JP2006084898A - Sound input device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a sound input device capable of removing diffusive noise with a slight increase in calculation quantity as compared with SBE applied to a general frequency domain ICA. <P>SOLUTION: The sound input device detects a sound including a target sound and a non-target sound by a plurality of microphones 10-1 to 10-n to obtain a plurality of sound signals including target sound signals and non-target sound signals and implements an independent component analyzing method of acquiring, by repetitive learning, a filter separating one target sound signal from the sound signals. A sound frequency band is divided into a plurality of frequency division bands at a band division stage 30, discrimination levels for evaluating target sound separation performance of the filter obtained by the learning are calculated by the frequency division bands at an evaluation stage 50, and a frequency division band for learning calculation is determined at a determination stage 60 among the frequency division bands based upon the discrimination levels. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to a voice input device.

  2. Description of the Related Art In recent years, a voice input system in a passenger compartment has been widely used for in-vehicle device operation by voice recognition, hands-free telephone, and the like. A factor that hinders the realization of these technologies is the presence of sound from a sound source other than the voice input user in the passenger compartment. As a method of separating the sound from the sound input user from the sound from other sound sources, each sound signal is acquired from a plurality of acoustic sensors, and only the plurality of acquired sound signals are used, Independent component analysis (hereinafter referred to as ICA) has been developed as a method for obtaining a filter for separating a speech signal by learning.

JP 2003-271166 A "Basics of blind source separation using array signal processing" Technical report of IEICE, EA2001-7. “What is independent component analysis?” Computer Today, pp. 38-43, 1998.9, No. 87, “Application to fMRI image analysis” Computer Today, pp. 60-67, 2001.1 No. 95.

  However, problems in the target signal separation processing based on the ICA include the following.

  First, the statistical independence between signals sent from the signal source is used, but in the actual environment, it is difficult to accurately estimate the statistics due to the signal transfer characteristics, background noise, etc. Separation accuracy deteriorates.

  Also, a diffusive signal source is very difficult to separate because it is difficult to consider it as one signal source.

  With respect to the above problem, Patent Document 1 proposes a method for removing the influence of a diffusive signal source in the ICA calculation process. In this method, the accuracy of the sound source separation process is predicted based on the size of the cost function calculated for each frequency in the ICA calculation process, and the filter response is reduced at frequencies where the accuracy of the sound source separation process is low ( Hereinafter, this is referred to as SBE (Sub-Band Eliminate). In SBE, since the process of determining whether or not the accuracy of the sound source separation process exceeds the threshold value for each frequency, the amount of calculation is larger than that of a general frequency domain ICA.

  An object of the present invention is to provide a voice input device that improves this point, and requires only a slight increase in the amount of calculation compared to SBE applied to a general frequency domain ICA, and can remove diffuse noise. .

  Execute the independent component analysis method that acquires each sound signal from multiple acoustic sensors and learns the filter that separates the target audio signal from the sound signal by using only the acquired multiple sound signals. In the voice input device, the acoustic frequency band is divided into a plurality of frequency division bands, and an identification level for evaluating the target voice separation performance of the filter obtained by the learning is determined for each frequency division band in the repetition process of the learning. A speech input device is configured to calculate and determine a frequency division band for performing the learning from the frequency division band based on the identification level.

  By implementing the present invention, the number of times of learning is reduced in the frequency division band where the learning result does not improve, and a slight increase in the amount of calculation is required compared to SBE applied to a general frequency domain ICA, thereby eliminating diffusive noise. It is possible to provide a voice input device that can be used.

  The case where the learning method for obtaining the filter, which is characteristic of the voice input device according to the present invention, is applied to an example of ICA will be described below.

  For example, as a signal source, in addition to receiving sound signals with K microphones (sensors), it is the same as a microphone by utilizing that the sound signals coming from each sound source are statistically independent. K or less than K sound sources can be separated. Initially, the sound source separation method using ICA was difficult to apply to a microphone array because the time difference between incoming sounds from each sound source was not considered. However, in recent years, many methods have been proposed in which a time difference is taken into account and a plurality of sound signals are observed using a microphone array to obtain an inverse transformation of the mixing process in the frequency domain.

