JPH09160594A - Noise removing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To remove noise components taking into consideration the noise characteristics to be superposed on voice and music or the like by multiplying the noise pattern by noise removal factor, and reducing the pattern from the pattern of the voice section judged as voice. SOLUTION: A noise pattern preparation part 3 obtains noise pattern N(f) in the noise pattern preparation section which is judged not to be voice section. A dispersion calculating part 4 obtains the dispersion Φi (f) based on the square error Θ between the noise pattern Ni-l (f) at a time in the noise pattern preparation section and the analytical result Xi at the present time. A noise removal factor control parameter calculating part 5 calculates the removal factor control parameter D to control the noise removal factor α as a mean value for the time and every frequency bandwidth of the dispersion ΦTS-1 (f). In addition, a removal factor calculating part 6 obtains the noise removal factor αbased on the value of the removal factor control parameter D. A noise removing part 7 removes the noise using the noise removal factor α.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise removing device for improving voice recognition performance by removing a noise component from a voice uttered in a noise environment.

[0002]

2. Description of the Related Art In order to put a system using voice recognition technology into practical use, it is necessary to improve the voice recognition performance in a noisy environment such as a home or a store.

Conventionally, in order to improve the recognition performance in such a noisy environment, a spectral subtraction method for removing a noise component from a speech on which noise is superimposed has been used.

The basic concept of the present spectral subtraction method will be described below with reference to FIG.

FIG. 6 shows the result of (a) bandpass filter analysis performed on a voice waveform on which noise is superimposed.
(B) shows the relationship between the estimated noise spectrum and (c) the noise-free speech spectrum.

The spectral subtraction method is to estimate the speech spectrum before the noise is superimposed by subtracting the noise spectrum from the speech spectrum on which the noise is superimposed.

For example, consider the case where noise n (f) is superimposed on the spectrum s (f) of a true voice without noise. In this case, the observable spectrum is only the spectrum X (f) in which noise is superimposed on the voice. f represents a frequency, and corresponds to the channel number of the bandpass filter here. Also, X (f) = s (f) + n (f) ((1.1)
Formula, and FIG. 6 (a).

[0008]

[Equation 1]

When the noise spectrum n (f) is known, the true speech spectrum s (f) can be obtained by subtracting n (f) from X (f) ((1.
See formula 2)).

However, in most cases, it is impossible to know the superimposed noise spectrum n (f).
As the noise spectrum to be subtracted, the estimated value N (f) of the noise spectrum instead of n (f) is subtracted from the speech spectrum X (f) on which noise is superimposed, so that the spectrum s (f) of the true speech is estimated spectrum S (f). f) will be obtained (equation (1.3) and FIGS. 6 (b) and 6 (c)).
reference).

Therefore, as the estimated noise spectrum N (f) is closer to the true noise spectrum n (f), the estimated spectrum S (f) closer to the true speech spectrum s (f) can be obtained.

By the way, FIG. 7 is a schematic block diagram of a conventional noise eliminator.

FIG. 8 is a diagram showing the relationship between the voice section and the noise pattern creating section in the conventional noise removing apparatus on the time axis. Further, FIG. 8 is a simulation diagram in which the frequency axis of FIG. 6A is expressed in one dimension.

An outline of a conventional speech recognition apparatus using the spectral subtraction method will be described below with reference to FIGS. 7 and 8.

In FIG. 7, reference numeral 1 is a voice analysis unit for analyzing a voice on which noise is superimposed, and 2 is an analysis result X _{i of the} voice analysis unit 1.
A voice section determination unit that determines the voice section by detecting the start time T _s and the end time T _e of the voice using the power and the amount of change in the spectrum obtained from (f). Here, i represents time.

3 is a voice section in the voice section determination unit 2.
Start time T of the section determined to exist _sPrevious analysis results,
In other words, it is estimated from the analysis result of the noise section before the speech section.
A noise pattern creating unit that creates a constant noise pattern N (f).
You.

Basically, the noise interval (from time i = 1 to T
_s-1 up) analysis results of X _i mean noise pattern N of (f)
It can be used as However, when considering a case where the characteristics of noise change gently with time, etc., the spectrum X _i (f) is closer in time to the start time T _{s of the} voice than the past noise pattern N _i-1 (f). It may be better to attach importance to.

Therefore, to create the noise pattern N (f),
Number 2

[0019]

(Equation 2)

In some cases, the weighting factors β and γ are used to control the reliability of the past noise pattern N _i-1 (f) and the current (time i) spectrum X _i (f) as shown in FIG. is there.

