JP2012181233A - Speech enhancement device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech enhancement device, a speech enhancement method and a speech enhancement program which can effectively suppress noise and enhance speech.SOLUTION: A speech enhancement device includes a cumulant estimation section for observation signal 21 to estimate cumulant of observation signal, a noise estimation section 11 to estimate a noise component, a cumulant estimation section for estimated noise 22 to estimate cumulant of estimated noise, a cumulant estimation section for speech component 24 to estimate cumulant of speech component based on the cumulant of the observation signal and the cumulant of the estimated noise, a kurtosis estimation section 23 to estimate kurtosis of speech component based on the cumulant of the speech component, a subtraction coefficient calculation part to calculate subtraction coefficient based on the kurtosis of the speech component, and a noise subtraction section 12 to perform noise subtraction on the observation signal using the subtraction coefficient calculated by the subtraction coefficient calculation section.

Description

  The present invention relates to a speech enhancement device, a speech enhancement method, and a speech enhancement program for enhancing the speech of an observation signal including a noise component and a speech component.

  In recent years, with an increase in applications that use voice, there is an increasing demand for extracting only the target voice from the presence of noise. For example, assume that a speaker speaks in an environment as shown in FIG. Speech from the speaker is collected by the microphone 1. At this time, ambient noise is also collected by the microphone 1. Therefore, the observation signal X (f, t) acquired by the microphone 1 includes the target audio signal S (f, t) and the noise signal N (f, t). That is, X (f, t) = S (f, t) + N (f, t).

  Then, noise estimation is performed from the acquired observation signal X (f, t). An estimated noise signal (estimated noise spectrum) is estimated from the observed signal X (f, t). In FIG. 1, N (f, t) with a hat indicating an estimated value is an estimated noise signal. By performing noise subtraction using the estimated noise signal, the output signal Y (f, t) can be obtained.

  Specific noise estimation methods include the following two. The first is a method for estimating a silent section of user voice. This method assumes that the noise is stationary. Then, an estimated noise spectrum is calculated by determining a section based on kurtosis (kurtosis), a power threshold value, and the like.

  The second method is a method using a microphone array. In this method, the radiated sound from the user is assumed to be the point sound source closest to the microphone. Then, a blind spot is formed in the user orientation, and a noise estimation spectrum is calculated.

Noise is subtracted using the estimated noise spectrum. In many nonlinear noise suppression processes, a filter coefficient H (f, t) is applied to the observation signal X (f, t) converted into the time-frequency domain. Specifically, the output signal Y (f, t) can be obtained by the following equation (1).

  Although the design of the filter coefficient H (f, t) varies depending on the method, the filter coefficient H (f, t) is generated by the observation signal X (f, t), the noise estimation signal, and the subtraction coefficient β. Specific design methods include, for example, (a) spectral subtraction (SS) method, (b) generalized spectral subtraction (GSS) method, and (c) Wiener filter (WF). (D) Parametric Wiener Filter (PWF) method. The filter coefficients H (f, t) are respectively expressed by equations (2) to (5).

(A)

(B)

(C)

(D)

  In the above methods (a) to (d), the noise suppression performance and sound quality performance vary depending on the setting of the subtraction coefficient β. For example, FIG. 4 shows a simulation result of the relationship between the subtraction coefficient and each performance. As shown in FIG. 4, when the subtraction coefficient β is set large, the suppression performance increases, but the sound quality performance decreases. On the other hand, when the subtraction coefficient β is set small, the suppression performance is low, but the sound quality performance is high.

  In a real environment, the degree of noise and voice mixing varies from frequency to frequency. For this reason, the optimum value of the subtraction coefficient β varies. Further, in the actual environment, since the degree of noise and voice mixing is unknown, even the graph as shown in FIG. 4 cannot be drawn. Therefore, it is difficult to obtain the optimum subtraction coefficient β.

JP 2000-330597 A JP 2007-330597 A

EUSIPCO 2010 pp. 994-998

  Patent Document 1, Patent Document 2, and Non-Patent Document 1 disclose other methods for enhancing speech by suppressing noise. In Patent Document 1, a subtraction coefficient data table storing a plurality of subtraction coefficients for estimating the S / N ratio of an input audio signal and controlling the amount of noise suppression is provided. The subtraction coefficient is determined based on the S / N ratio from the subtraction coefficient data table.

  In Patent Document 2, the signal-to-noise ratio (SNR) of the input signal is calculated for each frequency bin. When the SNR is low (when it is determined that there is a lot of noise and the voice is low), the subtraction coefficient is corrected to increase the subtraction amount. Thereby, suppression of an input signal can be strengthened more. On the other hand, when the SNR is high (when it is determined that there is little noise and there are many voices), the subtraction coefficient is corrected to reduce the subtraction amount. Thereby, suppression of an input signal can be made smaller.

  However, in the case of a spoken dialogue system, it is necessary to install the system in the operating environment in order to calculate the correction coefficient and subtraction coefficient data tables. In that environment, noise and voice data are measured in advance. The speech recognition rate at each subtraction coefficient value is calculated, and the subtraction coefficient value and correction coefficient of the subtraction coefficient data table must be determined. In an actual product, such advance processing is difficult. Further, when equipment such as a microphone or an AD converter is changed, the subtraction coefficient value of the subtraction coefficient data table must be determined in the same manner. When a subtraction coefficient set in advance in another environment is used, the value is not an optimum value. Therefore, excessive subtraction or undersubtraction of noise components occurs. Deterioration of audio components and residual noise components (occurrence of musical noise) occur. As a result, the speech recognition rate is lowered and the sound quality is deteriorated.

  In Non-Patent Document 1, automatic estimation of noise suppression rate (NRR) before and after processing and fluctuations in the distribution shape of the noise interval before and after processing are calculated as a “cartesis ratio”. Then, the subtraction coefficient is adaptively selected so that the value of the cartosis ratio is kept below the set value. In this way, oversubtraction or undersubtraction of the non-speech section is controlled.