  In general, when sound signals coming from a plurality of L sound sources are linearly mixed and observed by K microphones, the observed sound signal can be written as follows at a certain frequency f.

X (f) = A (f) S (f) (1)
Where S (f) is the sound signal vector transmitted from each sound source, X (f) is the observed signal vector observed at the microphone array that is the sound receiving point, and A (f) is the sound source and sound receiving point. Is a mixing matrix for the spatial acoustic system of, which can be written as follows:

S (f) = [S ₁ (f), ..., S _L (f)] ^T (2)
X (f) = [X ₁ (f), ..., X _L (f)] ^T (3)

ここで上添字^Ｔはベクトルの転置を表す。このとき、混合行列A(f)が既知であれば、受音点での観測信号ベクトルX(f)を用いて、
S(f) ＝ A(f)⁻X(f) (5)
（ただし、A(f)⁻は行列A(f)の一般逆行列を表す）のようにA(f)の一般逆行列A(f)⁻を計算することで音源から送出される音信号S(f)を計算することができる。しかし一般にA(f)は未知であり、X(f)だけを利用することで音信号S(f)を求めなければならない。

Here, the superscript ^T represents transposition of the vector. At this time, if the mixing matrix A (f) is known, using the observation signal vector X (f) at the sound receiving point,
S (f) = A (f) ⁻ X (f) (5)
(However, A (f) ^- the matrix A (f) represents a general inverse matrix) of the generalized inverse matrix A A (f) (f) as ^- the sound signal is sent from the sound source by calculating S (f) can be calculated. However, in general, A (f) is unknown, and the sound signal S (f) must be obtained by using only X (f).

  In order to solve this problem, it is assumed that the sound signal S (f) is generated stochastically and that all components of S (f) are all independent of each other. At this time, since the observation signal X (f) is a mixed signal, the distribution of each component of X (f) is not independent. Therefore, consider searching for an independent component included in the observation signal by ICA. That is, the matrix W (f) (hereinafter referred to as the inverse mixing matrix) that converts the observed signal X (f) into independent components is calculated, and the inverse mixing matrix W (f) is applied to the observed signal X (f) (matrix multiplication). ) To obtain an approximate signal to the sound signal S (f) transmitted from the sound source.

  As processing for obtaining the inverse transformation of the mixing process by ICA, a method of analyzing in the time domain and a method of analyzing in the frequency domain have been proposed. Here, a method for calculating in the frequency domain will be described as an example.

  A learning method for obtaining a filter, which is a feature of the voice input device according to the present invention, will be described with reference to FIG. As shown in the figure, the microphones 10-1 to 10-n, which are a plurality of acoustic sensors, detect the sound in which the target sound and the non-target sound are mixed (detection process 20). A plurality of acoustic signals in which the target audio signal and the non-target sound signal obtained in the detection process 20 are mixed are divided into a plurality of frequency division bands in the band division process 30, and a filter learning process in which a filter is acquired by repetition of learning It is thrown into 40. In the learning method for obtaining a filter, which is characterized by the speech input device according to the present invention, the identification level for evaluating the target speech separation performance of the filter in the middle of the iteration is set to the frequency in the evaluation step in the learning iteration process. The frequency division band is calculated for each division band, and the frequency division band for performing learning is determined from the frequency division bands in the determination step 60 based on the calculated identification level. The determination result is fed back to the filter learning process 40, and learning according to the determination result is executed. An example of the identification level and its calculation method will be described later.

First, a short-time frame analysis is performed on the signal observed by each microphone using an appropriate orthogonal transform. At this time, it is considered as a time series by plotting the complex spectrum value at a specific frequency bin at one microphone input. Here, the frequency bin indicates, for example, an individual complex component in a signal vector frequency-converted by short-time discrete Fourier transform. Similarly, the same operation is performed for other microphone inputs. The time-frequency signal sequence obtained here is
X (f, t) = [X ₁ (f, t), ..., X _K (f, t)] ^T (6)
Can be described. Next, signal separation is performed using the inverse mixing matrix W (f). This process is shown as follows.

