JPH0944186A - Noise suppressor - Google Patents

Abstract

(57) [Abstract] [Purpose] To provide a noise suppression device that suppresses noise from an acoustic signal in which noise is mixed. A signal time division means 100 divides an input signal into short time segments, and a pitch extraction means 200 extracts the pitch of the input signal. The frequency band dividing means 400 frequency-divides the short time signal, and the energy or effective value calculating means 600 calculates the energy or effective value of the output of the frequency dividing means. The feature amount calculation means 800 calculates the feature amount, and the weight value calculation means 900 calculates the signal weight value. The multiplying means 910 multiplies the frequency of the input signal by the weight value, and the summing means 930 sums the output of the multiplying means 910 over time and frequency.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for extracting a target sound from an audio signal containing noise by weighting the frequency components of the audio signal such as voice using human auditory characteristics.

[0002]

2. Description of the Related Art There are several types of techniques for extracting only one acoustic signal from acoustic signals in which a plurality of signals are mixed and acoustic signals in which noise is mixed. For example, there is a method in which a plurality of microphones are arranged and the signals are separated by utilizing the time difference, phase difference, and level difference of the signals reaching the microphones. Japanese Patent Laid-Open No. 1-118900 is one example. FIG. 1 is a block diagram of this conventional noise suppressor, in which 1 is a close-talking microphone for inputting a voice signal, 2 is a sensor microphone for inputting a noise component signal, 3 and 4 are low-pass filters, and 5 is a band. Pass filter bank, 10i to 10N are noise elimination circuits, 11i
To 11N are phase difference detection and correction circuits, and 12i to 12N are level difference detection and correction circuits. Noise is removed by using a time signal analyzed by a bandpass filter bank and a spectrum signal obtained by rectifying and smoothing the time signal. However, this method cannot separate signals when a plurality of acoustic signals are mixed in a monaural signal. In addition, when noise properties and signal properties are known, signal separation using comb filters and autocorrelation has also been attempted. However, when the mixed signal is not known, this method cannot separate the signals.

By the way, recently, attention has been paid to the perceptual aspect of sound to attempt separation of acoustic signals. Such a method is effective even when a plurality of signals are mixed in a monaural signal, and the characteristics of the signal are unknown, and it is an effective means for separating a signal having a harmonic structure such as an instrument sound. The physical features of acoustic signals for perceptual sound integration are Bregma
It is examined psychologically by N. et al. When humans group sounds, it turns out that features such as the presence or absence of harmonic relationships of frequency components, deviation of commonality of amplitude changes of frequency components, and temporal continuity of sounds are important clues. Have (ASBregman Auditory Scene Analysis, Mit Press (19
90)).

Brown et al. Attempted signal separation by utilizing this feature (GJ Brown MPCooke (1992) A
Computational model of auditory scene analysis, P
roceedings of International Conference on Spoken L
anguage Processing (ICSLP)). The system first performs a process corresponding to frequency analysis using a model simulating human auditory characteristics. Next, frequency components are extracted, major regions are grouped using heuristics based on the characteristics of their rising times and harmonics, and this is regarded as the target voice, and the acoustic signal is resynthesized. It has become. However, in this method, only the unclear voice signal can be extracted because the signal is dropped when the target signal to be extracted and the signal to be removed overlap in frequency.
Moreover, in order to compensate for the loss of the signal, signal interpolation using DP (MPCooke (1993) Computational auditory sc
ene analysis: Exploiting principles of perceived c
ontinuity Speech Communication 13 391-399) and HM
Interpolation using M (MPCooke (1994) Learning to rec
ognise speech in noisy environments, Proceedings o
f ATR Workshop on A Biological Framework for Speec
h Perception and Producton) has been tried, but the effect is small even though enormous calculation is required to reproduce the information once dropped.

[0005]

However, in the above-mentioned conventional configuration, (1) the nature of the signal to be extracted must be known. (2) Multiple microphones are required. (3) Even if the S / N is improved, the acoustic signal important for voice recognition is also lost. (4) Since the characteristics of the signal must be robustly extracted, the amount of calculation becomes enormous. Had the problem.

The present invention solves the above-mentioned conventional problems. Even if the parameter extraction is uncertain, the input signal is weighted according to the likelihood that the signal to be extracted is included. It is intended to realize noise reduction without impairing the intelligibility of the voice without the signal to be lost.

[0007]

In order to achieve this object, a noise suppressing apparatus of the present invention is a signal time division means for time division of an acoustic signal, and a signal of a short time segment by the time division means of the signal. Pitch extraction means for extracting a pitch, first storage means for storing the pitch obtained by the pitch extraction means, and frequency band division means for dividing the output signal of the time division means for the signal into a plurality of frequency bands A second for storing the output of the frequency band dividing means
Storage means, energy calculation means for calculating the energy of the output of the first storage means, third storage means for storing the output of the energy calculation means, third storage means and the first storage means. A feature amount calculation unit that calculates a feature amount from the output of the storage unit, a weight value calculation unit that calculates a weight value of a signal in a certain frequency band in a certain time section, and an output of the third storage unit. It has multiplication means for calculating the product of the weight value calculation means, fourth storage means for storing the output of the multiplication means, and summing means for calculating the sum of the outputs of the fourth storage means. .

[0008]

According to the present invention, noise is mixed by weighting the frequency of the input signal by using the physical characteristics of the sound used by the human auditory sense to suppress the noise in hearing. A target acoustic signal can be extracted by suppressing noise from the acoustic signal.

[0009]

DESCRIPTION OF THE PREFERRED EMBODIMENTS First, FIG. 2 shows a configuration diagram common to the embodiments of the present invention.

In FIG. 2, 100 is a time dividing means for dividing the input signal into short time segments, 200 is a pitch extracting means for extracting the pitch of the input signal, 400 is a frequency band dividing means for frequency dividing the short time signal, Reference numeral 600 is energy or effective value calculation means, 800 is a characteristic amount calculation means for calculating a characteristic amount, 900 is a weight value calculation means, 91
0 is a multiplication means for multiplying the frequency of the input signal by the weight value, 9
30 is a summing means for summing the outputs of the multiplying means 910 over time and frequency.

Each specific embodiment will be described below. (Embodiment 1) FIG. 3 is a configuration diagram showing a configuration of a noise suppression device according to a first embodiment of the present invention.

In FIG. 3, 1000 is a signal time division means for performing time division of a signal, 1001 is a pitch extraction means, 1002 is a first storage means, 1003 is a frequency band division filter, and 1004 is a second storage means. 1005
Is energy calculation means, 1006 is third storage means,
Reference numeral 1007 is a feature amount calculation means, 1008 is a weight value calculation means, 1009 is a multiplication means, 1010 is a fourth storage means, and 1011 is a summing means.

The operation of the noise suppressor configured as above will be described. The signal time division means 1000 divides the acoustic signal into time segments of δt for each time T.
Since the input signal is a voice signal, δt is 13 ms and T is 26 ms. If the shape of the signal time division means 1000 is a Hanning window, the signal time division means W (t
+ KT) can be expressed by (Equation 1).

[0014]

[Equation 1]

The output signal of the signal time division means W (t + kT) is Sig (t + kT) (t = 0 to δt). If the input voice is Speech (t + kT), Si
g (t + kT) is represented by (Equation 2).

[0016]

[Equation 2]

The pitch extracting means 1001 uses Sig (t + k
Extract the pitch of T). The pitch is the cepstrum method,
It is extracted using an autocorrelation method or the like. Extracted pitch
F _p(KT). The first storage means 1002 is
F obtained by the pitch extracting means 1001_pRank (kT)
Pay. Further, the frequency band division filter 1003 is made of Si.
Divide g (t + kT) into N frequency bands of 20 Hz.
The output of the frequency band division filter 1003 is Sig.
(T + kT, f_m)) (M = 1, 2, ...
N). f_mIs the center frequency of the band. N narrowband signals
No. Sig (t + kT, f_m) Is the second storage means 1004
Stored in. In the energy calculation means 1005, each S
ig (t + kT, f_m) Energy is calculated. Said
The N outputs of the energy calculation means 1005 are Ene
(KT, f_m). Ene (kT, f_m) Is (number
Calculated in 3).

[0018]

(Equation 3)

N pieces of Ene (kT, f _m ) are stored in the third storage means 1006. Feature amount calculation means 1007
Calculates the feature amount of Sig (t + kT, f _m ). In the present embodiment, pitch, onset, onset synchronism, offset, offset synchronism, continuity, and harmonicity are calculated as feature amounts. Of these, the feature amount excluding the pitch is calculated by the feature amount calculation means 1007.

FIG. 4 is a detailed diagram of the feature amount calculating means 1007 in this embodiment. 1101 is an onset calculation means, 1102 is an offset calculation means, 1103 is a continuity calculation means, 1104 is a harmonicity calculation means, 110
Reference numeral 5 is an onset synchronism calculating means, and 1106 is an offset synchronism calculating means.

The onset calculation means 1101 determines the onset based on the temporal change of the output of the third storage means 1006. That is, when the energy at the current time has increased by a certain threshold value or more with respect to the energy at the time, it is determined to be onset.

The offset calculation means 1102 determines the offset based on the temporal change of the output of the third storage means. That is, when the energy at the current time has decreased by a certain threshold value or more with respect to the energy at the time, it is determined to be an offset.

The continuity calculating means 1103 determines whether the signal at the previous time and the signal at the present time are continuous based on the temporal change of the output of the third storage means 1006. That is, it is determined that the energy at the current time is continuous when the energy at the current time increases or decreases below a certain threshold with respect to the energy at the time.

In the harmonic calculation means 1104, Sig (t
+ KT, f _m ) is harmonic when it has a harmonic relationship with the pitch f _p (kT). That is, it is determined that f _m, which corresponds to a frequency that is an integral multiple of the pitch, is harmonic to the pitch. In the onset synchronism calculation means 1105, the onset at Sig (t + kT, f _m ) is performed at the pitch f.
It is assumed that they are synchronous when they are simultaneously based on the onset time of _p (kT). Offset synchronism calculation means 11
In 06, it is assumed that the offsets at Sig (t + kT, f _m ) are synchronous when they are simultaneously based on the offset time of the pitch f _p (kT).

The weight value calculating means 1008 shown in FIG. 3 combines Sig (t + kT, pitch, onset, offset, synchronism of onset, synchronism of offset, harmonicity, continuity) of each feature quantity. weight value w (k at f _m)
T, f _m ) is calculated. The feature quantities are classified into those relating to the continuous features and those relating to the frequency features, and the respective features are integrated. The onset, the synchronism of the onset, the offset, the synchronism of the offset, and the continuity are integrated into the continuous characteristic. Further, the pitch and the harmonic property are integrated to form the frequency characteristic.

An example of a method of integrating the feature amounts in Sig (t + kT, f _m ) will be shown below. The frequency characteristic and the successive characteristic can be expressed by logical expressions (4) and (5).

[0027]

(Equation 4)

[0028]

(Equation 5)

At this time, the weight value w (kT, f _m ) of S (t + kT, f _m ) becomes (Equation 6).

[0030]

(Equation 6)

Logical values are assigned as in (Table 1).

[0032]

[Table 1]

When a logical value as shown in (Table 2) is assigned to a certain S (t + kT, f _m ), the weight value w
(KT, f _m ) becomes (Equation 7).

[0034]

[Table 2]

[0035]

(Equation 7)

The multiplication means 1009 shown in FIG. 3 uses Sig (t + k
T, f _m ) is multiplied by the weight value w (kT, f _m ). The output of the multiplication means 1009 is stored in the fourth storage means 1010. The summing means 1011 is w for all k and m.
(KT, f _m ) × Sig (t + kT, f _m ) is summed.
The output of the summing means 1011 is an acoustic signal in which noise is suppressed.

It should be noted that it is not necessary to use all the feature amounts if each feature amount includes harmonicity or continuity.
Further, the feature amount (pitch, onset, offset,
A feature amount other than onset synchronism, offset synchronism, harmonicity, and continuity) may be newly combined. The time width δt of the signal time division means 1000 and the time period T of division are not limited to δt = 13 ms and T = 26 ms, and the shape is not limited to the Hanning window. The pitch extracting means 1001 is not limited to the cepstrum method and the autocorrelation method. In addition, the frequency band of the frequency band division filter 1003 is 20.
It is not limited to Hz. Further, in the harmonic calculation unit 1104, what is determined to be harmonic is not limited to an integer multiple of the pitch, and it may be an even multiple or an odd multiple.
Moreover, you may give width | variety centering on those frequencies. Further, in the onset synchronism calculation means 1105, the time lag for determining that the synchronization is synchronous may not be zero.
That is, it may be determined that the onset is synchronous even before and after the time when the onset is detected in the pitch band. Further, in the offset synchronism calculation means 1106, the time lag for determining that it is synchronous may not be zero. That is, it may be determined that the pitch band is synchronous before and after the time when the offset is detected. Further, in the weight value calculation means 1008, each feature amount may be expressed by a value other than the logical value (0, 1), and the method of calculating the weight value by integrating the feature amounts is not limited to the logical product or the logical sum. Not a thing. Also,
The calculation of the difference in energy used for determining the onset, offset, and continuity is not limited to the comparison of adjacent time segments, and the time may be separated or a plurality of time segments may be compared.

As in Embodiment 9 described later, the frequency band division filter 1003 may use Fourier transform, and the energy calculating means 1006 may use effective value calculating means.

(Embodiment 2) An embodiment of the signal separation device of the present invention will be described below with reference to the drawings.

The main block diagram of this embodiment is FIG. 3 as in the case of the first embodiment. The operation of the noise suppression device configured as above will be described.

The signal time division means 1000 divides the acoustic signal into time segments of δt for each time T. Since the input signal is a voice signal, δt is 13 ms, T is 26 ms
And When the shape of the time division means of the signal is a Hanning window, the time division means W (t + kT) of the signal can be expressed by (Equation 1).

The output signal of the signal time division means W (t + kT) is Sig (t + kT) (t = 0 to δt). If the input voice is Speech (t + kT), Si
g (t + kT) is represented by (Equation 2).

The pitch extraction means 1001 extracts the pitch of Sig (t + kT) using a plurality of extraction methods. FIG.
FIG. 4 is a detailed view of the pitch extracting means 1001 in this embodiment. In FIG. 5, reference numeral 1200 is a first pitch extracting means, 1201 is a second pitch extracting means, 1202 is an nth pitch extracting means (n = 2, 3, ...), 12
Reference numeral 03 is a means for calculating the estimated value of the pitch. The pitch f _p1 (kT) is calculated by the first pitch extracting means 1200 as the second pitch
By the pitch extracting means 1201 of the pitch f _p2 (kT)
But the pitch f _pn by the pitch extraction unit 1202 of the n
It is assumed that (kT) is calculated. At this time, the estimated pitch value f _p (k
T) is calculated by (Equation 8).

[0044]

(Equation 8)

In FIG. 3, the first storage means 1002
Is the pitch f obtained by the pitch extracting means 1001.
Store _p (kT).

