediction coefficient a (m-1, i;
n) is the prediction coefficient value in the (m-1) th order, and the reflection coefficient γ _m (n) is the reflection coefficient value in the mth order. The prediction coefficient calculator 5 generates the i-th prediction coefficient a (m, i; n) in the m-th order based on the reflection coefficient γ _m (n), and outputs it to the prediction coefficient-frequency characteristic converter 6 give. The prediction residual calculator 6 calculates the reflection coefficient γ _m (n) and the input signal power P
Input _O (n) and predict residual power P for m-th order
_m (n) is output and given to the prediction coefficient-frequency characteristic converter 6. Calculations in the reflection coefficient calculator 4, the prediction coefficient calculator 5, and the prediction residual calculator 6 recursively have an order m =
Repeat from 1 to order m = M (where M is the optimum order). Prediction coefficient a (M,
i; n) and the prediction residual power P _M (n) are feature quantities expressing the spectrum, and the spectrum is obtained from this. Therefore, the prediction coefficient-frequency characteristic converter 6 uses the prediction coefficient a (M, i; n) and the prediction residual power P _M (n) to determine the Mth-order spectrum A of the time series signal X (n).
Generate and output _M (f, n). Spectrum A
_M (f, n) indicates the amplitude at frequency f and time n when M-th order spectrum analysis is performed.

[0004]

However, the spectrum analysis method having the above configuration has the following problems (i) and (ii). (I) The processes in the reflection coefficient calculator 4, the prediction coefficient calculator 5, and the prediction residual calculator 6 are recursive, and when the optimum order M is increased, the reflection coefficient calculator 4, the prediction coefficient calculator 5, and the number of multiplications in the prediction residual calculator 6 increases with the cubic expression of M. When the time series signal to be analyzed is affected by noise or the like, or when analyzing a signal having a relatively wide signal band, it is necessary to increase the optimum order M, and the amount of data processing increases rapidly. To do. That is, the processing time increases. (Ii) The smoothing degree (resolution) of the frequency spectrum in the time direction is determined by the characteristics of the low-pass filter used in the distortion remover 2, and the smoothing degree becomes constant over the entire frequency band. It is necessary to select the degree of smoothness, that is, the characteristic of the distortion removal filter, so as to be optimum depending on the speed of variation in the frequency characteristic of the analysis target signal in the time direction. That is, when the time variation of the frequency characteristic of the signal to be analyzed is fast, it is necessary to weaken the degree of smoothing so that the fast variation can be analyzed. When the time variation of the frequency characteristic of the signal is slow, it is necessary to increase the degree of smoothing so that the slow variation can be analyzed. However, in the case of an actual signal such as voice or sound, the time variation of the frequency characteristic of the signal tends to be relatively fast in the high frequency region, and the time variation of the frequency characteristic of the signal tends to be relatively slow in the low frequency region. There are many. Therefore, when the above-mentioned spectrum analysis method is applied to a signal having such characteristics, optimum analysis cannot be performed for all bands, and analysis accuracy deteriorates.

[0005]

According to the present invention, after calculating a signal product of a time series signal, a high frequency component of the signal product is removed,
In the spectrum analysis method for calculating the frequency spectrum of the time-series signal based on the covariance value or the autocorrelation value by adding the signal products after removing the high-frequency components and calculating the covariance value or the autocorrelation value, I am taking steps. That is, in the spectrum analysis method of the present invention, the time-series signal is divided into a plurality of bands, and band division processing for extracting each time-series signal of each band, and intra-frame in the time-series signal of each band are performed. A covariance calculation for obtaining a covariance value or an autocorrelation for each band by removing the high-frequency components of the signal product after adding And processing. Furthermore, based on each of the covariance values or autocorrelation values, a spectrum calculation process for respectively obtaining the frequency spectrum of the time-series signal for each band, by synthesizing each frequency spectrum obtained by the spectrum calculation process, A spectrum synthesizing process for obtaining a spectrum in the entire frequency band is performed.

