JP6473066B2 - Noise suppression device, method and program thereof - Google Patents

Description

  The present invention relates to a noise suppression technique for suppressing a noise component included in a certain collected sound signal by using another collected sound signal. In particular, the present invention relates to a noise suppression technique for suppressing a noise component contained in one of a plurality of collected sound signals obtained from a plurality of close-talking microphones.

  When sound is picked up by a microphone, it is an inevitable event that sound of the surrounding environment is picked up along with the sound. Therefore, when a sound including a target sound component and a noise component is collected by a microphone, techniques for removing or suppressing the noise component by some method have been studied so far.

  For example, conventionally, when noise suppression is performed using a microphone, the spectral subtraction method has been generally used because of a small amount of calculation (see Non-Patent Document 1). In the spectral subtraction method, the noise power of a noise interval (that is, a time interval in which a voice (target sound) to be collected (target sound) is not included) is estimated from a signal collected by a close-talking microphone or the like. This is a technique for suppressing noise by subtracting, on the frequency spectrum, a noise component to be superimposed on a collected sound signal in a speech section (time section including the target sound) using the estimated noise power.

  In addition to the spectral subtraction method for monaural microphones, two microphones are mounted on the smartphone, microphone array processing is performed from the signals collected by the two microphones, and the signals collected by the sub-microphones arranged on the back are Noise suppression processing is performed in which noise suppression is performed by removing the component from the signal of the main microphone arranged so as to be positioned near the mouth during a call (see Non-Patent Document 2 and Non-Patent Document 3). . The premise for this processing method is that the characteristics of the two microphones are the same to some extent, the sub microphones only pick up noise, and the main microphone picks up both the target sound and noise.

  The noise source superimposed on the target sound exists at a position distant from the target sound source, and is more greatly affected by the transfer characteristics between the noise source and the microphone. In addition, since its characteristics are generally unknown, it is necessary to estimate it. Therefore, system identification by an adaptive filter is performed with the transmission process as an unknown system, and the target sound is extracted by subtracting the filter output obtained by multiplying the sound collection signal of the sub microphone by the adaptive filter from the sound collection signal of the main microphone.

  At this time, it is desirable that the distance between the two microphones is not too large and not too small. This is because, if the distance between the two microphones is too large, noise components having different characteristics will be collected, and the erroneous noise component will be subtracted in the simple spectral subtraction method. On the other hand, if the distance between the two microphones is too small, the correlation between the noise components of each microphone will increase, but the target sound component that should not be removed will be collected simultaneously with the noise component by the sub microphone. The premise that microphones only collect noise is broken. In other words, the spectral subtraction method using two microphones has the contradictory acoustic characteristics that the target sound must be picked up only by the main microphone while keeping the correlation while picking up the noise with the two microphones. Applied as an ideal. However, in reality, since the characteristics of the two microphones are the same, it is difficult to prevent the target sound that has been turned around from being collected by the sub microphone.

BOLL S. F., "Suppression of acoustic noise in speech using spectral subtraction.Acoustics", Speech and Signal Processing, 1979, IEEE Transactions on, Volume: 27, Issue: 2, pp.113-120. Jian Zhang et. Al. "A FAST TWO-MICROPHONE NOISE REDUCTION ALGORITHM BASED ON POWER LEVEL RATIO FOR MOBILE PHONE", Kowloon: Chinese Spoken Language Processing (ISCSLP), 2012, 8th International Symposium on, pp.206-209. Isao Nakanishi, “Knowledge Forest”, Group 1 (Signal / System)-Volume 9 (Digital Signal Processing) Chapter 3, Adaptive Signal Processing, [online], IEICE, “Knowledge Forest”, [2014 Search October 23], Internet <http://www.ieice-hbkb.org/files/01/01gun_09hen_03m.pdf>

  When using a plurality of adjacent microphones, the sub-microphone picks up sound only by performing the spectral subtraction method that removes the noise component from the sound pickup signal of the main microphone using the sound pickup signal of one sub-microphone. There is also a problem that even the target sound is removed from the collected sound signal of the main microphone, and the target sound is deteriorated, which is known as musical noise, and distortion of the sound occurs.

  Therefore, in order to suppress deterioration and distortion of speech, system identification is performed by an adaptive filter only in a noise interval that is a non-speech interval. By using the adaptive filter estimated in the noise section, it is possible to eliminate the noise while leaving the target sound. In addition, the noise cancellation performance is made more effective by the filtering process utilizing the positional relationship of close talk.

  In the present invention, the process of promoting learning of the adaptive filter only in the non-speech period is realized by modifying the expression of the adaptive filter. An object of the present invention is to provide a noise suppression apparatus, a method and a program for suppressing noise by suppressing deterioration and distortion of a voice without requiring a new apparatus such as VAD (voice activity detection).

In order to solve the above problem, according to an aspect of the present invention, a noise suppression device suppresses a noise component included in a first main audio signal using a first reference signal and a second reference signal. The noise suppression device performs filtering using the first adaptive filter for the first reference signal and filtering using the first adaptive filter unit for obtaining the first filtered signal and the second adaptive filter for the second reference signal. A second adaptive filter unit for obtaining a second filtered signal, and a subtracting unit for obtaining a value obtained by subtracting the first filtered signal and the second filtered signal from the first main audio signal as an error signal. The first adaptive filter unit sequentially updates the first adaptive filter using the error signal and the first reference signal so that the error signal is minimized, and is the ratio of the error signal to the first reference signal. If the absolute value of the error ratio beta ₁ is in a predetermined threshold value a ₁ or less, and updates the first adaptive filter by a 1-1 updating amount based on monotonically increasing function with respect to error rate beta ₁ If the absolute value of the error ratio beta ₁ is greater than a predetermined threshold value a ₁ is the 2-1st update amount based on monotonically increasing function weight increase is smaller than the 1-1 update amount relative error ratio beta ₁ The first adaptive filter is updated, and the second adaptive filter unit sequentially updates the second adaptive filter using the error signal and the second reference signal so that the error signal is minimized. If the absolute value of the error ratio β ₂ , which is the ratio of the error signal with respect to, is less than or equal to the predetermined threshold a ₂ , the second adaptive filter is updated with the 1-2 update amount based on the monotonically increasing function for the error ratio β ₂ . , if the absolute value of the error ratio beta ₂ is greater than a predetermined threshold value a _2, the 2-2nd update amount based on monotonically increasing function weight increase is smaller than the 1-2 update amount relative error ratio beta ₂ To update the second adaptive filter.

