JPS6211360B2 - - Google Patents

Abstract

PURPOSE:To enable to discriminate voice and voicelessness accurately, by obtaining the spectrum of audio with the linear forescasting analysis and the pitch with the autocorrelation analysis method, to compress the deviation of the distribution of the maximum value of K parameter and autocorrelation. CONSTITUTION:The K parameter pick up unit 102 receives audio input signal from the terminal 102, picks up the K parameter being the linear forecasting coefficient, called partial autocorrelation, to a plurality of orders with the linear forecasting analysis method, and outputs the K parameter of a plurality of orders piked up to the log area ratio converter 103. The converter 103 converts the K parameter of primary, secondary order,... into the log area ratio L1, L2... by using the value memorized in the read only memory in advance and outputs it to the voice and voicelessess discriminator 104. The discriminator 104 discriminates voice or voicelessness and outputs it, depending whether the value of W1L1+W2L2+... is greater than the threshold value for discrimination predetermined or smaller. Further, W1 and W2 are constants obtained in advance.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a voiced/unvoiced discriminator for discriminating whether speech is voiced or unvoiced from a speech waveform, and more particularly to a voiced/unvoiced discriminator using a linear discriminant. The present invention relates to a voiced/unvoiced discriminator that discriminates between voiced and unvoiced speech using a linear discriminant that combines autocorrelation coefficients used for pattern pitch extraction, K parameters representing spectral envelope information, etc. as they are or after simple conversion.

It is known that voiced/unvoiced discrimination information in speech is an important parameter in speech analysis and synthesis. For example, in a speech analysis and synthesis system, voiced/unvoiced discrimination information discriminated by an analysis section has a great influence on the quality of synthesized speech synthesized by a synthesis section. This influence cannot be ignored, as it often fatally deteriorates the quality of synthesized speech, regardless of whether a voiced part is determined to be unvoiced or an unvoiced part is determined to be voiced. For example, if a voiced sound part is determined to be unvoiced, the synthesized sound will have a so-called "scattered" sound, greatly impairing its naturalness. Furthermore, if an unvoiced sound portion is determined to be voiced, the synthesized sound will have a so-called "choppy" sound, which may greatly impair naturalness and have a serious impact on intelligibility.

Conventionally, parameters that can distinguish between voiced and unvoiced to some extent are short-time average power, which takes advantage of the fact that energy differs between voiced and unvoiced, and relatively low frequency range, which takes advantage of the fact that the density function of the frequency power spectrum is different between the two. The ratio to the short-time average power of Autocorrelation coefficients that express formant information well, such as autocorrelation coefficients at close delay times, and the maximum value of autocorrelation coefficients at delay times in a range where pitch period delay times are considered to be distributed (hereinafter referred to as δ _MAX ), various parameters determined by linear predictive analysis method, for example, the so-called α
Parameters that are found as direct solutions to linear equations known as parameters, K parameters that indicate the so-called reflection coefficient in the vocal tract, or known as partial autocorrelation coefficients or partial autocorrelation coefficients.
Many parameters are known, such as parameters related to the α parameter, or parameters known as the so-called cepstrum.

However, none of these parameters alone has sufficient voiced/unvoiced discrimination characteristics. In other words, all of the above parameters can be expected to have some effect on voiced/unvoiced discrimination, but they are not perfect.

Therefore, voiced/unvoiced discrimination devices have conventionally performed voiced/unvoiced discrimination by combining various parameters such as those described above. Generally, the following three methods are used to combine parameters.

The first method is to determine voicedness in advance for each parameter used in the combination, for example, using a threshold value that can clearly determine that it is voiced, and if at least one of the parameters is determined to be voiced, it is determined that voiced overall. This method uses logical operations such as determination, that is, operations using logical sums and logical products. The second method is to use a linear discriminant in which each parameter used in the combination is a variable.
The third method is a mixed method of the first method and the second method. Of these three methods, the second method is known to be particularly suitable for computer-like devices. As for the parameters to be selected, it goes without saying that it is advantageous to use, as much as possible, parameters extracted from the spectrum information analysis device and the pitch, which are used simultaneously with the voiced/unvoiced discriminator.

α parameter extracted from linear predictive analysis method,
Parameters such as the K parameter are parameters that can distinguish between voiced and unvoiced to some extent, and in particular, the distribution of low-order K parameters such as the first-order K parameter is relatively well separated between voiced and unvoiced cases. Are known.