  However, Non-Patent Document 1 controls oversubtraction or undersubtraction of non-voice sections. In other words, over-subtraction or under-subtraction of the speech section is not evaluated. Therefore, there is a possibility that oversubtraction or undersubtraction of the speech section, which is the purpose of speech recognition, occurs. As described above, in Patent Document 1, Patent Document 2, and Non-Patent Document 3, it is difficult to effectively enhance speech.

  The present invention has been made in view of the above-described problems, and an object thereof is to provide a speech enhancement device, a speech enhancement method, and a speech enhancement program that can effectively enhance speech.

  A speech enhancement apparatus according to an aspect of the present invention is a speech enhancement apparatus that enhances speech with respect to an observation signal acquired by a microphone unit, and estimates a cumulant of an observation signal including a noise component and a speech component. A first cumulant estimating unit; a noise estimating unit for estimating a noise component included in the observed signal; a second cumulant estimating unit for estimating a cumulant of the estimated noise estimated by the noise estimating unit; and the observed signal , A third cumulant estimator for estimating the cumulant of the speech component based on the cumulant of the estimated noise, and a first quatsis estimator for estimating the custosis of the speech component based on the cumulant of the speech component A subtraction coefficient adapting unit that calculates a subtraction coefficient based on the speech component cartesis, and the subtraction coefficient adapting unit Using the calculated subtraction coefficient is obtained by and a noise subtracting unit for noise subtracted from the observed signal.

  The speech enhancement apparatus includes: a fourth cumulant estimation unit that estimates a cumulant of the output signal output from the noise subtraction unit; and a cartosis estimation unit that estimates the kurtosis of the output signal based on the cumulant of the output signal. , And the subtraction coefficient adaptation unit may calculate a subtraction coefficient based on the cartesis of the output signal.

  In the above speech enhancement device, the cumulant of the speech component may be estimated based on a difference between the cumulant of the observed signal and the cumulant of the estimated noise.

  In the speech enhancement apparatus, the microphone unit may include a microphone array having a plurality of microphones, and the noise estimation unit may estimate the estimated noise by a microphone array process.

  A speech enhancement method according to an aspect of the present invention is a speech enhancement method for enhancing speech with respect to an observation signal acquired by a microphone unit, and calculates a cumulant of an observation signal including a noise component and a speech component. A step of estimating a noise included in the observed signal, a step of calculating a cumulant of the estimated estimated noise, a cumulant of the observed signal, and a cumulant of the speech component based on the cumulant of the estimated noise. A step of calculating, based on the cumulant of the speech component, a step of estimating a speech component cartesis, a step of calculating a subtraction coefficient based on the speech component cartesis, and using the subtraction factor, the observation signal Noise subtracting.

  The speech enhancement method further includes the steps of: calculating a cumulant of the output signal; and calculating a kurtosis of the output signal based on the cumulant of the output signal; Based on the above, the subtraction coefficient may be calculated.

  In the speech enhancement method, the speech component cumulant may be estimated based on a difference between the observed signal cumulant and the estimated noise cumulant.

  In the speech enhancement method, the microphone unit may include a microphone array having a plurality of microphones, and estimated noise may be estimated by microphone array processing.

  A speech enhancement program according to an aspect of the present invention is a speech enhancement program that enhances speech with respect to an observation signal acquired by a microphone unit, and includes an observation signal including a noise component and a speech component for a computer. A step of calculating a cumulant of the observed signal, a step of estimating a noise included in the observed signal, a step of calculating a cumulant of the estimated estimated noise, a cumulant of the observed signal, and a cumulant of the estimated noise, Using a step of calculating a cumulant of an audio component, a step of estimating a kurtosis of an audio component based on the cumulant of the audio component, a step of calculating a subtraction coefficient based on the audio component kurtosis, Subtracting noise from the observed signal , It is those with a.

  The speech enhancement program further comprises: causing the computer to calculate a cumulant of the output signal; and causing the computer to calculate a kurtosis of the output signal based on the cumulant of the output signal; The subtraction coefficient may be calculated on the basis of the categorization of the audio component.

  In the speech enhancement program, the speech component cumulant may be estimated based on a difference between the observed signal cumulant and the estimated noise cumulant.

  In the speech enhancement method, the microphone unit may include a microphone array having a plurality of microphones, and the estimated noise may be estimated by a microphone array process.

  According to the present invention, it is possible to provide a voice enhancement device, a voice enhancement method, and a voice enhancement program that can effectively enhance a voice.

1 is a block diagram showing a configuration of a speech enhancement apparatus according to a first exemplary embodiment. It is a block diagram which shows the structure of the audio | voice emphasis apparatus concerning Embodiment 2. FIG. It is a figure which shows a general noise subtraction process. It is a simulation result which shows the relationship between the subtraction coefficient in noise subtraction processing, and performance.

  Hereinafter, embodiments of a moving body according to the present invention will be described in detail with reference to the drawings. However, the present invention is not limited to the following embodiments. In addition, for clarity of explanation, the following description and drawings are simplified as appropriate.

Embodiment 1 FIG.
First, a speech enhancement apparatus according to Embodiment 1 of the present invention will be described with reference to FIG. FIG. 1 is a block diagram showing a system configuration of the speech enhancement apparatus. The microphone 1 collects sound generated around and outputs an observation signal based on the sound. The observation signal includes a voice component and a noise component. The voice component is a signal of the voice of the speaker who is the object of voice recognition, and the noise component is a signal other than the voice of the speaker. A microphone 1 is connected to the speech enhancement device 2. Therefore, the observation signal collected by the microphone 1 is input to the speech enhancement device 2.