For example, in the equation 2, as the value of γ is increased, the noise pattern N _i (f) is weighted (reliable) on the current (time i) spectrum X _i (f). .

The reference numeral 10 subtracts the noise pattern N (f) from the analysis result X _i (f) of the section (portion from time T _s to T _e ) determined to be the voice section by the voice section determination unit 2 according to the equation 3. This is a noise removal unit that removes noise.

[0023]

(Equation 3)

If the noise is not stationary, the estimation accuracy of the estimated noise pattern N (f) becomes low. If N (f) is subtracted from X _i (f) as it is, the voice component is reduced and the distortion is generated. It may become large.

In such a case, Equation 4

[0026]

(Equation 4)

As shown in, the estimated noise pattern N
In some cases, the removal amount may be reduced by multiplying (f) by the removal coefficient α (α ≦ 1).

For the value of the removal coefficient α, various noises and voices are used, and a representative value is set in advance so that the voice recognition rate and the voice distortion caused by the noise removal are minimized, or it is set as a voice section. A method of making a decision according to the voice power at each time in the determined section has also been proposed.

[0029]

However, as described above, as the noise removal coefficient α, the constant based on the constant determined by preliminary experiments and the voice power at each time is used without considering the characteristics of noise. However, there is a problem that it is difficult to obtain sufficient noise removal performance.

Furthermore, even when considering the characteristics of a certain type of noise, such noise has a stationary frequency region and a non-stationary frequency region. Since the same value is used for all frequency regions of the spectrum, there is a problem that noise cannot be removed with high precision.

[0031]

Therefore, the present invention has been made in view of the above problems, and a pattern of noise created based on a noise section determined to be noise is determined as voice. In a noise removal device that removes noise by subtracting from the pattern of the voice section, the noise pattern is multiplied by the noise removal coefficient α determined based on the variation of the noise component in the noise section, and this pattern is determined as the voice. It is characterized in that it is subtracted from the pattern of the generated voice section.

According to the present invention, the noise section is divided at a certain time t within the noise section, and the noise pattern obtained at the time t is used to determine the noise pattern from the time t + 1 to the time t + N belonging to the noise section. A feature is that the variation of the noise component is obtained and the noise removal coefficient α is determined based on the variation.

The present invention is the noise removing apparatus for removing noise by subtracting a noise pattern created based on a noise section determined to be noise from a pattern of a voice section determined to be speech. It is characterized in that the noise removal coefficient α (f), which is a ratio for reducing the noise pattern for each frequency band, is obtained based on the variation of the noise component value for each frequency band in the noise section.

According to the present invention, the noise section is divided at a certain time t within the noise section, and the noise pattern obtained at the time t is used to determine the noise pattern from the time t + 1 to the time t + N belonging to the noise section. The variation of the noise component for each frequency band is obtained, and the noise removal coefficient α, which is a ratio for reducing the noise pattern for each frequency band based on the variation.
It is characterized by obtaining (f).

The present invention provides the noise removing device for removing noise by subtracting a noise pattern created based on a noise section determined to be noise from a pattern of a speech section determined to be speech. Based on the value of the noise component for each frequency band in the noise section, a noise reduction coefficient α (f), which is a ratio for reducing the noise pattern, is calculated for each frequency band, and the noise removal coefficient α (f) is calculated as described above. It is characterized in that the product of the noise pattern is subtracted from the pattern of the voice section determined as voice.

The variation is obtained from the difference between the noise pattern at time i-1 belonging to the noise section and the input pattern at time i belonging to the noise section.

Further, the variation is the square of the difference between the noise pattern at time i-1 belonging to the noise section and the input pattern at time i belonging to the noise section, which is 2 of the noise pattern at time i-1. It is characterized in that it is obtained from what is normalized by the value of the power.

[0038]

1 is a block diagram of a noise elimination apparatus according to the present invention.

The operation of the noise eliminator according to the first embodiment of the present invention will be described below with reference to FIG.

In the configuration diagram of FIG. 1, the same components as those of the conventional noise eliminator are designated by the same reference numerals.

The configuration of the noise eliminator in the first embodiment of the present invention is different from that of the conventional one in three points.

The first point is that the variation calculation section 4 is provided in connection with the speech analysis section 1 and the noise pattern creation section 3, and the second point is that the variation calculation section 4 is connected with the removal coefficient. That is, the control parameter calculator 5 is provided.

Further, the third point is that a removal coefficient calculation unit 6 is provided in connection with the removal coefficient control parameter calculation unit 5 and the noise removal unit 7.