In order to solve the above problem, according to another aspect of the present invention, a noise suppression method suppresses a noise component included in a first main audio signal using a first reference signal and a second reference signal. . In the noise suppression method, the first adaptive filter unit performs filtering using the first adaptive filter on the first reference signal to obtain a first filtered signal, and the second adaptive filter unit includes A second adaptive filter step for filtering the second reference signal using a second adaptive filter to obtain a second filtered signal; and a subtracting unit that performs the first filtered signal and the second filtered signal from the first main audio signal. Subtracting a value obtained by subtracting as an error signal, and in the first adaptive filter step, the error signal and the first reference signal are used to sequentially minimize the error signal. update an adaptive filter, when the absolute value of the error ratio beta ₁ is the ratio of the error signal of a predetermined threshold value a ₁ or less for the first reference signal, the error split A first adaptive filter updates the first 1-1 updating amount based on monotonically increasing function with respect to engagement beta _1, when the absolute value of the error ratio beta ₁ is greater than a predetermined threshold value a _1, to the error ratio beta ₁ The first adaptive filter is updated with the 2-1 update amount based on the monotonically increasing function whose increase amount is smaller than the 1-1 update amount, and the error signal and the second reference signal are used in the second adaptive filter step. The second adaptive filter is sequentially updated so that the error signal is minimized, and when the absolute value of the error ratio β ₂ , which is the ratio of the error signal to the second reference signal, is equal to or less than the predetermined threshold a ₂ , a second adaptive filter updates the first-second update amount based on monotonically increasing function with respect to error rate beta _2, when the absolute value of the error ratio beta ₂ is greater than a predetermined threshold value a _2, compared error ratio beta ₂ Monotonically increasing function with an increase smaller than the 1-2 update amount The second adaptive filter is updated with the 2-2 update amount based on the number.

  According to the present invention, there is an effect that it is possible to suppress deterioration and distortion of voice and suppress noise without newly requiring a device such as VAD (voice activity detection).

The functional block diagram of the noise suppression apparatus which concerns on 1st embodiment. The figure which shows the example of the processing flow of the noise suppression apparatus which concerns on 1st embodiment. . The figure for demonstrating the positional relationship of a main microphone and a submicrophone. The figure which shows the example of the limiting function f ((beta) ₁ ), f ((beta) ₂ ). The figure which shows the example of the limiting function f ((beta) ₁ ), f ((beta) ₂ ). The functional block diagram of the noise suppression apparatus which concerns on 2nd embodiment. The figure which shows the example of the processing flow of the noise suppression apparatus which concerns on 2nd embodiment. The figure for demonstrating the positional relationship of a main microphone and a submicrophone.

  Hereinafter, embodiments of the present invention will be described. In the drawings used for the following description, constituent parts having the same function and steps for performing the same process are denoted by the same reference numerals, and redundant description is omitted. Further, the processing performed for each element of a vector or matrix is applied to all elements of the vector or matrix unless otherwise specified.

  The noise suppression device is equipped with two main microphones and one sub microphone. For example, the noise suppression device has a configuration including a close-talking microphone (headset microphone). Due to its characteristics, the position where the microphone can be placed is limited to the periphery of the face, so it is difficult to separate the microphones beyond a predetermined interval. . Therefore, in the present embodiment, it is assumed that noise is collected with the same sound pressure (similar sound pressure) in the main microphone and the sub microphone. In the present embodiment, the sound pressure of the target sound included in the sound collection signal collected by the main microphone is larger than the sound pressure of the target sound contained in the sound collection signal collected by the sub microphone. A main microphone and a sub microphone are arranged. For example, the main microphone is arranged so as to be located near the mouth, and the sub microphone is arranged so that the target sound is farthest from the mouth and the target sound is difficult to enter. More specifically, for example, the first main microphone is placed further on the tip (near the mouth) of the tip of the microphone arm of the headset, and the second main microphone is also on the slightly root side (a little from the mouth). Place the sub-microphone near the back of the head or ear, such as at the base of the headset headband. Based on this premise, an adaptive filter learning method is performed by filtering a voice signal (sound pickup signal) and using a limit function that places different limits on the update amount of the adaptive filter in a voice interval and a non-voice interval. The point of the present invention is to suppress noise by stabilizing the adaptive filter, suppressing deterioration of the target sound and distortion of the sound.

  In the present embodiment, noise suppression and speech enhancement are performed using an adaptive filter using three microphones attached to the close-talking headset.

  If the arrangement is as described above, the relationship between the collected sound signals between the microphones is as follows. The target sound is picked up by the main microphone with a higher sound pressure than the sub microphone. In addition, noise is picked up with the same sound pressure in both the main microphone and the sub microphone. Using this property, noise is suppressed by an adaptive filter and the target sound is emphasized.

<Noise Suppression Device 100 according to First Embodiment>
FIG. 1 is a functional block diagram of a noise suppression apparatus 100 according to the first embodiment, and FIG. 2 shows a processing flow thereof. Note that the functional block diagram of FIG. 1 is for clearly showing the processing circuit, and the circuit configuration is actually built in the noise suppression apparatus 100.

  The noise suppression apparatus 100 includes a main microphone 101, 102, a sub microphone 103, a proximity sound enhancement unit 104, a proximity sound suppression unit 105, a first adaptive filter unit 110, a second adaptive filter unit 111, a subtraction unit 120, and a filter design unit. 130 and a spectrum filter unit 140.

<Main microphones 101 and 102 and sub microphone 103>
The main microphones 101 and 102 collect the target sound and noise, respectively (S101, S102), and output the first sound collection signal s ₁ (n) and the second sound collection signal s ₂ (n). The sub microphone 103 picks up the target sound and noise (S103) and outputs the second reference signal x ₂ (n). Note that n is an index representing time. The positional relationship between the main microphones 101 and 102 and the sub microphone 103 is shown in FIG. The two main microphones 101 and 102 are arranged at the mouth of the speaker, and the sub microphone 103 is arranged at a position far from the mouth. The three microphones are fixed by wires (arms) or the like, and the main body of the noise suppression device 100 (the main microphones 101 and 102 and the sub microphone 103 are connected from the noise suppression device 100) via a wiring cable, wireless communication, or the like. To each other). However, the main body of the noise suppression device 100 may be downsized and integrated with a microphone fixing arm or the like.

  For example, the main microphones 101 and 102 and the sub microphone 103 are all omnidirectional microphones, and the main microphones 101 and 102 have the same frequency characteristics of microphone sensitivity. It is also assumed that the frequency characteristics of the microphone sensitivity of the sub microphone 103 are the same as the other two. However, the present invention does not limit the characteristics of each of the main microphone and the sub microphone.

  The main microphones 101 and 102 are arranged so as to approach the user's mouth when the noise suppression apparatus 100 is used as a transmission / reception apparatus or a voice input apparatus. The sub microphone 103 is arranged at a position for recording ambient noise having a high correlation with the ambient noise collected by the main microphones 101 and 102 while being away from the arrangement position of the main microphones 101 and 102. Denies that the user's voice is transmitted to the sub-microphone 103 side by being transmitted through the space, transmitted through vibrations of the user's bones, muscles, and housing, or reflected by the surrounding acoustic environment. It is not a thing.

