JP2001022368A - Speech discrimination device and method therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a speech discrimination device capable of discriminating speech independently of the level of background noise with less of discrimination error. SOLUTION: This device is provided with a noise level discriminating means 5 which divides an input speech frame into plural subframes and obtains noise level for every subframe based on the comparison between input speech power and noise power, a threshold level calculating means 9 which receives the noise level obtained for every subframe and obtains a threshold to detect a speech interval for every subframe, and a collating means 11 which determines whether the input speech frame is the speech interval or not by using the threshold obtained for every subframe.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) used for digitally transmitting or storing speech, and to a discriminating method thereof.

[0002]

2. Description of the Related Art A conventional voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) of this kind is disclosed in, for example, Japanese Patent Application Laid-Open No. 61-2780.
In the above-mentioned apparatus, the sum of the low-order terms of the cepstrum is used as a discrimination parameter for voiced and unvoiced sounds, and the discrimination result is a binary value of a voiced sound and an unvoiced sound.

FIG. 6 shows a conventional voiced sound shown in the above document.
FIG. 7 is a block diagram illustrating a configuration of an unvoiced sound discriminating apparatus (speech discriminating apparatus). FIG. 7 is a diagram illustrating a distribution of discriminating parameters 20 of the voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) in FIG.
In the figure, 18 is a cepstrum, 19 is an addition circuit, 20 is a discrimination parameter, 21 is a threshold value comparison circuit, and 22 is a discrimination result.

The operation of the voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) shown in FIG. 6 will be described below with reference to FIGS. First, the addition circuit 19 receives the input cepstrum 18.
, And outputs the sum as a discrimination parameter 20. The threshold comparing circuit 21 determines that the input discrimination parameter 20 is unvoiced when it is less than a predetermined fixed threshold, and that it is voiced when it is not less than the fixed threshold, and outputs a discrimination result 22.

FIG. 7 shows the case where the power of the background noise is so large as to be insignificant as compared with the power of the speech signal, that is, the noise level is high, and the case where the power of the background noise is so small as to be negligible as compared with the power of the speech signal, that is, the noise. 9 shows a model of the distribution of the discrimination parameters 20 when the level is low.
In the figure, a curve A is an unvoiced sound when the noise level is low, and a curve C
Is the unvoiced sound when the noise level is high, curve D is the distribution of curves A and C combined, and curve B is the distribution of voiced sounds. The distribution of the discrimination parameter 20 in the voiced sound does not change significantly depending on the level of the noise level. Now, when the fixed threshold for optimally separating the unvoiced sound A and the voiced sound B is E1 when the noise level is low, the number of erroneous determinations for discriminating the unvoiced sound C from the voiced sound B when the noise level is high increases. on the other hand,
When the fixed threshold for optimally separating the unvoiced sound C and the voiced sound B is E2 when the noise level is high, the number of errors in determining the voiced sound B as the unvoiced sound A increases when the noise level is low. Further, if E3 is a fixed threshold value for optimally separating the unvoiced sound D and the voiced sound B, when the noise level is low, E1 is
It is clear that the discrimination error increases when the noise level is large as compared with the case where E2 is used as the fixed threshold. Also,
Regardless of which threshold value is used, when the discrimination parameter 20 is a value near the threshold value, there are many discrimination errors and the reliability is low.

[0006]

The conventional voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) is configured as described above, and since only the sum of the low-order terms of the cepstrum is used as the discriminating parameter, the discriminating parameter is When the value is close to the discrimination threshold, there are many discrimination errors, and when a discrimination threshold for discriminating voiced and unvoiced sounds is set, the discrimination error increases when the speech has a noise level different from the background noise level assumed. There are issues.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problems, and an object of the present invention is to provide a voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) which does not depend on the level of the background noise level and has few discrimination errors, and a discriminating method thereof. The purpose is to get.