[0006]

According to the present invention, since the spectrum analysis method is configured as described above, the input time-series signal is divided into a plurality of bands by the band division processing, and the time-series signal of each of the bands is respectively divided. To be extracted. By the covariance calculation process,
The signal product in the frame of the time-series signal for each band is obtained, the high frequency component of the signal product is removed, and the distortion depending on the frame period is removed. Then, the signal products from which the high frequency components have been removed are added to obtain the covariance value or the autocorrelation. Here, since the processing is advanced in the state of band division, if the number of band divisions is properly selected, the total amount of processing is greatly reduced. Also, when removing the distortion that depends on the frame period,
It is possible to select each band. Based on each covariance value or autocorrelation value, the frequency spectrum of the time-series signal for each band is obtained by the spectrum calculation process. The frequency spectra obtained by the spectrum calculation process are combined by the spectrum combining process to obtain the spectrum in the entire frequency band.

[0007]

FIG. 1 is a functional block diagram of a spectrum analyzer showing an embodiment of the present invention. This spectrum analyzer performs band-splitting spectrum analysis by the MEM method using covariance, and Nb performs band-splitting and down-sampling on a digitized time-series signal X (n) such as voice or sound. Band signal extraction units 10 ₁ , 10
₂ , ..., 10 _Nb . Each band signal extraction unit 10 ₁
Band-divided time series signals X are output on the output side of 10 to 10 Nb.
_{From k} (q) (k = 1, 2, ... Nb), a covariance φ _{k, m} (i, j; q) having an integer (i, j) in _{the m-} th order and 0
Nb distortion elimination covariance calculation units 20 ₁ , 20 ₂ , ..., 20 _Nb for obtaining the input signal power P _{k, 0} (q) at the next time are connected respectively. Each distortion removal covariance calculation unit 20 ₁ ,
On the output side of 20 ₂ , ..., 20 _Nb , the spectrum for each band is calculated for the covariance after distortion removal for each band N
b spectrum calculation units 30 ₁ , 30 ₂ , ..., 30
_Nb is connected respectively. Spectrum calculation unit 30
The output side of ₁ , 30 ₂ , ..., 30 _Nb is connected to a spectrum synthesizing unit 40 that synthesizes spectra for each band to calculate spectra for all bands. That is, the present apparatus is configured such that each unit for performing spectrum analysis on the time-series signal X _k (q) for each band is connected in parallel in Nb stages, and the spectrum synthesizing unit 40 synthesizes those outputs. The configuration will be described below, focusing on the system that processes the k-th band signal.

FIG. 3 is a functional block diagram showing the k-th band signal extracting section in FIG. k-th band signal extraction unit 10
_k is for performing band division processing, and is provided with a band pass filter 11 _k , a frequency shifter 12 _k, and a decimation processor 13 _k configured by a digital signal processor or individual arithmetic circuits. Band pass filter 1
Output of 1 _k is connected to the frequency shifter 12 _k, the output side of the frequency shifter 12 _k is connected to the thinning processor 13 _k. Here, k is an arbitrary integer from 1 to Nb, and Nb represents the number of band divisions. bandpass filter 11 _k of the k-th band signal extractor in 10 _k may have a function of filtering the sequence signal X (n) when the input and extracts the k-th band signal component XB _k (n) There is. The frequency shifter 12 _k has a function of shifting the frequency of the input XB _k (n) to the low frequency side and calculating the frequency-shifted signal X _k (n). The decimation processor 13 _k decimates X _k (n) so that the sampling frequency is determined by the bandwidth of the bandpass filter, and the decimation signal x _k (q) is subjected to the distortion removal covariance calculation unit 2
It outputs to 0 _k .

FIG. 4 is a functional block diagram showing the distortion removal covariance calculation unit in FIG. Distortion removal covariance calculation unit 20
_k is for performing covariance calculation processing for each band, and includes a signal product calculator 21 _k and a distortion removal filter 2 configured by a digital signal processor or individual arithmetic circuits.
It is provided with 2 _k and an adder 23 _k . Signal product calculator 21
The output side of the _k is connected to the distortion removing filter 22 _k, the output side of the distortion removing filter 22 _k is connected to the adder 23 _k. The signal product calculator 21 _k has a function of calculating the signal product V _k (τ, q) of the time-series signal x _k (q) after band division. The signal x _k (q) and the signal product V _k (τ,
In q), q is time, τ = j-i = 0,1, ...
M _k , i, j are integers. The distortion removal filter 22 _k is
The signal product V _k (τ, q) is input and filtered to remove the high frequency component, that is, the distortion component, V ″.
_It has a function of outputting _k (τ, q), and is configured by a non-recursive filter (FIR), a recursive filter (IIR), or the like. The adder 23 _k inputs the signal product V ″ _k (τ, q) and receives the covariance φ _{k, m} (i, j; q) and the input signal power P.
_It has a function of outputting _{k, 0} (q) and giving them to the subsequent spectrum calculation unit 30 _k . Note that the covariance φ _{k, m}
In (i, j; q), the orders m = 1, 2, ..., M _k ,
, M _k , integer j = 0, 1, ..., M _k
Is.