<Near tone enhancement unit 104>
The proximity sound enhancement unit 104 receives the first sound collection signal s ₁ (n) and the second sound collection signal s ₂ (n), and receives the first proximity sound enhancement filter f ^- ₁ (n) and the second proximity sound enhancement signal. Using the filter f ^- ₂ (n), the filtered first filtered sound signal f ^- ₁ ^H (n) s ^- ₁ (n) and the filtered second collected signal f ^- ₂ ^{H _{(n) s - 2 (n}} ) and the sum of the calculated as the first main audio signal d (n) of (S104), and outputs. The processing of the proximity sound enhancement unit 104 is expressed by the following equation.
d (n) = f ^- ₁ ^H (n) s ^- ₁ (n) + f ^- ₂ ^H (n) s ^- ₂ (n)
Where f ^- ₁ (n) = [f _{1_0} (n), f _{1_1} (n),…, f _{1_M-1} (n)] ^T , f ^- ₂ (n) = [f _{2_0} (n), f _{2_1} (n), ..., f _{2_M-1} (n)] ^T , s ^- ₁ (n) = [s ₁ (n), s ₁ (n-1), ..., s ₁ (n-M + 1)] ^T , s ^- ₂ (n) = [s ₂ (n), s ₂ (n-1), ..., s ₂ (n-M + 1)] ^T , ^T is transpose, ^H is complex conjugate transpose Represent. The first proximity sound enhancement filter f ^- ₁ (n) and the second closest tone enhancement filter f ^- for performing convolution with ₂ (n), has a length of the tap size M, respectively first collected signal s ^- ₁ (n) and the second voice collecting signal s ^- ₂ (n) and used for calculation.

For example, f ^- ₁ (n) and _{f 1_0 (n) = 1,} f 1_m (n) = 0 (m = 1,2, ..., M-1) and, f ^- ₂ a (n) f _{2_0} (n ) =-1 and f _{2_m} (n) = 0 (m = 1, 2,..., M-1). In this case, the first main audio signal d (n) is the difference between s ₁ (n) and s ₂ (n). That is, d (n) = s ₁ (n) −s ₂ (n).

If the sound source of the target sound is very close to the main microphone 101 compared to the main microphone 102 and the sound pressure difference collected by the main microphones 101 and 102 is large (in the case of the voice of a close speaker), the first main audio signal d (n) is close to the first sound pickup signal s ₁ (n). On the other hand, when the sound source is far from the main microphone 101 (in the case of ambient noise), the sound pressures of the first sound collection signal s ₁ (n) and the second sound collection signal s ₂ (n) are almost the same, The first main audio signal d (n) approaches zero. Therefore, these coefficients (the first proximity sound enhancement filter f ^- ₁ (n) and the second proximity sound enhancement filter f ^- ₂ (n)) are one of the filters that emphasize the proximity sound. Note that the voice of the mouth can also be emphasized by the method of generating a directional beam at the mouth, but there is a problem that a slight deviation of the microphone position greatly affects the directivity angle at the mouth. By setting the filter as described above, if the main microphone 101 is closer to the mouth than the main microphone 102 without being affected by some positional / direction deviation of the close-up microphone, the proximity sound can be generated. There is an advantage that it can be emphasized. This is an example in which the processing is simplified. As another example, the frequency characteristics (amplitude and phase) of the first proximity sound enhancement filter f ^- ₁ (n) and the second proximity sound enhancement filter f ^- ₂ (n) are included. ) Correction filter may be used. Note that the first proximity sound enhancement filter f ^- ₁ (n) and the second proximity sound enhancement filter f ^- ₂ (n) do not depend on the time n, and the first proximity sound enhancement filter f ^- ₁ and the second proximity sound enhancement filter f ^- ₁ (n) It may be expressed as a proximity sound enhancement filter f ^- ₂ .

<Proximity Sound Suppression Unit 105>
Proximity sound suppressing unit 105, the first voice collecting signal s ₁ (n) and receiving a second voice collecting signal s ₂ (n), the first proximity sound suppressing filter g ^- ₁ (n) and the suppression second closest sound Filtered using the filter g ^- ₂ (n), respectively, the filtered first collected sound signal g ^- ₁ ^H (n) s ^- ₁ (n) and the filtered second collected sound signal g ^- ₂ ^H ( n) The sum of s ⁻ ₂ (n) is obtained as the first reference signal x ₁ (n) and output. The processing of the close sound suppression unit 105 is expressed by the following equation.
x ₁ (n) = g ^- ₁ ^H (n) s ^- ₁ (n) + g ^- ₂ ^H (n) s ^- ₂ (n)
Where g ^- ₁ (n) = [g _{1_0} (n), g _{1_1} (n),…, g _{1_M-1} (n)] ^T , g ^- ₂ (n) = [g _{2_0} (n), g _{2_1} (n), ..., and _{g 2_M-1 (n)]} T. The first proximity sound suppressing filter g ^- ₁ (n) and the second proximity sound suppressing filter g ^- for performing convolution with ₂ (n), has a length of the tap size M, respectively first collected signal s ^- ₁ (n) and the second voice collecting signal s ^- ₂ (n) and used for calculation. For example, the voice of the close speaker input to the main microphone 102 has a difference between the delay of the sample a and the gain b compared to the sound of the close speaker input to the main microphone 101 according to the distance attenuation. Be observed. However, a is an integer of 1 or more, and b is a real number of 0 ≦ b ≦ 1. Then g ^- ₁ (n) is g _{1_a} (n) = b, g _{1_m} (n) = 0 (m = 0,1, ..., M-1, where m ≠ a) and g ^- ₂ ( the _{n) g 2_0 (n) =} - 1, g 2_m (n) = 0 (m = 1, ..., M-1) to. In this case, the first reference signal x ₁ (n) is a difference between b × s ₁ (na) and s ₂ (n). That is, x ₁ (n) = b × s ₁ (na) −s ₂ (n).

Compared to the voice of the near speaker input to the main microphone 101, the voice of the close speaker input to the main microphone 102 is observed with a difference between the delay of the sample a and the gain b. When a person speaks, the first reference signal x ₁ (n) = b × s ₁ (na) −s ₂ (a) includes a voice of a nearby speaker that is reduced, and the first reference signal x ₁ ( n) approaches 0. In addition, ambient noise is hardly reduced by this processing. As a result, an effect of suppressing only the voice of the close speaker can be obtained. Therefore, this coefficient (the first proximity sound suppression filter g ^- ₁ (n) and the second proximity sound suppression filter g ^- ₂ (n)) is one of the filters for suppressing the proximity sound. In _{^{addition, g - 1 (n),}} g - good using the correction filter of the frequency characteristic ₂ (n). The first proximity sound suppression filter g ^- ₁ (n) and the second proximity sound suppression filter g ^- ₂ (n) do not depend on the time n, and the first proximity sound suppression filter g ^- ₁ and the second proximity sound suppression filter g ^- ₂ (n) The suppression filter g ^- ₂ may be used.

<First Adaptive Filter Unit 110>
First adaptive filter unit 110, first reference signal x ₁ (n) and receives the error signal e (n), a first adaptive filter h in the first reference signal x ₁ (n) ^- a ₁ (n) using performs filtering Te (S110), the filtered signal ^{_{h - 1 H (n) x}} - 1 obtains (n), and outputs. Where h ^- ₁ (n) = [h _{1_0} (n), h _{1_1} (n),…, h _{1_M-1} (n)] ^T , x ^- ₁ (n) = [x ₁ (n), x ₁ (n-1), ..., x ₁ (n-M + 1)] ^T. First adaptive filter h ^- ₁ (n) is for performing a convolution operation, has a length of the tap size M, the first reference signal x ^- using ₁ (n) to operation.