Furthermore, the maximum value of the autocorrelation coefficient in the range of delay times that is considered to be approximately distributed in pitch cycle delay times, that is, δ _MAX , extracted from the autocorrelation analysis method, also has a distribution that is relatively different between voiced and unvoiced cases. known to be well separated. Using the above properties, for example, the first-order K parameter and δ
A method is known in which voiced/unvoiced discrimination is performed using a linear discriminant with _MAX as a variable. (For example, published patent publication, JP 51-149705, "Drive sound source signal analysis method") However, the above method does not apply to the distribution example of the first-order K parameter (for example, Fumihiro Yato, Matsushita Kure, "Speech analysis and synthesis system analysis method"). "A Study on Voiced/Unvoiced Determination", Proceedings of the 1975 National Conference of the Information Division of the Institute of Electronics and Communication Engineers,
As is clear from Figure 1) on page 205, the distribution of the first-order K parameter is highly biased. When determining a linear discriminant, it is difficult to select the coefficients of the discriminant and the threshold for discrimination that provide a good degree of separation between voiced and unvoiced states.
Furthermore, since δ _MAX is also biased, it has a similar drawback.

That is, the threshold value of the linear discriminant is determined on the premise that the distribution of each class of parameters used for discrimination is a normal distribution. however,
As mentioned above, the distribution of parameters conventionally used to determine voiced/unvoiced, such as K ₁ , K _{2 ,} etc., is significantly deviated from the normal distribution. For this reason, conventional methods that directly use K ₁ , K ₂ , etc. as discrimination parameters have the disadvantage that they often result in erroneous discrimination.

SUMMARY OF THE INVENTION An object of the present invention is to provide a voiced/unvoiced discriminating device which greatly improves the above-mentioned drawbacks and enables precise voiced/unvoiced discrimination.

Specifically, by performing a certain transformation on the parameters obtained by using the linear prediction analysis method for the voice spectrum information and the autocorrelation analysis method for the pitch information, for example, the bias in the distribution of the K parameter and δ _MAX can be compressed. The present invention provides a voiced/unvoiced discrimination device that enables precise discrimination between voiced and unvoiced.

According to the present invention, voiced/unvoiced discrimination is performed by comparing the value of a discriminant having a predetermined coefficient and having a predetermined characteristic parameter of an input audio signal as a variable with a predetermined threshold value to determine voiced/unvoiced. In the apparatus, at least one of the variables has a partial autocorrelation coefficient (K parameter) up to a predetermined order.
A voiced/unvoiced discriminator is obtained, which is characterized in that this partial autocorrelation coefficient is converted into a log area ratio and used as a variable in the discriminant.

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

First, to explain the basic concept of the present invention, the first-order K parameter and the second-order K parameter, which are generally obtained by using the linear predictive analysis method of voice spectrum information, have a large bias in distribution as described above. When these two parameters are quantized for the purpose of transmission or storage, for example, they may be quantized after being converted into log area ratios (logarithmic values of vocal tract cross-sectional area ratios). For example, R.VISWANATHAN,
JOHN MAKHOUL, “Quantigation Properties
of Transmission Parameters in Linear
Prediction Systems”, IEEE TRANSACTIONS
ON ACOUSTICS, SPEECH, AND SI−GNAL
PROCESSING.VOL.ASSP−23, NO.3
Based on equation (31) etc. of JUNE1975 pp309-321.
One of the reasons for converting the K parameter into a log area ratio is that the bias in the distribution is compressed by this conversion. Generally, the cross-sectional area of each part of the vocal tract changes continuously over time. Therefore, during voice analysis, the cross-sectional area of the vocal tract is determined by a cross-sectional area ratio representing each cross-sectional area on both sides of a sampling point in the vocal tract (for example, the average value of the vocal tract cross-sectional area between sampling points is used as the representative). ) is used. As mentioned above, since the K parameter means the reflection coefficient within the vocal tract, the cross-sectional area ratio can be expressed as (1+Kn)/(1-Kn). Here, n indicates the order of K. Therefore, the log area ratio can be expressed as log(1+Kn)/(1-Kn). More generally speaking, the vocal tract cross-sectional area ratio of the n-th vocal tract corresponds to the vocal tract cross-sectional area of nVoTo from the aperture, where Vo is the sound velocity and To is the sampling period.