The voice analysis unit 1 analyzes the input signal. Such a signal has a noise-only portion and a voice portion on which noise is superimposed.

From the analysis result analyzed by the voice analysis unit 1, the voice section determination unit 2 determines the voice section in which the voice exists (time i = T _s
˜T _e ) is detected.

Further, as shown in FIG. 2, the noise pattern creating unit 3 sets a section determined not to be a voice section as a noise pattern creating section, and the noise pattern N is calculated by the recurrence formula shown in the equation 2 in the noise pattern creating section. Find (f).

The variation calculation unit 4 is in the noise pattern creation section.
Noise pattern N at time i-1 of _i-1(F) and the current time
Analysis result X of tick i_iVariation Ψ based on the squared error Θ with_i
(F) is obtained by the recurrence formula of Equation 5 or Equation 6.

[0048]

(Equation 5)

[0049]

(Equation 6)

The noise removal coefficient control parameter calculation unit 5
The removal coefficient control parameter D for controlling the noise removal coefficient α as an average for each time and the frequency Ψ _Ts-1 (f) at the time T _s-1 calculated by the variation calculation unit 4 is calculated by _Equation 7. calculate.

[0051]

(Equation 7)

The removal coefficient calculation unit 6 obtains the noise removal coefficient α based on the value of the removal coefficient control parameter D calculated by the removal coefficient control parameter calculation unit 5.

The noise removing section 7 removes noise according to equation 4 using the obtained noise removing coefficient α.

The second embodiment of the present invention will be described below.

In the second embodiment, instead of the noise removal coefficient α in the first embodiment, the noise removal for each frequency component in which the variation of the noise component value for each frequency band of the noise spectrum is independently considered. Noise removal is performed based on the coefficient α (f).

Further, in the second embodiment of the present invention, only the functions of the removal coefficient control parameter calculation unit 5, the removal coefficient calculation unit 6 and the noise removal unit 7 in the first embodiment are different. The configuration of the noise eliminator in the second embodiment has the same configuration as that in the first embodiment.

Therefore, only the functions of the removal coefficient control parameter calculation unit 5, the removal coefficient calculation unit 6, and the noise removal unit 7 will be described below. (A) Removal Coefficient Control Parameter Calculation Unit 5 The calculation unit 5 includes the variation Ψ _ts-1 at the time T _s-1 calculated by the variation calculation unit 4 according to Equation 5 or Equation 6.
As the time average of (f), the removal coefficient control parameter D (f) for each frequency band is calculated by Equation 8.

[0058]

(Equation 8)

(B) Removal Coefficient Calculation Unit 6 The calculation unit 6 uses the removal coefficient control parameter D (f) value calculated by the removal coefficient control parameter calculation unit 5 to remove the noise removal coefficient α (f) for each frequency band. ). (C) Noise Removal Unit 7 The noise removal unit 7 removes noise according to Equation 9 using the obtained noise removal coefficient α (f).

[0060]

(Equation 9)

The third embodiment of the present invention will be described below.

In the third embodiment of the present invention, only the functions of the variation calculation unit 4 and the removal coefficient control parameter calculation unit 5 in the first embodiment are different, so that the noise in the third embodiment is reduced. The structure of the removing device has the same structure as that in the first embodiment.

Therefore, only the functions of the variation calculator 4 and the removal coefficient control parameter calculator 5 will be described below. (A) Variation Calculation Unit 4 The variation calculation unit 4 uses Equation 10 or Equation 11 instead of Equations 5 and 6 shown in the first embodiment to obtain variation Ψ.
Calculate _i (f).

[0064]

(Equation 10)

[0065]

[Equation 11]

Equations 10 and 11 are variations Ψ
_When updating _i (f), the weighting factors a and b are used to control the reliability of the past variation Ψ _i-1 (f) and the square error Θ of the present (time i).

For example, as the value of b is increased, the variation Ψ _i (f) in which the square error Θ of the current time (time i) is weighted (reliability) is calculated. (B) Removal Coefficient Control Parameter Calculation Unit 5 This calculation unit 5 calculates the time T calculated by the variation calculation unit 4.
_The removal coefficient control parameter D is calculated by Equation 12 as the average of the variation Ψ _{T s-1} (f) in _s-1 for each frequency band.

[0068]

(Equation 12)

The removal coefficient calculation unit 6 finds the noise removal coefficient α on the basis of the value of the removal coefficient control parameter D, and the noise removal unit 7 removes noise according to the equation 4 using the found noise removal coefficient α.

The fourth embodiment of the present invention will be described below.