  The voice of the observation signal is emphasized by the voice enhancement device 2. Then, the output signal in which the voice is emphasized is output to the output side device 3. The output side device 3 is a voice recognition system, a communication device, or the like, and performs predetermined processing on the output signal. For example, in the case of a speech recognition system, speech recognition processing is performed on the output signal.

  The speech enhancement device 2 is an arithmetic processing device having a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), a communication interface, and the like. A computer (PC) or the like. The speech enhancement device 2 has a removable HDD, optical disk, magneto-optical disk, etc., stores various programs and control parameters, and supplies the programs and data to a memory (not shown) as necessary. To do. Of course, the speech enhancement device 2 is not limited to one physical configuration. The voice emphasizing device 2 performs voice processing on sound data collected by the microphone 1.

  The speech enhancement device 2 includes a noise estimation unit 11, a noise subtraction unit 12, a cartesis calculation unit 20, and a subtraction coefficient calculation unit 30. The cartesis calculation unit 20 includes an observation signal cumulant estimation unit 21, an estimated noise cumulant estimation unit 22, a cartosis estimation unit 23, and a speech component cumulant estimation unit 24. The subtraction coefficient calculation unit 30 includes a subtraction coefficient adaptor 31, an output signal cumulant estimation unit 32, and an output signal cartesis estimation unit 33.

  The observation signal from the microphone 1 is input to the noise estimation unit 11, the noise subtraction unit 12, and the observation signal cumulant estimation unit 21. Note that the input observation signal X (f, t) is a signal in the time-frequency domain by the preprocessing before performing the speech enhancement processing. Specifically, an observation signal for a predetermined time is stored in a buffer, and the observation signal is divided into k frames (k is an integer of 2 or more). Here, in the time domain, the frames are divided by half shift so that adjacent frames are overlapped by half. Further, the frame may be divided using a window function. Further, the observation signal divided into frames is subjected to discrete Fourier transform. Thereby, the observation signal X (f, t) in the time-frequency domain can be obtained. Note that this pre-processing may be performed by the speech enhancement device 2 or may be performed by another device, for example, a microphone unit having the microphone 1.

  The noise estimation unit 11 performs noise estimation on the observation signal X (f, t). As a result, a noise estimation signal N (f, t) is generated. In FIG. 1, N (f, t) with a hat indicating estimation is shown in the noise estimation signal. However, in the description in the specification, the noise estimation signal is appropriately simplified to N (f, t ). The noise estimation signal N (f, t) is input to the noise subtraction unit 12 and the estimated noise cumulant estimation unit 22. The estimated noise cumulant estimation unit 22 estimates the cumulant of the estimated noise signal based on the noise estimation signal. Further, the observation signal cumulant estimation unit 21 estimates the cumulant of the observation signal X (f, t). The speech component cumulant estimation unit 24 estimates the speech component cumulant from the observed signal X (f, t) cumulant and the estimated noise cumulant. Since the cumulant is additive, the cumulant of the speech component is indicated by the difference between the observed signal cumulant and the noise estimation cumulant. The kurtosis estimation unit 23 estimates the kurtosis of the voice component based on the cumulant of the voice component.

  The subtraction coefficient adaptor 31 is inputted with the categorization of the voice component. The subtraction coefficient adaptor 31 adapts the subtraction coefficient β on the basis of the speech component cartesis. Then, the subtraction coefficient β obtained by the subtraction coefficient adaptor 31 is input to the noise subtraction unit 12. The noise subtraction unit 12 performs noise subtraction processing using the subtraction coefficient β. The noise subtracting unit 12 outputs an output signal Y (f, t) from which noise has been subtracted. Further, the output signal Y (f, t) is input to the output signal cumulant estimation unit 32. The output signal cumulant estimation unit 32 estimates the cumulant of the output signal Y (f, t). The output signal cartesis estimation unit 33 estimates the output signal cartesis from the cumulant of the output signal Y (f, t). The output signal cartesis is input to the subtraction coefficient adaptor 31.

  The subtraction coefficient adaptor 31 calculates a subtraction coefficient β based on the output signal cartesis and the speech component cartesis. For example, the calculation is repeated until the difference between the output signal cartesis and the speech component cartesis converges. That is, the subtraction coefficient β is calculated so that the difference between the output signal cartesis and the speech component cartesis converges. Then, based on the subtraction coefficient β, the noise subtraction unit 12 performs a noise subtraction process. For the noise subtraction, the above methods (a) to (d), that is, the equations (2) to (5) can be used. The subtraction coefficient β is an adaptive coefficient, and is determined according to the input observation signal X (f, t). That is, the filter for noise subtraction is self-adapted based on the input observation signal.

Next, the speech enhancement method in the speech enhancement apparatus 2 will be described in detail.
(Step 1)
First, the time domain observation signal x (t) acquired by the microphone 1 is divided into frames, and then discrete Fourier transform is performed. Thereby, the observation signal X (f, t) in the time-frequency domain can be obtained.
(Step 2)
The noise estimation unit 11 performs a noise estimation process. Here, a speech / non-speech interval is determined for the observed signal X (f, t), and the non-speech interval is used as a noise estimation signal. In addition, about the process of step 1 and step 2, a well-known method can be used and it does not specifically limit.
(Step 3)
Set frequency bin f = 0.
(Step 4)
An observation signal cumulant estimator 21, estimated noise for cumulant estimator 22 calculates the cumulant C _{noise signal} cumulant C _{observed signal} and the estimated noise signal N of the observation signals X (f, t) (f , t). Therefore, first, the moments of the observed signal X (f, t) and the estimated noise signal N (f, t) are obtained. For example, the second moment M _{2 of} the observed signal X (f, t) _{, the observed signal} , the fourth moment M _{4, the observed signal} , the sixth moment M _{6, the observed signal} , the eighth moment M _{8 and the observed signal} Can be obtained by the following equation (6).