The frequency band division filter 1003 divides Sig (t + kT) into N frequency bands of 20 Hz. The output of the frequency band division filter 1003 is Si
It is described as g (t + kT, f _m ) (m = 1, 2, ...
・ N). f _m is the center frequency of the band. The N narrow band signals Sig (t + kT, f _m ) are stored in the second storage means 100.
Stored in 4. The energy calculating means 1005 calculates the energy of each Sig (t + kT, f _m ). The N outputs of the energy calculating means 1005 are Ene
Described as (kT, f _m ). Ene (kT, f _m ) is calculated by (Equation 3).

N pieces of Ene (kT, f _m ) are stored in the third storage means 1006. Feature amount calculation means 1007
Calculates the feature amount of Sig (t + kT, f _m ). In the present embodiment, pitch, onset, onset synchronism, offset, offset synchronism, continuity, and harmonicity are calculated as feature amounts. Of these, the feature amount excluding the pitch is calculated by the feature amount calculation means 1007.

FIG. 4 is a detailed diagram of the feature amount calculating means 1007. 1101 is an on-set calculation means, 11
Reference numeral 02 is an offset calculation means, 1103 is a continuity calculation means, 1104 is a harmonicity calculation means, 1105 is an onset synchronism calculation means, 1106 is an offset synchronism calculation means.

The onset calculation means 1101 determines the onset based on the temporal change of the output of the third storage means 1006. That is, when the energy at the current time has increased by a certain threshold value or more with respect to the energy at the time, it is determined to be onset.

The offset calculating means 1102 determines the offset based on the temporal change of the output of the third storage means. That is, when the energy at the current time has decreased by a certain threshold value or more with respect to the energy at the time, it is determined to be an offset.

The continuity calculating means 1103 determines whether the signal at the previous time and the signal at the current time are continuous based on the temporal change of the output of the third storage means 1006. That is, it is determined that the energy at the current time is continuous when the energy at the current time increases or decreases below a certain threshold with respect to the energy at the time.

In the harmonic calculation means 1104, Sig (t
+ KT, f _m ) is harmonic when it has a harmonic relationship with the pitch f _p (kT). That is, it is determined that f _m, which corresponds to a frequency that is an integral multiple of the pitch, is harmonic with respect to the pitch.

In the onset synchronism calculation means 1105, the onset at Sig (t + kT, f _m ) is performed at the pitch f _p.
It is assumed to be synchronous when they are simultaneously based on the onset time of (kT).

In the offset synchronism calculation means 1106, the offset at Sig (t + kT, f _m ) is the pitch f _p.
It is assumed that they are synchronous when they are simultaneously based on the offset time of (kT).

The weight value calculating means 1008 in FIG. 3 combines the respective feature amounts (pitch, onset, offset, onset synchronism, offset synchronism, harmonicity, continuity) and Sig (t + kT, weight value w (t at f _m)
+ KT, f _m ) is calculated. The feature quantities are classified into those relating to the continuous features and those relating to the frequency features, and the respective features are integrated. The onset, the synchronism of the onset, the offset, the synchronism of the offset, and the continuity are integrated into the continuous characteristic. Further, the pitch and the harmonic property are integrated to form the frequency characteristic.

An example of a method of integrating the above feature amounts in Sig (t + kT, f _m ) will be shown below.

The frequency characteristic and the successive characteristic can be expressed by logical expressions (Equation 4) and (Equation 5).

At this time, the weight value w (kT, f _m ) of Sig (t + kT, f _m ) becomes (Equation 6).

Logical values are assigned as in (Table 1). When a logical value as shown in (Table 2) is assigned to a certain Sig (t + kT, f _m ), a weight value w (kT, f _m ).
Becomes (Equation 7).

The multiplication means 1009 in FIG. 3 uses Sig (t + k
T, f _m ) is multiplied by the weight value w (kT, f _m ). The output of the multiplication means 1009 is stored in the fourth storage means 1010. The summing means 1011 is w for all k and m.
(KT, f _m ) × Sig (t + kT, f _m ) is summed.
The output of the summing means 1011 is an acoustic signal in which noise is suppressed.

It should be noted that it is not necessary to use all the feature quantities if the harmonic content or the continuity is included in each of the feature quantities. Further, the characteristic amount (pitch, onset,
Offset, onset synchronism, offset synchronism, harmonicity, continuity) may be newly combined. Further, the time width δt of the time division means 1000 of the signal and the time period T of the division are δt = 13 ms, T
= 26 ms and the shape is not limited to the Hanning window. Also, the pitch extracting means 1001
Is not limited to the cepstrum method and the autocorrelation method. The frequency band of the frequency band division filter 1003 is not limited to 20 Hz. Further, what is determined to be harmonic by the harmonic calculation unit 1104 is not limited to an integral multiple of the pitch, and may be an even multiple or an odd multiple of the pitch. You may give it. Further, there is no time lag when the onset synchronism calculation means 1105 determines that it is synchronous.
It does not have to be. That is, it may be determined that the onset is synchronous even before and after the time when the onset is detected in the pitch band. Also, the offset synchronism calculation means 1106.
In the above, the time lag determined to be synchronous does not have to be 0. That is, it may be determined that the pitch band is synchronous before and after the time when the offset is detected. Further, each feature amount may be expressed by the weight value calculating means 1008 other than the logical value (0, 1), and the method of calculating the weight value by integrating the feature amounts is not limited to the logical product or the logical sum. Not a thing.

Further, the calculation of the energy difference used for determining the onset, the offset, and the continuity is not limited to the comparison of the adjacent time segments, and the time may be separated or a plurality of time segments may be compared.

As in Embodiment 9 described later, the frequency band division filter 1003 may use Fourier transform, and the energy calculating means 1005 may use effective value calculating means.

(Embodiment 3) An embodiment of the signal separation device of the present invention will be described below with reference to the drawings.

Also in this embodiment, the main configuration is shown in FIG. The operation of this embodiment will be described below.

The signal time division means 1000 divides the acoustic signal into time segments of δt for each time T. Since the input signal is a voice signal, δt is 13 ms and T is 26 ms. When the shape of the signal time division means 1000 is a Hanning window, the signal time division means W (t + kT) can be expressed by (Equation 1).

The output signal of the signal time division means W (t + kT) is Sig (t + kT) (t = 0 to δt). If the input voice is Speech (t + kT), S
ig (t + kT) is represented by (Equation 2).

The pitch extracting means 1001 uses Sig (t + k
Extract the pitch of T). The pitch is extracted using a cepstrum method, an autocorrelation method, or the like. First storage means 10
02 stores the pitch f _p (kT) obtained by the pitch extracting means 1001. Further, the frequency band division filter 1003 divides Sig (t + kT) into N frequency bands of 20 Hz. The frequency band division filter 1003
Is described as Sig (t + kT, f _m ) (m =
1, 2, ... N). f _m is the center frequency of the band. The N narrowband signals Sig (t + kT, f _m ) are the second
Stored in the storage means 1004. The energy calculating means 1005 calculates the energy of each Sig (t + kT, f _m ). N of the energy calculating means 1005
Each output is described as Ene (kT, f _m ). Ene
(KT, f _m ) is calculated by (Equation 3).

N pieces of Ene (kT, f _m ) are stored in the third storage means 1006. Feature amount calculation means 1007
Calculates the feature amount of Sig (t + kT, f _m ). In the present embodiment, pitch, onset, onset synchronism, offset, offset synchronism, continuity, and harmonicity are calculated as feature amounts.

FIG. 6 is a detailed diagram of the feature amount calculating means 1007 in this embodiment. In FIG. 6, 1301 is a pitch evaluation measure calculation means, 1302 is an onset evaluation measure calculation means, 1303 is an offset evaluation measure calculation means, 1304 is a continuity evaluation measure calculation means, 1
Reference numeral 305 is a harmonic evaluation evaluation measure calculation means, 1306 is an onset synchronization evaluation evaluation measure calculation means, and 1307 is an offset synchronization evaluation evaluation measure calculation means. In addition, the probability that each feature quantity is present in Sig (t + kT, f _m ) is used as the evaluation measure.

The means for calculating the evaluation measure of each feature will be described in detail below. Pitch evaluation measure calculation means 13
In 01, the range in which the pitch can exist with respect to f _p (kT) is stochastically defined, and the probability that the pitch exists in Sig (kT, f _m ) is calculated. FIG. 7 is an example of a probability distribution of frequencies in which pitch may exist. If the probability that the pitch exists for f _p (kT) has a Gaussian distribution, the probability that the pitch exists for Sig (kT, f _m ) _{Rel_pitch} (kT, f _m ) is the distribution variance σ. It is given by (Equation 9).

[0072]

[Equation 9]

FIG. 8 is a spectrum diagram when a harmonic compound sound composed of two components continues from time k = 2 to k = 6. With respect to this signal, the following means for calculating each likelihood will be described.

FIG. 9 is a schematic diagram relating to auditory perception with respect to changes in loudness. When another signal is added to one signal as shown in FIG. 9A, there is a change in the loudness of the sound. In the onset evaluation measure calculation means 1302, the human perception is assured that this change is not due to the time change of one sound but due to the addition of another different signal as shown in FIG. 9B. Based on the above, this is the onset certainty. The probability of onset is calculated from the output of the third storage means 1006. An example of the calculation method will be given below. The size of the sound that changes just the criterion that psychologically exceeds the continuity or whether a new sound is added is being investigated (Proceedings of the Autumn Meeting of the Acoustical Society of Japan, 1994 2-7-18). Using the output of the energy calculating means 1005, the probability of the onset is determined from the average of the energy at the current observation time and the energy up to the time before it. FIG. 10 is an explanatory diagram for a method of calculating the probability of onset.
The average of the energies up to the previous time is about 200 in the current observation frequency and two frequency bands close to that frequency.
The average of the segments in the ms section (16 segments) is used. Sig (kT, f _m) the probability Rel _{_on} the onset of (E _Dif) is calculated by the equation (10).

[0075]

(Equation 10)

FIG. 11 is a diagrammatic representation of (Equation 10). The onset certainty is calculated only in the pitch band until the onset is detected in the pitch band.
10 and 11 are explanatory views for explaining the certainty of the onset with respect to the signal of FIG. In FIG. 12, at time k = 2, the probability of the onset in the pitch band becomes 1.0 and exceeds the threshold value 0.8. Therefore, from time k = 2, the calculation of the probability of the onset in the entire frequency band is started. Next, it is assumed that the onset certainty factor 1.0 can be calculated by (k, m) = (2, 1) and (2, 3).
At this time, as shown in FIG. 13, (k, m) = (2, 1) and (2,
The probability of onset 1.0 is assigned to 3).

Offset evaluation measure calculation means 1303
At, the accuracy of the offset is calculated. There is a change in loudness when one signal is removed from the two signals. Therefore, the offset is determined by using the output of the energy calculating means 1005. The offset probability _{Rel_off} (E _Dif ) is _{given by} ( _Equation 11) as in the method of calculating the onset reliability.

[0078]

[Equation 11]

The accuracy of the offset is calculated only in the pitch frequency band from the detection of the onset to the detection of the offset in the pitch band. FIG. 14 and FIG.
FIG. 5 is an explanatory diagram for explaining the probability of the offset with respect to the signal of FIG. In FIG. 14, the probability of the offset in the pitch band is 1.0 at time k = 6, which exceeds a certain threshold value 0.9. Therefore, at time k = 6, the accuracy of the offset is calculated for all frequency bands. If the probability of offset 1.0 can be calculated by (k, m) = (6, 1) and (6, 3), the probability of offset is assigned as shown in FIG.

In the continuity evaluation measure calculating means 1304, when the loudness of the sound changes with time, the certainty that it is the change with time of one signal is used as an index. The probability of continuity is calculated from the output of the third storage unit 1006. FIG. 16 is an explanatory diagram for the method of calculating the probability of continuity. In FIG. 16, the average of the energy of the current observation segment in about 200 ms interval (16 segments) before the current observation time including the current observation time, the current observation frequency, and the current observation time in two frequency bands adjacent to the frequency are included. About 200m before the current observation time
The certainty of the continuity is determined from the average of the energy in the s section (16 segments). The probability of continuity _{Rel_con} (E _Dif ) is defined by ( _Equation 12).

[0081]

(Equation 12)

FIG. 17 is a schematic representation of (Equation 12). The likelihood of continuity is calculated until the offset is detected after the onset is detected. FIG. 18 is an explanatory diagram for explaining the certainty of the continuity with respect to the signal of FIG. When the onset is detected at the time k = 2 and the offset is detected at the time k = 6, the certainty of the continuity is calculated from the time k = 3 to 5, and the certainty of the continuity at each (k, m) is calculated. Likelihood is calculated as shown in FIG.

In the calculation unit 1305 of the evaluation measure of harmonic properties, the probability that Sig (t + kT, f _m ) is perceptually harmonic with respect to f _p (kT) is regarded as the certainty of harmonic properties. A frequency distribution that subjectively senses harmonics is determined from a human psychological measurement function centering on a frequency section that is an integral multiple of the pitch. Psychological measurement functions are published in IEICE Transactions '94 / 5Vo
l. Determine based on the data published in J77-A No. 5 731-740.
The pitch f _p (kT) at the center frequency f _m of the frequency band
The deviation u from the frequency of p times (p is an integer) is (Equation 13)
Is required.

[0084]

(Equation 13)

Distribution function Rel of the above-mentioned probability of harmonicity
_harm (u) is defined using u as in (Equation 14).

[0086]

[Equation 14]

FIG. 19 is a diagrammatic representation of (Equation 14). FIG. 20 is an explanatory diagram for explaining the certainty of the harmonic property with respect to the signal of FIG. In FIG. 20, the probability of pitch is 1.

The onset synchronicity evaluation measure calculating means 1306 judges that the onset at Sig (t + kT, f _m ) is subjectively simultaneous on the basis of the onset time of f _p (kT). The synchronicity is determined based on the time distribution. Perceptual onset synchronism is defined with reference to the data published in IEICE Transactions '94 / 5 Vol.J77-A No.5 731-740. When the difference between the onset time at which the onset is detected at Sig (t + kT, f _m ) and the closest pitch is v (ms), the certainty of the onset synchronism Rel_sync (kT, f _m ) is It can be described by (Equation 15).

[0089]

(Equation 15)

FIG. 21 is a diagrammatic representation of (Equation 15). FIG. 22 is an explanatory diagram for explaining the certainty of the onset synchronism with respect to the signal of FIG.

The offset synchronism evaluation measure calculating means 1307 can be calculated by the same method as the onset synchronism evaluation measure. The synchronism of the offset is determined based on the time distribution in which the offset at Sig (t + kT, f _m ) is subjectively judged to be simultaneous based on the offset time of f _p (kT). FIG. 23 is an explanatory diagram for explaining the certainty of the synchronization of the offset with respect to the signal of FIG.

In the weight value calculating means 1008 shown in FIG. 3, the above-mentioned features (pitch, onset, offset,
Enter the certainty about onset synchrony, offset synchrony, harmonicity, and continuity, and enter Sig (t + k
The weight value w (t + kT, f _m ) at T, f _m ) is calculated.