In the fourth embodiment, instead of the noise removal coefficient α in the third embodiment, the noise removal for each frequency component in which the variation of the noise component value for each frequency band of the noise spectrum is independently considered. Noise removal is performed based on the coefficient α (f).

Further, in the fourth embodiment of the present invention, only the functions of the removal coefficient control parameter calculation unit 5, the removal coefficient calculation unit 6 and the noise removal unit 7 in the third embodiment are different. The configuration of the noise eliminator in the fourth embodiment has the same configuration as that in the third embodiment.

Therefore, only the functions of the removal coefficient control parameter calculation unit 5, the removal coefficient calculation unit 6, and the noise removal unit 7 will be described below. (A) Removal coefficient control parameter calculation unit 5
Variation Ψ _{ts-1 at} time T _s-1 calculated based on
Let (f) be the removal coefficient control parameter D (f) (Equation 1)
3).

[0074]

(Equation 13)

(B) Removal Coefficient Calculation Unit 6 The calculation unit 6 uses the removal coefficient control parameter D (f) value calculated by the removal coefficient control parameter calculation unit 5 to remove the noise removal coefficient α (f ). (C) Noise Removal Unit 7 The noise removal unit 7 removes noise according to Equation 9 using the obtained noise removal coefficient α (f).

The fifth embodiment of the noise eliminator according to the present invention will be described below with reference to FIG.

In the configuration diagram of FIG. 3, the same components as those of the conventional noise eliminator are designated by the same reference numerals.

The difference of the configuration of the noise elimination device in the fifth embodiment of the present invention from that of the conventional one is that it is connected to the voice analysis unit 1, the noise pattern generation unit 3 and the noise elimination unit 7 to calculate the second elimination coefficient. The part 8 is provided.

In the second removal coefficient calculation section 8, the speech section determination section 2 determines the speech section from time T _s to T _{e as} shown in FIG.
If it is determined that the noise pattern N created at the time t (1 ≦ t <T _s ) in the noise pattern creation unit 3
_The noise removal coefficient α that minimizes the error ε between _t (f) and the noise spectrum X _i (f) in the section from time t + 1 to T _s-1 is calculated.

Here, the equation 14 is used to calculate the error ε.

[0081]

[Equation 14]

Alternatively, the noise elimination coefficient α (f) that minimizes the error ε (f) is calculated for each frequency independently according to the equation 15.

[0083]

(Equation 15)

Further, a plurality of points of time t may be provided and the noise removal coefficient α or the noise removal coefficient α (f) that minimizes the average of the errors ε calculated for each time t may be calculated.

The noise removing section 7 removes noise according to the equation 4 or the equation 9 using the removal coefficient α or α (f).

The sixth embodiment of the present invention will be described below in detail.

In the sixth embodiment of the present invention, only the function of the second removal coefficient calculation unit 8 in the fifth embodiment is different. Therefore, the configuration of the noise removal device in the sixth embodiment is The same configuration as that in the fifth embodiment is adopted.

The second removal coefficient calculation unit 8 causes the noise pattern generation unit 3 to perform the determination when the voice section determination unit 2 determines that the voice section is from time T _s to T _e as shown in FIG. Noise pattern N created at time t (1 ≦ t <T _s )
The error ε between _t (f) and the noise spectrum X _i (f) in the section from time t + 1 to T _s-1 is calculated using Equation 16,
According to number 17

[0089]

(Equation 16)

[0090]

[Equation 17]

When the error ε is larger than the threshold Th, the noise removal coefficient α is set to zero. When the error ε is less than the threshold Th, the function F that monotonically decreases the noise removal coefficient α with respect to ε. Set by (ε).

Alternatively, a predetermined constant may be used instead of the function F (ε).

Further, a plurality of points of time t may be provided, and the average of the errors ε calculated for each time t may be the above error ε.

Alternatively, the number 18

[0095]

(Equation 18)

The error ε (f) for each frequency is calculated according to

[0097]

[Equation 19]

Accordingly, the noise removal coefficient α (f) is set by the function F (ε (f)) that monotonically decreases with respect to ε (f).

Further, a predetermined constant may be used instead of the function F (ε (f)).

The noise removing section 7 removes noise according to the equation 4 or the equation 9 using the removal coefficient α or α (f).

The seventh embodiment of the present invention will be described below.

The configuration of the noise eliminator according to the present invention is different from that of the conventional one. The noise pattern generator 3 and the noise eliminator 7 are different.
That is, the third removal coefficient calculation unit 9 is connected to the above.

The third removal coefficient calculation unit 9, which is a typical constituent feature of the present invention, will be described below in detail with reference to FIG.