Note that <> _t indicates a time average in a frame. The cumulant C _{observation signal} of the _{observation signal} is _obtained from the moment M _{observation signal} of the observation signal. The second-order cumulant C _{2 of the} observation signal X (f, t) _{, the observation signal} , the fourth-order cumulant C _{4, the observation signal} , the sixth-order cumulant C _{6, the observation signal} , the eighth-order cumulant C _{8, and the observation signal} are: It can obtain | require by the following formula | equation (7).

Similarly, the second order moment M ₂ of the estimated noise signal N (f, x) _{, the noise signal} , the fourth order moment M _{4, the noise signal} , the sixth order moment M _{6, the noise signal} , the eighth order moment M _{8, The noise signal} can be obtained by the following equation (8).

The cumulant C _{noise signal} of the estimated noise signal is obtained from the moment M _{noise signal} of the estimated noise signal. The second order cumulant C _{2, noise signal} , fourth order cumulant C _{4, noise signal} , sixth order cumulant C _{6, noise signal} , eighth order cumulant C _{8, noise signal} of the estimated noise signal N (f, t) are The following equation (9) can be obtained.

In this way, the cumulants C _{observed signal} and the estimated noise signal N (f, t) of the observation signal X (f, t) can be calculated cumulant C _{noise signal.} In a time domain signal, assuming that the probability density function of the signal is 0 on average and symmetric, the odd-order moment and the cumulant are 0. Therefore, odd-order moments and cumulants need not be calculated. Furthermore, in the above description, the second-order, fourth-order, sixth-order, and eighth-order moments and cumulants are obtained, but the obtained order is not limited to this.

(Step 5)
The speech component cumulant estimation unit 24 calculates the cumulant C _{speech component} of the speech component in the observation signal X (f, t). Since the cumulant is additive, the cumulant of the speech component is indicated by the difference between the observed signal cumulant and the estimated noise cumulant. Thus, second order cumulant C _{2, audio components} of the sound component, the fourth-order cumulant C _{4, audio components,} sixth-order cumulant C _{6, audio components,} 8-order cumulant C _{8, the audio component} has the following formula ( 10).

(Step 6)
The kurtosis estimation unit 23 estimates the kurtosis K _{audio component} of the _{audio component} from the cumulant C _{audio component} of the audio component. The estimation of the cartesis is not particularly limited, but for example, Equation (11) can be used. Thereby, the Cartis K _{audio component} in the power spectrum region of the _{audio component} can be calculated.

(Step 7)
The initial value of the subtraction coefficient β is set in the subtraction coefficient adaptor 31, and the number of updates i = 0 is set. An appropriate value can be selected for the initial value of the subtraction coefficient β.
(Step 8)
Then, using the initial value of the subtraction coefficient β, the noise subtraction unit 12 performs a noise subtraction process on the observation signal X (f, t). For the noise subtraction process, for example, any one of the methods (a) to (d) can be used. Therefore, any one of the expressions (2) to (5) is adopted, and the initial value of the subtraction coefficient β is substituted into the expression. Thereby, the filter coefficient H (f, t) can be calculated. Then, an output signal Y (f, t) is calculated from the filter coefficient H (f, t) and the observation signal X (f, t). Specifically, Y (f, t) = H (f, t) X (f, t).

(Step 9)
The output signal cumulant estimation unit 32 estimates the cumulant C _{output signal} of the output signal Y (f, t). Therefore, first, a moment M _{output signal} of the output signal Y (f, t) is obtained. For example, the second moment M _{2 of} the output signal Y (f, t) _{, the output signal} , the fourth moment M _{4, the output signal} , the sixth moment M _{6, the output signal} , the eighth moment M _{8 and the output signal.} Can be obtained by the following equation (12).

The cumulant C _{output signal} of the _{output signal} is obtained from these moments. Output signal Y (f, t) second order cumulant C2 _{, output signal} , fourth order cumulant C4 _{, output signal} , sixth order cumulant C6 _{, output signal} , eighth order cumulant C8 _{, output signal} are It can obtain | require by the following formula | equation (13).

(Step 10)
Based on the cumulant C _{output signal} , the _{output signal} cartesis estimation unit 33 calculates a cartesis K _{output signal} of the _{output signal} . The estimation of the cartesis is not particularly limited, but for example, Equation (14) can be used. As a result, the Cartesis K _{output signal} in the power spectrum region of the audio component can be calculated.

(Step 11)
The subtraction coefficient adaptor 31 updates the subtraction coefficient β by comparing the Cartis K _{output signal} of the _{output signal} with the Cartis K _{sound component} of the _{sound component} . For example, the difference between the Cartis K _{output signal} of the _{output signal} and the Cartis K _{sound component} of the _{sound component} is obtained. Then, the subtraction coefficient β is updated according to the difference between the cartesis. Specifically, the subtraction coefficient β is updated using the following equation (15).

Threshold is a threshold value for determining whether or not the subtraction coefficient β has converged, and an arbitrary value can be set. Δβ is an increment value of the subtraction coefficient β in the loop calculation for converging β, and can be an arbitrary value. In addition, Δβ may be changed according to the difference in cartesis. As described above, when the Cartis K _{audio component} of the _{audio component} is larger than the Cartis K _{output signal of the output signal} , the subtraction coefficient adaptor 31 determines that the noise subtraction is small and increases the subtraction coefficient β. When the absolute value of the Cartis difference is smaller than the threshold value, the subtraction coefficient adaptor 31 determines that the subtraction coefficient β has converged. Accordingly, the frequency bin f is incremented and the process proceeds to (14) described later.