Of the characteristic amounts, the certainty values are classified into those relating to the temporal certainty and those relating to the frequency certainty, and the respective certainty values are integrated. The accuracy of the onset, the accuracy of the synchronization of the onset, the accuracy of the offset, the accuracy of the synchronization of the offset, and the accuracy of the continuity are integrated into the continuous accuracy. . Further, the certainty of the pitch and the certainty of the harmonic property are integrated to obtain the certainty in terms of frequency.

In FIG. 3, the multiplication means 1009 is Sig.
(T + kT, f _m ) is multiplied by the weight value w (kT, f _m ). The output of the multiplication means 1009 is the fourth storage means 10
10 is stored. The summing means 1011 sums w (kT, f _m ) × Sig (t + kT, f _m ) for all k, m. The output of the summing means 1011 is an acoustic signal in which noise is suppressed.

Note that it is not necessary to use all the feature quantities if the harmonic or the continuity is included in each of the feature quantities. Further, a feature amount other than the feature amount may be newly combined. Also, the signal time division means 10
The time width δt of 00 and the division time period T are δt = 13.
ms, T = 26 ms, and the shape is not limited to the Hanning window. The pitch extracting means 1001 is not limited to the cepstrum method and the autocorrelation method. The frequency band of the frequency band division filter 1003 is not limited to 20 Hz. Further, in the calculation of the probability that the pitch exists, it may be defined by a distribution other than the Gaussian distribution. Further, the threshold value used for determining the onset time may be other than 0.8. Further, the threshold value used for determining the offset time may be other than 0.9. In addition, the calculation method of the probability of the onset is (Equation 1
It is not limited to 0). The method of calculating the certainty of the offset is not limited to (Equation 11). The method of calculating the certainty of continuity is not limited to (Equation 12). Also,
The method of calculating the probability of harmonicity is not limited to (Equation 13) and (Equation 14). Further, the method of calculating the certainty of onset synchronism and the method of calculating the certainty of offset synchronism are not limited to (Equation 15).

The frequency band division filter 1003 may use Fourier transform, and the energy calculating means 1006 may use effective value calculating means, as in the ninth embodiment described later.

(Embodiment 4) An embodiment of the signal separation device of the present invention will be described below with reference to the drawings.

The main block diagram of this embodiment is also shown in FIG. The operation of this embodiment will be described.

The signal time division means 1000 divides the acoustic signal into time segments of δt for each time T. Since the input signal is a voice signal, δt is 13 ms and T is 26 ms. When the shape of the signal time division means 1000 is a Hanning window, the signal time division means W (t + kT) can be expressed by (Equation 1).

The output signal of the signal time division means W (t + kT) is Sig (t + kT) (t = 0 to δt). If the input voice is Speech (t + kT), S
ig (t + kT) is represented by (Equation 2).

Pitch extraction means 1001 uses a plurality of extraction methods
Is used to extract the pitch of Sig (t + kT). FIG.
4 is a detailed view of the pitch extracting means 1001. FIG.
4, 3810 is the first pitch extracting means, 382.
0 is the second pitch extraction means, 3830 is the estimated pitch value
Is a calculation means of. By the first pitch extracting means 3810
The pitch f_p1(KT) is the second pitch extraction means 38
20 by pitch f _p2It is assumed that (kT) is calculated.
At this time, the pitch estimation value calculation means 3830
Estimated value f of pitch_p(KT) is calculated by (Equation 16)
You.

[0102]

(Equation 16)

In FIG. 3, the first storage means 1002 stores f _p1 (kT) obtained by the pitch extraction means 1001.
And f _p2 (kT) and f _p (kT) are stored.

The frequency band division filter 1003 divides Sig (t + kT) into N frequency bands of 20 Hz. The output of the frequency band division filter 1003 is Si
It is described as g (t + kT, f _m ) (m = 1, 2, ...
・ N). f _m is the center frequency of the band. The N narrow band signals Sig (t + kT, f _m ) are stored in the second storage means 100.
Stored in 4. The energy calculating means 1005 calculates the energy of each Sig (t + kT, f _m ). The N outputs of the energy calculating means 1005 are Ene
Described as (kT, f _m ). Ene (kT, f _m ) is calculated by (Equation 3).

The N pieces of Ene (kT, f _m ) are stored in the third storage means 1006. Feature amount calculation means 10
07 calculates the feature amount of Sig (t + kT, f _m ).
In the present embodiment, pitch, onset, onset synchronism, offset, offset synchronism, continuity, and harmonicity are calculated as feature amounts.

FIG. 6 is a detailed diagram of the feature amount calculation means 1007 in this embodiment. In FIG. 6, 1301 is a pitch evaluation measure calculation means, 1302 is an onset evaluation measure calculation means, 1303 is an offset evaluation measure calculation means, 1304 is a continuity evaluation measure calculation means, 1
Reference numeral 305 is a harmonic evaluation evaluation measure calculation means, 1306 is an onset synchronization evaluation evaluation measure calculation means, and 1307 is an offset synchronization evaluation evaluation measure calculation means. In addition, the probability that each feature quantity is present in Sig (t + kT, f _m ) is used as the evaluation measure.

The means for calculating the feature quantity evaluation measure will be described in detail below. Pitch evaluation measure calculation means 1
301 is f calculated by the pitch extracting means 1001.
_From the difference between _p1 (kT) and f _p2 (kT), the pitch f _p (kT)
Determine the certainty of. f _p1 (kT) and f _p2 (kT)
When the difference between the two is 0, it is assumed that the pitch is accurately extracted and the pitch f _p
The probability of (kT) is 1. f _p1 (kT) and f _p2
The probability decreases as the difference in (kT) increases. The probability of pitch is f _p1 (kT) − as shown in FIG. 3900.
If the Gaussian distribution is used according to the value of f _p2 (kT), the pitch accuracy Rel (f _p1 (kT) −f _p2 (kT))
Is determined by (Equation 17).

[0108]

[Equation 17]

At the same time, the range in which a certain pitch can exist is stochastically defined with respect to the estimated value f _p (kT) of the pitch obtained as shown in FIG. 26, and it is confirmed that the pitch exists in Sig (kT, f _m ). Calculate the likelihood. Estimated value of pitch f _p (k
If the probability that the pitch exists with respect to T) has a Gaussian distribution, the probability that the pitch exists in Sig (kT, f _m ) _{Rel_pitch} (kT, f _m ) is σ
Then, it is given by (Equation 9).

FIG. 8 is a spectrum diagram when a harmonic compound sound composed of two components continues from time k = 2 to k = 6. With respect to this signal, the following means for calculating each likelihood will be described.

FIG. 9 is a schematic diagram relating to auditory perception with respect to changes in loudness. When another signal is added to one signal as shown in FIG. 9A, there is a change in the loudness of the sound. In the onset evaluation measure calculation means 1302, the human perception is assured that this change is not due to the time change of one sound but due to the addition of another different signal as shown in FIG. 9B. Based on the above, this is the onset certainty. The probability of the onset is calculated from the output of the third storage unit 1006. An example of the calculation method will be given below. The loudness of the sound that changes just the psychologically continuous judgment or whether a new sound is added is being investigated (19
Proceedings of the 94th Autumn Meeting of ASJ 2-7-18). Using the output of the energy calculating means 1005, the probability of the onset is determined from the average of the energy at the current observation time and the energy up to the time before it. FIG.
FIG. 6 is an explanatory diagram for a method of calculating the probability of onset. The average of the energy up to the previous time is the average of the segment of about 200 ms section (16 segments) in the current observation frequency and two frequency bands close to the current frequency. Sig (kT, f _m) the probability Rel _{_on} the onset of (E _Dif) is calculated by the equation (10).

FIG. 11 is a diagrammatic representation of (Equation 10). Onset certainty is calculated only in the pitch band until the onset is detected in the pitch band. FIG.
2 and 13 are explanatory views for explaining the certainty of the onset with respect to the signal of FIG. In FIG. 12, at time k = 2, the probability of the onset in the pitch band becomes 1.0 and exceeds the threshold value 0.8. Therefore, from time k = 2, the calculation of the probability of the onset in the entire frequency band is started. It is assumed that the onset certainty factor 1.0 can be calculated by (k, m) = (2, 1) and (2, 3). At this time, as shown in FIG. 2700, (k, m) = (2,1) and (2,3)
Is assigned to the onset certainty of 1.0.

Offset evaluation measure calculation means 1303
At, the accuracy of the offset is calculated. There is a change in loudness when one signal is removed from the two signals. Therefore, the offset is determined by using the output of the energy calculating means 1005. The offset probability _{Rel_off} (E _Dif ) is _{given by} ( _Equation 11) as in the method of calculating the onset reliability.

The probability of the offset is calculated only in the pitch frequency band from the detection of the onset to the detection of the offset in the pitch band. FIG. 14 and FIG.
FIG. 5 is an explanatory diagram for explaining the probability of the offset with respect to the signal of FIG. In FIG. 14, the probability of the offset in the pitch band is 1.0 at time k = 6, which exceeds a certain threshold value 0.9. Therefore, at time k = 6, the accuracy of the offset is calculated for all frequency bands. If the probability of offset 1.0 can be calculated by (k, m) = (6, 1) and (6, 3), the probability of offset is assigned as shown in FIG.

The continuity evaluation measure calculating means 1304 uses as an index the certainty that, when the loudness of the sound changes with time, it is the change with time of one signal. The certainty of the continuity is calculated by the output of the third storage unit 1006. FIG. 16 is an explanatory diagram for a method of calculating the probability of onset. In FIG. 16, the average of the energy of the current observation segment in about 200 ms interval (16 segments) before the current observation time including the current observation time, the current observation frequency, and the current observation time in two frequency bands adjacent to the frequency are included. The certainty of the continuity is determined from the average of the energies of about 200 ms section (16 segments) before the current observation time which is not present. The probability of continuity _{Rel_con} (E _Dif ) is defined by ( _Equation 12).

FIG. 17 is a diagrammatic representation of (Equation 12). The likelihood of continuity is calculated until the offset is detected after the onset is detected. FIG. 18 is an explanatory diagram for explaining the certainty of the continuity with respect to the signal of FIG. When the onset is detected at the time k = 2 and the offset is detected at the time k = 6, the certainty of the continuity is calculated from the time k = 3 to 5, and the continuity of each (k, m) is calculated. Probability is calculated as shown in FIG.

In the calculation unit 1305 of the evaluation measure of the harmonic property, the probability that the Sig (t + kT, f _m ) has a perceptual harmonic relationship with the f _p (kT) is regarded as the certainty of the harmonic property. . A frequency distribution that subjectively senses harmonics is determined from a human psychological measurement function centering on a frequency section that is an integral multiple of the pitch. The psychometric function is determined by referring to the data published in IEICE Transactions '94 / 5 Vol.J77-A No.5 731-740. A deviation u from a frequency that is p times (p is an integer) the pitch f _p (kT) at the center frequency f _m of the frequency band is obtained by (Equation 13).

Distribution function Rel of the above-mentioned probability of harmonicity
_harm (u) is defined using u as in (Equation 14).

FIG. 19 is a diagrammatic representation of (Equation 14). FIG. 20 is an explanatory diagram for explaining the certainty of the harmonic property with respect to the signal of FIG. In FIG. 20, the probability of pitch is 1.

The onset synchronicity evaluation measure calculating means 1306 judges that the onset at Sig (t + kT, f _m ) is subjectively simultaneous with the onset time of f _p (kT) as a reference. The synchronicity is determined based on the time distribution. Perceptual onset synchronism is defined with reference to the data published in IEICE Transactions '94 / 5 Vol.J77-A No.5 731-740. When the difference between the onset time at which the onset is detected at Sig (t + kT, f _m ) and the closest pitch is v (ms), the certainty of the onset synchronism Rel_sync (kT, f _m ) is ,
It can be described by (Equation 15).

FIG. 21 is a diagrammatic representation of (Equation 15). FIG. 22 is an explanatory diagram for explaining the certainty of the onset synchronism with respect to the signal of FIG.

The offset synchronism evaluation measure calculating means 1307 can be calculated by the same method as the onset synchronism evaluation measure. The synchronism of the offset is determined based on the time distribution in which the offset at Sig (t + kT, f _m ) is subjectively judged to be simultaneous based on the offset time of f _p (kT). FIG. 23 is an explanatory diagram for explaining the certainty of the synchronization of the offset with respect to the signal of FIG.

In the weight value calculating means 1008 shown in FIG. 3, the reliability of each feature (pitch, onset, offset, onset synchronism, offset synchronism, harmonicity, continuity) is assigned to the evaluation function. Input Sig (t + kT,
weight value w (t + kT at f _m), calculates the f _m).

Among the feature amounts, the ones are classified into those relating to the continuous likelihood and those relating to the frequency certainty, and the respective likelihoods are integrated. The accuracy of the onset, the accuracy of the synchronization of the onset, the accuracy of the offset, the accuracy of the synchronization of the offset, and the accuracy of the continuity are integrated into the continuous accuracy. . Further, the certainty of the pitch, the certainty of the range of variation of the pitch, and the certainty of the harmonic property are integrated into the certainty of the frequency.

In FIG. 3, the multiplication means 1009 is Sig.
(T + kT, f _m ) is multiplied by the weight value w (kT, f _m ). The output of the multiplication means 1009 is the fourth storage means 10
10 is stored. The summing means 1011 sums w (kT, f _m ) × Sig (t + kT, f _m ) for all k, m. The output of the summing means 1011 is an acoustic signal in which noise is suppressed.

It is not necessary to use all of the feature quantities if the harmonic content or the continuity is included in the feature quantities. Further, a feature amount other than the feature amount may be newly combined. Also, the signal time division means 100
The time width δt of 0 and the time period T of division are δt = 13 m.
s, T = 26 ms, and the shape is not limited to the Hanning window. Also, the pitch extracting means 1
001 is not limited to the cepstrum method and the autocorrelation method. The frequency band of the frequency band division filter 1003 is not limited to 20 Hz. Further, in the calculation of the probability that the pitch exists, the distribution may be defined by a distribution other than the Gaussian distribution. Further, the threshold value used for determining the onset time may be other than 0.8. Further, the threshold value used for determining the offset time may be other than 0.9. Also,
The method of calculating the certainty of the onset is not limited to (Equation 10). The method of calculating the certainty of the offset is not limited to (Equation 11). The method of calculating the certainty of continuity is not limited to (Equation 12). In addition, the calculation method of the above-mentioned probability of harmonicity is expressed by (Equation 13) and (Equation 1)
It is not limited to 4). Further, the method of calculating the certainty of onset synchronism and the method of calculating the certainty of offset synchronism are not limited to (Equation 15).

As in the ninth embodiment, the frequency band dividing filter 1003 may use Fourier transform,
Further, the energy calculating means 1006 may use an effective value calculating means.

(Embodiment 5) An embodiment of the signal separating apparatus of the present invention will be described below with reference to the drawings.

The main block diagram of the fifth embodiment of the present invention is shown in FIG. The operation of the noise suppression device configured as above will be described.

