JPH0731512B2 - Formant extractor - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a formant extractor, and more particularly to a band-division type formant extractor which performs stable formant extraction in real time using band division.

[Conventional technology]

It is known that a high-order equation whose LPC coefficient is a constant is solved by the Newton-Raphson method and the formants are extracted from the solution, and the extraction accuracy is also good. .

However, there is no algebraic method for solving high-order equations in the conventional formant extraction means, and the higher the order, the more difficult it becomes. For this reason, a solution was usually found numerically while requiring a huge amount of calculation, but it was impossible to obtain a solution stably in real time.

On the other hand, the input voice signal is divided into multiple frequency bands,
A so-called band-division formant extractor, which extracts formants based on LPC coefficients obtained by LPC analysis for each divided frequency band, is being proposed.

This band-division formant extractor has attracted attention as a device capable of performing stable formant extraction in real time with high accuracy, as the order of a higher-order equation is significantly reduced according to the number of divided bands.

For example, if the 12th-order LPC coefficient is used, the 6th-order LPC coefficient is extracted in each of the 2 bands and the higher-order equation is solved. It becomes the next order, and the solution can be obtained much easier than the 12th order which should be solved originally.

[Problems to be solved by the invention]

However, such a band-division type formant extractor also has the following basic problems.

That is, when a formant frequency exists at or near a band division point between a plurality of frequency bands, formant extraction is performed in a state in which all voice energy related to the formant is not used, and thus the extraction accuracy naturally deteriorates. There is a drawback that it ends up.

In order to eliminate this drawback, a method provided with an expanded LPC analysis means for overlappingly analyzing frequency components near a band division point including a band division point in two frequency bands has been proposed. However, this method still has the following drawbacks. That is, the proposed method selects poles to be analyzed redundantly in two frequency bands by using at most one boundary frequency. Further, generally, the poles analyzed in each band are slightly different from the true values depending on the analysis accuracy. for that reason,
When analyzing the poles existing very close to the boundary frequency, each of the poles analyzed in the two frequency bands may exceed the boundary frequency, and as a result, the pole frequency may not be detected.

Therefore, an object of the present invention is to provide a formant extractor capable of reliably detecting the pole frequency even when the poles existing very close to the boundary frequency are redundantly analyzed in two frequency bands.

[Means for Solving Problems] An apparatus of the present invention divides an input audio signal into a plurality of frequency bands, and LPC (Linear Predic
(ction coding, linear predictive coding), and an expanded LPC analysis means for expanding and LPCing each of the plurality of frequency bands including the formant extraction frequency band, and overlapping analysis in adjacent frequency bands. And means for generating a frequency window for eliminating poles analyzed in adjacent frequency bands in the overlapping analysis frequency section.

〔Example〕

The present invention will now be described in detail with reference to the drawings. FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention.

One embodiment shown in FIG. 1 is LPF (Low Pass Filter) 1, A / D
Converter 2, band division LPC analyzer 3, equation solver 4-1
.About.4-I, pole frequency / band radiation calculators 5-1 to 5-I, formant determiner 6 and the like.

The input audio signal input via the input line 101 is supplied to the A / D converter 2 after removing unnecessary high frequency of 3.4 kHz or more by the LPF1.

The A / D converter 2 samples the input voice signal at a predetermined sampling frequency, for example, 8 kHz, and quantizes it with a predetermined number of bits, for example, 12 bits, and the quantized voice signal is temporarily stored in an internal memory for a certain time length, for example, 32 mSEC. That is, 256 samples are stored for each, a window function process for multiplying this by a window function such as a Hamming function or a rectangular function is performed for every 10 mSEC of the basic analysis frame period, and the output is supplied to the band separation LPC analyzer 3.

FIG. 2 is a block diagram showing in detail a part of the band division LPC analyzer 3.

The band division LPC analyzer 3 includes a Fourier transformer 31, a power spectrum calculator 32, and a plurality I of autocorrelation coefficient measuring devices 33-1.
.About.33-I, a plurality of I LPC analyzers 34-1 to 34-I, and the like.