The third removal coefficient calculation unit 9 calculates the noise removal coefficient for each frequency band based on the magnitude of the noise component value for each frequency band in the noise pattern N (f) created by the noise pattern creation unit. Find α (f). That is, in the noise pattern N (f), the value of the noise removal coefficient α (f) related to the frequency band having a large noise component value for each frequency band is large, and the value of the noise component for each frequency band is small for a frequency band. The value of the noise removal coefficient α (f) is determined so that the related noise removal coefficient α (f) is reduced.

The noise removing section 7 receives the removal coefficient α (f).
Is used to remove noise according to Equation 9.

[0106]

As is apparent from the above description, according to the present invention, it is possible to remove a noise component in consideration of the characteristics of noise superimposed on voice, music and the like. Further, according to the present invention, the noise removal coefficient can be determined for each of the stationary frequency region and the non-stationary frequency region of noise superimposed on voice or music, so that highly accurate noise removal can be performed. You can do it.

[Brief description of the drawings]

FIG. 1 is a schematic configuration diagram of a first embodiment to a fourth embodiment of a noise removing device according to the present invention.

FIG. 2 is an explanatory diagram of a voice section and a noise pattern creation section in the first to fourth embodiments.

FIG. 3 is a schematic configuration diagram of fifth to sixth embodiments of a noise eliminator according to the present invention.

FIG. 4 is an explanatory diagram of a voice section and a noise pattern creation section in the fifth to sixth embodiments.

FIG. 5 is a schematic configuration diagram of a seventh embodiment of a noise eliminator according to the present invention.

FIG. 6 is an explanatory diagram of a spectral subtraction method.

FIG. 7 is a schematic configuration diagram of a conventional noise eliminator.

FIG. 8 is an explanatory diagram of a voice section and a noise pattern creation section in the conventional noise removal device.

[Explanation of symbols]

 1 ... Voice analysis unit 2 ... Voice section determination unit 3 ... Noise pattern creation unit 4 ... Variation calculation unit 5 ... Elimination coefficient control parameter calculation unit 6 ... Rejection coefficient calculation unit 7 ... Noise removal unit 8 ... Second removal coefficient calculation unit 9 ... Third removal coefficient calculation unit 10 ... Noise removal unit

Claims (7)

[Claims] 1. A noise removing device for removing noise by subtracting a noise pattern created based on a noise interval determined to be noise from a pattern of a voice interval determined to be voice, A noise removal device characterized by multiplying the noise pattern by a noise removal coefficient α determined based on the variation of the noise component in, and subtracting this pattern from the pattern of the voice section determined as voice. 2. The noise section according to claim 1, wherein the noise section is divided at a certain time t within the noise section, and a noise pattern obtained at the time t is used to belong to the noise section at a time t +.
A noise eliminator characterized in that a noise component variation of a noise pattern at time t + N is obtained from 1 and the noise elimination coefficient α is determined based on the variation. 3. A noise removing device for removing noise by subtracting a pattern of noise created based on a noise section determined to be noise from a pattern of a voice section determined to be voice, 2. A noise elimination device characterized by obtaining a noise elimination coefficient α (f), which is a ratio for reducing the noise pattern for each frequency band, based on the variation in the noise component value for each frequency band in. 4. The noise section according to claim 3, wherein the noise section is divided at a certain time t within the noise section, and the noise pattern obtained at the time t is used to belong to the noise section at time t +.
From 1 to obtain the variation of the noise component for each frequency band of the noise pattern at time t + N, and obtain the noise removal coefficient α (f) which is the ratio of reducing the noise pattern for each frequency band based on the variation. Characteristic noise elimination device. 5. A noise removing device for removing noise by subtracting a pattern of noise created based on a noise section determined to be noise from a pattern of a voice section determined to be voice, Based on the value of the noise component for each frequency band in, the noise removal coefficient α which is a ratio for reducing the noise pattern for each frequency band
(F) is obtained, and the noise removal coefficient α (f) multiplied by the noise pattern is subtracted from the pattern of the voice section determined as the voice. 6. The method according to claim 2, wherein the variation is obtained from a difference between a noise pattern at time i−1 belonging to the noise section and an input pattern at time i belonging to the noise section. Item 4. A noise elimination device according to item 4. 7. The variation is a square of a difference between a noise pattern at time i-1 belonging to the noise section and an input pattern at time i belonging to the noise section, which is expressed as time i-
5. The value obtained by normalizing with the squared value of the noise pattern in 1.
The noise removal device described.