The signal time division means 1000 divides the acoustic signal into time segments of δt for each time T. Since the input signal is a voice signal, δt is 13 ms and T is 26 ms. When the shape of the signal time division means 1000 is a Hanning window, the signal time division means W (t + kT) can be expressed by (Equation 1).

The output signal of the signal time division means W (t + kT) is Sig (t + kT) (t = 0 to δt). If the input voice is Speech (t + kT), Si
g (t + kT) is represented by (Equation 2).

The pitch extraction means 1001 extracts the pitch of Sig (t + kT) using a plurality of extraction methods. FIG.
FIG. 4 is a detailed view of the pitch extracting means 1001 in this embodiment. In FIG. 5, 1200 is the first pitch extracting means, 1201 is the second pitch extracting means, 1202 is the nth pitch extracting means, 1203 is the estimated pitch value f.
It is a means for calculating _p (t). First pitch extracting means 12
00 by the pitch f _p1 (kT) is the pitch f _p2 the second pitch extracting means 1201 (kT) is the pitch f _pn (kT) is calculated by the pitch extraction unit 1202 of the n. At this time, the estimated pitch value f _p (t)
The pitch estimation value f _p (k
T) is calculated by (Equation 8).

Also, the variance V _pitch (kT) of f _pn (kT) is calculated from f _p1 (kT) by (Equation 18).

[0134]

(Equation 18)

In FIG. 3, the first storage means 1002 stores f _p (kT) obtained by the pitch extraction means 1001.
And the variance V _pitch (kT).

The frequency band division filter 1003 divides Sig (t + kT) into N frequency bands of 20 Hz. The output of the frequency band division filter 1003 is Si
It is described as g (t + kT, f _m ) (m = 1, 2, ...
・ N). f _m is the center frequency of the band. The N narrow band signals Sig (t + kT, f _m ) are stored in the second storage means 100.
Stored in 4. In the energy calculation means 1005, each S
The energy of ig (t + kT, f _m ) is calculated. The N outputs of the energy calculating means 1005 are converted into Ene (k
T, f _m ). Ene (kT, f _m ) is (Equation 3)
Is calculated.

The N pieces of Ene (kT, f _m ) are stored in the third storage means 1006. Feature amount calculation means 10
07 calculates the feature amount of Sig (t + kT, f _m ).
In this embodiment, as the feature amount, pitch, onset, onset synchronism, offset, offset synchronism,
Calculate continuity and harmonicity.

FIG. 6 is a detailed diagram of the feature amount calculation means 1007 of FIG. In FIG. 6, 1301 is a pitch evaluation measure calculation unit, 1302 is an onset evaluation measure calculation unit, 1303 is an offset evaluation measure calculation unit,
Reference numeral 1304 denotes a continuity evaluation metric calculation unit, 1305 denotes a harmonic evaluation metric calculation unit, 1306 denotes an onset synchronicity evaluation metric calculation unit, and 1307 denotes an offset synchronism evaluation metric calculation unit. is there. In addition, the probability that each feature quantity is present in Sig (t + kT, f _m ) is used as the evaluation measure.

The means for calculating the evaluation measure of the feature quantity will be described in detail below. The means for calculating the pitch evaluation measure is the variance V calculated by the pitch extracting means 1001.
_The probability of the pitch f _p (kT) estimated from the _pitch (kT) is determined. When V _pitch (kT) is 0, it is assumed that the pitch is accurately extracted, and the probability of the pitch f _p (kT) is 1. The probability decreases as V _pitch (kT) increases. If the pitch _accuracy is Gaussian distributed according to the value of V _pitch (kT) as shown in FIG. 4100, the _pitch accuracy Rel (V _pitch (kT)) is determined by (Equation 19).

[0140]

[Equation 19]

At the same time, the range in which a certain pitch can exist is stochastically defined with respect to the estimated value f _p (kT) of the pitch obtained as shown in FIG. 28, and the probability that Sig (kT, f _m ) is the pitch To calculate. If the probability that the pitch exists with respect to the estimated value f _p (kT) of the pitch has a Gaussian distribution, the probability that the pitch exists in Sig (kT, f _m ) _{Rel_pitch} (kT, f _m ) is ( _Equation 20). ) Is given.

[0142]

(Equation 20)

FIG. 8 is a spectrum diagram when the harmonic compound sound consisting of two components continues from time k = 2 to k = 6. With respect to this signal, the following means for calculating each likelihood will be described.

FIG. 9 is a schematic diagram relating to auditory perception with respect to changes in loudness. When another signal is added to one signal as shown in FIG. 9A, there is a change in the loudness of the sound. In the onset evaluation measure calculation means 1302, the human perception is assured that this change is not due to the time change of one sound but due to the addition of another different signal as shown in FIG. 9B. Based on the above, this is the onset certainty. The probability of the onset is calculated from the output of the third storage unit 1006. An example of the calculation method will be given below. The loudness of the sound that changes just the psychologically continuous judgment or whether a new sound is added is being investigated (19
Proceedings of the 94th Autumn Meeting of ASJ 2-7-18). Using the output of the energy calculation means 1005, the probability of the onset is determined from the average of the energy at the current observation time and the energy before the current time. FIG. 10 is an explanatory diagram for a method of calculating the probability of onset. The average of the energies up to the previous time is about 2 in the current observation frequency and two frequency bands close to that frequency.
The average of the segments in the 00 ms section (16 segments) is used. Sig (kT, f _m) the probability Rel _{_on} the onset of (E _Dif) is calculated by the equation (10).

FIG. 11 is a diagrammatic representation of (Equation 10). The onset certainty is calculated only in the pitch band until an onset is detected in the pitch band. 12 and 13 are explanatory views for explaining the certainty of the onset for the signal of FIG. In FIG. 12, at time k = 2, the probability of the onset in the pitch band becomes 1.0 and exceeds the threshold value 0.8. Therefore, from time k = 2, the calculation of the probability of the onset in the entire frequency band is started. It is assumed that the onset certainty factor 1.0 can be calculated by (k, m) = (2, 1) and (2, 3). This time,
As shown in FIG. 13, the probability of onset 1.0 is assigned to (k, m) = (2, 1) and (2, 3).

The offset evaluation measure calculating means 13
At 03, the probability of offset is calculated. There is a change in loudness when one signal is removed from the two signals. Therefore, the energy calculating means 1005
The output is used to determine the offset. The offset probability _{Rel_off} (E _Dif ) is _{given by} ( _Equation 11) as in the method of calculating the onset reliability.

The probability of the offset is calculated only in the pitch frequency band from the detection of the onset to the detection of the offset in the pitch band. FIG. 14 and FIG.
FIG. 5 is an explanatory diagram for explaining the probability of the offset with respect to the signal of FIG. In FIG. 14, the probability of the offset in the pitch band is 1.0 at time k = 6, which exceeds a certain threshold value 0.9. Therefore, at time k = 6, the accuracy of the offset is calculated for all frequency bands. If the probability of offset 1.0 can be calculated by (k, m) = (6, 1) and (6, 3), the probability of offset is assigned as shown in FIG.

The continuity evaluation measure calculating means 1304 uses as an index the certainty that, when the loudness of the sound changes with time, it is the change with time of one signal. The certainty of the continuity is calculated by the output of the third storage unit 1006. FIG. 16 is an explanatory diagram for the method of calculating the probability of continuity. In FIG. 16, the average of the energy of the current observation segment in about 200 ms interval (16 segments) before the current observation time including the current observation time, the current observation frequency, and the current observation time in two frequency bands adjacent to the frequency are included. The certainty of the continuity is determined from the average of the energies of about 200 ms section (16 segments) before the current observation time which is not present. The probability of continuity _{Rel_con} (E _Dif ) is defined by ( _Equation 12).

FIG. 17 is a diagrammatic representation of (Equation 12). The likelihood of continuity is calculated until the offset is detected after the onset is detected. FIG. 18 is an explanatory diagram for explaining the certainty of the continuity with respect to the signal of FIG. When the onset at time k = 2 and the offset is detected at time k = 6, the probability of continuity is calculated from time k = 3 to 5 and the probability of continuity at each (k, m) is calculated. Is calculated as shown in FIG.

In the calculation unit 1305 of the evaluation measure of harmonic properties, the probability that Sig (t + kT, f _m ) is perceptually harmonic with respect to f _p (kT) is regarded as the certainty of harmonic properties. A frequency distribution that subjectively senses harmonics is determined from a human psychological measurement function centering on a frequency section that is an integral multiple of the pitch. Psychological measurement functions are published in IEICE Transactions '94 /
5Vol.J77-A No.5 Determined by referring to the data published in 731-740. The pitch f _p (k at the center frequency f _m of the frequency band
The deviation u from the frequency that is p times (T is p) (p is an integer) is (Equation 1)
Required in 3).

Distribution function Rel of the above-mentioned harmonic probability
_harm (u) is defined using u as in (Equation 14).

FIG. 19 is a diagrammatic representation of (Equation 14). FIG. 20 is an explanatory diagram for explaining the certainty of the harmonic property with respect to the signal of FIG. In FIG. 20, the probability of pitch is 1.

The onset synchronicity evaluation measure calculating means 1306 judges that the onset at Sig (t + kT, f _m ) is subjectively simultaneous on the basis of the onset time of f _p (kT). The synchronicity is determined based on the time distribution. The perceptual onset synchrony is based on the data published in IEICE Transactions '94 / 5 Vol.J77-A No.5 731-740.
c (kT, f _m ) is defined by (Equation 15). Sig (t +
When the onset is detected at kT, f _m ) and the difference in the onset time of the pitch closest to it is v (ms), the certainty of the onset synchronism Rel_sync
(KT, f _m ) can be described by (Equation 15).

FIG. 21 is a diagrammatic representation of (Equation 15). FIG. 22 is an explanatory diagram for explaining the certainty of the onset synchronism with respect to the signal of FIG.

The offset synchronism evaluation measure calculating means 1307 can be calculated by the same method as the onset synchronism evaluation measure. The synchronism of the offset is determined based on the time distribution in which the offset at Sig (t + kT, f _m ) is subjectively judged to be simultaneous based on the offset time of f _p (kT). FIG. 23 is an explanatory diagram for explaining the certainty of the synchronization of the offset with respect to the signal of FIG.

In the weight value calculation means 1008 shown in FIG. 3, the reliability of each feature (pitch, onset, offset, onset synchronism, offset synchronism, harmonicity, continuity) is assigned to the evaluation function. Input, Sig (t + k
The weight value w (t + kT, f _m ) at T, f _m ) is calculated.

[0157] Of the above feature amounts, the ones are classified into those related to the continuous certainty and those related to the frequency certainty, and the respective certainties are integrated. The onset accuracy, the onset synchronization accuracy, the offset accuracy, the offset synchronization accuracy, and the power continuity accuracy are integrated into the continuous accuracy. And Further, the certainty of the pitch, the certainty of the range of variation of the pitch, and the certainty of the harmonic property are integrated into the certainty of the frequency.

In FIG. 3, the multiplication means 1009 is S
ig (t + kT, f _m ) is multiplied by the weight value w (kT, f _m ). The output of the multiplication means 1009 is the fourth storage means 1
It is stored in 010. The summing means 1011 is for all k and m
With respect to w (kT, f _m ) × Sig (t + kT, f _m ).
Sum up. The output of the summing means 1011 is an acoustic signal in which noise is suppressed.

Note that it is not necessary to use all the feature amounts if the harmonic or the continuity is included in each of the feature amounts. Further, a feature amount other than the feature amount may be newly combined. Also, the signal time division means 100
The time width δt of 0 and the time period T of division are δt = 13 m.
It is not limited to s, T = 26 ms, and the shape is not limited to the Hanning window. The pitch extracting means 1001 is not limited to the cepstrum method and the autocorrelation method. The frequency band of the frequency band division filter 1003 is not limited to 20 Hz. Further, in the calculation of the probability that the pitch exists, it may be defined by a distribution other than the Gaussian distribution. Further, the threshold value used for determining the onset time may be other than 0.8. Further, the threshold value used for determining the offset time may be other than 0.9. In addition, the calculation method of the probability of the onset is (Equation 1
It is not limited to 0). The method of calculating the certainty of the offset is not limited to (Equation 11). The method of calculating the certainty of continuity is not limited to (Equation 12). Also,
The method of calculating the probability of harmonicity is not limited to (Equation 13) and (Equation 14). Further, the method of calculating the certainty of onset synchronism and the method of calculating the certainty of offset synchronism are not limited to (Equation 15).

The frequency band division filter 1003 may use Fourier transform as in the ninth embodiment.
Further, the energy calculating means 1006 may use an effective value calculating means.

(Embodiment 6) An embodiment of the signal separation device of the present invention will be described below with reference to the drawings.

FIG. 3 shows the main configuration of the sixth embodiment.
It is. The operation of the noise suppression device configured as above will be described. The signal time division means 1000 divides the acoustic signal into time segments of δt for each time T. Since the input signal is a voice signal, δt is 13 ms, T is 26 m
Let s. If the shape of the signal time division means 1000 is a Hanning window, the signal time division means W (t + k).
T) can be represented by (Equation 1).

The output signal of the signal time division means W (t + kT) is Sig (t + kT) (t = 0 to δt). If the input voice is Speech (t + kT), Si
g (t + kT) is represented by (Equation 2).

The pitch extracting means 1001 uses Sig (t + k
Extract the pitch of T). The pitch is extracted using a cepstrum method, an autocorrelation method, or the like. First storage means 10
02 stores the pitch f _p (kT) obtained by the pitch extracting means 1001. The frequency band division filter 1003 divides Sig (t + kT) into N frequency bands of 20 Hz. The frequency band division filter 10
The output of 03 is described as Sig (t + kT, f _m ) (m
= 1, 2, ... N). f _m is the center frequency of the band. The N narrowband signals Sig (t + kT, f _m ) are the second
Stored in the storage means 1004. The energy calculating means 1005 calculates the energy of each Sig (t + kT, f _m ). N of the energy calculating means 1005
Each output is described as Ene (kT, f _m ). Ene
(KT, f _m ) is calculated by (Equation 3).

The N pieces of Ene (kT, f _m ) are stored in the third storage means 1006. Feature amount calculation means 10
07 calculates the feature amount of Sig (t + kT, f _m ).
In this embodiment, as the feature amount, pitch, onset, onset synchronism, offset, offset synchronism,
Calculate continuity and harmonicity.

FIG. 6 is a detailed diagram of the feature amount calculation means 1007 in this embodiment. 1301 is a pitch evaluation measure calculation means, 1302 is an onset evaluation measure calculation means, 1303 is an offset evaluation measure calculation means, 13
Reference numeral 04 is a continuity evaluation measure calculation means, 1305 is a harmonic evaluation measure calculation means, 1306 is an onset synchronization evaluation measure calculation means, and 1307 is an offset synchronization evaluation measure calculation means. is there. In addition, as an evaluation measure, S
Probability that each feature quantity is present is used as ig (t + kT, f _m ).