Y (f, t) = [ Y 1 (f, t), ..., Y L (f, t)] T = W (f) X (f, t) (7)
Here, the demixing matrix W (f) is optimized so that the L time series outputs Y (f, t) are independent of each other. These processes are performed for all frequency bins. Finally, inverse orthogonal transformation is applied to the separated time series Y (f, t) to reconstruct the sound source signal time waveform.

  As an independence evaluation and demixing matrix optimization method, an unsupervised learning algorithm based on the minimization of Kullback-Leibler divergence and an algorithm for decorrelating a second-order or higher-order correlation have been proposed (see above). Non-patent document 1).

  Note that ICA is not limited to sound signal processing, for example, separates signals that have arrived due to crosstalk in mobile communication, etc., or separates signals generated at various locations inside the brain into electroencephalographs or magnetoencephalographs. This is used to separate and extract a target signal from measurement signals when measured from the outside using fMRI (Functional Magnetic Resonance Imaging) or the like (see Non-Patent Document 2 above). ).

  In the following, an example of an identification level and a calculation method in a learning method for obtaining a filter, which is a feature of the voice input device according to the present invention, will be described.

  In frequency bands where it is difficult to separate signals using ICA, the identification level, that is, the value of separation accuracy (for example, cosine distance) is often not improved even after several tens of learning. Since the calculation for learning in such a band is redundant, it is only necessary to detect a band with a small change in the identification level, that is, the separation accuracy, and end the calculation of the frequency domain ICA.

  In the following, a determination function for determining the end of calculation of the frequency domain ICA is introduced.

_First , a time-frequency signal sequence collected by each microphone and subjected to short-time frame analysis is represented by X (f, t) = [X ₁ (f, t), ..., X _K (f, t)] ^T Is described. Next, signal separation is performed using an inverse mixing matrix W (f) optimized by ICA. This process is shown as the above equation (7). Here, Y (f, t) is a separated signal subjected to sound source separation. At this time, W (f) is a matrix of L rows and K columns, and represents the attenuation characteristics of a filter that separates the target audio signal from a plurality of acoustic signals.

  As a method for detecting a band with a small change in separation accuracy of a filter that separates target speech signals from a plurality of acoustic signals, as a discrimination level for evaluating the target speech separation performance of the filter after completion of learning by ICA, the separation signals are independent. A cost function for evaluating the performance is defined, and a band with a small change in separation accuracy is determined based on the rate of change of the cost function. For this cost function, for example, a higher-order correlation value between separated signals, a cosine distance, or the like may be used. In particular, the cosine distance is efficient with a small amount of calculation. Below, the cost function based on the cosine distance of two sound sources is shown.

ここで、記号〈
〉_ｔは局所時間区間、たとえば、時刻t-200msから時刻tまでの時間区間において時間に関する平均をとることを表し、記号＊は複素共役を表す。式(8)の右辺は、二つの分離信号Y_１(f,t)、Y_２(f,t)のその局所時間区間における相関係数を表し、これがこの場合の識別レベルとなっている。式(8)の左辺のtは、上記のt、すなわち、局所時間区間の上端(時間が左から右に流れるとしたときの右端)を表していて、右辺におけるtとは意味が異なる。

Where the symbol <
> _T represents taking an average with respect to time in a local time interval, for example, a time interval from time t-200 ms to time t, and symbol * represents a complex conjugate. The right side of Equation (8) represents the correlation coefficient of the two separated signals Y ₁ (f, t) and Y ₂ (f, t) in the local time interval, which is the discrimination level in this case. T on the left side of Equation (8) represents the above t, that is, the upper end of the local time interval (the right end when time flows from left to right), and has a different meaning from t on the right side.