[0008]

A speech discriminating apparatus according to the present invention divides an input speech frame into a plurality of subframes, and obtains a noise level for each subframe based on a comparison between the input speech power and the noise power. Noise level determination means and a noise level determined for each subframe are input,
Threshold calculation means for determining a threshold for detecting a voice section for each subframe, and matching means for determining whether an input voice frame is a voice section using each threshold obtained for each subframe And characterized in that:

The noise level determination means outputs a binarized noise level, and the threshold value calculation means stores two constants corresponding to each value of the binary determination in advance, and stores the two constants. , A constant corresponding to the binarized noise level value is selected, and a threshold value is calculated using the selected constant.

The noise level judging means outputs three or more multi-valued noise levels, and the threshold value calculating means previously stores a plurality of constants corresponding to each value of the multi-value judgment. A constant corresponding to the multilevel noise level value is selected from a plurality of constants, and a threshold value is calculated using the selected constant.

[0011] A speech discriminating method according to the present invention includes: a noise level discriminating step of dividing an input speech frame into a plurality of subframes and obtaining a noise level for each subframe based on a comparison between the input speech power and the noise power; By inputting the noise level obtained for each sub-frame, and for each sub-frame, a threshold calculation step of obtaining a threshold for detecting a voice section, and using each threshold obtained for each sub-frame, A collation step of determining whether or not the section is a voice section.

The noise level determination step outputs a binarized noise level. The threshold value calculation step stores two constants corresponding to each value of the binary determination in advance, and stores the two constants. , A constant corresponding to the binarized noise level value is selected, and a threshold value is calculated using the selected constant.

The noise level determining step outputs a multi-valued noise level of 3 or more, and the threshold value calculating step stores a plurality of constants corresponding to each value of the multi-level determination in advance. A constant corresponding to the multilevel noise level value is selected from a plurality of constants, and a threshold value is calculated using the selected constant.

[0014]

In the following embodiment of the present invention, in a voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) for discriminating voiced sound and unvoiced sound of a voice signal, a discrimination parameter value obtained by analyzing an input voice frame is determined. Based on the voiced sound, unvoiced sound, and a plurality of different discriminating conditions for discriminating between noisy conditions, a power, a normalized autocorrelation peak value, a number of zero crossings, First-order linear prediction coefficients, the result of discriminating past speech frames,
By using at least one of the lower-order terms of the cepstrum as a discrimination parameter, collating with a predetermined threshold, and performing the above-mentioned voiced sound, unvoiced sound, and a collation unit that outputs a result of the discrimination between voiced, voiced sound, unvoiced sound, It is possible to reduce the erroneous determination of the silent determination.

In an embodiment described below of the present invention, in a voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) for discriminating between a voiced sound and an unvoiced sound of a voice signal, a discrimination parameter obtained by analyzing an input voice frame is a voiced sound. , Unvoiced sound, collate with the discrimination conditions of discrimination, and output voiced sound, unvoiced sound, or no sound as a discrimination result if it falls into any of the categories, and if it does not fall into any of the categories, voiced sound characteristics The matching means operates to output a determination result as a quasi-voiced sound when it has a sound, and as a quasi-silent when it has a silent feature, so that an intermediate quasi-voiced sound, a quasi-voiced sound, Silence can be output as a determination result.

In an embodiment described below of the present invention, in a voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) for discriminating voiced sound and unvoiced sound of a voice signal, a background noise level of an input voice frame is obtained and output as a noise level. The noise level determining means operates to select a voiced sound, an unvoiced sound, or a non-voiced discriminating condition from a plurality of different discriminating conditions based on the value of the noise level, and analyze a frame of the input voice. Voiced sound, unvoiced sound, by operating the matching means to perform a silent determination by comparing the obtained determination parameter and a predetermined threshold, by the value of the noise level,
The threshold for voiced sound, unvoiced sound, and silence discrimination can be changed.

In an embodiment described below of the present invention, the noise level determining means determines at least one of a discrimination result, power, and a peak value of normalized autocorrelation between an input speech frame and a past speech frame as a discrimination parameter. Used as
By comparing with a predetermined threshold, a section corresponding to a voiced section and a silent section is determined for the input voice frame and the past voice frame, and the average of the power of the voiced section and the silent section is calculated, and By comparing the voiced sound average power and the silent average power as the voiced sound average power and the silent average power, the noise level can be determined and output.