(Step 12)
The update count i is incremented.
(Step 13)
It is determined whether or not the number of updates i exceeds I. Thus, it is determined whether or not the loop calculation for obtaining the subtraction coefficient β has been performed a sufficient number of times. If the number of updates i is smaller than I, the process returns to step 8. On the other hand, if the number of updates i is equal to or greater than I, the frequency bin f is incremented and the process proceeds to the next step 14. That is, when the subtraction coefficient β does not converge, the loop calculation from step 8 to step 12 is repeated until the number of updates i reaches I. Of course, as described above, when the subtraction coefficient β converges, the loop calculation may be skipped before the number of updates i reaches I.
(Step 14)
It is determined whether or not the subtraction coefficient β has been calculated for all frequency bins. Specifically, when the frequency bin f is smaller than F, the process returns to step 4 to obtain the subtraction coefficient β of the next frequency bin. Note that F is the number of frequency bins. On the other hand, if the frequency bin f is greater than F, a time domain output signal is obtained. Specifically, the output signal Y (f, t) calculated by the noise subtracting unit 12 is subjected to inverse Fourier transform. Then, the output signal subjected to inverse Fourier transform is windowed, and data in the time domain is obtained by overlap add. As a result, the output signal y (t) in the time domain is output to the output side device 3. That is, until the frequency bin f reaches F, the loop calculation from step 4 to step 13 is repeated. Note that the processing in step 14 may be performed by the speech enhancement device 2 or may be performed by another device, for example, the output side device 3.

  In this way, the blinding of the cartesis of the speech component in the observed signal is estimated. Here, the cumulant of the speech component is calculated from the difference between the cumulant of the observed signal and the estimated noise using the additive property of the cumulant. Then, the cartesis is calculated from the cumulant of the speech component. As a result, it is possible to estimate the speech component cartesis without performing complicated operations such as convolution. Over / under noise subtraction is evaluated by comparing the kurtosis estimated value of the speech component and the kurtosis estimated value of the output signal. The subtraction coefficient is adjusted according to the evaluation result. That is, if the spectral subtraction is excessive, the subtraction coefficient is reduced, and if it is excessive, the subtraction coefficient is increased. Thereby, noise can be appropriately suppressed. In particular, the subtraction coefficient is adaptively calculated by using the speech component cartesis relating to the speech evaluation quality. In other words, by calculating the voice section cartesis, the fluctuation of the voice section distribution shape before and after the processing can be obtained. Thereby, distortion (for example, cepstrum distortion) of an audio component can be suppressed. It is possible to accurately calculate the kurtosis of the sound component from the observation signal in which noise and speech are mixed. The accuracy of the speech recognition process in the output side device 3 can be improved.

  In the above description, the subtraction coefficient β is calculated for each frequency bin. Thereby, noise can be suppressed more appropriately, and the accuracy of the speech recognition process can be improved. Of course, the subtraction coefficient β may be obtained by batch processing. Hereinafter, a method of calculating the subtraction coefficient β collectively will be described as a first modification.

Modification 1 of Embodiment 1
Since the basic method of noise subtraction is the above-described processing and method, description thereof is omitted. In the first modification, the difference from the method according to the first embodiment will be mainly described.

In the first modification, the subtraction coefficient β for a plurality of frequency bins is calculated all at once. Therefore, the increment of the frequency bin f in step 3 and step 13 and the determination of the frequency bin f in step 14 are not required. Furthermore, the moment calculation formulas in step 4 and step 9 are different. Specifically, the following formula (16), formula (17), and formula (18) are used instead of formula (6), formula (8), and formula (12). By doing so, the moment M _{observation signal of} the _{observation signal} , the moment M _{noise component of} the _{noise component} , and the moment M _{output signal} of the _{output signal} can be obtained, respectively.

  The cumulant is estimated using the above moment. Then, the cartesis is calculated from the cumulant. Note that the calculation processing for obtaining cumulant and cartesis is the same as in the first embodiment, and a description thereof will be omitted. In the first modification, the subtraction coefficients can be calculated collectively from the 0th frequency bin to the (F-1) th frequency bin. That is, a common subtraction coefficient β is used for F frequency bins. Thereby, calculation time can be shortened compared with Embodiment 1. FIG. In addition, since the voice section cartesis is used, the voice can be effectively enhanced.

Embodiment 2. FIG.
The speech enhancement apparatus 2 according to the second embodiment will be described with reference to FIG. In the second embodiment, a microphone array 5 provided with six microphones 1 is used. Then, the observation signals X _{0 to} X ₅ acquired by the microphone array 5 are input to the speech enhancement device 2. Further, the noise estimation process is different from that of the second embodiment. Since basic processes other than these are the same as those in the first embodiment or the first modification, description thereof will be omitted.

  When a microphone array 5 composed of a plurality of microphones 1 is used, it is possible to suppress noise with respect to a nearby point sound source by phase difference control. Therefore, in this embodiment, noise estimation is performed by phase difference control. As shown in the first embodiment, it is not necessary to estimate the speech / non-speech interval in the frame interval. That is, even when there is a voice, it is possible to cancel the voice. Therefore, the above-described voice segment detection process can be eliminated.

The observation signal from the microphone array 5 is input to the noise estimation processing unit 41, the speech estimation processing unit 42, and the observation signal cumulant calculation unit 51. Here, as in the first embodiment, the observation signal is converted into a time-frequency domain signal by pre-processing before noise reduction. The noise estimation processing unit 41 performs noise estimation processing on the plurality of observation signals X ₀ (f, t) to X ₅ (f, t). In the present embodiment, microphone array processing is performed, and an estimated noise signal is generated for each of the observation signals X ₀ (f, t) to X ₅ (f, t). For example, the position of a sound source is estimated by microphone array processing, and sound other than the sound source is estimated as noise. Specifically, noise estimation is performed by adaptive array processing using a null beam former or ICA (independent component analysis). Note that noise estimation by the microphone array 5 is not particularly limited, and a known method can be used.