Therefore, if this is applied to voiced/unvoiced discrimination, especially for the first-order K parameters and second-order K parameters whose distributions are highly biased, the above parameters will not be directly used in the linear discriminant for voiced/unvoiced discrimination. ,
More reliable voiced/unvoiced discrimination can be performed.

Furthermore, the log area ratio has usually been calculated for the quantization of the K parameter, and therefore, the use of the log area ratio has the advantage that it generally does not impose a burden on the amount of calculation. It is clear that the third-order and subsequent K parameters have relatively small deviations in distribution and can be used directly in the linear discriminant for voiced/unvoiced discrimination. In addition, the distribution of δ _MAX has some bias (for example, Figure 2 of the paper by Tanito et al. mentioned above), and by performing a nonlinear transformation such as the transformation using the following formula aδ _MAX / (b−cδ _MAX ), , it is clear that it becomes more effective as a parameter used in the linear discriminant for voicing/unvoicing discrimination. In the above formula, a, b, and c are all constants.

Note that the nonlinear transformation generally involves an increase in the amount of calculations, and unlike the case where the first-order K parameter or the second-order K parameter is directly used in the linear discriminant for voiced/unvoiced discrimination, δ _MAX is directly used for the linear discriminant. Considerable effects can be expected even when used as a parameter in the equation, so if a slight error in voiced/unvoiced discrimination can be tolerated, the above δ _MAX can be used directly as a parameter of the linear discriminant, or the above δ _MAX can be used as a parameter in the linear _discriminant . It is necessary to select whether to use it after nonlinear transformation depending on the requirement for accuracy of voiced and unvoiced.

The present invention is based on the first order log area ratio, the second order log area ratio, the third order and subsequent K parameters, and δ _MAX.
The voiced/unvoiced discrimination is performed using linear discriminants based on various combinations including necessarily a first-order log area ratio.

Next, a specific configuration example of the voiced/unvoiced discriminator according to the present invention will be explained.

FIG. 1 is a block diagram showing a first embodiment.

An audio waveform input signal is supplied to a K-parameter extractor 102 via a waveform input terminal 101 . K
The parameter extractor 102 extracts K parameters, which are a type of linear prediction coefficient called partial autocorrelation coefficient or partial autocorrelation coefficient, from the audio waveform input signal up to multiple orders, at least up to second order, for example.
AUTOCORRELATION method or REGUL-AR
LATTICE method or COVARI-ANCE
LATICE method (For the above three analytical methods, see JOHN MAKHOUL, “Stable and Efficient
“Lattice Methods for Linear Prediction”
IEEE TRANSACTIONS ON ACOUSTICS,
SPEECH, AND SIGNAL PROCESSING,
VOL.ASSP−25, NO.5 OCTOBER1977, pp423
~428), etc., and the extracted at least first-order K parameter and second-order K
parameters to the log area ratio converter 103.

The log area ratio converter 103 converts the primary K parameters and secondary K parameters supplied from the K parameter extractor 102 into log area ratios using, for example, values stored in advance in a read-only memory. and outputs the conversion result to the voiced/unvoiced discriminator 104. Voiced/unvoiced discriminator 10
4 means that the value shown by the following formula W ₁ L ₁ + W ₂ L ₂ is larger than the predetermined discrimination threshold, where L ₁ is the first-order log area ratio and L ₂ is the second-order log area ratio. Whether it is voiced or unvoiced is determined by whether it is small or small. Note that the discrimination threshold and W ₁ and W ₂ are constants that are determined in advance using a method such as multivariate analysis, for example. Furthermore, the voiced/unvoiced discriminator 104 outputs a voiced/unvoiced signal output terminal 1 as a voiced/unvoiced discrimination signal.
Output to 05.

FIG. 2 is a block diagram for explaining the second embodiment.

An audio waveform input signal is supplied to a K-parameter extractor 202 via a waveform input terminal 201 . The K-parameter extractor 202 extracts K-parameters of at least 3rd-order or higher order from the audio waveform input signal by a linear predictive analysis method such as the AUTOCORRELATION method, and extracts 1st-order and 2nd-order K parameters among the extracted K parameters. The K parameters from the 3rd to the Nth order are sent to the log area ratio converter 203.
The parameters are output to the voiced/unvoiced discriminator 204, respectively. Note that N is an integer of 3 or more.