In the embodiment described below of the present invention, the noise level determining means updates the average of the power of the frame of the input voice larger than the average power of the past voice frame for each frame of the input voice. Calculated as voiced sound average power, and calculated while updating the average power of frames of input voice frames smaller than the average power of past voice frames for each frame of input voice as silent average power, By comparing the voiced sound average power and the silent average power,
The level of the noise level can be determined and output.

[0019]

[Embodiment 1] FIG. 1 is a configuration block diagram showing a voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) and a discriminating method according to a first embodiment of the present invention. In FIG. 1, 1 is power as a discrimination parameter obtained by analyzing a frame of an input voice, 2 is a peak value of normalized autocorrelation, 3 is the number of zero crossings, 4 is a first-order linear prediction coefficient, and 5 is noise. Level determining means, 6 is noise level, 7 is silent average power, 8 is voiced sound average power, 9 is threshold calculating means, 10 is power determination threshold, 11 is matching means, 12 is determination result, 13 is register, 14 is The power of the past speech frame, 15 is the normalized autocorrelation peak value of the past speech frame, 16 is the result of the discrimination of the past speech frame, and 17 is the lower order term of the cepstrum.

The operation of the voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) shown in FIG. 1 and the operation of the discriminating method will be described below with reference to the drawings. First, the noise level determination means 5 calculates the normalized auto-correlation peak value 2 of the frame of the input voice, the past normalized auto-correlation peak value 15 stored in the register 13, and the determination result 16 of the past voice frame. On the other hand, a predetermined condition for determining a silent section (for example, the normalized autocorrelation peak value 2 is set to a predetermined threshold P1 for 10 consecutive frames)
Is determined from the power 1 of the input speech frame and the power 14 of the past speech frame, and is output as the silence average power 7. On the other hand, the condition for determining the average power voiced sound section of the section satisfying the voiced sound section determination condition (for example, the peak value of the normalized autocorrelation is equal to or greater than the predetermined threshold value P2 for five consecutive frames) is defined as the input voice frame. , And the power 14 of the past speech frame, and outputs as the voiced sound average power 8. The difference between the silent average power 7 and the voiced average power 8 is equal to a predetermined threshold D1.
If it is smaller, it is determined that the noise level is high, and the noise level 6 output from the noise level determination means 5 is “1”.
If the noise level is larger than the predetermined threshold value D1, it is determined that the noise level is low.
"0" is output as the noise level 6 which is the output of.

Next, in the threshold value calculating means 9, when the input noise level 6 is “0”, the noise level is input from the noise level determining means 5 according to equation (1). Averaged silent power 7 is PUV, voiced average power 8
Is used as the PV, a threshold for determining the power of the frame of the input voice is determined, and the power determination threshold 10 output from the threshold calculator 9 shown in the equations (1) and (2) is compared with the comparator 11.
To send to.

[0022]

(Equation 1)

Here, TH1, TH2, and TH3 are power determination thresholds, PUV is a silent average power, and PV is a voiced average power.

Next, the collating means 11 outputs the power 1 of the input speech frame, the peak value 2 of the normalized autocorrelation, the number of zero crossings 3, the first-order linear prediction coefficient 4, the sum 17 of the low-order terms of the cepstrum, the noise Noise level 6 from level determination means 5, power determination threshold value 10 from threshold value calculation means 9, register 13
As an input, the determination result 16 of the past speech frame from is input. For example, first, any of the following categories a or b to e is selected. In the case of a, that is, when any of the logical products of the equation (3) of the determination condition is satisfied, the voice signal is determined to be unvoiced and the determination result 12 is output. In the case of b to e, be to e
Which of the sections is selected is determined by the threshold value calculating means 9.
Is determined based on the magnitude relationship between the value of TH, which is the power discrimination threshold value 10, and the value of POW, which is power 1 of the input speech frame. In the above, when a can be determined to be unvoiced,
b corresponds to a case where the probability of a voiced sound is high, c corresponds to a case where the probability of a voiced sound is slightly high, d corresponds to a case where the probability of silence is slightly high, and e corresponds to a case where the probability of silence is high.