  The observation signal cumulant calculation unit 51 calculates the cumulant of the observation signal, similarly to the observation signal cumulant estimation unit 21. Further, the estimated noise cumulant calculating unit 52 estimates the cumulant of the estimated noise signal in the same manner as the estimated noise cumulant estimating unit 22. Similar to the speech component cumulant estimation unit 24, the speech component cumulant calculation unit 53 calculates the speech component cumulant based on the observed signal cumulant and the estimated noise signal cumulant. Specifically, the cumulant of the speech component can be obtained from the difference between the cumulant of the observed signal and the cumulant of the estimated noise signal. Similar to the cartosis estimation unit 23, the cartesis calculation unit 54 calculates the speech component cartesis from the speech component cumulant. Similar to the subtraction coefficient adaptor 31, the subtraction parameter adaptation determiner 55 calculates a subtraction coefficient based on the kurtosis of the audio component.

  The subtraction coefficient β calculated by the subtraction parameter adaptive determination unit 55 is input to the speech estimation processing unit 42. Then, the speech estimation processing unit 42 performs speech estimation processing using this subtraction coefficient β. That is, the speech estimation processing unit 42 performs the noise subtraction process using the filter coefficient and the subtraction coefficient β, similarly to the noise subtraction unit 12. As a result, an output signal in which the voice is emphasized is generated. The output signal is input to the output signal cumulant calculation unit 57.

Similarly to the output signal cumulant estimation unit 32, the output signal cumulant calculation unit 57 calculates the cumulant of the output signal. Similar to the output signal cartesis estimation unit 33, the cartesis calculation unit 56 calculates the output signal cartesis based on the cumulant of the output signal. The cartesis of the output signal is input to the subtraction parameter adaptive determination unit 55. Similar to the subtraction coefficient adaptor 31, the subtraction parameter adaptation determiner 55 calculates a subtraction coefficient β from the output signal cartesis and the speech component cartesis. Then, the subtraction coefficient β calculated by the subtraction parameter adaptive determination unit 55 is input to the speech estimation processing unit 42. The speech estimation processing unit 42 performs speech estimation processing based on the updated subtraction coefficient β. That is, the speech estimation processing unit 42 performs noise subtraction processing using the filter coefficient and the subtraction coefficient in the same manner as the noise subtraction unit 12. Thereby, the voice is estimated and an output signal in which the voice is emphasized is output. Then, the output signal is output to an external device, for example, the output side device 3 (not shown in FIG. 2) shown in the first embodiment. The above processing can be calculated by the mathematical formula shown in the first embodiment or the first modification. Thereby, output signals Y ₀ (f, t) to Y ₅ (f, t) can be obtained for each microphone 1.

Next, the speech enhancement method in speech enhancement apparatus 2 of the present embodiment will be described in detail. In the following description, the kurtosis is calculated for each frequency bin as in the first embodiment. However, the kurtosis may be calculated collectively as in the first modification.
(Step 101)
First, the time domain observation signals x ₀ (t) to x ₅ (t) acquired by the microphone array 5 are divided into frames, and then discrete Fourier transform is performed. Thereby, observation signals X ₀ (f, t) to X ₅ (f, t) in the time-frequency domain can be obtained.
(Step 102)
Set frequency bin f = 0.
(Step 103)
The noise estimation processing unit 41 executes noise estimation processing. Here, a noise estimation signal can be obtained by performing microphone array processing. A noise estimation signal is generated for each microphone 1. In the example of FIG. 2, since there are six microphones 1, noise estimation signals N ₀ (f, t) to N ₅ (f, t) are calculated. In addition, about the process of step 101 and step 103, since a well-known method can be used, it is not specifically limited. Of course, the number of microphones 1 is not limited to six.
(Step 104)
The number n = 0 of the microphone 1 is set. That is, a process for calculating cumulant and cartosis is performed on the observation signal X ₀ (f, t) acquired by the first microphone 1 and its noise estimation signal N ₀ (f, t). .

(Step 105)
The observed signal cumulant calculating unit 51 and the estimated noise cumulant calculating unit 52 calculate the cumulant C _{observed signal n of the} observed signal X _n (f, t) and the cumulant C noise signal _n of the estimated noise signal N _n (f, t), _respectively . calculate. Therefore, first, the moments of the observed signal X _n (f, t) and the estimated noise signal N _n (f, t) are obtained. For example, the second moment M _{2 of} the observation signal X (f, t) _{, the observation signal n} , the fourth moment M _{4, the observation signal n} 6, the sixth moment M _{6, the observation signal n 1} , and the eighth moment M _{8 , The observation signal n} can be obtained by an equation similar to the above equation (6). The observation signal cumulant calculation unit 51 obtains the cumulant C _{observation signal n} of the _{observation signal} from the moment M _{observation signal n} of the observation signal. Observation signal X _n (f, t) second-order cumulant C _{2, observation signal n} 4th-order cumulant C _{4, observation signal n} 6th-order cumulant C _{6, observation signal n 1} , eighth-order cumulant C _{8, The observation signal n} can be obtained by the above equation (7).

Similarly, the estimated noise cumulant calculation unit 52 performs the second moment M _{2, the noise signal n} 4, the fourth moment M _{4, the noise signal n 1} , and the sixth moment M M of the estimated noise signal N _n (f, x). _{6. The noise signal n} , the eighth-order moment M8 _{, and the noise signal n} can be obtained by the above equation (8). The estimated noise cumulant calculating unit 52 obtains a cumulant C _{noise signal n} of the estimated noise signal from the moment M _{noise signal n} of the estimated noise signal. Second order cumulant C _{2 of} estimated noise signal N _n (f, t) _{, noise signal n} 4th order cumulant C _{4, noise signal n} 6th order cumulant C _{6, noise signal n 6th} order cumulant C _{8 , Noise signal n} can be obtained by the above equation (9). In this way, it is possible to calculate the cumulant C _{noise signal n} of the observation signal _X n (f, t) cumulant C _{observed signal n} and the estimated noise signal _N n (f, t) of the.
(Step 106)
The voice component cumulant calculation unit 53 calculates the cumulant C _{voice component n} of the voice component in the observation signal X _n (f, t). Since the cumulant is additive, the cumulant of the speech component is indicated by the difference between the observed signal cumulant and the estimated noise cumulant. Therefore, the second-order cumulant C _{2, the voice component n} 4, the fourth-order cumulant C _{4, the voice component n} 6, the sixth-order cumulant C _{6, the voice component n} 8, the eighth-order cumulant C _{8, and the voice component n} are It is shown by the above formula (10).