The log area ratio converter 203 converts the primary K parameters and secondary K parameters supplied from the K parameter extractor 202 into log area ratios using, for example, values recorded in advance in a read-only memory. and outputs the conversion result to the voiced/unvoiced discriminator 204. Voiced/unvoiced classifier 20
4 is the first-order log area ratio as L ₁ , the second-order log area ratio as L ₂ , the third-order K parameter as K ₃ ,
Assuming the Nth-order K parameter as K _N , the following formula Voiced or unvoiced is determined based on whether the value indicated by is larger or smaller than a predetermined discrimination threshold. Note that V ₁ , V _{2 .} . . , and V _N are constants. Furthermore, the voiced/unvoiced discriminator 204 outputs the voiced/unvoiced discrimination result to the voiced/unvoiced signal output terminal 205 as a voiced/unvoiced discrimination signal.

FIG. 3 is a block diagram for explaining another example of the configuration of the second embodiment.

An audio waveform input signal is supplied to a K-parameter extractor 302 via a waveform input terminal 301 . The K-parameter extractor 302 has means for measuring the ratio of predicted residual power to input signal power, that is, normalized predicted residual power, from the waveform input signal, and has a function of controlling the order of the K-parameter based on this measurement result. It is a vessel.

The K parameter extractor 302 is connected to the waveform input terminal 30
1, and extracts K-parameters of multiple orders until the normalized prediction residual power becomes equal to or less than a preset value.

Regarding the K-parameters of multiple orders, the maximum order is usually limited to, for example, the 8th order. Furthermore, it is empirically known that the minimum order is 2 if an appropriate set value of the normalized prediction residual power is selected. Furthermore, the K-parameter extractor 302 sends the extracted primary and secondary K-parameters to the K-parameter transmission path 1,303,
The analytical order signal indicating N1 is transmitted through the analytical order signal transmission path 3.
Output to 05. If N1 is 3 or more, the K parameter extractor 302 further outputs one or more K parameters from the third order to the N1 order to the K parameter transmission line 2,304. The log area ratio converter 306 converts the primary and secondary K parameters supplied via the K parameter transmission lines 1 and 303 into log area ratios, respectively, and outputs the conversion results to the log area ratio transmission line 307. .

The voiced/unvoiced discriminator 308 is connected to the analysis order signal transmission path 3
It is controlled by an analysis order signal indicating the order N1 of the K parameter supplied from 05. When N1 is 2, the voiced/unvoiced discriminator 308 determines the primary log area ratio L ₁ and the secondary log area ratio L ₂ supplied via the log area ratio transmission line 307.
, voiced/unvoiced discrimination is performed using the following formula W ₁ L ₁ +W ₂ L ₂ . In addition, when N1 is 3 or more, the K parameters K ₃ , ..., K _N1 from the third order to the N1 order supplied via the L ₁ , L ₂ and the K parameter transmission line 2, 304 are used. , below formula is used to determine voiced/unvoiced. Note that the above W ₁ ,
W ₂ and Ui ^N1 (i=1,...N1) are constants.
Furthermore, the voiced/unvoiced discriminator 308 outputs the voiced/unvoiced discrimination result to the voiced/unvoiced discrimination signal output terminal 309 .

FIG. 4 is a block diagram for explaining the third embodiment.

An audio waveform input signal band-limited to, for example, 3400 Hz is supplied to an A\D converter 402 via a waveform input terminal 401 . The A/D converter 402 samples the audio waveform input signal at, for example, 8000 Hz and outputs it to the autocorrelation number measuring device 403. The autocorrelation coefficient measuring device 403 calculates the ratio δ ₁ of the autocorrelation coefficient at a delay time relative to one sampling period of the sampled audio waveform input signal, for example, 1/8000 SEC, to the autocorrelation coefficient at zero delay. measure.

Note that, as is well known, the above δ ₁ coincides with the first-order K parameter K ₁ .