  In actual application, since the value of the identification level depends on the time cut-out position in the short-time frame analysis, a significant discontinuity may occur between frequencies. An example of such a frequency discontinuity phenomenon of the cost function is shown by a line that changes drastically in FIG. In order to avoid this, as an example, it is conceivable to use a smoothed cost function obtained by taking a moving average of the cost function of Equation (8) with a certain frequency bandwidth, that is, a smoothed discrimination level. This can be written as follows (see solid line with little change in FIG. 9):

ここでＢは平滑化幅を与えるパラメタである。すなわち、この場合の平滑化は、局所周波数区間において平均しておこなわれる。この平滑化されたコスト関数J_Ｓ(f,t)は、分離された信号が独立なものであれば値は小さくなり、非独立なものであれば値は大きくなる。また、その最大値は１である。

Here, B is a parameter that gives a smoothing width. That is, the smoothing in this case is performed by averaging in the local frequency section. The smoothed cost function J _S (f, t) has a small value if the separated signal is independent, and a large value if the separated signal is non-independent. The maximum value is 1.

Furthermore, by using the change rate ΔJ of the cost function, it is possible to detect a band where the separation accuracy is not improved. This ΔJ corresponds to the change rate between the identification levels described in claim 1. For example, at time t, the rate of change ΔJ _S (f, t) in one frequency division band is the cost function J _S (f, t) of the frequency division band at time t and m> 0, Using the cost function J _S (f, tm)
ΔJ _S (f, t) = ‖J _S (f, t) −J _S (f, tm) ‖ (10)
Can be expressed as The right side of equation (10) represents the norm of the difference vector J _S (f, t) −J _S (f, tm) between the two smoothed discrimination levels, for example, J _S (f, t) −J _S (f , tm) squared with respect to frequency in this frequency division band. F on the left side of Equation (10) is a frequency for displaying this frequency division band, for example, the center frequency of this frequency division band, and has a different meaning from f on the right side.

  At this time, the decision function B (f) is

と与えることができる。ただし、J_Ｔは判定のために予め設定された閾値である。B(f)=1ならば、fで表示される周波数分割帯域における学習計算を行い、B(f)=0ならば、その学習計算を行わないと決定する。

And can be given. However, _JT is a threshold set in advance for determination. If B (f) = 1, learning calculation is performed in the frequency division band indicated by f. If B (f) = 0, it is determined that the learning calculation is not performed.

Expression (11) enables automatic detection of a band that may improve separation accuracy without using information related to a sound source in advance. Note that ΔJ _S (f, t) is smoothed like J _S (f, t), considering the influence of multiple bands, and instead of J _S (f, t), (f, t) may be used as it is.

  The above description is for the case of two sound sources. However, even when the number of sound sources is three or more, the effect of the present invention can be obtained by applying the above method between the sound sources. It is done.

  Hereinafter, the configuration of the present invention will be described with reference to embodiments.

(Embodiment 1)
FIG. 1 is a block diagram of a filter update process in the first embodiment. In the figure, reference numerals 10-1 to 10-n denote a plurality of acoustic sensors that detect the sound in which the target sound and the non-target sound are mixed and output the sound as a plurality of sound signals in which the target sound signal and the non-target sound signal are mixed. 20 is a detection process in which an acoustic signal that is an output of the microphones 10-1 to 10-n is detected and converted into a discrete signal, 30 is a decomposition process of the discrete signal into a frequency, and This is a band division process for dividing the frequency division band. The transform for decomposing the signal into frequencies is generally FFT, but any transform may be used as long as it is an orthogonal transform system such as a wavelet or Z transform.

  The filter learning process 40 is a process of acquiring a filter that separates at least one target speech signal from a plurality of acoustic signals divided into frequency division bands by repeating learning. Any method may be used for this process as long as it is a learning method for each frequency, but in this embodiment, a frequency domain ICA is used.