The Fourier transformer 31 is a DFT (Discrete Foe) of the quantized audio signal for each basic analysis frame input from the A / D converter 2.
urier Transform (discrete Fourier transform) and transform it into a value in the frequency domain, which is then calculated by the power spectrum calculator 32
Supply to.

The power spectrum calculator 32 calculates the power spectrum of each input frequency spectrum component through arithmetic operations such as square addition of the real and imaginary parts, and temporarily stores them in the internal memory.

Now, the autocorrelation coefficient measuring devices 33-1 to 33-I read the power spectra stored in the power spectrum calculating device 32 for each preset divided frequency band, and read the data as follows. IDFT (Inve
rse DFT, inverse discrete Fourier transform).

The power spectrum is a scalar quantity and does not include the imaginary part,
Therefore, this Fourier inverse variable is an operation for only the cosine coefficient relation part of the real part. The inverse Fourier transform is performed for each frequency band such that the phase difference of the cosine coefficient for each frequency component becomes zero at the low frequency end of each band.

By the way, each frequency band of the autocorrelation coefficient measuring devices 33-1 to 33-I is expanded wider than the original frequency band including the formant extraction frequency band in the case of the present embodiment, and the division point of the original frequency band is expanded. Or, we are trying to avoid the problem when the formant frequency exists in the vicinity.

FIG. 3 is a band division explanatory diagram showing the characteristics of the band division in the method provided with the conventional expanded LPC analysis means.

Although FIG. 3 shows an example in which the number of divided bands is two,
Basically, the same applies to the case of the above division number.

In FIG. 3, S is the spectrum envelope of the input voice signal, and the conventional band-division type formant extractor is provided with two frequency bands B ₁ and B ₂ which are set so that they do not overlap each other as shown by the solid line. , The formants are obtained using the LPC coefficients extracted for each band.

Considering the case where there is a formant, for example, a second formant at the band division point P in such a band division, the energy of the second formant is not correctly evaluated in either of the frequency bands B ₁ and B _2. it is obvious. Therefore, in the conventional method, the frequency band B ₁ is W ₁
, And B ₂ extends its frequency band to W ₂ as shown by the alternate long and short dash line. In other words, by expanding and converting to a frequency band wider than the original frequency band including the formant extraction frequency band, the second formant is completely included in the expanded frequency band W ₁ , and the formant exists at or near the band division point. The problem of being present is basically eliminated.

The extent to which such band expansion is performed can be easily set in advance based on many audio materials and experience.

As is clear from the contents described above, in the case of the example shown in FIG. 3, the phase of the frequency at Q point in the _first divided frequency band W _{1 and the} frequency at R point in the second divided frequency band. Each phase is a phase reference point where the phase angle of the COSINE coefficient becomes zero.

The autocorrelation coefficient measuring units 33-1 to 33-I perform the above-mentioned inverse Fourier transform on the respective frequency bands set in this way. The result of the inverse Fourier transform performed in this way is nothing but the autocorrelation coefficient itself, and the autocorrelation coefficient measuring units 33-1 to 33-I correspond to respective frequency bands and the like and have orders of preset values. The autocorrelation coefficient is measured and these are respectively measured by LPC analyzers 34-1 to 34-
Supply to I.

Each of the LPC analyzers 34-1 to 34-I extracts the α parameter as an LPC coefficient by a known means using the autocorrelation coefficient thus supplied. These extracted α parameters are α parameters of the order corresponding to the order of the input autocorrelation coefficient.

In this way, when the α parameters output from the LPC analyzers 34-1 to 34-I are originally of a single frequency band when extracted as those of the same order, the nth order is the number of bands I Corresponding to, each is significantly reduced to n / I order. For example, if a 12th-order α parameter is extracted in the case of a single frequency band, the 6th-order α parameter is significantly reduced in each of only two bands.