Further, the first adaptive filter unit 110 sequentially uses the error signal e (n) and the first reference signal x ₁ (n) so that the error signal e (n) is minimized. h ^- ₁ (n) is updated, and the absolute value of the error signal e (n) relative to the first reference signal x ₁ (n) (hereinafter also referred to as “error ratio”) β ₁ is a predetermined threshold (this embodiment) in the case of 1 to) less the threshold value, the first adaptive filter h by 1-1 updating amount based on monotonically increasing function with respect to error rate beta ₁ ^- ₁ (n) to update the absolute error ratio beta ₁ value is greater than a predetermined threshold value, the error rate β by 2-1st update amount based on monotonically increasing function weight increase is smaller than the 1-1 update amount relative to ₁ the first adaptive filter h ^- ₁ ( Update n).

<Second Adaptive Filter Unit 111>
The second adaptive filter portion 111 includes a second reference signal x ₂ (n) receives the error signal e (n), the second reference signal x ^- a ₂ (n) second adaptive filter h ^- ₂ (n) of used to perform filtering by (S 111), the filtered signal ^{_{h - 2 H (n) x}} - seeking ₂ (n), and outputs. Where h ^- ₂ (n) = [h _{2_0} (n), h _{2_1} (n), ..., h _{2_M-1} (n)] ^T , x ^- ₂ (n) = [x ₂ (n), x ₂ (n-1), ..., x 2 (n-M + 1)] and ^T. Since the second adaptive filter h ^- ₂ (n) performs the convolution operation, the second adaptive filter h ^- ₂ (n) has a length of the tap size M and uses the second reference signal x ^- ₂ (n) for the operation.

Further, the second adaptive filter unit 111 sequentially uses the error signal e (n) and the second reference signal x ₂ (n) so as to minimize the error signal e (n). h ^- ₂ (n) is updated, and the absolute value of the error signal e (n) relative to the second reference signal x ₂ (n) (hereinafter also referred to as “error ratio”) β ₂ is a predetermined threshold (this embodiment) in the case 1 to) below threshold, the 1-2 second adaptive filter by updating amount h based on monotonically increasing function with respect to error rate β ₂ ^- ₂ (n) to update the absolute error ratio beta ₂ If the value is greater than a predetermined threshold value, the error rate β by 2-2nd update amount based on monotonically increasing function weight increase is smaller than for the ₂ 1-2 update amount second adaptive filter h ^- ₂ ( Update n).

<Adaptive filter design method>
The adaptive filter design method will be described below. The first main audio signal d (n) is a signal s ₁ (n), s ₂ (n) collected by the two main microphones 101 and 102 arranged for collecting the target sound, Although the signal emphasizes the target sound, it is a signal that is suppressed but includes noise. The first reference signal x ₁ (n) emphasizes the signals s ₁ (n) and s ₂ (n) collected by the two main microphones 101 and 102 with respect to sounds other than the target sound (here, noise). Although it is a signal, it is a signal that includes a target sound that is suppressed. The second reference signal x ₂ (n) is based on a signal picked up by one sub microphone 103 arranged to pick up the ambient noise correlated with the ambient noise included in the first main audio signal. This is a signal obtained (in this embodiment, the signal itself picked up by the sub microphone 103). The first reference signal x ₁ (n) includes noise having substantially the same characteristics (delay and frequency characteristics) as the noise included in the first main audio signal d (n). On the other hand, the second reference signal x ₂ (n) is slightly different from the noise included in the first main speech signal d (n) (noise arrival time difference, frequency response). Therefore, the target sound is observed less than the first reference signal x ₁ (n). Noise using the first reference signal signal x ₁ (n) and the second reference signal signal x ₂ (n), the first adaptive filter h ^- ₁ (n) and the second adaptive filter h ^- ₂ (n), Remove. In the present embodiment, a normalized LMS (NLMS: Normalized least mean square) method is used to update the adaptive filter (see cited document 3).

The first adaptive filter h ^- ₁ (n) and the second adaptive filter h ^- ₂ (n) are filtered from the first main audio signal d (n) to the filtered signal h ^- ₁ ^H (n) x ^- ₁ (n). the filtered signal h ^- performing ₂ (n) and a value obtained by subtracting an error signal e (n) so that a minimum filter design ^- ₂ ^H (n) x.
e (n) = d (n) -h ^- ₁ ^H (n) x ^- ₁ (n) -h ^- ₂ ^H (n) x ^- ₂ (n) (1)
The first adaptive filter h ^- ₁ (n) and the second adaptive filter h ^- ₂ (n) are updated sequentially. In normal NLMS, the update formula is as follows.

Where || x ₁ (n) || and || x ₂ (n) || are the norms of the first reference signal x ₁ (n) and the second reference signal x ₂ (n), respectively, and the adaptation constant μ is This is a step size parameter that determines the update amount of the update formula. The adaptive coefficient μ takes a constant value regardless of the error signal e (n) during the system operation, and the value range is a real number of 0 <μ <2. This update formula is decomposed as follows.

Wherein beta ₁ and beta _2, respectively first reference signal x ^- ₁ Ratio (ratio) and the second reference signal of the error signal e (n) with respect to the norm of (n) x ^- error signal with respect to the norm of ₂ (n) It represents the ratio (ratio) of e (n).

In the state where the learning of the first adaptive filter h ^- ₁ (n) and the second adaptive filter h ^- ₂ (n) has converged to some extent, filtering is performed in the non-speech section from the positional relationship of the three microphones attached to the headset. The noise component of the sum of the post-signal h ^- ₁ ^H (n) x ^- ₁ (n) and the post-filtering signal h ^- ₂ ^H (n) x ^- ₂ (n) and the first main audio signal d (n) Can be set to the same sound pressure. Therefore, after filtering of the first adaptive filter h ^- ₁ (n) and the second adaptive filter h ^- ₂ (n), the filtered signal h ^- ₁ ^H (n) from the first main audio signal d (n) is filtered. The error signal e (n), which is the value obtained by subtracting x ^- ₁ (n) and h ^- ₂ ^H (n) x ^- ₂ (n), becomes smaller and the first reference signal x ^- ₁ (n) and a second reference signal x ^- is also reduced ratio of the error signal e (n) for ₂ (n), respectively norm, -1 <β _{1 <1,} the -1 <β ₂ <1.

On the other hand, since the main microphones 101 and 102 are arranged near the mouth of the speaker, the sound pressure of the sound picked up by the main microphones 101 and 102 increases in the voice in the voice section. Then, the error signal e (n) includes many target sound components emitted by the speaker, and the absolute value of the error signal e (n) is the first reference signal x ^- ₁ (n) and the second reference signal x ^- _2. (n) Each norm || x ^- ₁ (n) ||, || x ^- ₂ (n) || is larger than β ₁ <-1, 1 <β ₁ , β ₂ <-1, 1 <a β _2. In the present embodiment, the amount of update of the filter in the speech interval is reduced by using nonlinear limit functions f (β ₁ ) and f (β ₂ ) for β ₁ and β ₂ , respectively.

The limiting functions f (β ₁ ) and f (β ₂ ) are respectively | β ₁ | ≧ 1 (β ₁ ≦ -1, β ₁ ≧ 1), | β ₂ | ≧ 1 (β ₂ ≦ -1, β ₂ ≧ It is a nonlinear function that takes a small value in 1). For example, it is represented by the following formula.