  The evaluation process 50 is a process of calculating the identification level for evaluating the target speech separation performance of the filter acquired in the filter learning process 40 for each frequency division band. The determination process 60 is a process of determining a frequency division band in which learning calculation is performed based on the identification level calculated in the evaluation process 50, and the determination result is fed back to the filter learning process 40. Note that the evaluation process 50 and the determination process 60 do not have to be executed each time learning is repeated in the filter learning process 40. For example, the operation may be performed once in 10 learning operations, or may be performed every time learning is started, and may be performed once in 10 when learning of most frequency division bands is completed.

  As a specific method for determining the frequency division band for performing the learning calculation, for example, the rate of change between the identification level calculated at time t and the identification level calculated at time tm when m> 0 is set in advance. In a frequency division band that does not exceed a set threshold, there is a method of determining not to perform learning calculation in repetition of learning after time t. At this time, in the frequency division band in which the change rate between the identification level calculated at time t and the identification level calculated at time tm is greater than a preset threshold, frequency division for performing learning calculation after time t When the determination of the band is performed anew, it is determined that the learning calculation in the repetition of learning is performed from time t until the determination is performed, and the frequency division band for performing the learning calculation after time t is not determined. When the learning repetition process ends, it is determined to perform learning calculation in the repetition of learning from time t to the end.

  After completion of learning, the filter acquired in the filter learning process 40 is used as the content of the attenuation process 45 in FIG.

  As described above, the amount of calculation at the time of filter learning can be reduced by omitting useless calculation in adaptive learning for each frequency division band.

As the discrimination level, for example, J (f, t) represented by equation (8) can be used. As the rate of change between the identification levels, for example, ΔJ _S (f, t represented by equation (10) can be used. t) can be used.

  FIG. 2 is a block diagram of the filtering process. The acoustic signals, which are the outputs of the microphones 10-1 to 10-n, in which the target sound signal and the non-target sound signal are mixed, are converted into discrete signals in the detection process 20, and the filter acquired in the filter learning process 40 The signal is output as a target audio signal (signal R100) through an attenuation process 45. The attenuation process 45 extracts the target sound signal from the input acoustic signal or suppresses the non-target sound signal.

  FIG. 3 is a block diagram of the filter update system. As the microphones 110-1 to 110-n, general microphones can be used. The detection unit 120 corresponds to the filter (anti-aliasing filter) 220, the AD converter 230, and the arithmetic device 240 in FIG. 5, and is configured by combining general operation circuits such as a CPU, MPU, DSP, and FPGA. . The band dividing means 130 corresponds to the arithmetic device 240 and the storage device 250 in FIG. The filter learning unit 140 corresponds to the arithmetic device 240 and the storage device 250 in FIG. Evaluation means 150 corresponds to arithmetic device 240 and storage device 250 in FIG. The determining means 160 corresponds to the arithmetic device 240 and the storage device 250 in FIG.

  FIG. 4 is a block diagram of the filter processing system. The microphones 110-1 to 110-n and the detection means 120 are the same as those shown in FIG. The attenuation means 145 corresponds to the arithmetic device 240 and the storage device 250 in FIG. The storage unit 170 corresponds to the storage device 250 in FIG. 5 and is configured by a general storage medium such as a cache memory, a main memory, an HDD, a CD, an MD, a DVD, an optical disk, and an FDD.

  FIG. 5 is a block diagram illustrating an example of a system configuration. The acoustic signals that are the outputs of the microphones 210-1 to 210-n are input to the AD converter 230 through the filter 220, and after AD conversion, are input to the arithmetic device 240 and are subjected to arithmetic processing. The filter 220 is used to remove noise included in the acoustic signal.

  FIG. 6 is a flowchart of the filter learning procedure. S100 to S170 represent individual steps.

  In S100, the system is initialized and the memory is read.

  Sound input is detected in S110. If detected, the process proceeds to S120.

  In S120, input signal band division processing is performed. The bandwidth for each frequency bin may be fixed or variable.

  In S130, learning and updating of the filter are performed. For example, a frequency domain ICA is used.