In addition, the autocorrelation coefficient measuring device 403 calculates the ratio δ _MAX of the maximum value of the autocorrelation coefficient to the autocorrelation coefficient at zero delay in a range of delay times in which pitch period delay times are considered to be approximately distributed, for example, from 2mSEC to 18mSEC. measure. Furthermore, the autocorrelation coefficient measuring device 403 sends the δ ₁ to a log area ratio converter 404 and sends the δ _MAX to a nonlinear converter 405.
Output to each. Log area ratio converter 4
04 converts δ ₁ supplied from the autocorrelation coefficient measuring device 403 into a first-order log area ratio L 1 as a first-order K parameter K ₁ and outputs _{the L 1 to} the voiced/unvoiced discriminator 406 . The nonlinear converter 405 receives the δ _MA supplied from the autocorrelation coefficient measuring device 403.
For example, _X is converted to δ' _MAX by the following formula δ' _MAX = aδ _MAX / (b-cδ _MAX ), and the δ' _MAX is output to the voiced/unvoiced discriminator 406. However, a, b, and c are constants. The voiced/unvoiced discriminator 406 performs voiced/unvoiced discrimination using the following formula T ₁ L ₁ +T ₂ δ _MAX from the first-order log area ratio L ₁ and δ _MAX after nonlinear transformation, that is, δ ′ _MAX .

Note that T ₁ and T ₂ above are constants. Furthermore, the voiced/unvoiced discriminator 406 outputs the voiced/unvoiced discrimination result to the voiced/unvoiced discrimination signal output terminal 407 as a voiced/unvoiced discrimination signal. It is clear that in the third embodiment, a configuration in which the nonlinear converter 405 in FIG. 4 is removed can be implemented.

FIG. 5 is a block diagram for explaining the fourth embodiment.

An audio waveform input signal band-limited to, for example, 3400 Hz is supplied to an A/D converter 502 via a waveform input terminal 501 . The A/D converter 502 samples the audio waveform input signal at, for example, 8000Hz and converts it into K
Parameter extractor 503 and autocorrelation coefficient measuring device 5
Output to 04. A K parameter extractor 503 extracts the first and second order from the sampled audio waveform input signal.
The following K parameters, _K1 and _K2 , are extracted and output to the log area ratio converter 505.

When the log area ratio converter 505 is supplied from the K parameter extractor 503, the first-order and second-order K
Convert the parameters, namely K ₁ and K ₂ , into linear and quadratic log area ratios L ₁ and L ₂ , and calculate the above L ₁ and
L ₂ is output to the voiced/unvoiced discriminator 507. The autocorrelation coefficient measuring device 504 measures the ratio δ _MAX of the maximum value of the autocorrelation coefficient at a delay time in the range of 2 mSEC to 18 mSEC from the sampled audio waveform input signal to the autocorrelation coefficient at zero delay, The δ _MAX is output to a nonlinear converter 506 .

Nonlinear converter 506 is autocorrelation coefficient measuring device 50
The δ _MAX supplied from 4 is converted into δ ′ MAX by the following formula, for example, δ ′ _MAX = a δ _MAX / (b−c δ _MAX ), and _{the δ ′ MAX is} output to the voiced/unvoiced discriminator 507 . However, a, b, and c are constants. The voiced/unvoiced discriminator 507 uses the following formula S ₁ L 1 +S ₂ L 2 + _{S 3 δ} ′ from the first-order log area ratio L ₁ , the second-order log area ratio L ₂ , _and δ _MAX after nonlinear _{transformation} , that is, δ ′ MAX. Voiced/unvoiced discrimination is performed using _MAX .

Note that the above S ₁ , S ₂ , and S ₃ are constants. Furthermore, the voiced/unvoiced discrimination result is outputted to the voiced/unvoiced discrimination signal output terminal 508 as a voiced/unvoiced discrimination signal. It is clear that a configuration in which the nonlinear converter 506 in FIG. 5 is removed can be implemented in the fourth embodiment.

FIG. 6 is a block diagram for explaining another configuration example in the fifth embodiment.

An audio waveform input signal band-limited to, for example, 3400 Hz is supplied to an A/D converter 602 via a waveform input terminal 601 . The A/D converter 602 samples the audio waveform input signal at, for example, 8000Hz and converts it into K
Parameter extractor 603 and autocorrelation coefficient measuring device 6
Output to 04. A K parameter extractor 603 extracts K parameters up to the Nth order from the sampled audio waveform input signal, and among the K parameters, the first and second order K parameters are sent to a log area ratio converter 605. The K parameters from the 3rd order to the Nth order are output to the voiced/unvoiced discriminator 607, respectively. However, N is an integer of 3 or more.