The band division type LPC coefficient analysis is described in detail, for example, in JP-A-58-211797, "Band division type vocoder" or JP-A-58-220199, "Band division type vocoder".

The band 1α parameter to the band Iα parameter output from the LPC analyzers 34-1 to 34-I are respectively the equation solver 4
-1 to equation solver 4-I.

It is known to obtain the pole frequency and its bandwidth from the LPC coefficient. First, the α parameter as the LPC coefficient is set as a constant, and the complex conjugate solution of the polynomial equation of the same order as the α parameter is obtained. The basic means is to find the pole frequency and its bandwidth from. in this case,
If the polynomial equation is quadratic, the complex conjugate obtained by solving the polynomial equation is one set at the maximum, and one pole is determined corresponding to this one set of complex conjugate solutions. Therefore, for example, a 6th-order polynomial equation formed by inputting 6th-order LPC coefficients relating to a band obtained by dividing the original frequency into approximately 2 parts is expressed as a combination of 3 quadratic equations. When possible, three sets of complex conjugate solutions, and thus three poles will be determined. That is, even if the number of divided bands is two, the solution of a twelfth-order polynomial equation based on the twelfth-order LPC coefficient is reduced to the sixth order, and the processing becomes extremely easy.

The solution of the polynomial equation will be described in more detail as follows.

Now, assuming that the transfer function of the speech synthesis filter using the all-pole digital filter on the synthesis side is H (Z ^-1 ), H
(Z ^-1 ) = 1 / Ap (Z ^-1 ) Here, Ap (Z ⁻¹ ) is represented by the following equation (1).

Ap (Z ^-1 ) = 1 + α ₁ Z ⁻¹ + α ₂ Z ⁻² ＋ −−−− αpZ ⁻ p…
(1) In equation (1), Z = exp (jλ), λ = 2πΔTf, ΔT is the sampling period, f is the frequency, p is the order of the speech synthesis filter,
α _{1 to} αp are p-order α parameters.

To find the pole is to find the root of Ap (Z ^-1 ) = 0. For example, in the case of quadratic, the root of 1 + α ₁ Z ^-1 + α ₂ Z ^-2 = 0 or Z ² + α ₁ Z + α ₂ = 0 Is to ask.

In the case of the quadratic, it always has a set of complex conjugate solutions, and in general, multi-dimensional equations are processed by iteratively performing numerical solution so that they can be converted and expressed into the adjoint of a plurality of quadratic equations. With the solution of
Band-divided and lower-ordered complex conjugate solutions of the lower-order polynomial equations are obtained and are respectively calculated as pole frequency / bandwidth calculators 5-1 to 5-
Output to I.

Now, regarding the method of obtaining the pole frequency and its bandwidth from the complex conjugate solution of the multi-dimensional equation with α as the coefficient obtained in this way, for example, Ryozo Saito, Kazuo Nakata, “Basics of Speech Information Processing”, Ohmsha, It is described in detail in other documents and its outline is as follows.

The conjugate complex solution obtained from the quadratic equation is generally expressed by the following equation (2).

Z = re ^jθ , = re ^−jθ (2) Further, Z is shown by the following equation (3) on the complex plane.

^{Z = e ST = e (-p} + jω) T = e - p T · e jωT = re
^j (3) From equations (2) and (3), the pole frequency f and its bandwidth b can be expressed by the following equations (4) and (5), respectively.

The above description is for the case where the α parameter is quadratic.
It is obvious that the poles can be almost arbitrarily determined from the complex conjugate solutions by the idea that the quadratic equations formed by the quadratic α parameter are also combined in the case of the following or higher order.

The formant determiner 6 inputs the pole frequencies and the bandwidths of the respective divided bands thus output, and determines the formants contained in these pole frequencies based on a known method. The formant calculation range in this formant determination is a conventional method that decomposes all frequency bands without overlapping with each other, as in the case of B ₁ and B ₂ shown in FIG. The target is the divided frequency band of the method. This is the method adopted because the expanded frequency band is used only for the purpose of accurately extracting the formant and thus determining the pole frequency.