  In S150, the sound source separation accuracy at time t, that is, the discrimination level Jt is calculated, and further, the rate of change ΔJ between Jt and the discrimination level Jt-m at time t−m (m> 0) is calculated. The identification level Jt at time t is stored.

  In S160, the update flag is determined and the flag is set. That is, when ΔJ ≦ threshold, the flag of the divided band is turned off.

  In S170, the update flag is checked. If all the update flags are OFF, the learning ends. If there is an ON flag, the process returns to S140.

  When the learning is completed, the completed filter is used as the filter of the attenuation process 45 in FIG.

  FIG. 7 is a flowchart of the filter learning procedure (contents of S150 and S160).

  In S151, the discrimination level Jt (for example, the cosine distance of signals of two or more sound sources) at time t is calculated.

  In S152, the change rate ΔJ of the discrimination level is calculated from the time t-m (m> 0) and the discrimination level Jt-m at time t in the divided band ω.

  In S161, if ΔJ ≦ threshold in the divided band ω, the process proceeds to S163, and if not, the process proceeds to S162.

  In S162, the divided band ω is updated, and the process proceeds to S164.

  In S163, the update flag is turned OFF.

  In S164, when the divided band ω does not exist, this flow is ended, and when it exists, the process returns to S152.

  In the above flow, in the frequency division band where the rate of change ΔJ between the identification level Jt calculated at time t and the identification level Jt-m calculated at time tm does not exceed a preset threshold, the time t Thereafter, it is determined not to perform learning calculation in repeated learning.

  Furthermore, in the above flow, in the frequency division band in which the change rate ΔJ between the identification level Jt calculated at time t and the identification level Jt-m calculated at time tm is larger than a preset threshold value. When the frequency division band for which learning calculation is performed after time t is performed again, it is determined to perform learning calculation in the repetition of learning from time t until the determination is performed, and learning calculation is performed after time t When the repetition process of learning ends without determining the frequency division band, it is determined to perform learning calculation in the repetition of learning from time t to the end.

  FIG. 8 is a flowchart of a filtering process procedure for an input signal.

  In S100, the system is initialized and the filter is read into the memory.

  In S110, sound input is detected. If detected, the process proceeds to S180.

  In S180, the input signal is filtered, and the resulting target audio signal is sent out.

  FIG. 9 is a diagram illustrating a calculation result example (cosine distance) of a cost function which is an example of an identification level. In the figure, the horizontal axis represents frequency, and the vertical axis represents a cost function (for example, cosine distance). A rapidly changing line (dotted line) is a calculated value of the cost function, and another line (solid line) is a value after smoothing (smoothing) the calculated value of the cost function. Obtained through several tens to several hundreds of learning sessions. The closer the value of the cost function is to 1, the higher the possibility that separation will not be successful. For example, in the frequency division band of less than 100 Hz or 3500 Hz or more, the cost function value is high, so that a reduction in separation accuracy is predicted. The initial value shows a value close to 1 in all frequency division bands, and the value of the cost function decreases with learning.

  FIG. 10 is a conceptual diagram of the relationship between the cost function and the number of learning times. The horizontal axis represents frequency, and the vertical axis represents a cost function (for example, cosine distance). L110 conceptually shows the value of the cost function when the frequency domain ICA is learned 10 times, and L120 conceptually shows the value of the cost function when the frequency domain ICA is learned 20 times. The rate of change from L110 to L120 is large near 1 kHz, and the rate of change from L110 to L120 is small near 100 Hzz and 3 kHz.

Figure 11 is a conceptual diagram relating to displacement of the cost function J when finished learning when ΔJ does not exceed the threshold value J _T. L210 (dashed line) is the J value for each frequency when learning 10 times, L220 (dotted line) is the J value for each frequency when learning 20 times, and L230 (solid line) is when learning 30 times The value of J for each frequency. Change rate .DELTA.J upon transition from L210 to L220 (20 times) does not exceed the threshold value J _T frequency division band i.e. .DELTA.J (20 times) ≦ J _T a is frequency division band (above 100Hz and less than 3 kHz) is 20 times Then, the learning is terminated, and the learning calculation is not performed thereafter. Rate of change during the transition from L220 to L230 .DELTA.J (30 times) the threshold J does not exceed _T frequency division bands That .DELTA.J (30 times) ≦ J _T a is frequency division band (above 500Hz and less than 2 kHz) is a Learning The learning calculation is not performed thereafter.