The log area ratio converter 605 converts the primary and secondary K parameters, that is, K ₁ and K ₂ , supplied from the K parameter extractor 603 into primary and secondary log area ratios L ₁ and L _2. Convert L ₁ and L ₂
is output to the voiced/unvoiced discriminator 607. The autocorrelation coefficient measuring device 604 measures the so-called δ _MAX extracted from the sampled audio waveform input signal at a delay time in the range of, for example, 2 mSEC to 18 mSEC, and outputs the δ _MAX to the nonlinear converter 606 . The nonlinear converter 606 converts the δ _MAX to δ ′ MAX using the following formula, for example, δ ′ _MAX = aδ _MAX / (b−c δ _MAX ), and outputs _{the δ ′ MAX to} the voiced/unvoiced discriminator 607 . However, a, b, and c are constants.

The voiced/unvoiced discriminator 607 uses the first order log area ratio L ₁ , the second order log area ratio L ₂ , and the third order to N
From the K parameters K ₃ , ......, K _N up to the next, and δ _MAX after nonlinear transformation, that is, δ ′ _MAX , the following formula is used to determine voiced/unvoiced. In addition, _Q1 above...
...Q _N+1 is a constant. Furthermore, voiced/unvoiced discriminator 6
07 outputs the voiced/unvoiced discrimination result to the voiced/unvoiced discrimination signal output terminal 608 as a voiced/unvoiced discrimination signal. It is clear that a configuration in which the nonlinear converter 606 in FIG. 6 is removed from the fifth embodiment described above can be implemented.

FIG. 7 is a block diagram for explaining another example of the configuration of the fifth embodiment.

An audio waveform input signal band-limited to, for example, 3400 Hz is supplied to an A/D converter 702 via a waveform input terminal 701 . The A/D converter 702 samples the audio waveform input signal at, for example, 8000Hz and converts it into K
Parameter extractor 703 and autocorrelation coefficient measuring device 7
Output to 04.

The K parameter extractor 703 is an analyzer that has means for measuring the normalized prediction residual power from the sampled audio waveform input signal, and has a function of controlling the order of the K parameter based on the measurement result. K
The parameter extractor 703 analyzes the sampled audio waveform input signal, and extracts K parameters of multiple orders until the normalized predicted residual power becomes equal to or less than a preset value. Usually, there are restrictions such as Furthermore, it is empirically known that if an appropriate set value of the normalized prediction residual power is selected, the minimum order is 2.

Furthermore, the K parameter extractor 703 extracts the extracted 1
The next and second-order K parameters are output to the K parameter transmission line 1,705, and the analytical order signal indicating the order N1 of the K parameter is output to the analytical order signal transmission line 707. If N1 is 3 or more, the K parameter extractor 703 further extracts one or more K parameters from the 3rd order to the N1 order to the K parameter transmission paths 2, 70.
Output to 6. The log area ratio converter 708 converts the primary and secondary K parameters supplied via the K parameter transmission lines 1 and 705 into log area ratios, respectively, and outputs the conversion results to the log area transmission line 709. Autocorrelation coefficient measuring device 704
is the sampled audio waveform input signal supplied from the A/D converter 702, for example, from 2 mSEC to 18
The so-called δ _MAX extracted at the delay time in the range of mSEC is measured, and the δ _MAX is converted to δ ′ _MAX by the following formula, for example, δ ′ _MAX = aδ _MAX / (b−cδ _MAX ), and the δ _MAX is δ transmission line 71
Output to 1. The voiced/unvoiced discriminator 712 is controlled by an analysis order signal indicating the order N1 of the K parameter supplied via the analysis order signal transmission line 707. When N1 is 2, the voiced/unvoiced discriminator 712 sets the primary log area ratio supplied via the log area ratio transmission path 709 as L ₁ , the secondary log area ratio as L ₂ , and the δ transmission path 711 Voiced/unvoiced discrimination is performed using the following formula S ₁ L ₁ +S ₂ L ₂ +S ₃ δ ' _MAX , with the value supplied via .delta.' _MAX . Note that the above S ₁ ,
S ₂ and S ₃ are constants.

In addition, when N1 is 3 or more, the above-mentioned L ₁ , L ₂ , δ
' K parameters from 3rd order to N1 order supplied via _MAX and K parameter transmission line 2,706
Using K ₃ ,...K _N1 , the following formula is used to determine voiced/unvoiced.