However, as described above, the poles analyzed in each band are slightly different from the true values depending on the analysis accuracy. Therefore, when the poles existing very close to the boundary frequency are analyzed, each of the poles analyzed in the two frequency bands exceeds the boundary frequency, and as a result, the pole frequency may not be detected.

FIG. 4 is a block diagram showing the configuration of the formant determiner 6 described in detail in FIG. The formant determiner 6 includes band discriminators 601-1 to 60-1, and frequency window signal generators 602-1 to I-1.
603-1 to I-1, polar discriminators 604-1 to I-1, and formman and estimator 605.

The pole frequency / bandwidth calculator 5-1 supplies the pole discriminator 601-1 with the polar wave number and bandwidth data of the band 1, that is, the lowest band of the Sub frequency band divided into I pieces. For ease of explanation, the lower limit of the frequency range in which band 1 is assigned as a divided frequency band is _BL (1) and the upper limit is
_BH (1), band 1 is the lower limit of the frequency range assigned as the expanded frequency band _EL (1), the upper limit is
_EH (1), similarly set the lower limit of the divided frequency band of band 2 to _BL
(2), the upper limit is _BH (2), the lower limit of the expanded frequency band is
_EL (2), upper limit is _EH (2), ..., lower limit of divided frequency band of band I is _BL (I), upper limit is _BH (I), lower limit of expanded frequency band is _EL (I), upper limit is _EH It is defined as (I). These _BL (i), _BH (i), _EL (i),
_EH (i), (i = 1, 2, ..., I) has the relationship of the following expression (6). _EL (i) < _BL (i) < _EL (i + 1) < _BH (i) = _BL (i + 1) < _EH (i) (6) The band discriminator 601-1 converts the supplied data into polar frequencies. Discriminate as follows. That is, the pole frequency is
Remove poles smaller than _BL (1). Then the pole frequency is
_The poles larger than _BL (1) and smaller than _EL (2) are selected, and this data is output to the formant estimator 605 via the output line 606-1. Further, the pole whose pole frequency exceeds _EL (2) is selected, and this data is output to the frequency window signal generator 602-2 and the pole discriminator 604-1 via the output line 607-1.

The polar frequency / bandwidth calculator 5-2 (not shown in FIG. 1) outputs the polar frequency and bandwidth data of band 2 to the band discriminator 60.
It is supplied to 1-2. The band discriminator 601-2 discriminates the supplied data according to the pole frequency as follows. The pole frequency is
_A pole smaller than _EH (1) is selected, and this data is output to the frequency window signal generator 603-1 and the pole discriminator 604-1 via the output line 608-2. Also, a pole whose pole frequency exceeds _EH (1) and is smaller than _EL (3) is selected, and this data is output to the formant estimator 605 via the output line 606-2. Further, a pole whose pole frequency exceeds _EL (3) is selected, and this data is output to the frequency window signal generator 602-2 and the pole discriminator 604-2 via the output line 607-2.

Band discriminators 601-3 to 601- (I-1) are band discriminators 601.
-2, the same processing as the pole frequency / bandwidth calculator 5-3 to 5
-(I-1) (not shown in FIG. 1).

The pole frequency / bandwidth calculator 5-I supplies the pole frequency and bandwidth data of the band I to the band discriminator 601-1.
The band discriminator 601-I discriminates the supplied data according to the pole frequency as follows. A pole having a pole frequency smaller than _EH (I-1) is selected, and this data is output to the frequency window signal generator 603- (I-1) and the pole discriminator 6 via the output line 608-I.
Output to 04- (I-1). Also, the pole frequency is _EH (I
Poles above -1) and smaller than _BH (I) are selected, and this data is output to formant estimator 6 via output line 606-I.
Output to 05. The poles whose pole frequency exceeds _BH (I) are removed.

In the above description, regarding band 1 and I, _EL (1)
< _BL (1), _BH (I) < _EH (I)
This is _EL (1) = _BL (1), _BH (I) = f
_{Even EH} (I) will not make any difference.