Next, FIG. 2, FIG. 3, FIG. 4, FIG.
According to the determination flow, any of voiced sound, semi-voiced sound, semi-silent sound, and silent sound is determined, and a determination result 12 is output. In addition,
In the categories a to e, the discriminating conditions that can be distinguished from voiced sound, unvoiced sound, and no sound are different from each other. Therefore, the discriminating conditions need to be set individually in each section.
This determination condition is experimentally determined. Here, quasi-voiced sound is defined as the case where some of the conditions that are determined to be voiced are missing, and quasi-silent is defined as the case where some of the conditions that are determined to be silent are missing. I do.

[0026]

(Equation 2)

B: When POW> TH1, the determination is made according to FIG. c: When TH1 ≧ POW> TH2, the determination is made according to FIG. d: When TH2 ≧ POW> TH3, the determination is made according to FIG. e: In the case of POW ≦ TH3, determination is made according to FIG. However, the discriminant of the above-mentioned section a, FIGS.
In 4 and 5, TH1, TH2, and TH3 are power discrimination thresholds 10 (however, TH1>TH2> TH3), PUV is silent average power 7, PV is voiced average power 8, POW.
Is the power 1, AC is the peak value 2 of the normalized autocorrelation, C is the sum 17 of the low-order terms of the cepstrum, CMIN is the determination threshold of the sum of the low-order terms of the cepstrum, Z is the number of zero crossings 3, and A1 is the first order. Linear prediction coefficient 4, NL: noise level 6, VO: past speech frame discrimination result 16, T1, T11, T12, T2, T21,
T22, T23, T24, T3, T31, T32, T33, T34, T4, T41, T42, T43, T44
All represent fixed thresholds.

Next, the register 13 stores the power 1 of the input speech frame, the peak value 2 of the normalized autocorrelation, the accumulated power of the past 10 frames, the normalized autocorrelation peak value, and the result of the discrimination by the matching means. Update.

Embodiment 2 FIG. In the first embodiment, the power determination threshold value is determined based on the silent average power and the voiced average power. However, the power determination threshold value is determined based on, for example, Expression (4) from the maximum value of the power of the past voice frame. Is also possible.

[0030]

(Equation 3)

However, in the equation (4), TH1, TH
2, TH3 are power determination thresholds, and Pmax is, for example, the maximum value of power in the past 30 frames. It is also possible to use the maximum value of the power in the past voice frame to correct the power discrimination threshold calculated from the average unvoiced power and the average voiced voice power, or to correct the voiced, unvoiced, and non-voiced discrimination results. is there.

Embodiment 3 FIG. In the first embodiment, silence is determined based on the peak value of the normalized auto-correlation function, the determination result of the past speech frame, and the noise level according to FIG. 2. For example, it is determined that no sound has occurred in the past by using a low-order term of the cepstrum coefficient. It is also possible to obtain a spectrum outline of the obtained frame and determine silence based on the spectrum outline and the distance of the spectrum of the frame of the input voice.

Embodiment 4 FIG. In the first embodiment, the discrimination is performed using the discrimination parameters obtained by analyzing each frame of the input voice. However, the frame of the input voice is divided into a plurality of subframes, and the analysis is performed for each subframe. It is also possible to make a determination using the parameters obtained or to correct the result of the determination.

Embodiment 5 FIG. In the first embodiment, the power of the frame of the input speech is used as the discrimination parameter to classify the discrimination conditions. However, the sum of the low-order terms of the cepstrum may be used.

Embodiment 6 FIG. In the first embodiment, the noise level is set to 2
Although the value is determined, it can be a multi-valued or continuous numerical value.

Embodiment 7 FIG. In the first embodiment, the maximum amplitude value in the frame can be included in the determination parameter.