  Thus, in the present invention, it is possible to appropriately decide not to perform learning calculation for the frequency division band that does not improve the learning effect, and to reduce the amount of calculation at the time of filter learning.

(Embodiment 2)
Specific examples of executing the evaluation process 50 of FIG. 1 include the following cases.

(Case 1)
The evaluation process 50 is executed every time the learning is repeated. That is, the identification level is calculated every time learning is repeated, and the evaluation is performed by calculating the rate of change between the current and previous identification levels. Although the learning band can be determined most precisely, the calculation amount increases. If the CPU speed is sufficiently high, it can be practically executed.

(Case 2)
The evaluation process 50 is performed once every 10 learning iterations. That is, when the length of time for one repetition of learning is u and i = 10u, the evaluation process 50 executed at time t is executed again at time t + i. Compared to Case 1, the amount of calculation can be greatly reduced. The number of times the evaluation process 50 is performed once, that is, how many times i is greater than u, may be determined in consideration of the environment and the amount of calculation allowed.

(Case 3)
The above time m is set to 5u in length for 5 repetitions of learning. If m = u, there is a possibility that the learning operation of the frequency division band that has a learning effect may be erroneously stopped, but such an error is reduced by increasing the time m in this way. That is, a calculation that is robust (not easily affected) with respect to a minute change is possible.

(Case 4)
Set the learning stop time to the learning timing several times later. When the setting of the threshold value J _T is inappropriate may be repeated learning will take place many times. In order to prevent this, a procedure for setting the learning stop time is inserted in the learning repetition flow. Thereby, the calculation amount regarding evaluation / determination can be reduced. Most simply, the learning operation may be terminated when the number of learning repetitions reaches a specified value. In this case, as shown in FIG. 12, a stop process 80 for ending the learning is performed as shown in FIG. 13, including the completion of learning when the maximum number of learning times is reached in step S170. Provide.

(Case 5)
As the values of Jt and Jt-m, the total or average of the values of the first, second, and previous times of the repetition of learning is adopted. By doing so, it is possible to perform a calculation that is robust (not easily influenced) with respect to a minute change.

(Case 6)
In cases 1 to 4, J _S (j, t) and J _S (j, tm) represented by Equation (9) are used as Jt and Jt-m, and ΔJ _S represented by Equation (10) is represented as ΔJ. (j, t) is used, or J (j, t) and J (j, tm) represented by the formula (8) are used as Jt and Jt-m, and ΔJ is represented by the following formula as ΔJ. Use (j, t).

ΔJ (f, t) = ‖J (f, t) −J (f, tm) ‖ (12)
Here, the meaning of the expression on the right side is the same as that in Expression (10).

(Embodiment 3)
The frequency division band is divided into a plurality of blocks, and learning calculation in repetition of learning is not performed in all frequency division bands belonging to the block including the frequency division band determined not to perform learning calculation in repetition of learning. decide. An example is shown in FIG. In the figure, the frequency division band (not shown) is divided into five blocks from B1 to B5, where the number of frequency division bands ≧ the number of blocks. B1 and B5 end with 10 learnings, B2 and B4 end with 20 learnings, and B3 has progressed to 30 times. By such block division, the amount of calculation related to evaluation / determination can be reduced.