The means for calculating the feature quantity evaluation measure will be described in detail below. Pitch evaluation measure calculation means 1
In 301, a range in which a pitch can exist with respect to f _p (kT) is stochastically defined, and the probability that Sig (kT, f _m ) is the pitch is calculated. FIG. 7 is an example of a probability distribution of frequencies in which pitch may exist. If the probability that the pitch exists for f _p (kT) has a Gaussian distribution, the probability that the pitch exists for Sig (kT, f _m ) _{Rel_pitch} (kT, f _m ) is given by ( _Equation 9). .

FIG. 8 is a spectrum diagram when a harmonic compound sound composed of two components continues from time k = 2 to k = 6. With respect to this signal, the following means for calculating each likelihood will be described.

FIG. 9 is a schematic diagram relating to auditory perception with respect to changes in loudness. When another signal is added to one signal as shown in FIG. 9A, there is a change in the loudness of the sound. In the onset evaluation measure calculation means 1302, the human perception is assured that this change is not due to the time change of one sound but due to the addition of another different signal as shown in FIG. 9B. Based on the above, this is the onset certainty. The probability of the onset is calculated from the output of the third storage unit 1006. An example of the calculation method will be given below. The loudness of the sound that changes just the psychologically continuous judgment or whether a new sound is added is being investigated (19
Proceedings of the 94th Autumn Meeting of ASJ 2-7-18). Using the output of the energy calculating means 1005, the probability of the onset is determined from the average of the energy at the current observation time and the energy up to the time before it. FIG.
FIG. 6 is an explanatory diagram for a method of calculating the probability of onset. The average of the energy up to the previous time is the average of the segments of about 200 ms section (16 segments) in the current observation frequency and two frequency bands close to the current frequency. Sig (kT, f _m) the probability Rel _{_on} the onset of (E _Dif) is calculated by the equation (10).

FIG. 11 is a diagrammatic representation of (Equation 10). The onset certainty is calculated only in the pitch band until an onset is detected in the pitch band. 12 and 13 are explanatory views for explaining the certainty of the onset for the signal of FIG. In FIG. 12, at time k = 2, the probability of the onset in the pitch band becomes 1.0 and exceeds the threshold value 0.8. Therefore, from time k = 2, the calculation of the probability of the onset in the entire frequency band is started. It is assumed that the onset certainty factor 1.0 can be calculated by (k, m) = (2, 1) and (2, 3). This time,
As shown in FIG. 13, the probability of onset 1.0 is assigned to (k, m) = (2, 1) and (2, 3).

Offset evaluation measure calculation means 1303
At, the accuracy of the offset is calculated. There is a change in loudness when one signal is removed from the two signals. Therefore, the offset is determined by using the output of the energy calculating means 1005. The offset probability _{Rel_off} (E _Dif ) is _{given by} ( _Equation 11) as in the method of calculating the onset reliability.

The probability of offset is calculated only in the pitch frequency band from the detection of the onset to the detection of the offset in the pitch band. 14 and 15 are explanatory views for explaining the certainty of the offset with respect to the signal of FIG. In FIG. 14, the probability of the offset in the pitch band is 1.0 at time k = 6, which exceeds a certain threshold value 0.9. Therefore, at time k = 6, the accuracy of the offset is calculated for all frequency bands.
If the probability of offset 1.0 can be calculated by (k, m) = (6, 1) and (6, 3), the probability of offset is assigned as shown in FIG.

The continuity evaluation measure calculating means 1304 uses as an index the certainty that when a loudness of a sound changes with time, it is the change with time of one signal. The certainty of the continuity is calculated by the output of the third storage unit 1006. FIG. 16 is an explanatory diagram for a method of calculating the likelihood of continuity. In FIG. 16, the average of the energy of the current observation segment in about 200 ms interval (16 segments) before the current observation time including the current observation time, the current observation frequency, and the current observation time in two frequency bands adjacent to the frequency are included. About 20 before the current observation time
The probability of the continuity is determined from the average of the energies in the 0 ms section (16 segments). The probability of continuity _{Rel_con} (E _Dif ) is defined by ( _Equation 12).

FIG. 17 is a diagrammatic representation of (Equation 12). The likelihood of continuity is calculated until the offset is detected after the onset is detected. FIG. 18 is an explanatory diagram for explaining the certainty of the continuity with respect to the signal of FIG. When the onset is detected at the time k = 2 and the offset is detected at the time k = 6, the certainty of the continuity is calculated from the time k = 3 to 5, and the continuity of each (k, m) is calculated. Probability is calculated as shown in FIG.

In the calculation unit 1305 of the evaluation measure of harmonic property, the likelihood of Sig (t + kT, f _m ) being a perceptual harmonic relationship with f _p (kT) is regarded as the certainty of harmonic property. A frequency distribution that subjectively senses harmonics is determined from a human psychological measurement function centering on a frequency section that is an integral multiple of the pitch. Psychological measurement functions are published in IEICE Transactions '94 /
5Vol.J77-A No.5 Determined by referring to the data published in 731-740. Pitch f _p (kT) at the center frequency f _m of the frequency band
The deviation u from the frequency of p times (p is an integer) is (Equation 13)
Is required.

Distribution function Rel of the above-mentioned probability of harmonicity
_harm (u) is defined using u as in (Equation 14).

FIG. 19 is a diagrammatic representation of (Equation 14). FIG. 20 is an explanatory diagram for explaining the certainty of the harmonic property with respect to the signal of FIG. In FIG. 20, the probability of pitch is 1.

The onset synchronicity evaluation measure calculating means 1306 judges that the onset at Sig (t + kT, f _m ) is subjectively simultaneous on the basis of the onset time of f _p (kT). The synchronicity is determined based on the time distribution. Perceptual onset synchronism is the distribution of certainty of synchronism Rel_sync based on the data published in IEICE Transactions '94 / 5 Vol.J77-A No.5 731-740.
(KT, f _m ) is defined by (Equation 15). Sig (t +
When the onset is detected at kT, f _m ) and the difference in the onset time of the pitch closest to it is v (ms), the certainty of the onset synchronism Rel_sync
(KT, f _m ) can be described by (Equation 15).

FIG. 21 is a diagrammatic representation of (Equation 15). FIG. 22 is an explanatory diagram for explaining the certainty of the onset synchronism with respect to the signal of FIG.

The offset synchronism evaluation measure calculating means 1307 can be calculated by the same method as the onset synchronism evaluation measure. The synchronism of the offset is determined based on the time distribution in which the offset at Sig (t + kT, f _m ) is subjectively judged to be simultaneous based on the offset time of f _p (kT). FIG. 23 is an explanatory diagram for explaining the certainty of the synchronization of the offset with respect to the signal of FIG.

In the weight value calculating means 1008 shown in FIG. 3, the reliability of each feature (pitch, onset, offset, onset synchronism, offset synchronism, harmonicity, continuity) is assigned to the evaluation function. Input, Sig (t + k
The weight value w (t + kT, f _m ) at T, f _m ) is calculated.

Of the above feature amounts, the certainty values are classified into those relating to continuous certainty and those relating to frequency certainty, and the respective certaintyes are integrated. The onset accuracy, the onset synchronization accuracy, the offset accuracy, the offset synchronization accuracy, and the power continuity accuracy are integrated into the continuous accuracy. And Further, the certainty of the pitch, the certainty of the range of variation of the pitch, and the certainty of the harmonic property are integrated into the certainty of the frequency.

FIG. 29 is a flowchart relating to the calculation of the evaluation value. At the time when the onset is detected, the continuous certainty is given by the product of the certainty of the onset and the certainty of the synchronism, and is calculated by (Equation 21).

[0184]

(Equation 21)

In the time from the detection of the onset to the detection of the offset, the continuous probability Rel
_{_t} (kT, f _m ) is given by the product of the continuous certainty of the previous time and the certainty of the continuity of the previous time, and is calculated by (Equation 22).

[0186]

(Equation 22)

FIG. 30 shows an example of the distribution of the successive probability. Also, the frequency accuracy _{Rel_f} (k
T, f _m ) can be expressed by (Equation 23).

[0188]

(Equation 23)

The importance of the continuous certainty and the frequency certainty is g_t (kT, f _m ), g_f (kT,
f _m) to.

In this embodiment, λ-fuzzy integration is used as the evaluation function. At this time, the comprehensive evaluation value w (kT, f _m ) can be calculated by (Equation 24).

[0191]

(Equation 24)

FIG. 31 shows an example of the calculation result of the comprehensive evaluation value. _{However, g_t (kT, f m)} = g_f (kT, f m),
λ = 0.

The multiplication means 1009 of FIG. 3 uses Sig (t + k
T, f _m ) is multiplied by the evaluation value w (kT, f _m ). The output of the multiplication means 190 is stored in the fourth storage means 1010. The summing means 1011 is w for all k and m.
(KT, f _m ) × Sig (t + kT, f _m ) is summed.
The output of the summing means 1011 is an acoustic signal in which noise is suppressed.

Note that it is not necessary to use all the feature quantities if the harmonic content or the continuity is included in the feature quantities. Further, a feature amount other than the feature amount may be newly combined. Also, the signal time division means 10
The time width δt of 00 and the division time period T are δt = 13.
ms, T = 26 ms, and the shape is not limited to the Hanning window. The pitch extracting means 1001 is not limited to the cepstrum method and the autocorrelation method. In addition, the frequency band division filter 1003
The frequency band of is not limited to 20 Hz. Also,
The threshold of the probability of the onset is not limited to 0.8. Further, the threshold of the certainty of the offset is 0.9
Not limited to The method of calculating the probability of the onset is not limited to (Equation 7). The method of calculating the certainty of the offset is not limited to (Equation 11). The method of calculating the certainty of continuity is not limited to (Equation 12). In addition, the method of calculating the above-mentioned probability of harmonicity is
It is not limited to 3) and (Equation 14). Further, the method of calculating the certainty of onset synchronism and the method of calculating the certainty of offset synchronism are not limited to (Equation 15). Also,
The method for calculating the comprehensive evaluation value is not limited to λ-fuzzy integration, and other means such as Sugeno integration, Choquet integration, and inverse Sugeno integration may be used.

As in the ninth embodiment, the frequency band division filter 1003 may use Fourier transform, and the energy calculation means 1006 may use effective value calculation means.

(Embodiment 7) An embodiment of the signal separation device of the present invention will be described below with reference to the drawings.

The main block diagram of the seventh embodiment is shown in FIG.
It is. The operation of the noise suppression device configured as above will be described. The signal time division means 1000 divides the acoustic signal into time segments of δt for each time T. Since the input signal is a voice signal, δt is 13 ms and T is 26
ms. If the shape of the signal time division means 1000 is a Hanning window, the signal time division means W (t +).
kT) can be expressed by (Equation 1).

The output signal of the signal time division means W (t + kT) is Sig (t + kT) (t = 0 to δt). If the input voice is Speech (t + kT), S
ig (t + kT) is represented by (Equation 2).

The pitch extraction means 1001 extracts the pitch of Sig (t + kT) using a plurality of extraction methods. FIG.
4 is a detailed view of the pitch extracting means 1001. FIG.
4, 3810 is the first pitch extracting means, 382.
Reference numeral 0 is a second pitch extracting means, and 3830 is a means for calculating an estimated pitch value f _p (t). First pitch extracting means 3
Pitch f _p1 (kT) by 810, the pitch f _p2 the second pitch extracting means 3820 (kT) is calculated. At this time, the calculating means 3830 of the estimated value of the pitch f _p (t), the estimated value of the pitch f _p (kT) is calculated by equation (8).

In FIG. 3, the first storage means 1002
Is f _p1 (k obtained by the pitch extracting means 1001)
T), f _p2 (kT) and f _p (kT) are stored.

The frequency band division filter 1003 divides Sig (t + kT) into N frequency bands of 20 Hz. The output of the frequency band division filter 1003 is Sig.
It is described as (t + kT, f _m ) (m = 1, 2, ...
N). f _m is the center frequency of the band. The N narrow band signals Sig (t + kT, f _m ) are stored in the second storage means 100.
Stored in 4. In the energy calculation means 1005, each S
The energy of ig (t + kT, f _m ) is calculated. The N outputs of the energy calculating means 1005 are converted into Ene (k
T, f _m ). Ene (kT, f _m ) is (Equation 3)
Is calculated.

The N pieces of Ene (kT, f _m ) are stored in the third storage means 1006. Feature amount calculation means 10
07 calculates the feature amount of Sig (t + kT, f _m ).
In this embodiment, as the feature amount, pitch, onset, onset synchronism, offset, offset synchronism,
Calculate continuity and harmonicity.

FIG. 6 is a diagram. FIG. 6 is a diagram showing the feature quantity calculating means 10 of this embodiment.
It is a detailed view of 07. 1301 is a pitch evaluation measure calculation means; 1302 is an onset evaluation measure calculation means;
Reference numeral 1303 denotes a means for calculating an offset evaluation measure, 1304.
Is a continuity evaluation metric calculation means, 1305 is a harmonic evaluation metric calculation means, 1306 is an onset synchronicity evaluation metric calculation means, and 1307 is an offset synchronicity evaluation metric calculation means. . In addition, as an evaluation measure, Sig
The probability that each feature quantity exists at (t + kT, f _m ) is used. The calculation means of the evaluation measure of each feature will be described in detail below.

The means for calculating the pitch evaluation measure is f _p1 (kT) and f _p2 (k) calculated by the pitch extracting means 1001.
The probability of the pitch f _p (kT) is determined from the difference of T). When the difference between f _p1 (kT) and f _p2 (kT) is 0, the pitch is accurately extracted, and the probability of the pitch f _p (kT) is 1. The probability decreases as the difference between f _p1 (kT) and f _p2 (kT) increases. If the probability of pitch is Gaussian distributed by the value of f _p1 (kT) −f _p2 (kT) as shown in FIG. 25, the probability of pitch Rel (f
_p1 (kT) -f _p2 (kT)) is determined by (Equation 17).

At the same time, as shown in FIG. 26, the range in which a certain pitch can exist is stochastically defined with respect to the obtained estimated value f _p (kT) of the pitch, and Sig (kT, f _m ) is the pitch. Calculate the likelihood. Estimated value of pitch f _p (kT)
_{Assuming that the} probability that the pitch exists has a Gaussian distribution, the probability that the pitch exists in Sig (kT, f _m ) _{Rel_pitch} (kT, f _m ) is given by ( _Equation 9).

FIG. 8 is a spectrum diagram when the harmonic compound sound consisting of two components continues from time k = 2 to k = 6. With respect to this signal, the following means for calculating each likelihood will be described.

FIG. 9 is a schematic diagram relating to auditory perception with respect to changes in loudness. When another signal is added to one signal as shown in FIG. 9A, there is a change in the loudness of the sound. The onset evaluation measure calculation means 1302 confirms that the change is not due to the time change of one sound but due to the addition of another different signal as shown in FIG. 9B. Determined based on perception, this is the onset certainty. The probability of the onset is calculated from the output of the third storage unit 1006. An example of the calculation method will be given below.
The volume of the sound that changes just the criterion for psychological continuity or whether a new sound is added is being investigated (Proceedings of the Autumn Meeting of the Acoustical Society of Japan, 1994 2-7-18).
Using the output of the energy calculation means 1005, the probability of the onset is determined from the average of the energy at the current observation time and the energy before the current time. FIG.
0 is an explanatory diagram for a method of calculating the probability of onset. The average of the energy up to the previous time is the average of the segment of about 200 ms section (16 segments) in the current observation frequency and two frequency bands close to the current frequency. Sig (kT, f _m) the probability Rel _{_on} the onset of (E _Dif) is calculated by the equation (10).