Note that the above U ^N1 _j (j=1, N1+1) is a constant. Further, the voiced/unvoiced discriminator 712 outputs the voiced/unvoiced discrimination result to the voiced/unvoiced discrimination signal output terminal 713 as a voiced/unvoiced discrimination signal. It is clear that a configuration in which the nonlinear converter 710 in FIG. 7 is removed can be implemented in the fifth embodiment.

The above-described embodiments can be summarized as follows. That is, (1) In a voiced/unvoiced discriminator for discriminating whether speech is voiced or unvoiced, the logarithm of the primary vocal tract cross-sectional area ratio and the logarithmic value of the secondary vocal tract cross-sectional area ratio are used as variables. (2) A voiced/unvoiced discriminator having means for discriminating voiced/unvoiced speech using a linear discriminant; , the logarithm of the second-order vocal tract cross-sectional area ratio, and the third-order or multiple partial autocorrelation coefficients after the third-order are used as variables, or the first-order The order is determined from the logarithm of the vocal tract cross-sectional area ratio of the second order, the logarithm of the vocal tract cross-sectional area ratio of the second order, and the normalized predicted residual power value, and multiple orders excluding the first order and the second order. A voiced/unvoiced discriminator having means for discriminating whether speech is voiced or unvoiced using multiple types of linear discriminants using partial autocorrelation coefficients as variables; In the device, the logarithm of the first-order vocal tract cross-sectional area ratio and the maximum value of the autocorrelation coefficient within a finite delay time range excluding the vicinity of zero delay time, or the nonlinear transformation value of the maximum value. (4) A voiced/unvoiced discriminator having means for discriminating whether speech is voiced or unvoiced using a linear discriminant as a variable; The logarithmic value of the area ratio, the logarithmic value of the quadratic vocal tract cross-sectional area ratio, and the maximum value of the autocorrelation coefficient in a finite delay time range excluding the vicinity of zero delay time, or a nonlinear conversion value of the maximum value. (5) A voiced/unvoiced discriminator having a means for discriminating whether speech is voiced or unvoiced using a linear discriminant using a linear discriminant as a variable; The logarithm of the vocal tract cross-sectional area ratio, the logarithmic value of the quadratic vocal tract cross-sectional area ratio, and the maximum value of the autocorrelation coefficient within a finite delay time range excluding the vicinity of zero delay time, or the nonlinearity of the maximum value. Using one type of linear discriminant with variables consisting of three variables consisting of the converted value and a plurality of cubic or post-cubic partial autocorrelation coefficients, or using the three variables and the normalized predicted residual It has means for determining whether speech is voiced or unvoiced using various types of linear discriminant formulas whose order is determined from the power value and whose variables are partial autocorrelation coefficients of multiple orders excluding first order and second order. This is a voiced/unvoiced discrimination device.

As explained above, among the parameters used for voiced/unvoiced discrimination, the distribution of the parameters is large, and if used as variables in the linear discriminant, the degree of separation between voiced and unvoiced is poor.
The next K parameter is converted in advance to a parameter that has the effect of compressing the bias of the distribution, such as the log area ratio, and is used as a parameter of the linear discriminant after conversion. Therefore, there is an effect that the voiced/unvoiced discrimination accuracy is significantly improved compared to the conventional voiced/unvoiced discrimination device.

[Brief explanation of the drawing]

1 to 7 are schematic block diagrams showing an embodiment of a voiced/unvoiced discriminating device according to the present invention. 102, 202, 302, 503, 603, 7
03...K parameter extractor, 402, 502,
602,702...A/D converter, 103,20
3,306,404,505,605,708...
...Log area ratio converter, 104, 204, 3
08,406,507,607,712...Voiced/unvoiced discriminator, 403,504,604,704...
...Autocorrelation coefficient measuring device, 405, 506, 60
6,710...Nonlinear converter.

Claims (1)

[Claims] 1. Voiced/unvoiced discrimination that discriminates voiced/unvoiced by comparing the value of a discriminant having a predetermined coefficient and using a predetermined characteristic parameter of an input audio signal as a variable with a predetermined threshold value. In the apparatus, at least one of the variables is a partial autocorrelation coefficient (K parameter) up to a predetermined order, and this partial autocorrelation coefficient is converted into a log area ratio and used as a variable in the discriminant. A voiced/unvoiced discrimination device. 2. The voiced/unvoiced discrimination device according to claim 1, wherein the predetermined order is 1. 3. The voiced/unvoiced discrimination device according to claim 1, wherein the predetermined order is 2.