(Step 107)
The kurtosis calculation unit 54 estimates the kurtosis K _{audio component n} of the _{audio component} from the cumulant C _{audio component n} of the audio component. The estimation of the cartesis is not particularly limited, but for example, the above formula (11) can be used. As a result, the Cartis K _{audio component n} in the power spectrum region of the _{audio component} can be calculated.

(Step 108)
Further, the initial value of the subtraction coefficient β is set in the subtraction parameter adaptive determination unit 55, and the update count i = 0 is set. An appropriate value can be selected for the initial value of the subtraction coefficient β.
(Step 109)
Then, using the initial value of the subtraction coefficient β, the speech estimation processing unit 42 performs noise subtraction processing on the observation signal X _n (f, t). For the noise subtraction process, for example, any one of the methods (a) to (d) can be used. Therefore, any one of the expressions (2) to (5) is adopted, and the initial value of the subtraction coefficient β is substituted into the expression. Thereby, the filter coefficient H (f, t) can be calculated. Then, an output signal Y _n (f, t) is calculated from the filter coefficient H _n (f, t) and the observation signal X (f, t). Specifically, Y _n (f, t) = H _n (f, t) X _n (f, t).

(Step 110)
The output signal cumulant calculation unit 57 estimates the cumulant C _{output signal} of the output signal Y _n (f, t). Therefore, first, the moment M _{output signal n} of the output signal Y _n (f, t) is obtained. For example, the second moment M _{2 of} the output signal Y (f, t) _{, the output signal n} 4, the fourth moment M _{4, the output signal n} 6, the sixth moment M _{6, the output signal n 6,} and the eighth moment M _{8 The output signal n} can be obtained by the above equation (12). From these moments, the cumulant C _{output signal n} of the _{output signal} is obtained. Output signal Y _n (f, t) second-order cumulant C _{2, output signal n} 4th-order cumulant C _{4, output signal n} 6th-order cumulant C _{6, output signal n} 8th-order cumulant C _{8, The output signal n} can be obtained by the above equation (13).

(Step 111)
Based on the cumulant C _{output signal} , the cartesis calculation unit 56 calculates the cartosis K _{output signal n} of the _{output signal} . The estimation of the cartesis is not particularly limited, but for example, the above formula (14) can be used. As a result, the cartosis K _{output signal n} in the power spectrum region of the audio component can be calculated.

(Step 112)
The subtraction parameter adaptive determination unit 55 updates the subtraction coefficient β and increments the number of updates i. The subtraction parameter adaptive determination unit 55 compares the output signal kartsys K _{output signal n} and the sound component kartsys K _{sound component n} to calculate a subtraction coefficient β. For example, the difference between the Cartis K _{output signal n of the output signal} and the Cartis K _{sound component n} of the _{sound component} is obtained. Then, the subtraction coefficient β is updated according to the difference between the cartesis. Specifically, the subtraction coefficient β is updated using the above equation (15). Further, the update count i is incremented.

(Step 113)
It is determined whether or not the number of updates i exceeds I. Thus, it is determined whether or not the loop calculation for obtaining the subtraction coefficient β has been performed a sufficient number of times. If the number of updates i is smaller than I, the process returns to step 109. On the other hand, if the number of updates i is greater than or equal to I, the frequency bin f is incremented and the process proceeds to the next step 114. That is, when the subtraction coefficient β does not converge, the processes of Step 8 to Step 11 are repeated until the number of updates i reaches I. If the β subtraction coefficient converges before the number of updates i reaches I, the loop calculation may be skipped and the process proceeds to the next step 114. For example, when the cartesis difference or ratio is smaller than the threshold Threshold, the loop calculation from Step 109 to Step 112 may be skipped.

(Step 114)
It is determined whether or not the subtraction coefficient β has been calculated for all the microphones 1. For example, when the number of microphones 1 included in the microphone array 5 is M, it is determined whether or not the microphone number n is M or more. If the microphone number n is smaller than M, the process returns to step 105. If the microphone number n is greater than or equal to M, the frequency bin f is incremented and the process proceeds to the next step 115.
(Step 115)
It is determined whether or not the subtraction coefficient β has been calculated for all frequency bins. Specifically, when the frequency bin f is smaller than F, the process returns to step 104 to obtain the subtraction coefficient β of the next frequency bin f. Note that F is the number of frequency bins. On the other hand, when the frequency bin f is F or more, a time domain output signal is obtained. Specifically, the output signal Y (f, t) calculated by the speech estimation processing unit 42 is subjected to inverse Fourier transform. Then, the output signal subjected to inverse Fourier transform is windowed (Hamming window), and time domain data is obtained by overlap addition. As a result, the time-domain output signal y (t) is output from the speech enhancement device 2. Note that the processing of step 115 may be performed by the speech enhancement device 2 or another device.

  By doing in this way, the noise in an observation signal is reduced effectively similarly to Embodiment 1. Therefore, the voice in the observation signal can be emphasized, and the accuracy of the voice recognition process in the subsequent voice recognition system can be improved. Further, a microphone array 5 is used as a microphone unit for acquiring an audio signal. For this reason, noise estimation can be performed effectively. In the first to second embodiments, the loop calculation for calculating the subtraction coefficient may be performed on the same observation signal, or may be performed using the observation signal acquired as needed. good. That is, the latest observation signal may be used for each loop calculation.