For example, when L = 5, the function shown in FIG. 4 is obtained. For example, the limiting functions f (β ₁ ) and f (β ₂ ) may be sigmoid functions represented by the following expressions.

For example, when L = 5, the function shown in FIG. 5 is obtained. 4 and 5, when the absolute values of the error ratios β ₁ and β ₂ are 1 or less, the 1-1 update amount based on the monotonically increasing function for each of the error ratios β ₁ and β ₂ , 1-2 When the absolute values of the error ratios β ₁ and β ₂ are larger than ₁ using the update amount, the error rates β ₁ and β ₂ are larger than the 1-1 update amount and the 1-2 update amount, respectively. A 2-1 update amount and a 2-2 update amount based on a monotonically increasing function with a small amount are used. Therefore, the update amount (1-1 update amount, 1-2 update amount) when the absolute values of the error ratios β ₁ and β ₂ are 1 (threshold) or less is the absolute value of the error ratios β ₁ and β ₂ . Is larger than the update amount (the 2-1 update amount, the 2-2 update amount) when 1 is greater than 1 (threshold). The first adaptive filter h under section 1-1 updates the amount or the 2-1 update amount by the equation (5-1) ^- Updates the ₁ (n), the by formula (5-2) 1-2 The second adaptive filter h ^- ₂ (n) is updated based on the update amount or the 2-2nd update amount.

Function of constraints beta ₁ = 1, the boundary of beta ₁ = -1, the absolute value decreases the beta ₁ value, beta ₂ = 1, the boundary of beta ₂ = -1, beta absolute value of ₂ The value decreases. In the noise section, both the main microphones 101 and 102 and the sub microphone 103 are observed with the same sound pressure. Therefore, if the main filter signal is not suppressed at all by the adaptive filter, β ₁ and β ₂ take 1, and suppression is performed. If it can be done, | β ₁ | <1 and | β ₂ | <1. Next, in the speech interval, the error signal e (n) is observed larger than the first reference signal x ₁ (n) and the second reference signal x ₂ (n), so β ₁ > 1 and β ₂ > 1. Become. If the error signal e (n) is learned to be small in the speech section, the target speech is suppressed. In order to avoid this, the design of the limiting function is such that when β ₁ > 1 and β ₂ > 1, the update amount of the filter is reduced by taking a small value.

In other words, there is a large difference in sound pressure that can be observed between the first main audio signal d (n), the first reference signal x ₁ (n), and the second reference signal x ₂ (n) in the voice section, Since the target sound component remains in the error signal e (n), the ratio of the error signal e (n) to the norms of the first reference signal x ₁ (n) and the second reference signal x ₂ (n) also increases. Therefore, | β ₁ |> 1 and | β ₂ |> 1. _{Conversely, | β 1 |> 1,} | β 2 |> 1 and comprising the section is likely speech _{period, | β 1 |> 1,} | to limit at> 1 become section | beta ₂ Thus, the step size of the adaptive filter in the speech section can be suppressed. In other words, the values of f (β ₁ ) and f (β ₂ ) can be made smaller than β ₁ and β ₂ in the sections where | β ₁ |> 1 and | β ₂ |> 1, respectively. In addition, according to this method, the step size is not set to 0 in | β ₁ |> 1, | β ₂ |> 1 (voice section), so that the noise source exists near the sub microphone 103 or the noise source is When the transfer function from the noise source to the sub microphone 103 and the main microphones 101 and 102 changes greatly, the update of the filter stops when | β ₁ |> 1 and | β ₂ |> 1. Can be prevented. By using the limiting functions f (β ₁ ) and f (β ₂ ), it is possible to suppress the processing of the adaptive filter that proceeds in the direction of erasing the speech signal that is the target sound existing in the speech section. In addition, it is possible to obtain an effect of improving the stability of the filter by relaxing the learning of the filter in a speech section having a characteristic different from that of noise.

<Subtraction unit 120>
Subtraction unit 120, the first main audio signal d (n) and the filtered signal _{^{h - 1 H (n) x}} - 1 and _{^{(n) h - 2 H (}} n) x - receive ₂ and (n), the Subtract the filtered signal h ^- ₁ ^H (n) x ^- ₁ (n) and h ^- ₂ ^H (n) x ^- ₂ (n) from the primary audio signal d (n), and the difference d (n) _{^{-h - 1 H (n) x}} - 1 (n) -h - 2 H (n) x - seeking ₂ (n) as an error signal e (n) (S120), and outputs.

<Filter design unit 130>
The filter design unit 130 receives the first reference signal x ₁ (n), the second reference signal x ₂ (n), and the error signal e (n), and suppresses the noise component that has not been erased by the subtraction unit 120. Is designed (S130) and output.

Although there are various filter design methods, for example, the filter design is performed using a method using a noise removal technique based on PSD (power-spectrum density) estimation described in Reference Document 1.
(Reference 1) Kenta Niwa, Yusuke Hioka, Kazunori Kobayashi, Noriyoshi Kamado, “Implementation of a microphone array to improve speech recognition under noisy conditions”, Proc. Of the Acoustical Society of Japan, 2014, pp .717-718

For example, the filter design unit 130 uses the first reference signal x ₁ (n), the second reference signal x ₂ (n), and the error signal e (n) as a frequency domain first reference signal X _1. (ω, τ), frequency domain second reference signal X ₂ (ω, τ) and frequency domain error signal E (ω, τ) are converted. From the ratio of the error signal to the first reference signal and the second reference signal, | E (ω, τ) | / | X ₁ (ω, τ) |> 1 and | E (ω, τ) | / | X ₂ ( The frequency domain error signal E (ω, τ) when ω, τ) |> 1 is the spectrum Es (ω, τ) of the target sound, and | E (ω, τ) | / | X ₁ (ω, τ ) | ≤ 1 or | E (ω, τ) | / | X ₂ (ω, τ) | ≤ 1 when the frequency domain error signal E (ω, τ) is the noise spectrum En (ω, τ). To do. At this time, the post filter G (ω) is designed based on the Wiener method according to the following equation.

Xs (ω) = E [| Es (ω, τ) | ² ], Xn (ω) = E [| En (ω, τ) ² |]. Here, ω is an index representing a frequency, τ is an index representing a frame, and E [] is an average value of the frame τ. For example, the spectrum may be calculated by converting a time domain signal into a frequency domain signal by fast Fourier transform (FFT).

<Spectral filter unit 140>
The spectrum filter unit 140 receives the error signal e (n) and the filter G, and filters the error signal e (n) using the filter G (S140). In order to suppress the unerased noise component included in the error signal e (n), the post filter G (ω) is multiplied.
Y (ω, τ) = G (ω) E (ω, τ) (9)
Finally, Y (ω, τ) is subjected to inverse fast Fourier transform (IFFT) to obtain an output signal y (n).