[0037]

As described above, according to the present invention, there is little discrimination error near the discrimination threshold of a discrimination parameter obtained by analyzing a frame of an input voice, and it does not depend on the background noise level. A voiced / unvoiced sound discriminating device (speech discriminating device) with few discrimination errors and a discriminating method thereof can be obtained. Further, it is possible to obtain a voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) capable of discriminating a voice frame in an intermediate state having both voiced and unvoiced features, and a discriminating method thereof.

[Brief description of the drawings]

FIG. 1 is a configuration block diagram of a voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) according to a first embodiment of the present invention.

FIG. 2 is a diagram exemplifying discrimination conditions of a voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) in FIG. 1;

FIG. 3 is a diagram exemplifying discrimination conditions of a voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) in FIG. 1;

FIG. 4 is a diagram exemplifying discrimination conditions of a voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) of FIG. 1;

FIG. 5 is a diagram exemplifying discrimination conditions of a voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus) in FIG. 1;

FIG. 6 is a configuration diagram showing a conventional voiced / unvoiced sound discriminating apparatus (speech discriminating apparatus).

FIG. 7 is a diagram showing distribution of discrimination parameters of the voiced / unvoiced sound discrimination device (speech discrimination device) of FIG. 6;

[Explanation of symbols]

1 Input speech frame power, 2 Peak value of normalized autocorrelation, 3 Zero crossing number, 4th linear prediction coefficient, 5 Noise level judgment means, 6 Noise level, 7 Silent average power, 8 Voiced sound average Power, 9 threshold calculating means, 10 power discriminating threshold, 11 matching means, 12 discrimination result, 13 register, 14 power of past speech frame, 15 normalized autocorrelation peak value of past speech frame, 16 past speech frame Determination result, 17
Low order term of cepstrum, 18 cepstrum, 19 addition circuit, 20 discrimination parameter, 21 threshold value comparison circuit,
2 Determination result.

Claims (6)

[Claims] An input speech frame is divided into a plurality of subframes, and a noise level determining means for obtaining a noise level for each subframe based on a comparison between the input speech power and the noise power; A threshold level calculating means for inputting a noise level and determining a threshold for detecting a voice section for each sub-frame; and using each threshold determined for each sub-frame to determine whether or not the input voice frame is a voice section. A voice discriminating device, comprising: a matching unit for determining 2. The noise level judging means outputs a binarized noise level, and the threshold value calculating means outputs a binarized noise level corresponding to each value of the binary judgment.
2. A method according to claim 1, wherein two constants are stored in advance, a constant corresponding to a binarized noise level value is selected from the two constants, and a threshold value is calculated using the selected constant. The voice discriminating apparatus described in the above. 3. The noise level judging means outputs three or more multi-valued noise levels, and the threshold value calculating means stores a plurality of constants corresponding to each value of the multi-value judgment in advance. 2. The speech discriminating apparatus according to claim 1, wherein a constant corresponding to the multilevel noise level value is selected from the plurality of constants, and a threshold value is calculated using the selected constant. 4. A noise level determining step of dividing an input voice frame into a plurality of subframes and determining a noise level for each subframe based on a comparison between the input voice power and the noise power; A threshold calculation step of inputting a noise level and calculating a threshold for detecting a voice section for each subframe; and using each threshold calculated for each subframe to determine whether an input voice frame is a voice section. And a collating step for determining the sound. 5. The noise level determining step outputs a binarized noise level, and the threshold value calculating step includes a step of calculating a binary noise level corresponding to each value of the binary determination.
5. A method according to claim 4, wherein two constants are stored in advance, a constant corresponding to a binarized noise level value is selected from the two constants, and a threshold value is calculated using the selected constant. The described sound discrimination method. 6. The noise level determining step outputs a multi-level noise level of 3 or more, and the threshold calculating step stores a plurality of constants corresponding to each value of the multi-level determination in advance. 5. The voice discriminating method according to claim 4, wherein a constant corresponding to a multilevel noise level value is selected from the plurality of constants, and a threshold value is calculated using the selected constant.