FIG. 11 is a diagrammatic representation of (Equation 10). Onset certainty is calculated only in the pitch band until the onset is detected in the pitch band. FIG.
13 and FIG. 13 are explanatory diagrams for explaining the certainty of the onset with respect to the signal of FIG. Time k in FIG.
= 2, the probability of the onset of the pitch band is 1.
Since it becomes 0 and exceeds the threshold value 0.8, the calculation of the probability of the onset is started from the time k = 2 in the entire frequency band.
It is assumed that the onset certainty factor 1.0 can be calculated by (k, m) = (2, 1) and (2, 3). At this time, FIG.
As in 3, the onset certainty of 1.0 is assigned to (k, m) = (2, 1) and (2, 3).

Offset evaluation measure calculation means 1303
At, the accuracy of the offset is calculated. There is a change in loudness when one signal is removed from the two signals. Therefore, the offset is determined by using the output of the energy calculating means 1005. The offset probability _{Rel_off} (E _Dif ) is _{given by} ( _Equation 11) as in the method of calculating the onset reliability.

The accuracy of the offset is calculated only in the pitch frequency band from the detection of the onset to the detection of the offset in the pitch band. FIG. 14 and FIG.
FIG. 5 is an explanatory diagram for explaining the probability of the offset with respect to the signal of FIG. In FIG. 14, the probability of the offset in the pitch band is 1.0 at time k = 6, which exceeds a certain threshold value 0.9. Therefore, at time k = 6, the accuracy of the offset is calculated for all frequency bands. If the probability of offset 1.0 can be calculated by (k, m) = (6, 1) and (6, 3), the probability of offset is assigned as shown in FIG.

The continuity evaluation measure calculating means 1304 uses as an index the certainty that, when the loudness of the sound changes with time, it is the change with time of one signal. The certainty of the continuity is calculated by the output of the third storage unit 1006. FIG. 16 is an explanatory diagram for a method of calculating the likelihood of continuity. In FIG. 16, the average of the energy of the current observation segment in about 200 ms interval (16 segments) before the current observation time including the current observation time, the current observation frequency, and the current observation time in two frequency bands adjacent to the frequency are included. About 20 before the current observation time
The probability of the continuity is determined from the average of the energies in the 0 ms section (16 segments). The probability of continuity _{Rel_con} (E _Dif ) is defined by ( _Equation 12).

FIG. 17 is a diagrammatic representation of (Equation 12). The likelihood of continuity is calculated until the offset is detected after the onset is detected. FIG. 18 is an explanatory diagram for explaining the certainty of the continuity with respect to the signal of FIG. When the onset is detected at the time k = 2 and the offset is detected at the time k = 6, the certainty of the continuity is calculated from the time k = 3 to 5, and the continuity of each (k, m) is calculated. Probability is calculated as shown in FIG.

In the calculation unit 1305 of the evaluation measure of harmonic property, the probability that Sig (t + kT, f _m ) has a perceptual harmonic relationship with f _p (kT) is regarded as the certainty of harmonic property. A frequency distribution that subjectively senses harmonics is determined from a human psychological measurement function centering on a frequency section that is an integral multiple of the pitch. Psychological measurement functions are published in IEICE Transactions '94 /
5Vol.J77-A No.5 Determined by referring to the data published in 731-740. The pitch f _p (k at the center frequency f _m of the frequency band
The deviation u from the frequency that is p times (T is p) (p is an integer) is (Equation 1)
Required in 3).

Distribution function Rel of the above-mentioned harmonic accuracy
_harm (u) is defined using u as in (Equation 14).

FIG. 19 is a diagrammatic representation of (Equation 14). FIG. 20 is an explanatory diagram for explaining the certainty of the harmonic property with respect to the signal of FIG. In FIG. 20, the probability of pitch is 1.

Onset synchronicity evaluation measure calculation means 1306 determines that the onsets at Sig (t + kT, f _m ) are subjectively simultaneous based on the onset time of f _p (kT). The synchronicity is determined based on the time distribution. The perceptual onset synchronism is the distribution of the certainty of the synchronism, Rel_sync, with reference to the data published in IEICE Transactions '94 / 5 Vol.J77-A No.5731-740.
(KT, f _m ) is defined by (Equation 15). Sig (t +
When the onset is detected at kT, f _m ) and the difference in the onset time of the closest pitch is v (ms), the certainty of the onset synchronism Rel_sync (k
T, f _m ) can be described by (Equation 15).

FIG. 21 is a diagrammatic representation of (Equation 15). FIG. 22 is an explanatory diagram for explaining the certainty of the onset synchronism with respect to the signal of FIG.

The offset synchronism evaluation measure calculating means 1307 can be calculated by the same method as the onset synchronism evaluation measure. The synchronism of the offset is determined based on the time distribution in which the offset at Sig (t + kT, f _m ) is subjectively judged to be simultaneous based on the offset time of f _p (kT). FIG. 23 is an explanatory diagram for explaining the certainty of the synchronization of the offset with respect to the signal of FIG.

In the weight value calculation means 1008 shown in FIG. 3, the reliability of each feature (pitch, onset, offset, onset synchronism, offset synchronism, harmonicity, continuity) is assigned to the evaluation function. Input, Sig (t + k
The weight value w (t + kT, f _m ) at T, f _m ) is calculated.
Among the respective feature amounts, the certainty values are classified into those relating to the continuous certainty and those relating to the frequency certainty, and the respective certainty is integrated. The onset accuracy, the onset synchronization accuracy, the offset accuracy, the offset synchronization accuracy, and the power continuity accuracy are integrated into the continuous accuracy. And Further, the certainty of the pitch, the certainty of the range of variation of the pitch, and the certainty of the harmonic property are integrated into the certainty of the frequency.

FIG. 29 is a flow chart regarding the calculation of the evaluation value. At the time when the onset is detected, the continuous certainty is given by the product of the certainty of the onset and the certainty of the synchronism, and is calculated by (Equation 21).

In the time from the detection of the onset to the detection of the offset, the successive probability Rel
_{_t} (kT, f _m ) is given by the product of the continuous certainty of the previous time and the certainty of the continuity of the previous time, and is calculated by (Equation 22).

FIG. 30 shows an example of the distribution of the above-mentioned successive probability. Also, the frequency accuracy _{Rel_f} (k
T, f _m ) can be expressed by (Equation 23).

The importance of the continuous certainty and the frequency certainty is g_t (kT, f _m ), g_f (kT,
f _m) to.

In this embodiment, λ-fuzzy integration is used as the evaluation function. At this time, the comprehensive evaluation value w (kT, f _m ) can be calculated by (Equation 24).

FIG. 31 shows an example of the calculation result of the comprehensive evaluation value. _{However, g_t (kT, f m)} = g_f (kT, f m),
λ = 0.

The multiplication means 1009 of FIG. 3 uses Sig (t + k
T, f _m ) is multiplied by the evaluation value w (kT, f _m ). The output of the multiplication means 190 is stored in the fourth storage means 1010. The summing means 1011 is for all k and m,
Sum up w (kT, f _m ) × Sig (t + kT, f _m ). The output of the summing means 1011 is an acoustic signal in which noise is suppressed.

It is not necessary to use all of the feature quantities if the harmonic content or the continuity of the feature quantities is included. Further, a feature amount other than the feature amount may be newly combined. Also, the signal time division means 10
The time width δt of 00 and the division time period T are δt = 13.
ms, T = 26 ms, and the shape is not limited to the Hanning window. The pitch extracting means 1001 is not limited to the cepstrum method and the autocorrelation method. In addition, the frequency band division filter 100
The frequency band of 3 is not limited to 20 Hz. Further, the threshold of the probability of the onset is not limited to 0.8. Also, the threshold of the certainty of the offset is 0.
It is not limited to 9. The method of calculating the probability of onset is not limited to (Equation 10). The method of calculating the certainty of the offset is not limited to (Equation 11). The method of calculating the certainty of continuity is not limited to (Equation 12). In addition, the method of calculating the above-mentioned probability of harmonicity is
3) Not limited to (Equation 14). The method of calculating the certainty of onset synchronism and the method of calculating the certainty of offset synchronism are not limited to (Equation 15).

As in the ninth embodiment, the frequency band filter 1003 may use Fourier transform, and the energy calculating means 1006 may use effective value calculating means.

(Embodiment 8) An embodiment of the signal separation device of the present invention will be described below with reference to the drawings.

The main block diagram of the eighth embodiment is shown in FIG.
It is. The operation of the noise suppression device configured as above will be described. The signal time division means 1000 divides the acoustic signal into time segments of δt for each time T. Since the input signal is a voice signal, δt is 13 ms and T is 26
ms. If the shape of the signal time division means 1000 is a Hanning window, the signal time division means W (t +).
kT) can be expressed by (Equation 1).

The output signal of the signal time division means W (t + kT) is Sig (t + kT) (t = 0 to δt). If the input voice is Speech (t + kT), S
ig (t + kT) is represented by (Equation 2).

The pitch extraction means 1001 extracts the pitch of Sig (t + kT) using a plurality of extraction methods. FIG.
FIG. 3 is a detailed view of the pitch extracting means 1001. In FIG. 5, reference numeral 1200 is a first pitch extraction means, 1201 is a second pitch extraction means, 1202 is an nth pitch extraction means, 1203 is a pitch estimated value f _p (t) calculation means. The pitch f by the first pitch extracting means 1200.
_p1 (kT) is the second pitch extraction means 1201 and the pitch f _p2 (kT) is the nth pitch extraction means 1202.
Assume that the pitch f _pn (kT) is calculated by. At this time, the estimated value f _p calculation means 1203 (t) of the pitch, the estimated value of the pitch f _p (kT) is calculated by equation (8).

Also, the variance V _pitch (kT) of f _pn (kT) is calculated from f _p1 (kT) by (Equation 18).

In FIG. 3, the first storage means 1002 stores f _p (kT) obtained by the pitch extraction means 1001.
And the variance V _pitch (kT).

The frequency band dividing filter 1003 divides Sig (t + kT) into N frequency bands of 20 Hz. The output of the frequency band division filter 1003 is Si
It is described as g (t + kT, f _m ) (m = 1, 2, ...
・ N). f _m is the center frequency of the band. The N narrow band signals Sig (t + kT, f _m ) are stored in the second storage means 100.
Stored in 4. In the energy calculation means 1005, each S
The energy of ig (t + kT, f _m ) is calculated. The N outputs of the energy calculating means 1005 are converted into Ene (k
T, f _m ). Ene (kT, f _m ) is (Equation 3)
Is calculated.

The N pieces of Ene (kT, f _m ) are stored in the third storage means 1006. Feature amount calculation means 10
07 calculates the feature amount of Sig (t + kT, f _m ).
In the present embodiment, pitch, onset, onset synchronism, offset, offset synchronism, continuity, and harmonicity are calculated as feature amounts.

FIG. 6 is a detailed diagram of the feature amount calculating means 1007 of FIG. 1301 is a pitch evaluation measure calculation means, 1302 is an onset evaluation measure calculation means, 13
Reference numeral 03 is an offset evaluation measure calculation unit, 1304 is a continuity evaluation measure calculation unit, 1305 is a harmonic evaluation measure calculation unit, 1306 is an onset synchronization evaluation measure calculation unit, and 1307 is an offset. Is a means of calculating an evaluation measure of synchronism. Also, as an evaluation measure, Sig (t
+ KT, f _m ) The probability that each feature quantity exists is used.

The calculation means of the evaluation measure of each feature will be described in detail below. The pitch evaluation measure calculation means is the variance V calculated by the pitch extraction means 1001.
_The probability of the pitch f _p (kT) estimated from the _pitch (kT) is determined. When V _pitch (kT) is 0, it is assumed that the pitch is accurately extracted, and the probability of the pitch f _p (kT) is 1. The probability decreases as V _pitch (kT) increases. The pitch accuracy is V as shown in Fig. 27.
If Gaussian distribution is used depending on the value of _pitch (kT),
The certainty of the pitch Rel (V _pitch (kT)) is determined by (Equation 19).

At the same time, with respect to the estimated value f _p (kT) of the pitch obtained as shown in FIG. 28, the range in which a certain pitch can exist is stochastically defined, and the probability that Sig (kT, f _m ) is the pitch. To calculate. If the probability that the pitch exists with respect to the estimated value f _p (kT) of the pitch has a Gaussian distribution, the probability that the pitch exists in Sig (kT, f _m ) _{Rel_pitch} (kT, f _m ) is ( _Equation 20). ) Is given.

FIG. 8 is a spectrum diagram when a harmonic composite sound composed of two components continues from time k = 2 to k = 6. With respect to this signal, the following means for calculating each likelihood will be described.

FIG. 9 is a schematic diagram relating to auditory perception with respect to changes in loudness. When another signal is added to one signal as shown in FIG. 9A, there is a change in the loudness of the sound. In the onset evaluation measure calculation means 1302, the human perception is assured that this change is not due to the time change of one sound but due to the addition of another different signal as shown in FIG. 9B. Based on the above, this is the onset certainty. The probability of the onset is calculated from the output of the third storage unit 1006. An example of the calculation method will be given below. The loudness of the sound that changes just the psychologically continuous judgment or whether a new sound is added is being investigated (19
Proceedings of the 94th Autumn Meeting of ASJ 2-7-18). Using the output of the energy calculation means 1005, the probability of the onset is determined from the average of the energy at the current observation time and the energy before the current time. FIG. 10 is an explanatory diagram for a method of calculating the probability of onset. The average of the energies up to the previous time is about 2 in the current observation frequency and two frequency bands close to that frequency.
The average of the segments in the 00 ms section (16 segments) is used. Sig (kT, f _m) the probability Rel _{_on} the onset of (E _Dif) is calculated by the equation (10).

FIG. 11 is a diagrammatic representation of (Equation 10). The onset certainty is calculated only in the pitch band until an onset is detected in the pitch band. 12 and 13 are explanatory views for explaining the certainty of the onset for the signal of FIG. In FIG. 12, at time k = 2, the probability of the onset in the pitch band becomes 1.0 and exceeds the threshold value 0.8. Therefore, from time k = 2, the calculation of the probability of the onset in the entire frequency band is started. It is assumed that the onset certainty factor 1.0 can be calculated by (k, m) = (2, 1) and (2, 3). This time,
As shown in FIG. 13, the probability of onset 1.0 is assigned to (k, m) = (2, 1) and (2, 3).

Offset evaluation measure calculation means 1303
At, the accuracy of the offset is calculated. There is a change in loudness when one signal is removed from the two signals. Therefore, the offset is determined by using the output of the energy calculating means 1005. The offset probability _{Rel_off} (E _Dif ) is _{given by} ( _Equation 11) as in the method of calculating the onset reliability.