<Effect>
With such a configuration, it is possible to suppress deterioration and distortion of voice and suppress noise without requiring a new device such as VAD (voice activity detection). In this embodiment, when noise suppression is performed using the main microphone and the sub microphone, the speed of update of the adaptive filter is changed by a limiting function for each voice section and noise section. As a result, it is possible to create a filter that suppresses filter learning in the wrong direction in the speech section and eliminates only noise collected by two microphones with equivalent sound pressure. In addition, by relaxing the filter learning in the speech section, it is possible to prevent speech suppression and stabilize the filter. Also, the proximity sound emphasis processing by distance attenuation makes the close-up microphone mounting method robust, and the increase in noise due to the variation in frequency characteristics when various people wear it can be suppressed by the effect of the adaptive filter.

<Modification>
In this embodiment, the filter update amount is limited in β ₁ <−1, β ₁ > 1 or β ₂ <−1, β ₂ > 1, but β ₁ <−a ₁ , β ₁ > a The update amount of the filter may be limited in ₁ or β ₂ <-a ₂ and β ₂ > a ₂ . The values of a ₁ and a ₂ are a ₁ > 0 and a ₂ > 0. For example, the expressions (6-1) and (6-2) may be replaced with the following expressions.

In the above formulas (6′-1) and (6′-2), L ₁ and L ₂ are used in place of L in formulas (6-1) and (6-2), respectively.

In this embodiment, the same adaptation constant μ is used in the expressions (2-1) and (2-2), but different adaptation constants μ ₁ and μ ₂ (0 <μ ₁ <2,0 <μ ₂ <2) may be used.

Further, in the present embodiment, the first proximity sound enhancement filter f ^- ₁ (n), the second closest tone enhancing filter f ^- ₂ (n), the first proximity sound suppressing filter g ^- ₁ (n), the second proximity sound The filter lengths of the suppression filter g ^- ₂ (n), the first adaptive filter h ^- ₁ (n), and the second adaptive filter h ^- ₂ (n) are the same length M, but the first reference signal and the second reference The filter lengths may be different from each other within an appropriate range for obtaining the signal and the signal after filtering.

In the present embodiment, the processes of the first adaptive filter unit 110, the second adaptive filter unit 111, and the subtracting unit 120 are performed in the time domain, but may be performed in the frequency domain. For example, a frequency domain conversion unit (not shown) is provided, and the first main audio signal d (n), the first reference signal x ₁ (n), and the second reference signal x ₂ (n) are frequency domain signals, respectively. The first main audio signal D (ω, τ), the frequency domain first reference signal X ₁ (ω, τ), and the frequency domain second reference signal X ₂ (ω, τ) are converted.

The first adaptive filter unit 110 receives the frequency domain first reference signal X ₁ (ω, τ) and the frequency domain error signal E (ω, τ), and generates the frequency domain first reference signal X ₁ (ω, τ). Filtering is performed using the adaptive filter H ₁ (ω, τ) (S110), and a filtered signal H ₁ (ω, τ) X ₁ (ω, τ) is obtained and output.

The second adaptive filter unit 111 receives the frequency domain second reference signal X ₂ (ω, τ) and the frequency domain error signal E (ω, τ), and receives the frequency domain second reference signal X ₂ (ω, τ). ) Is filtered using the adaptive filter H ₂ (ω, τ) (S111), and a filtered signal H ₂ (ω, τ) X ₂ (ω, τ) is obtained and output.

  The first adaptive filter unit 110 and the second adaptive filter unit 111 update the filter according to the following equation.

When the frequency domain error signal E (ω, τ) contains many target sound components, the absolute value of the frequency domain error signal E (ω, τ) is the frequency domain first reference signal X ₁ (ω, τ), the frequency domain The norm of each of the second reference signals X ₂ (ω, τ) || X ₁ (ω, τ) || and || X ₂ (ω, τ) || is larger than β ₁ <-1, 1 <β ₁ , β ₂ <-1, 1 <β ₂ . Even in the case of this modification, the first adaptive filter unit 110 and the second adaptive filter unit 111 are similar to the first embodiment in that the limit functions f (β ₁ ), which are nonlinear with respect to the error ratios β ₁ and β ₂ , respectively. By using f (β ₂ ), it is possible to reduce the amount of filter update in the speech interval. In this description, the same limiting functions f (β ₁ ) and f (β ₂ ) are used for the non-linear limiting functions f (β ₁ ) and f (β ₂ ) in all frequency bands. Each may be configured to use different limiting functions f (β ₁ , ω) and f (β ₂ , ω).

In the subtracting unit 120, the frequency domain first main audio signal D (ω, τ) and the filtered signal H ₁ (ω, τ) X ₁ (ω, τ), H ₂ (ω, τ) X ₂ (ω, τ) ) And filtered signal H ₁ (ω, τ) X ₁ (ω, τ) and filtered signal H ₂ (ω, τ) X ₂ ( The value D (ω, τ) -H ₁ (ω, τ) X ₁ (ω, τ) -H ₂ (ω, τ) X ₂ (ω, τ) obtained by subtracting An area error signal E (ω, τ) is obtained (S120) and output. If the subsequent stage (filter design unit 130 and spectral filter unit 140) performs processing in the frequency domain, the frequency domain first reference signal X ₁ (ω, τ) and the frequency domain second reference signal X ₂ (ω, τ) and the frequency domain error signal E (ω, τ) may be used, and if a time domain signal is used, it may be converted into a frequency domain signal and output to the subsequent stage.

  Also, the point of the present embodiment is to use a restriction function that places different restrictions on the update amount of the adaptive filter in the voice section and the non-voice section. Therefore, the noise suppression apparatus 100 does not necessarily include the main microphones 101 and 102, the sub microphone 103, the filter design unit 130, and the spectrum filter unit 140.

<Noise Suppression Device 200 according to Second Embodiment>
A description will be given centering on differences from the first embodiment.

  FIG. 6 is a functional block diagram of the noise suppression apparatus 200 according to the second embodiment, and FIG. 7 shows a processing flow thereof.

  The noise suppression apparatus 100 includes P main microphones 10-p, Q sub microphones 10-P + q, a proximity sound enhancement unit 204, a proximity sound suppression unit 205, a noise enhancement unit 201, a first adaptive filter unit 110, a first Two adaptive filter units 111, a subtracting unit 120, a filter design unit 130, and a spectrum filter unit 140 are included. Here, p = 1, 2,..., P, q = 1, 2,..., Q, and P and Q are each an integer of 2 or more.

In the present embodiment, processing of the proximity sound enhancement unit 204 and the proximity sound suppression unit 205 is performed using the P main microphones 10-p, and the noise enhancement unit 201 is performed using the Q sub microphones 10-P + q. The signal with enhanced noise is defined as a second reference signal x ₂ (n) (see FIGS. 6 and 7). About the noise emphasis part 201, what is necessary is just to emphasize the direction away from the target audio | voice etc., for example using the delay sum beamformer etc. which are shown in the reference literature 2.
(Reference 2) “Knowledge Forest, Group 2 (Image / Sound / Language)-Volume 6 (Sound Signal Processing), Chapter 2 Sound Source Separation, 2-1 Beamforming”, [online], 2012, Electronic Information Communication Academic Society, [October 15, 2015 search], Internet <URL: http://www.ieice-hbkb.org/files/02/02gun_06hen_02.pdf>

In the present embodiment, the second reference signal x ₂ (n) is a signal obtained by emphasizing a signal collected by the Q sub microphones 10-P + q with respect to sounds other than the target sound. On the other hand, in the first embodiment, the second reference signal is a signal itself picked up by one sub microphone. In any case, it can be said that the second reference signal x ₂ (n) is a signal obtained based on the signal picked up by the sub microphone. If the second reference signal x ₂ (n) is a signal obtained by emphasizing signals other than the target sound collected by the Q sub microphones 10 -P + q, the sound is collected by one sub microphone 103. Compared to the case where the signal is used as it is, the proportion of the target sound included in the second reference signal x ₂ (n) is further reduced.