Next, the operation of the frequency window signal generators 602-1 to 602- (I-1) and 603-1 to 603- (I-1) will be described.

The frequency window signal generator 602-1 is supplied with data of a pole whose pole frequency exceeds _EL (2) from the band discriminator 601-1.
Of course, the highest frequency shown by this data is within _EH (1). Here, the pole frequency and bandwidth of the supplied data are _H (1,1), _H (1,2), ... _H (1, n _H1 ) b _H (1,1), b _H (1,2) ), ... and _{b H (1, n H1)} . Here, _H (1, i), b _H (1, i) (i = 1,2, ...
n _H1 ) indicates the pole frequency and bandwidth of the same pole. Also, _EL (2) < _H (1,1) < _H (1,2) <... < _H (1, n _H1 ) < _EH (1) ... (7), where n _H1 is the supplied data. The number of poles included.

The frequency window signal generator 602-1 outputs pole data _H (1, i), b
Generate a frequency window corresponding to _H (1, i). The center frequency of this frequency window is _H (1, i), and its window length is _WH (1, i).
i) is Is defined by Where k ₁₁ is an empirically optimized constant. The expression (8) means that the frequency window length _WH (1,
i) is a monotonically decreasing function of the corresponding pole bandwidth b _H (1, i), and is dependent on the pole center frequency _H (1, i) and the expanded band boundary frequency _EH (1) in the direction of the adjacent frequency band. It is a monotonically decreasing function. Note that _WH (1, i) does not necessarily have to be defined by equation (8), and b _H
A monotonically decreasing function of (1, i) and / or _H (1, i)
Any function that monotonically decreases in the direction of the adjacent frequency band of

The frequency window signal generator 602-1 outputs the generated frequency window signal to the pole discriminator 60-1.

The frequency window signal generators 602-2 to 602- (I-1) are the same as the 602
A frequency window signal is generated in the same manner as -1, and each pole discriminator 604
It outputs to -2-604- (I-1).

The frequency window signal generator 603-1 is supplied with pole data having a pole frequency smaller than _EH (1) from the band discriminator 602-1. Of course, the lowest frequency indicated by this data is _EL (2) or higher. Here, the pole frequency and bandwidth of the supplied data are _L (2,1), _L (2,2), ..., _L (2, n _L2 ) b _L (2,1), b _L (2,2 2), ..., b _L (2, n _L2 ). Here, _L (2, i), b _L (2, i) (i = 1,2, ...
n _L2 ) indicates the pole frequency and bandwidth of the same pole. Also, _EL (2) < _L (2,1) < _L (2,2) <... < _L (2, n _L2 ) < _EH (1) (9), where n _L2 is the supplied data. The number of poles included.

The center frequency of the frequency window signal generator 603-1 is _L (2, i)
And a window length _WL (2, i) of which is defined by the following equation (10) is generated and output to the pole discriminator 604-1.

The frequency window generators 605-2 to 603 (I-1) generate frequency window signals in the same manner as the frequency window generator 603-1, and each of them is a polar discriminator 604-2.
To 604 (I-1).

The pole discriminator 604-1 discriminates the poles that are redundantly analyzed in both the band 1 and the band 2, removes unnecessary poles, and outputs the result to the formant estimator 605. The pole discriminator 604-1 is a microprocessor and has three two-dimensional work areas. One of the regions is supplied from the band discriminator 601-1 through the output line 607-1, the pole data composed of the pole frequency and the bandwidth, and the frequency window signal generator 602-1. The frequency window signals are arranged in the order of polar frequencies as shown below.

Note that s _H (1, i) (i = 1,2 ... n _H1 ) is a status signal and is all set to “0”.

Further, in one of the other regions, the pole data constituted by the pole frequency and the bandwidth supplied from the band discriminator 601-2 via the output line 608-2, and the frequency window signal generator 603- 1
The frequency signals supplied from them are arranged in the order of polar frequencies as shown below. However, s _L (2,1) (i = 1,2 ... n _L2 )
Similarly, all are "0".