  The noise suppression processing described above may be realized by causing a computer including a DSP (Digital Signal Processor), MPU (Micro Processing Unit), CPU (Central Processing Unit), or a combination thereof to execute a program.

  In the above example, a program including a group of instructions for causing a computer to perform speech enhancement processing is stored using various types of non-transitory computer readable media and supplied to the computer. can do. Non-transitory computer readable media include various types of tangible storage media. Examples of non-transitory computer-readable media include magnetic recording media (for example, flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (for example, magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, CD-R / W, semiconductor memory (for example, mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM, RAM (Random Access Memory)) are included. The program may also be supplied to the computer by various types of transitory computer readable media. Examples of transitory computer readable media include electrical signals, optical signals, and electromagnetic waves. The temporary computer-readable medium can supply the program to the computer via a wired communication path such as an electric wire and an optical fiber, or a wireless communication path.

DESCRIPTION OF SYMBOLS 1 Microphone 2 Speech enhancement device 3 Output side device 5 Microphone array 11 Noise estimation unit 12 Noise subtraction unit 20 Cartis calculation unit 21 Observation signal cumulant estimation unit 22 Estimated noise cumulant estimation unit 23 Cartis estimation unit 24 Speech component cumulant estimation unit DESCRIPTION OF SYMBOLS 30 Subtraction coefficient calculation unit 31 Subtraction coefficient adaptor 32 Output signal cumulant estimation part 33 Output signal cartesis estimation part 41 Noise estimation processing part 42 Speech estimation processing part 51 Observation signal cumulant calculation part 52 Estimated noise cumulant calculation part 53 Speech Ingredient cumulant calculation unit 54 Cartis calculation unit 55 Subtractive parameter adaptive decision unit 56 Cartis calculation unit 57 Output signal cumulant calculation unit

Claims (12)

A speech enhancement device that enhances speech with respect to an observation signal acquired by a microphone unit,
A first cumulant estimation unit for estimating a cumulant of an observation signal including a noise component and a speech component;
A noise estimation unit for estimating a noise component included in the observation signal;
A second cumulant estimation unit for estimating a cumulant of the estimated noise estimated by the noise estimation unit;
A third cumulant estimation unit that estimates a cumulant of a speech component based on the cumulant of the observed signal and the cumulant of the estimated noise;
A first kurtosis estimator for estimating a kurtosis of an audio component based on the cumulant of the audio component;
A subtraction coefficient adaptation unit that calculates a subtraction coefficient based on the speech component cartesis;
A speech enhancement apparatus, comprising: a noise subtraction unit that performs noise subtraction on the observation signal using the subtraction coefficient calculated by the subtraction coefficient adaptation unit. A fourth cumulant estimation unit for estimating a cumulant of the output signal output from the noise subtraction unit;
A cartesis estimation unit that estimates the output signal cartesis based on the output signal cumulant, and
The speech enhancement apparatus according to claim 1, wherein the subtraction coefficient adaptation unit calculates a subtraction coefficient based on the cartesis of the output signal.   The speech enhancement apparatus according to claim 1, wherein a cumulant of the speech component is estimated based on a difference between the cumulant of the observation signal and the cumulant of the estimated noise. The microphone unit includes a microphone array having a plurality of microphones;
The speech enhancement apparatus according to claim 1, wherein the noise estimation unit estimates the estimated noise by microphone array processing. A speech enhancement method for enhancing speech with respect to an observation signal acquired by a microphone unit,
Calculating a cumulant of an observation signal including a noise component and a speech component;
Estimating noise contained in the observed signal;
Calculating an estimated noise cumulant;
Calculating a speech component cumulant based on the observed signal cumulant and the estimated noise cumulant;
Estimating the speech component cartesis based on the speech component cumulant;
Calculating a subtraction coefficient based on the speech component cartesis;
Subtracting noise from the observation signal using the subtraction coefficient. Calculating a cumulant of the output signal;
Calculating a cartesis of the output signal based on the cumulant of the output signal; and
The speech enhancement method according to claim 5, wherein the subtraction coefficient is calculated based on the output signal cartesis and the speech component cartesis.   The speech enhancement method according to claim 5 or 6, wherein the cumulant of the speech component is estimated based on a difference between the cumulant of the observation signal and the cumulant of the estimated noise. The microphone unit includes a microphone array having a plurality of microphones;
The speech enhancement method according to claim 5, wherein the estimated noise is estimated by a microphone array process. A speech enhancement program for enhancing speech with respect to an observation signal acquired by a microphone unit,
Against the computer,
Calculating a cumulant of an observation signal including a noise component and a speech component;
Estimating noise included in the observed signal;
Calculating the estimated noise cumulant,
Calculating a speech component cumulant based on the observed signal cumulant and the estimated noise cumulant;
Estimating the speech component cartesis based on the speech component cumulant;
Calculating a subtraction coefficient based on the speech component cartesis;
And a step of subtracting noise from the observation signal using the subtraction coefficient. Against the computer,
Calculating a cumulant of the output signal;
Calculating the output signal cartesis based on the cumulant of the output signal; and
The speech enhancement program according to claim 9, wherein the subtraction coefficient is calculated based on a cartesis of the output signal and a cartesis of the speech component.   The speech enhancement method according to claim 9 or 10, wherein a cumulant of the speech component is estimated based on a difference between the cumulant of the observation signal and the cumulant of the estimated noise. The microphone unit includes a microphone array having a plurality of microphones;
The speech enhancement program according to any one of claims 9 to 11, wherein the estimated noise is estimated by a microphone array process.