The certainty of the offset is calculated only in the pitch frequency band from the detection of the onset to the detection of the offset in the pitch band. FIGS. 214 and 15 are explanatory views for explaining the probability of the offset with respect to the signal of FIG.

In FIG. 14, at time k = 6, the probability of the offset in the pitch band is 1.0, which exceeds a certain threshold value 0.9. Therefore, at time k = 6, the accuracy of the offset is calculated for all frequency bands. . (K,
If m) = (6,1) and (6,3) can calculate the offset certainty of 1.0, the offset certainty is assigned as shown in FIG.

The continuity evaluation measure calculating means 1304 uses as an index the certainty that, when the loudness of the sound changes with time, it is the change with time of one signal. The certainty of the continuity is calculated by the output of the third storage unit 1006. FIG. 16 is an explanatory diagram for a method of calculating the likelihood of continuity. In FIG. 16, the average of the energy of the current observation segment in about 200 ms interval (16 segments) before the current observation time including the current observation time, the current observation frequency, and the current observation time in two frequency bands adjacent to the frequency are included. About 20 before the current observation time
The probability of the continuity is determined from the average of the energies in the 0 ms section (16 segments). The probability of continuity _{Rel_con} (E _Dif ) is defined by ( _Equation 12).

FIG. 17 is a diagrammatic representation of (Equation 12). The likelihood of continuity is calculated until the offset is detected after the onset is detected. FIG. 18 is an explanatory diagram for explaining the certainty of the continuity with respect to the signal of FIG. When the onset is detected at the time k = 2 and the offset is detected at the time k = 6, the certainty of the continuity is calculated from the time k = 3 to 5, and the continuity of each (k, m) is calculated. Probability is calculated as shown in FIG.

The calculation unit 1305 of the evaluation measure of the harmonic property determines the probability of the harmonic property by the probability that Sig (t + kT, f _m ) is perceptually harmonic with respect to the f _p (kT). A frequency distribution that subjectively senses harmonics is determined from a human psychological measurement function centering on a frequency section that is an integral multiple of the pitch. The psychometric function is determined by referring to the data published in IEICE Transactions '94 / 5 Vol.J77-A No.5 731-740. Deviation u from the frequency of p times (p is an integer) the pitch f _p (kT) at the center frequency f _m of the frequency band
Is calculated by (Equation 13).

Distribution function Rel of the above-mentioned harmonic accuracy
_harm (u) is defined using u as in (Equation 14).

FIG. 19 is a diagrammatic representation of (Equation 14). FIG. 20 is an explanatory diagram for explaining the certainty of the harmonic property with respect to the signal of FIG. In FIG. 20, the probability of pitch is 1.

Onset synchronicity evaluation measure calculation means 1306 determines that the onset at Sig (t + kT, f _m ) is subjectively simultaneous based on the onset time of f _p (kT). The synchronicity is determined based on the time distribution. The perceptual onset synchrony is based on the data published in IEICE Transactions '94 / 5 Vol.J77-A No.5 731-740.
c (kT, f _m ) is defined by (Equation 15). Sig (t +
When the onset is detected at kT, f _m ) and the difference in the onset time of the pitch closest to it is v (ms), the certainty of the onset synchronism Rel_sync
(KT, f _m ) can be described by (Equation 15).

FIG. 21 is a diagrammatic representation of (Equation 15). FIG. 22 is an explanatory diagram for explaining the certainty of the onset synchronism with respect to the signal of FIG.

The offset synchronism evaluation measure calculating means 1307 can be calculated by the same method as the onset synchronism evaluation measure. The synchronism of the offset is determined based on the time distribution in which the offset at Sig (t + kT, f _m ) is subjectively judged to be simultaneous based on the offset time of f _p (kT). FIG. 23 is an explanatory diagram for explaining the certainty of the synchronization of the offset with respect to the signal of FIG.

In the weight value calculating means 1008 shown in FIG. 3, the reliability of each feature (pitch, onset, offset, onset synchronism, offset synchronism, harmonicity, continuity) is assigned to the evaluation function. Input, Sig (t + k
The weight value w (t + kT, f _m ) at T, f _m ) is calculated.

[0255] Of the characteristic amounts, the ones are classified into those related to the continuous certainty and those related to the frequency certainty, and the respective certainties are integrated. The onset accuracy, the onset synchronization accuracy, the offset accuracy, the offset synchronization accuracy, and the power continuity accuracy are integrated into the continuous accuracy. And Further, the certainty of the pitch, the certainty of the range of variation of the pitch, and the certainty of the harmonic property are integrated into the certainty of the frequency.

FIG. 29 is a flow chart regarding the calculation of the evaluation value. At the time when the onset is detected, the continuous certainty is given by the product of the certainty of the onset and the certainty of the synchronism, and is calculated by (Equation 21).

In the time from the detection of the onset to the detection of the offset, the continuous certainty _{Rel_t is set.}
(KT, f _m ) is given by the product of the continuous probability of the previous time and the probability of the continuity of the previous time (Equation 2).
Calculated in 2).

FIG. 4400 is an example of the distribution of the above-mentioned successive probability. Also, the frequency accuracy Rel
_{_f} (kT, f _m) can be expressed by equation (23).

The importance of the continuous certainty and the frequency certainty is g_t (kT, f _m ), g_f (kT,
f _m) to.

In this embodiment, λ-fuzzy integration is used as the evaluation function. At this time, the comprehensive evaluation value w (kT, f _m ) can be calculated by (Equation 24).

FIG. 31 shows an example of the calculation result of the comprehensive evaluation value. _{However, g_t (kT, f m)} = g_f (kT, f m),
λ = 0.

The multiplication means 1009 of FIG. 3 uses Sig (t + k
T, f _m ) is multiplied by the evaluation value w (kT, f _m ). The output of the multiplication means 1009 is stored in the fourth storage means 1010. The summing means 1011 sums w (kT, f _m ) × Sig (t + kT, f _m ) for all k, m. The output of the summing means 1011 is an acoustic signal in which noise is suppressed.

Note that it is not necessary to use all the feature quantities if the harmonic content or the continuity is included in each of the feature quantities. Further, a feature amount other than the feature amount may be newly combined. Also, the pitch extracting means 1001
Is not limited to the cepstrum method and the autocorrelation method.
The frequency band of the frequency band division filter 1003 is not limited to 20 Hz. Further, the threshold of the probability of the onset is not limited to 0.8. Further, the threshold value of the certainty of the offset is not limited to 0.9.
In addition, the calculation method of the probability of the onset is (Equation 1
It is not limited to 0). The method of calculating the certainty of the offset is not limited to (Equation 11). Further, the method of calculating the probability of harmonicity is not limited to (Equation 13) and (Equation 14). In addition, the method of calculating the onset synchronicity probability and the method of calculating the offset synchronism probability are described in (Equation 1)
It is not limited to 5). Further, in the calculation of the probability that the pitch exists, it may be defined by a distribution other than the Gaussian distribution.

As in the ninth embodiment, the frequency band filter 1003 may use Fourier transform, and the energy calculating means 1006 may use an effective value calculating means.

(Embodiment 9) An embodiment of the signal separation device of the present invention will be described below with reference to the drawings.

FIG. 32 is a block diagram showing the structure of the noise suppressing apparatus according to the ninth embodiment of the present invention. In FIG. 32, reference numeral 9000 is a signal time division means for performing time division of a signal, 9001 is a pitch extraction means, 9002 is a first storage means, 9003 is a Fourier transform means, 9004 is a second storage means, and 9005 is an effective value. Calculation means, 9006 is third storage means, 9007 is feature amount calculation means, 900
8 is a weight value calculation means, 9009 is a multiplication means, 9010
Is the inverse Fourier transform means, 9011 is the fourth storage means, 9
012 is a summing means.

The operation of the noise suppressor configured as above will be described. The signal time division means 9000 divides the acoustic signal into time segments of δt for each time T.
Since the input signal is a voice signal, δt is 13 ms and T is 26 ms. If the shape of the signal time division means 9000 is a Hanning window, then the signal time division means W will be described.
(T + kT) can be expressed by (Equation 1).

The output signal of the signal time division means W (t + kT) is Sig (t + kT) (t = 0 to δt). If the input voice is Speech (t + kT), S
ig (t + kT) is represented by (Equation 2).

The pitch extracting means 9001 uses Sig (t + k
Extract the pitch of T). The pitch is the cepstrum method,
It is extracted using an autocorrelation method or the like. First storage means 9
002 is the pitch obtained by the pitch extraction means 9001.
Chi f_pStore (kT). Also, the Fourier transform means 9
003 Fourier transforms Sig (t + kT). this
The frequency resolution at that time is 20 Hz. Fourier transform
The output of the means 9003 is Sig (t + kT, f_m) And description
(M = 1, 2, ... N). f_mIs the frequency
You. N frequency sampling points Sig (t + kT, f_m)
Are stored in the second storage means 9004. RMS value calculation
In the means 9005, each Sig (t + kT, f _m) Effective
Calculate the value. N output of the effective value calculation means 9005
Force is Ene (kT, f_m). Ene (kT,
f_m) Is calculated by (Equation 25).

[0270]

(Equation 25)

The N pieces of Ene (kT, f _m ) are stored in the third storage means 9006. Feature amount calculation means 90
07 calculates the feature amount of Sig (t + kT, f _m ).
In the present embodiment, pitch, onset, onset synchronism, offset, offset synchronism, continuity, and harmonicity are calculated as feature amounts.

[0272]

(Equation 26)

FIG. 5 is a characteristic amount calculation means 900 of this embodiment.
FIG. 7 is a detailed view of FIG. 1201 is a pitch evaluation measure calculation means; 1202 is an onset evaluation measure calculation means;
Reference numeral 203 is an offset evaluation measure calculating means, 1204 is a continuity evaluation measure calculating means, 1205 is a harmonic evaluation measure calculating means, 1206 is an onset synchronization evaluation measure calculating means, 1207 is an offset. Of the evaluation measure of the synchronism of Sig (t + k)
The probability that each feature quantity exists in T, f _m ) is used.
The calculation means of the evaluation measure of each feature will be described in detail below.

The pitch evaluation measure calculation means 1201 stochastically defines the range in which the pitch can exist with respect to f _p (kT), and calculates the probability that Sig (kT, f _m ) is the pitch. FIG. 7 is an example of a probability distribution of frequencies in which pitch may exist. If the probability that the pitch exists with respect to f _p (kT) has a Gaussian distribution, the probability that the pitch exists with Sig (kT, f _m ) R
el_pitch (kT, fm) is given by (Equation 9).

FIG. 8 is a spectrum diagram when a harmonic compound sound composed of two components continues from time k = 2 to k = 6. With respect to this signal, the following means for calculating each likelihood will be described.

FIG. 9 is a schematic diagram relating to auditory perception with respect to changes in loudness. When another signal is added to one signal as shown in FIG. 9A, there is a change in the loudness of the sound. The onset evaluation measure calculation means 1202 confirms that the change is not caused by the time change of one sound but by the addition of another different signal as shown in FIG. 9B. Determined based on perception, this is the onset certainty. The probability of the onset is calculated by the output of the third storage means 9006. An example of the calculation method will be given below. The loudness of the sound that changes just the psychologically continuous judgment or whether a new sound is added is being investigated (19
Proceedings of the 94th Autumn Meeting of ASJ 2-7-18). Using the output of the effective value calculation means 9005, the certainty of the onset is determined from the average of the effective values at the current observation time and the effective values up to the previous time. FIG. 10 is an explanatory diagram for a method of calculating the probability of onset. The average of the effective values up to the previous time is about 200 ms interval (16 ms) in the current observation frequency and two frequency bands close to that frequency.
Number of segments). Sig
_{Probability of the onset of} (kT, f _m ) _{Rel_on}
(E _Dif ) is calculated by ( _Equation 10).

FIG. 11 is a diagrammatic representation of (Equation 10). Onset certainty is calculated only in the pitch band until the onset is detected in the pitch band. FIG.
13 and FIG. 13 are explanatory diagrams for explaining the certainty of the onset with respect to the signal of FIG. Time k in FIG.
= 2, the probability of the onset of the pitch band is 1.
Since it becomes 0 and exceeds the threshold value 0.8, the calculation of the probability of the onset is started from the time k = 2 in the entire frequency band.
It is assumed that the onset certainty factor 1.0 can be calculated by (k, m) = (2, 1) and (2, 3). Figure 27 at this time
As in 00, (k, m) = (2, 1) and (2, 3) are assigned the onset certainty of 1.0.

Offset evaluation measure calculation means 1203
At, the accuracy of the offset is calculated. There is a change in loudness when one signal is removed from the two signals. Therefore, the offset is determined by using the output of the effective value calculation means 9005. Likelihood Rel_ _off of the offset (E _Dif) is given by the same method of calculating the likelihood of the onset (11).

The probability of the offset is calculated only in the pitch frequency band from the detection of the onset to the detection of the offset in the pitch band. FIG. 14 and FIG.
FIG. 5 is an explanatory diagram for explaining the probability of the offset with respect to the signal of FIG. In FIG. 14, the probability of the offset in the pitch band is 1.0 at time k = 6, which exceeds a certain threshold value 0.9. Therefore, at time k = 6, the accuracy of the offset is calculated for all frequency bands. If the probability of offset 1.0 can be calculated by (k, m) = (6, 1) and (6, 3), the probability of offset is assigned as shown in FIG.

[0280] The continuity evaluation measure calculating means 1204 uses as an index the certainty that, when the loudness of the sound changes with time, it is the change with time of one signal. The certainty of the continuity is calculated from the output of the third storage unit 9006. FIG. 16 is an explanatory diagram for a method of calculating the likelihood of continuity. In FIG. 16, the average of the effective values of the current observation segment in the approximately 200 ms interval (16 segments) before the current observation time including the current observation time, the current observation frequency, and the current observation time in two frequency bands adjacent to the frequency are shown. About 200m before the current observation time not included
The certainty of the continuity is determined from the average of the effective values in the s section (16 segments). The probability of continuity _{Rel_con} (E _Dif ) is defined by ( _Equation 12).

FIG. 17 is a diagrammatic representation of (Equation 12). The likelihood of continuity is calculated until the offset is detected after the onset is detected. FIG. 18 is an explanatory diagram for explaining the certainty of the continuity with respect to the signal of FIG. When the onset is detected at the time k = 2 and the offset is detected at the time k = 6, the certainty of the continuity is calculated from the time k = 3 to 5, and the continuity of each (k, m) is calculated. Probability is calculated as shown in FIG.

In the harmonic measure evaluation measure calculation means 1205, the probability that Sig (t + kT, f _m ) has a perceptual harmonic relationship with f _p (kT) is regarded as the accuracy of harmonic. A frequency distribution that subjectively senses harmonics is determined from a human psychological measurement function centering on a frequency section that is an integral multiple of the pitch. Psychological measurement functions are published in IEICE Transactions '94 /
5 Determine based on the data published in Vol.J77-A No.5731-740. The pitch f _p (k at the center frequency f _m of the frequency band
The deviation u from the frequency that is p times (T is p) (p is an integer) is (Equation 1)
Required in 3).