FIG. 5 is a functional block diagram showing the spectrum calculation section in FIG. The spectrum calculation unit 30 _k performs spectrum calculation processing for each band, and includes a reflection coefficient calculator 31 _k and a prediction coefficient calculator 32 configured by a digital signal processor or an individual arithmetic circuit.
_k , a prediction residual calculator 33 _k, and a prediction coefficient-frequency characteristic converter 34 _k . The output side of the reflection coefficient calculator 31 _k is connected to the prediction coefficient calculator 32 _k and the prediction residual calculator 33 _k , and the prediction coefficient calculator 32 _k and the prediction residual calculator 33 _k are connected.
The output side of the _k, the prediction coefficient - is connected to the frequency characteristic converter 34 _k. A part of the output of the prediction coefficient calculator 32 _k is a structure which is input to the reflection coefficient calculator 31 _k.
The spectrum calculation unit 30 _k uses the covariance φ _{k, m} (i, j;
q) and the input signal power P _{k, 0} (q), it has a function of calculating the spectrum A _{k, m} (f, q) in the k-th band. The reflection coefficient calculator 31 _k determines the covariance φ _{k, m-1}
(I, j; q) and the prediction coefficient calculator 32 is output from the _k (m-1) i th prediction coefficient of the next time a _k (m -1,
From i; q), the m-th order reflection coefficient γ _{k, m} (q) is obtained,
It has a function of outputting it to the prediction coefficient calculator 32 _k and the prediction residual calculator 33 _k . The prediction coefficient calculator 32 _k is
Based on the reflection coefficient γ _{k, m} (q), the i-th prediction coefficient a _k (m, i; q) in the m-th order is calculated and supplied to the prediction coefficient-frequency characteristic converter 34 _k. Have The prediction residual calculator 33 _k calculates the m-th prediction residual power P from the reflection coefficient γ _{k, m} (q) and the input signal power P _{k, 0} (q).
Prediction coefficient-frequency characteristic converter 34 _{k by} obtaining _{k, m} (q)
F has the function of giving. The prediction coefficient-frequency characteristic converter 34 _k uses the prediction coefficient a _k (m, i; q) and the prediction residual power P _{k, m} (q) to generate a spectrum A _{k, m} (f,
q).

The spectrum synthesizing section 40 performs a spectrum synthesizing process, and is composed of a digital signal processor or an individual arithmetic circuit. The spectrum synthesizing unit 40 includes a spectrum calculating unit 30 _{1 for} each band.
It has a function of synthesizing all the spectra A _{k, m} (f, q) calculated from ˜30 _Nb and outputting the spectra A _m (f, q) of the entire band. Next, the operation of processing the k-th band signal by the spectrum analyzer of FIG. 1 will be described. The time series signal X (n) digitized by an analog / digital converter or the like is input to the bandpass filter 11 _k of the _kth band signal extraction unit 10 _k . Band pass filter 1
When 1 _k is composed of, for example, an acyclic filter, the filter performs the convolution operation of the following equation (1) to extract the signal VB _k (n) corresponding to the component of the k-th band B _k .