<Main microphone 10-p and sub microphone 10-P + q>
Each of the P main microphones 10-p collects a target sound and noise (S10-p), and outputs a _p- th collected signal sp (n). The Q sub-microphones 10-P + q collect the target sound and noise (S10-P + q) and output the P + q collected sound signal s _{P + q} (n). FIG. 8 shows the positional relationship between the P main microphones 10-p and the Q sub microphones 10-P + q. In FIG. 8, P = 3 and Q = 3. P main microphones 10-p are arranged at the mouth of the speaker, and Q sub-microphones 10-P + q are arranged at positions far from the mouth. P + Q microphones are fixed via wires (arms), headbands, and the like, and the noise suppression device 200 main body (P noises from the noise suppression device 200) is arranged outside via a wiring cable, wireless communication, or the like. To the main microphone 10-p and Q sub-microphones 10-P + q).

<Noise enhancement unit 201>
The noise enhancement unit 201 receives the Qth P + q collected sound signals s _{P + q} (n), obtains a second reference signal x ₂ (n) that emphasizes sounds other than the target sound (S201), and outputs To do. For example, using the technique of Reference Document 2, a beam is formed in a direction different from the direction of the user's mouth, and the second reference signal x ₂ (n) is obtained.

<Proximity Sound Enhancement Unit 204 and Proximity Sound Suppression Unit 205>
Proximity sound enhancement unit 204 receives the P-number of the p collected signal s _p (n), we obtain a first main audio signal d (n) which highlighted the target sound (S204), and outputs. For example, the difference (p = 2, 3,..., P) between the first collected signal s ₁ (n) and the pth collected signal s _p (n) is calculated, and the average of the P−1 difference signals is calculated. Thus, the first main audio signal d (n) is obtained.

Proximity sound suppressing unit 205 receives the P-number of the p collected signal s _p (n), the first reference signal x ₁ (n) and calculated (S205) which highlighted the sounds other than the target sound, and outputs. For example, using the technique of Reference Document 2, a beam is formed in a direction different from the direction of the user's mouth, and the first reference signal x ₁ (n) is obtained.

In the case of P = 2, the proximity sound emphasizing unit 204 and the proximity sound suppressing unit 205 perform the first main audio signal d (n) and the first reference signal x ₁ (n) by the same method as in the first embodiment. You may ask for.

With this configuration, the first main audio signal d (n), the first reference signal x ₁ (n), and the second reference signal x ₂ (n) have the same properties as in the first embodiment. . That is, the first main audio signal d (n) signal the signal s _p (n) picked up by the arranged number P of the main microphones 10-p in order to pick up a target sound highlighted the target sound However, although it is suppressed, it is a signal including noise. The first reference signal x ₁ (n) is the signal s _p picked up by the P number of main microphones 10-p (n), but (in this case noise) sounds other than the target sound is emphasized signal for suppression The target sound is also included in the signal. The second reference signal x ₂ (n) is collected by Q sub-microphones 10-Pq arranged to pick up ambient noise correlated with the ambient noise included in the first main audio signal. It is a signal obtained based on the received signal, and emphasized for sounds other than the target sound. The first reference signal x ₁ (n) includes noise having substantially the same characteristics (delay and frequency characteristics) as the noise included in the first main audio signal d (n). On the other hand, the second reference signal x ₂ (n) is slightly different from the noise included in the first main speech signal d (n) (noise arrival time difference, frequency response). Therefore, the target sound is observed less than the first reference signal x ₁ (n).

The subsequent processing is the same as in the first embodiment.
<Effect>
With the configuration described above, the same effects as those of the first embodiment can be obtained. In addition, you may combine the modification of 1st embodiment, and this embodiment.

<Other variations>
The present invention is not limited to the above-described embodiments and modifications. For example, the various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. In addition, it can change suitably in the range which does not deviate from the meaning of this invention.

<Program and recording medium>
In addition, various processing functions in each device described in the above embodiments and modifications may be realized by a computer. In that case, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on a computer, various processing functions in each of the above devices are realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its storage unit. When executing the process, this computer reads the program stored in its own storage unit and executes the process according to the read program. As another embodiment of this program, a computer may read a program directly from a portable recording medium and execute processing according to the program. Further, each time a program is transferred from the server computer to the computer, processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program includes information provided for processing by the electronic computer and equivalent to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In addition, although each device is configured by executing a predetermined program on a computer, at least a part of these processing contents may be realized by hardware.

Claims (9)