Distribution function Rel of the above-mentioned harmonic probability
_harm (u) is defined using u as in (Equation 14).

FIG. 19 is a diagrammatic representation of (Equation 14). FIG. 20 is an explanatory diagram for explaining the certainty of the harmonic property with respect to the signal of FIG. In FIG. 20, the probability of pitch is 1.

The onset synchronicity evaluation measure calculating means 1206 judges that the onset at Sig (t + kT, f _m ) is subjectively simultaneous on the basis of the onset time of f _p (kT). The synchronicity is determined based on the time distribution. The perceptual onset synchronism is the distribution of the certainty of the synchronism, Rel_sync, with reference to the data published in IEICE Transactions '94 / 5 Vol.J77-A No.5731-740.
(KT, f _m ) is defined by (Equation 15). Sig (t +
When the onset is detected at kT, f _m ) and the difference in the onset time of the pitch closest to it is v (ms), the certainty of the onset synchronism Rel_sync
(KT, f _m ) can be described by (Equation 15).

FIG. 21 is a diagrammatic representation of (Equation 15). FIG. 22 is an explanatory diagram for explaining the certainty of the onset synchronism with respect to the signal of FIG.

The offset synchronism evaluation measure calculating means 1207 can be calculated by the same method as the onset synchronism evaluation measure. The synchronism of the offset is determined based on the time distribution in which the offset at Sig (t + kT, f _m ) is subjectively judged to be simultaneous based on the offset time of f _p (kT). FIG. 23 is an explanatory diagram for explaining the certainty of the synchronization of the offset with respect to the signal of FIG.

The weight value calculation means 9008 shown in FIG.
Probability about each of the above characteristics (pitch, onset, offset, synchronism of onset, synchronism of offset, harmonicity, continuity) is input to the evaluation function, and Sig is set.
(T + kT, f _m) weight values _{w (t + kT, f m} ) at is calculated.

[0289] Of the feature amounts, the ones are classified into those related to the continuous certainty and those related to the frequency certainty, and the respective certainties are integrated. The on-set accuracy, the on-set synchronization accuracy, the offset accuracy, the offset synchronization accuracy, and the power continuity accuracy are integrated to obtain the continuous accuracy. Let's be like Further, the certainty of the pitch, the certainty of the range of variation of the pitch, and the certainty of the harmonic property are integrated to be the frequency certainty.

FIG. 29 is a flow chart regarding the calculation of the evaluation value. At the time when the onset is detected, the continuous certainty is given by the product of the certainty of the onset and the certainty of the synchronism, and is calculated by (Equation 21).

In the time from the detection of the onset to the detection of the offset, the continuous probability Rel
_{_t} (kT, f _m ) is given by the product of the continuous certainty of the previous time and the certainty of the continuity with the previous time,
It is calculated by (Equation 22).

FIG. 30 shows an example of the distribution of the successive probability. Also, the frequency accuracy _{Rel_f} (k
T, f _m ) can be expressed by (Equation 23).

[0293] The continuous accuracy and the frequency accuracy are
The importance of karasashi is g_t (kT, f _m), G_f (kT,
f_m).

In this embodiment, λ-fuzzy integration is used as the evaluation function. At this time, the comprehensive evaluation value w (kT, f _m ) can be calculated by (Equation 24).

FIG. 31 shows an example of the calculation result of the comprehensive evaluation value. _{However, g_t (kT, f m)} = g_f (kT, f m),
λ = 0.

The multiplication means 9009 in FIG. 32 is Sig (t +
kT, f _m) and the evaluation value w (kT, multiplying f _m).
The output of the multiplication means 190 is the fourth storage means 9010.
Stored in. The inverse Fourier transform means 9011 uses w (kT, f _m ) × Sig (t + k) for all k and m.
Inverse Fourier transform of T, f _m ). The output of the inverse Fourier transform means 9011 is an acoustic signal in which noise is suppressed.

Note that it is not necessary to use all the feature quantities if the harmonic or the continuity is included in the feature quantities. Further, a feature amount other than the feature amount may be newly combined. Also. Time division means 90 for the signal
The time width δt of 00 and the division time period T are δt = 13.
ms, T = 26 ms, and the shape is not limited to the Hanning window. The pitch extracting means 9001 is not limited to the cepstrum method and the autocorrelation method. The frequency resolution of the Fourier transform means 9003 and the inverse Fourier transform means 9010 is 20 Hz.
It is not limited to. Further, the threshold of the probability of the onset is not limited to 0.8. Further, the threshold of the probability of the onset is not limited to 0.9. The method of calculating the probability of onset is not limited to (Equation 10). The method of calculating the certainty of the offset is not limited to (Equation 11). The method of calculating the certainty of continuity is not limited to (Equation 12). In addition, the calculation method of the probability of the above-mentioned harmonic property is (Equation 13) and (Equation 14)
Not limited to Further, the method of calculating the certainty of onset synchronism and the method of calculating the certainty of offset synchronism are not limited to (Equation 15). Further, the calculation method of the comprehensive evaluation value is not limited to λ-fuzzy integration, but Sugeno integration, Choquet integration,
Other means such as inverse Sugano integral may be used.

FIG. 33 is a diagram showing the S / N ratio and the residual of the output signal obtained by the above configuration. 33 (A) is an S / N ratio and residual of the input signal, (B) is S / N ratio and residual when λ = −0.9, and (C) is λ = 0.9. S /
N ratio and residual. As shown in Fig. 10000 (B) (C)
In both cases, the S / N ratio is increased and the residual is decreased as compared with (A), and the effect of the above-mentioned configuration can be confirmed.

[0299]

As described below, the present invention provides the signal time division means, pitch extraction means, first storage means, frequency band division means, second storage means, energy or effective value calculation means, and third means. By providing the storage means, the feature amount calculation means, the weight value calculation means, the multiplication means, the fourth storage means, and the summation means, It realizes an excellent noise suppression device that can extract the target acoustic signal by suppressing the noise from the acoustic signal in which the noise is mixed by weighting the frequency of the input signal using physical characteristics. is there.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a noise suppression device in a conventional example.

FIG. 2 is a block diagram of a noise suppression device of the present invention.

FIG. 3 is a first embodiment and a second embodiment of the noise suppression device of the invention.
And configuration diagrams in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8.

FIG. 4 is a diagram showing an example of feature amount calculation means in Embodiments 1 and 2 of the noise suppression apparatus of the present invention.

FIG. 5 is a diagram showing an example of pitch extracting means in the noise suppressing device according to the second and fifth embodiments and the eighth and ninth embodiments of the present invention.

FIG. 6 is a diagram showing an example of a feature amount calculation means in a noise suppressing device according to the third and fourth embodiments, the fifth and sixth embodiments, and the seventh and eighth embodiments.

FIG. 7 is a diagram showing an example of a probability distribution of frequencies at which pitches exist in the noise suppressing device of the third embodiment, the sixth embodiment, and the ninth embodiment.

FIG. 8 is a diagram for explaining the operation of the noise suppression device of the present invention in Example 3 and Example 4, and Example 5 and Example 6 and Example 7 and Example 8 and Example 9;

FIG. 9 is a graph for explaining the criteria for determining the onset in the noise suppressing device of the present invention in the third and fourth embodiments, the fifth and sixth embodiments, the sixth and seventh embodiments, and the eighth and ninth embodiments. Schematic diagram of auditory perception for changes in loudness of human

FIG. 10 is an explanatory diagram of a method for calculating the probability of onset in the third and fifth embodiments of the present invention, the fifth embodiment, the sixth embodiment, the seventh embodiment, the eighth embodiment, and the ninth embodiment.

FIG. 11 is a diagram schematically showing (Equation 10) in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8 and Example 9 of the noise suppressing device of the present invention.

FIG. 12 is a diagram for explaining the operation of the noise suppression device of the present invention in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8 and Example 9;

FIG. 13 is a diagram for explaining the operation of the noise suppression device of the present invention in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8 and Example 9.

FIG. 14 is a diagram for explaining the operation of the noise suppression device of the present invention in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8 and Example 9.

FIG. 15 is a diagram for explaining the operation of the noise suppression device of the present invention in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8 and Example 9.

FIG. 16 is a diagram for explaining a method of calculating the likelihood of continuity in the third and fourth embodiments, the fifth embodiment, the sixth embodiment, the seventh embodiment, the eighth embodiment, and the ninth embodiment of the noise suppressing device of the present invention. Figure

FIG. 17 is a diagram schematically showing (Equation 12) in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8 and Example 9 of the noise suppressing device of the present invention.

FIG. 18 is a diagram for explaining the operation of the noise suppression device of the present invention in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8 and Example 9.

FIG. 19 is a diagram schematically showing (Equation 14) in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8 and Example 9 of the noise suppression device of the invention.

FIG. 20 is a diagram for explaining the operation of the noise suppression device of the present invention in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8 and Example 9;

FIG. 21 is a diagram schematically showing (Equation 15) in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8 and Example 9 of the noise suppression device of the invention.

FIG. 22 is a diagram for explaining the operation of the noise suppression device of the present invention in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8 and Example 9;

FIG. 23 is a diagram for explaining the operation of the noise suppression device of the present invention in Example 3 and Example 4 and Example 5 and Example 6 and Example 7 and Example 8 and Example 9;

FIG. 24 is a diagram showing an example of pitch extracting means in Embodiment 4 and Embodiment 7 of the noise suppressing device of the invention.

FIG. 25 is a diagram showing an example of a pitch certainty distribution in Embodiments 4 and 7 of the noise suppression apparatus of the present invention.

FIG. 26 is a diagram showing an example of a probability distribution in which a pitch exists with respect to an estimated pitch value in Embodiments 4 and 7 of the noise suppression apparatus of the present invention.

FIG. 27 is a diagram showing an example of a probability distribution of probability with respect to an estimated value of pitch in the fifth and eighth embodiments of the noise suppressing device of the present invention.

FIG. 28 is a diagram showing an example of a probability distribution in which the pitch exists with respect to the estimated value of the pitch in the fifth and eighth embodiments of the noise suppressing device of the present invention.

FIG. 29 is a flow chart showing an example of evaluation value calculation in Embodiment 6 and Embodiment 7 and Embodiment 8 and Embodiment 9 of the noise suppressing device of the present invention.

FIG. 30 is a diagram for explaining the operation of the noise suppression device of the present invention in the sixth and seventh embodiments and the eighth and ninth embodiments.

FIG. 31 is a diagram for explaining the operation of the noise suppression device of the present invention in Embodiment 6 and Embodiment 7 and Embodiment 8 and Embodiment 9;

FIG. 32 is a configuration diagram showing a configuration of a noise suppression device according to a ninth embodiment of the present invention.

FIG. 33 is a diagram showing an effect of the noise suppression device according to the ninth embodiment of the present invention.

[Explanation of symbols]

 1 Close-talking microphone for voice input 2 Sensor microphone for noise component signal input 3 Low-pass filter 4 Low-pass filter 5 Band-pass filter bank 10i to 10N Noise removal circuit 11i to 11N Phase difference detection correction circuit 12i to 12N Level difference detection correction circuit 100 signal time division means 200 pitch extraction means 300 first storage means 400 frequency band division means 500 second storage means 600 energy or effective value calculation means 700 third storage means 800 feature amount calculation means 900 weight values Calculation means 910 multiplication means 920 fourth storage means 930 summation means 3810 first pitch extraction means 3820 second pitch extraction means 3830 pitch estimated value calculation means 1000 signal time division means 1001 pitch extraction means 1002 first The memory hand 1003 frequency band division filter 1004 second storage means 1005 energy calculation means 1006 third storage means 1007 characteristic amount calculation means 1008 weight value calculation means 1009 multiplication means 1010 fourth storage means 1011 summation means 1101 onset Calculation means 1102 Offset calculation means 1103 Continuity calculation means 1104 Harmonicity calculation means 1105 Onset synchronization calculation means 1106 Offset synchronization calculation means 1200 First pitch extraction means 1201 Second pitch extraction Means 1202 Pitch extracting means 1203 Pitch estimation value calculating means 1301 Pitch evaluation measure calculating means 1302 Onset evaluation measure calculating means 1303 Offset evaluation measure calculating means 1304 Continuity evaluation measure calculation Means 1305 Harmonic evaluation measure calculation means 1306 Onset synchronicity evaluation measure calculation means 1307 Offset synchronism evaluation measure calculation means 9000 Signal time division means 9001 Pitch extraction means 9002 First storage means 9003 Fourier transform means 9004 Second storage means 9005 Effective value calculation means 9006 Third storage means 9007 Feature amount calculation means 9008 Weight value calculation means 9009 Multiplying means 9010 Fourier inverse transform means 9011 Fourth storage means 9012 Summing means

Claims (10)

[Claims] 1. A signal time division means for time-dividing an acoustic signal, a pitch extraction means for extracting a signal pitch of a short time segment by the signal time division means, and a pitch obtained by the pitch extraction means. Second storage means for storing, frequency band division means for dividing the output signal of the time division means for the signal into a plurality of frequency bands, and third storage means for storing the output of the frequency band division means,
Energy or effective value calculation means for calculating the energy or effective value of the output of the third storage means, fourth storage means for storing the output of the energy or effective value or effective value calculation means, and the fourth Storage means and the second
Of the characteristic amount from the output of the storage unit, a weight value calculating unit that calculates the weight value of the signal of a certain frequency band in a certain time section, and an output of the third storage unit. And multiplication means for calculating the product of the weight value calculation means, and fourth storage means for storing the output of the multiplication means,
And a summing means for calculating a sum of outputs of the fourth storage means. 2. The noise suppression device according to claim 1, wherein the weight value calculation means determines the weight value depending on the presence or absence of the feature amount. 3. The noise suppression device according to claim 1, wherein the feature quantity calculation means uses the likelihood of the feature quantity as an evaluation measure of the feature quantity. 4. The noise suppression device according to claim 2, wherein the weight value calculation means calculates the weight value by integrating the evaluation measures of the feature quantities by multiplication. 5. The noise suppression device according to claim 2, wherein the weight value calculation means calculates the weight value by integrating the evaluation measures of the feature quantity by fuzzy integration. 6. The noise suppressing apparatus according to claim 2, wherein the pitch extracting means uses a plurality of pitch extracting methods and uses an average value of a plurality of outputs as a pitch. 7. The noise suppressing apparatus according to claim 3, wherein the feature quantity calculating means calculates the pitch certainty from the outputs of the plurality of pitch extracting means. 8. A noise suppressing apparatus, wherein the pitch extracting means estimates the pitch from the pitches estimated by a plurality of different pitch estimating methods. 9. A noise suppression apparatus, wherein the feature amount calculation means obtains the probability of pitch estimated from the difference between pitches estimated by two different pitch estimation methods. 10. A noise suppression apparatus, wherein the feature amount calculation means obtains the probability of pitch estimated from the variance of pitches estimated by a plurality of different pitch estimation methods.