It is a figure explaining the learning method for obtaining the filter in the voice input device concerning the present invention. It is a block diagram of a filtering process in the voice input device according to the present invention. It is a block diagram of the filter update system in the voice input device concerning the present invention. It is a block diagram of the filter processing system in the voice input device concerning the present invention. It is a figure which shows one structural example of the audio | voice input apparatus which concerns on this invention. It is a flowchart of the filter update in the audio | voice input apparatus which concerns on this invention. It is a detailed flowchart of filter update in the voice input device according to the present invention. It is a flowchart of the filter process in the audio | voice input apparatus which concerns on this invention. It is a figure which shows the example of a calculation result of the cost function in a speech input device. It is a conceptual diagram of the relationship between the cost function and the learning frequency in a voice input device. It is a conceptual diagram of the relationship between the cost function and the learning frequency in the voice input device according to the present invention. It is a flowchart of the filter update when the restriction | limiting is provided in the frequency | count of learning in the audio | voice input apparatus which concerns on this invention. It is a figure explaining the learning method for obtaining the filter in the case where the restriction | limiting is provided in the frequency | count of learning in the audio | voice input apparatus which concerns on this invention. It is a conceptual diagram of the relationship between the cost function and the number of learnings when performing block division in the speech input device according to the present invention.

Explanation of symbols

  10-1 to 10-n: microphone, 20: detection process, 30: band division process, 40: filter learning process, 45: attenuation process, 50: evaluation process, 60: decision process, 80: stop process, 110-1 110-n: microphone, 120: detection means, 130: band division means, 140: filter learning means, 145: attenuation means, 150: evaluation means, 160: determination means, 170: storage means, 210-1 to 210- n: microphone, 220: filter, 230: AD converter, 240: arithmetic device, 250: storage device.

Claims (8)

A plurality of acoustic signals in which the target voice signal and the non-target sound signal are mixed are obtained by detecting the sound in which the target voice and the non-target sound are mixed with a plurality of acoustic sensors, and at least one of the plurality of acoustic signals is acquired from the plurality of acoustic signals. In a voice input device that executes an independent component analysis method for acquiring a filter that separates the target voice signal by repetition of learning,
An acoustic frequency band is divided into a plurality of frequency division bands, and an identification level for evaluating the target speech separation performance of the filter obtained by the learning is calculated for each frequency division band in the learning process, A voice input device that determines a frequency division band for performing learning calculation from the frequency division band based on the frequency division band.   In the process of determining the frequency division band for performing the learning calculation, a rate of change between the identification level calculated at time t and the identification level calculated at time tm is set in advance as m> 0. The speech input device according to claim 1, wherein in the frequency division band that does not exceed the threshold, it is determined not to perform learning calculation in repetition of the learning after time t.   The voice input device according to claim 1 or 2, wherein the determination of the frequency division band for performing the learning calculation performed at time t is performed again at time t + i with i> 0.   In the process of determining the frequency division band for performing the learning calculation, a threshold value at which a rate of change between the identification level calculated at time t and the identification level calculated at time tm is set as m> 0 is set in advance. In a frequency division band larger than that, when the determination of the frequency division band for performing the learning calculation after time t is performed again, the learning calculation in the repetition of learning is performed from time t until the determination is performed. When the learning repetition process ends without determining the frequency division band for performing the learning calculation after time t, it is determined that the learning calculation in the learning repetition is performed from time t to the end. The voice input device according to claim 1, 2, or 3.   5. The voice input device according to claim 1, wherein the learning operation is terminated when the number of repetitions of the learning reaches a specified value.   The identification level is a correlation coefficient in a local time interval between two separated signals obtained by filtering the plurality of acoustic signals with the filter, and a change rate between the two identification levels is 6. The voice input device according to claim 1, wherein the voice input device is a norm of a difference vector between two identification levels.   The identification level is a correlation coefficient in a local time interval between two separated signals obtained by filtering the plurality of acoustic signals with the filter, and a change rate between the two identification levels is 6. The discrimination level is a norm of a difference vector between the two smoothed discriminating levels by averaging two discriminating levels in the local frequency section. Voice input device.   The frequency division band is divided into a plurality of blocks, and the learning calculation in the repetition of learning is performed in all frequency division bands belonging to the block including the frequency division band determined not to perform the learning calculation in the learning repetition. The voice input device according to claim 1, wherein the voice input device is determined not to be used.

Images