[0012]

[Equation 1] However, i: integer hb _{k, i} ; tap coefficient Nh _k of i-th complex number that realizes the k-th bandpass filter Nh _k ; the number of tap coefficients. Here, the Kth band B _k is a normalized frequency of 0.
It is the k-th band when the band up to π is divided into Nb bands. The band is divided evenly. When 0 to π are equally divided into Nb pieces, and K = 1, 2, ... N from the low frequency band
When numbered b, the center frequency of the k-th band B _k
WC _k and bandwidth Wb _k are determined by the following equations (2) and (3). Wb _k = π / Nb (2) WC _k = (2k-1) / 2 · Wb _k (k = 1,2, ..., Nb) ··· (3) The frequency shifter 12 _k is input. Amplitude modulation is performed on XB _k (n), and the input signal XB _k (n) is k
Th only center frequency WC _k of band B _k to calculate the shifted signal to the low frequency, further signal X _k of the real part only (n)
Is output. That is, the signal X _k (n) of the following equation (4) is output. X _k (n) = Re [XB _k (n) · exp (−j · WC _k · n)] (4) where Re [,] is an operation for extracting the real part of the complex signal, and exp (·) Is a complex exponential function, and j is an imaginary unit.

The decimation processor 13 _k receives the input signal X
_k (n) is down-sampled at a sampling frequency determined by the bandwidth Wb _k of the k-th band B _k to obtain x _k (q) as follows. For example, the band from 0 to π is Nb
In the case of even division into X pieces, X _k (n) every Nb pieces,
Output as x _k (q). That is, Is. The signal product calculator 21 _k finds the product V _k (τ, q) within the frame of the time series signal x _k (q) by the following equation (5). V _k (τ, q) = x _k (q) · x _k (q−τ) (5) where τ = 0, 1, ..., M _k M _k ; analysis for k-th band signal component Order If the analysis order M _{k of} each band is optimum, it is considered that there is a relationship of the following expression (6).

[0014]

[Equation 2] However, M is the order of the signal X (n). Here, by setting the division number Nb to 1 or more, M _k becomes a positive number smaller than the analysis order M when analyzing the entire frequency, so that the processing amount in the following processing is reduced. When the distortion removal filter _22k is formed of, for example, a non-recursive filter, the filter performs a convolution operation of the following expression (7) to obtain a signal obtained by removing the distortion component from the signal product V _k (τ, q). The product V " _k (τ, q) is output.

[0015]

(Equation 3) However, i = 0, 1, ..., H−1 h _k (i); Tap coefficient of filter Here, by changing the tap coefficient h _{k of the} filter for each k-th band, the characteristics of the distortion removal filter are changed. It is possible to set for each band. The adder 23 _k inputs the signal product V ″ _k (τ, q), compares and judges the magnitude relation between the integers i and j, and from the following equations (8a) and (8b), the m-th order covariance φ _{k, m} (i, j; q) is _{obtained, and} the input signal power P _{k, 0} (q) of the equation (9) is _obtained .

[0016]

[Equation 4] Next, the reflection coefficient calculator 31 _k determines the covariance φ of the first order.
_{k, m-1} (i, j; q) and the prediction coefficient a _k (m-1, i;
q) and input the reflection coefficient γ from the following equation (10).
_{Find k, m} (q).

[0017]

(Equation 5) The prediction coefficient calculator 32 _k receives the reflection coefficient γ _{k, m} (q) and calculates the prediction coefficient a _k (m, i; q) from the following equation (11).
Ask for. In addition, the prediction residual calculator 33 _k uses the reflection coefficient γ
_{k, m} (q) and the input signal power P _{k, 0} (q) are input, and the prediction residual power P _{k, m} (q) is _obtained from the equation (12).

A _k (m, i; n) = a _k (m-1, i; n) + γ _{k, m} (n) · _ak (m-1, m−i; n) (11) ) P _{k, m} (n) = P _{k, m-1} (n) · {1- (γ _{k, m} (n)) ² } (12) Calculation of the above formulas (10) to (12) Is recursively m
It is repeated from = 1 to m = M _k . Also, (5),
The calculations of formulas (7), (8a), (8b), and (9) are performed at each time q, whereas the calculations of formulas (10) to (12) need to be performed at each time q. Absent. It may be performed once for each frame period (interval) that actually requires the spectrum. Since the obtained prediction coefficient a _k (m, i; q) and the prediction residual power P _{k, m} (q) are the feature quantities expressing the spectrum, these are used for voice coding transmission, synthesis or recognition, etc. It can also be used directly for signal processing. Next, if necessary, the prediction coefficient-frequency characteristic converter 34 _k uses the prediction coefficient a _k (m, i; q) and the prediction residual power P.
Input _{k, m} (q) and use the following formula (13) to calculate the spectrum A
_{Find k, m} (f, q).