A noise suppression device that suppresses a noise component included in a first main audio signal using a first reference signal and a second reference signal,
Filtering the first reference signal using a first adaptive filter, a first adaptive filter unit for obtaining a first filtered signal;
Filtering the second reference signal using a second adaptive filter, and obtaining a second filtered signal, a second adaptive filter unit;
A subtracting unit for obtaining, as an error signal, a value obtained by subtracting the first filtered signal and the second filtered signal from the first main audio signal;
The first adaptive filter unit sequentially updates the first adaptive filter using the error signal and the first reference signal so as to minimize the error signal, and the error relative to the first reference signal is determined. When the absolute value of the error ratio β ₁ , which is the signal ratio, is equal to or less than a predetermined threshold a ₁ , the first adaptive filter is updated with a 1-1 update amount based on a monotonically increasing function with respect to the error ratio β ₁ . When the absolute value of the error rate β ₁ is larger than a predetermined threshold a ₁ , a second increase based on a monotonically increasing function with an increase amount smaller than the 1-1 update amount with respect to the error rate β ₁ . Updating the first adaptive filter by one update amount;
The second adaptive filter unit uses the error signal and the second reference signal to sequentially update the second adaptive filter so that the error signal is minimized, and the error relative to the second reference signal When the absolute value of the error ratio β ₂ , which is a signal ratio, is equal to or less than a predetermined threshold value a ₂ , the second adaptive filter is updated with a 1-2 update amount based on a monotonically increasing function with respect to the error ratio β ₂ . When the absolute value of the error rate β ₂ is larger than a predetermined threshold value a _2, a _second value based on a monotonically increasing function that is smaller than the first-2 update amount with respect to the error rate β ₂ is obtained. Updating the second adaptive filter by two update amounts;
Noise suppression device. The noise suppression device of claim 1,
The first main audio signal is a signal in which a signal collected by two or more main microphones arranged to collect the target sound is emphasized for the target sound,
The first reference signal is a signal in which signals collected by two or more main microphones are emphasized for sounds other than the target sound,
The second reference signal is a signal obtained based on a signal picked up by one or more sub-microphones arranged to pick up ambient noise correlated with the ambient noise included in the first main audio signal. Is,
Noise suppression device. The noise suppression device of claim 1,
The first main audio signal is a signal picked up by two or more main microphones arranged at the mouth of the speaker, and a signal that picks up the target sound and the ambient noise that are the utterance of the speaker. , A signal that emphasizes the target sound,
The first reference signal is a signal picked up by two or more main microphones, and a signal obtained by picking up the target sound that is the utterance of the speaker and ambient noise is emphasized for sounds other than the target sound. Signal,
The second reference signal is a signal obtained based on a signal picked up by one or more sub-microphones,
In the sub microphone, the sound pressure of the target sound included in the signal collected by the sub microphone is smaller than the sound pressure of the target sound included in the signal collected by the main microphone, and is collected by the sub microphone. The sound pressure of the noise contained in the signal that was struck is arranged to be approximately the same as the sound pressure of the noise contained in the signal collected by the main microphone,
Noise suppression device. The noise suppression device according to any one of claims 1 to 3,
An index representing the time with the 1-1 update amount or the 2-1 update amount as f (β ₁ ) and the 1-2 update amount or the 2-2 update amount as f (β ₂ ). Is n, the filter length of the filter coefficient of the first adaptive filter is M _1, and the filter coefficient at time n is h ⁻ ₁ (n) = [h ₁ (n), h ₁ (n−1),. , h ₁ (nM ₁ +1)], the filter length of the filter coefficient of the second adaptive filter is M _2, and the filter coefficient at time n is h ⁻ ₂ (n) = [h ₂ (n), h ₂ (n−1),..., H ₂ (nM ₂ +1)], the adaptive constant of the first adaptive filter is μ ₁ , the adaptive constant of the second adaptive filter is μ _2, and the above-mentioned at time n The first reference signal is x ₁ (n), x ⁻ ₁ (n) = [x ₁ (n), x ₁ (n−1),..., X ₁ (n−M + 1)]. The norm of one reference signal x ^- ₁ (n) is || x ^- ₁ (n) ||, the second reference signal at time n is x ₂ (n), and x ^- ₂ (n) = [x _{2 (n), x 2 (n} -1), ..., x 2 (n-M + 1)] And, wherein the second reference signal x ^- each greater than 1 and ₂ (n) ||, the error signal at time n and e (n), L ₁ and L ₂ a ^- norm || x of ₂ (n) The update formula of the first adaptive filter and the second adaptive filter is a real number.

And

Or

And

Or

Is,
Noise suppression device. A noise suppression method for suppressing a noise component included in a first main audio signal using a first reference signal and a second reference signal,
A first adaptive filter unit that performs filtering using the first adaptive filter on the first reference signal to obtain a first filtered signal;
A second adaptive filter unit performs filtering using the second adaptive filter on the second reference signal, and obtains a second post-filtering signal; and
A subtracting step for subtracting a value obtained by subtracting the first filtered signal and the second filtered signal from the first main audio signal as an error signal;
In the first adaptive filter step, using the error signal and the first reference signal, the first adaptive filter is sequentially updated so that the error signal is minimized, and the error relative to the first reference signal is determined. When the absolute value of the error ratio β ₁ , which is the signal ratio, is equal to or less than a predetermined threshold a ₁ , the first adaptive filter is updated with a 1-1 update amount based on a monotonically increasing function with respect to the error ratio β ₁ . When the absolute value of the error rate β ₁ is larger than a predetermined threshold a ₁ , a second increase based on a monotonically increasing function with an increase amount smaller than the 1-1 update amount with respect to the error rate β ₁ . Updating the first adaptive filter by one update amount;
In the second adaptive filter step, using the error signal and the second reference signal, the second adaptive filter is sequentially updated so that the error signal is minimized, and the error relative to the second reference signal is determined. When the absolute value of the error ratio β ₂ , which is a signal ratio, is equal to or less than a predetermined threshold value a ₂ , the second adaptive filter is updated with a 1-2 update amount based on a monotonically increasing function with respect to the error ratio β ₂ . When the absolute value of the error rate β ₂ is larger than a predetermined threshold value a _2, a _second value based on a monotonically increasing function that is smaller than the first-2 update amount with respect to the error rate β ₂ is obtained. Updating the second adaptive filter by two update amounts;
Noise suppression method. The noise suppression method according to claim 5, comprising:
The first main audio signal is a signal in which a signal collected by two or more main microphones arranged to collect the target sound is emphasized for the target sound,
The first reference signal is a signal in which signals collected by two or more main microphones are emphasized for sounds other than the target sound,
The second reference signal is a signal obtained based on a signal picked up by one or more sub-microphones arranged to pick up ambient noise correlated with the ambient noise included in the first main audio signal. Is,
Noise suppression method. The noise suppression method according to claim 5, comprising:
The first main audio signal is a signal picked up by two or more main microphones arranged at the mouth of the speaker, and a signal that picks up the target sound and the ambient noise that are the utterance of the speaker. , A signal that emphasizes the target sound,
The first reference signal is a signal picked up by two or more main microphones, and a signal obtained by picking up the target sound that is the utterance of the speaker and ambient noise is emphasized for sounds other than the target sound. Signal,
The second reference signal is a signal obtained based on a signal picked up by one or more sub-microphones,
In the sub microphone, the sound pressure of the target sound included in the signal collected by the sub microphone is smaller than the sound pressure of the target sound included in the signal collected by the main microphone, and is collected by the sub microphone. The sound pressure of the noise contained in the signal that was struck is arranged to be approximately the same as the sound pressure of the noise contained in the signal collected by the main microphone,
Noise suppression method. The noise suppression method according to any one of claims 5 to 7,
An index representing the time with the 1-1 update amount or the 2-1 update amount as f (β ₁ ) and the 1-2 update amount or the 2-2 update amount as f (β ₂ ). Is n, the filter length of the filter coefficient of the first adaptive filter is M _1, and the filter coefficient at time n is h ⁻ ₁ (n) = [h ₁ (n), h ₁ (n−1),. , h ₁ (nM ₁ +1)], the filter length of the filter coefficient of the second adaptive filter is M _2, and the filter coefficient at time n is h ⁻ ₂ (n) = [h ₂ (n), h ₂ (n−1),..., H ₂ (nM ₂ +1)], the adaptive constant of the first adaptive filter is μ ₁ , the adaptive constant of the second adaptive filter is μ _2, and the above-mentioned at time n The first reference signal is x ₁ (n), x ⁻ ₁ (n) = [x ₁ (n), x ₁ (n−1),..., X ₁ (n−M + 1)]. The norm of one reference signal x ^- ₁ (n) is || x ^- ₁ (n) ||, the second reference signal at time n is x ₂ (n), and x ^- ₂ (n) = [x _{2 (n), x 2 (n} -1), ..., x 2 (n-M + 1)] And, wherein the second reference signal x ^- each greater than 1 and ₂ (n) ||, the error signal at time n and e (n), L ₁ and L ₂ a ^- norm || x of ₂ (n) The update formula of the first adaptive filter and the second adaptive filter is a real number.

And

Or

And

Or

Is,
Noise suppression method.   The program for functioning a computer as a noise suppression apparatus in any one of Claims 1-4.

Images