The polar discriminator 604-1 then calculates max { _WH (1, i), _WL (2,
j)} (i = 1, ..., n _H1 , j = 1, ..., n _L2 ). _WH
If (1, k) is the maximum, set s _H (1, k) to “1”, All column data of _L (2, l) for which is satisfied is removed from the area shown by the equation (12). Also, _WL (2, m) = max { _WH
In the case of (1, i), _WL (2, d)}, set s _L (2, m) to "1"
Set to All column data of _H (1, n) for which is satisfied are removed from the area shown by the equation (11).

Next, the same processing is performed by excluding the data whose status signal is "1". By performing such processing one after another, all status signals become "1". At this time, no other pole frequency exists in any frequency window.

Next, the pole discriminator 604-1 writes the poles existing in the two areas in the third work area in the order of frequency, and further writes the written data (expressed by the set of the center frequency and the bandwidth). Output to formant estimator 605. Polar discriminator 604-2 ~
The 604- (I-1) also performs the same operation. The formant estimator 605 estimates a formant from the supplied plurality of poles by a known method, and outputs the formant data output line 7
Output to.

〔The invention's effect〕

As described above, the present invention divides the frequency band into a plurality of frequency bands and performs LPC analysis on each band, and performs LPC analysis on a band expanded from the conventional frequency band division contents including the formant extraction frequency band. Formant extractor for extracting formants, comprising means for generating frequency windows for eliminating poles analyzed in mutually adjacent frequency bands in overlapping analysis frequency intervals for overlapping analysis in adjacent frequency bands Therefore, it is possible to solve the drawback that, when analyzing the poles existing very close to the boundary frequency, each of the poles analyzed in the two frequency bands exceeds the boundary frequency, and as a result, the pole frequency is not detected. .

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a block diagram showing in detail a portion of the band division LPC analyzer 3 of FIG. 1, and FIG. 3 shows the characteristics of the band division in the present invention. FIG. 4 is a block diagram of the formant determiner 6 showing the band division. 1 …… LPF, 2 …… A / D converter, 3 …… Band division LPC
Analyzer, 4-1 to 4-1 ... Equation solver, 5-1 to 5
-1 ... Pole frequency / bandwidth calculator, 6 ... Formant determiner, 31 ... Fourier transformer, 32 ... Power spectrum calculator, 33-1 to 33-1 ... Autocorrelation coefficient measuring device, 34
-1 to 34-1 LPC analyzer, 601-1 to I band discriminator,
602-1 to (I, 1), 603-1 to (I-1) ... Frequency window signal generator, 604-1 to (I-1) ... Pole discriminator, 605 ...
Formant estimator.

Claims (5)

[Claims] 1. A means for performing LPC (Linear Prediction Coding) analysis for each frequency band by dividing an input speech signal into a plurality of preset frequency bands without overlapping, and a formant extraction frequency band. Each of the plurality of frequency bands including a pre-expansion LPC analysis means for expanding the LPC by giving overlapping frequencies and overlapping frequency bands adjacent to each other in the overlapping analysis frequency section for overlapping analysis in adjacent frequency bands. Formant extractor, comprising means for generating a frequency window for removing the poles analyzed in. 2. A central frequency of the frequency window corresponds to a central frequency of the poles analyzed, and the window length of the frequency window is a function of the bandwidth of the poles analyzed. The formant extractor according to item (1). 3. The formant extractor according to claim 2, wherein the function is a monotonically decreasing function. 4. The center frequency of the frequency window corresponds to the center frequency of the poles analyzed, and the window length of the frequency window is a function of the center frequency of the poles analyzed and the band boundary frequency or expanded band frequency. The formant extractor according to claim (1), characterized in that. 5. The formant extractor according to claim 4, wherein the function is a function that monotonically decreases in a direction of an adjacent frequency band.