[0019]

(Equation 6) However, A _{k, m} (f, q); spectrum of k-th band f; frequency Δt; sampling interval The spectrum for each band obtained by the above method is (1
The spectrum is synthesized by the spectrum synthesizing unit 40 using the equation (4). _Am (f, q) = _{Ak, m} (f-1 (k-1) /Nb.fs,q), [(k-1) .fs / Nb <f <k.fs / Nb] (14) As described above, in the present embodiment, each band signal component x _k (q) is extracted from the time series signal X (n), and the signal product V _k is extracted.
(Tau, q) is calculated, and the product V _k (τ, q) signal products V to remove the distortion component from the _"k (τ, q) are determined. Then, the signal volume V" _k (tau, q ), The covariance φ _{k, m}
Since (i, j; q) and the human power signal power P _{k, 0} (n) are calculated, it is possible to reduce the optimal order when analyzing the time-series signal for each band, and to divide the number of band divisions. Nb
With proper selection, the overall throughput will be significantly reduced. That is, the processing time is shortened. Further, it becomes possible to select the optimum characteristics of the distortion removal filter for the signal product of each band for each band, and it becomes possible to perform analysis that matches the characteristics of the actual signal and improve the analysis accuracy. The present invention is not limited to the above embodiment, and various modifications can be made.
The following are examples of such modifications. In the adder of FIG. 1 and FIG. 5, the equation (8a) is changed to the following (1
By performing the calculation with equation (5), the autocorrelation value can be obtained instead of the covariance value. Even when the spectrum A _m (f, n) is calculated using the autocorrelation value, the same effect as in the above embodiment can be obtained.

[0020]

(Equation 7)

[0021]

As described in detail above, according to the present invention, the input time series signal is divided into bands by the band division process, and the covariance value of each band or The autocorrelation value is obtained, the frequency spectrum for each band is obtained by the spectrum calculation process, and the spectrum in the entire frequency band is obtained by the spectrum synthesis process. Therefore, it is possible to reduce the optimum order when analyzing the time-series signal for each band, and the overall processing amount is significantly reduced. Therefore, the processing time can be shortened. Also, it becomes possible to apply a distortion removal filter with optimal characteristics for the signal product of each band,
An analysis suitable for the characteristics of the actual signal becomes possible. for that reason,
Analysis accuracy is improved.

[Brief description of drawings]

FIG. 1 is a functional block diagram of a spectrum analyzer showing an embodiment of the present invention.

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

3 is a functional block diagram showing a k-th band signal extraction section in FIG. 1. FIG.

4 is a functional block diagram showing a distortion removal covariance calculation unit in FIG. 1. FIG.

5 is a functional block diagram showing a spectrum calculation unit in FIG. 1. FIG.

[Explanation of symbols]

10 ₁ to 10 _Nb 1 th ~Nb th band signal extracting unit 20 ₁ to 20 _Nb distortion removal covariance calculation unit 30 ₁ to 30 _Nb spectrum calculation section 40 spectrum combining unit

Claims (1)

[Claims] 1. A covariance value or an autocorrelation is calculated by calculating a signal product of a time-series signal, removing high frequency components of the signal product, and adding the signal products after the removal of the high frequency component. In a spectrum analysis method for obtaining a frequency spectrum of the time-series signal based on a variance value or an autocorrelation value, the time-series signal is divided into a plurality of bands, and a band division process for extracting the time-series signal for each band, respectively. , After removing the high-frequency component of the signal product after obtaining the signal product in the frame in the time-series signal for each band,
A covariance calculation process for obtaining a covariance value or an autocorrelation for each band by adding the signal products from which the high-frequency components have been removed, and each band based on the covariance value or the autocorrelation value. It is possible to perform a spectrum calculation process for obtaining the frequency spectrum of each time-series signal, and a spectrum synthesis process for obtaining a spectrum in the entire frequency band by combining the frequency spectra obtained by the spectrum calculation process. Characteristic spectrum analysis method.