JP2010049086A - Object signal section estimating device, method and program, and recording medium - Google Patents

Abstract

In a noise environment and in a situation where the direction of arrival of a target signal cannot be accurately determined, a target signal section is accurately estimated with a small amount of calculation.
Each signal observed by a plurality of sensors is cut out for each frame that is a predetermined time interval, the signal of each frame for each cut out sensor is converted into a frequency domain, and the frequency domain for each time frequency bin A signal is generated for each sensor. In addition, the fundamental frequency is estimated, and only the frequency domain signal corresponding to the reference sensor is normalized with respect to only the grid near the fundamental frequency or its harmonic component, and the frequency domain signals corresponding to the sensors other than the reference sensor are normalized. Generate a normalized signal value for each frequency bin. Then, an uneven distribution value indicating the uneven distribution of the normalized signal value is obtained for each grid, and an uneven distribution index value indicating the uneven distribution for each frame is calculated using them, and the uneven distribution index value is used as an index. It is determined whether or not it corresponds to the target signal section.
[Selection] Figure 1

Description

  The present invention relates to a signal processing technique, and more particularly, to a technique for estimating a section where a target signal exists from an observation signal including noise.

  In acoustic signal processing technologies such as encoding, processing for target signals such as audio signals and music signals, suppression of noise signals, dereverberation, and automatic speech recognition, the target signal is derived from an input acoustic signal containing multiple types of signals. It is necessary to estimate existing intervals. The accuracy of the target signal interval estimation greatly affects the subsequent signal processing performance.

  When a plurality of microphones can be used when estimating a target signal section under environmental noise, information on the difference in arrival time of signals can be used to estimate a section of the target acoustic signal. Conventionally, a method for estimating a target signal section by emphasizing a signal in the direction of arrival with a known direction of arrival of the target signal (Non-patent Document 1), and a threshold for acoustic features such as the number of zero crossings are used as the estimated direction of arrival of the target signal A method for determining the presence / absence of speech based on the presence / absence of a spatial spectrum peak (Non-Patent Document 3), and a section in which the estimated signal arrival direction is constant over time. There is a technique (Non-patent Document 4) that uses a section in which speech exists. However, in order to obtain sufficient accuracy by these methods, it is necessary that the arrival direction of the target signal is known or the surrounding environment is quiet.

  Also, there is a method of using a plurality of microphones and estimating a target signal section for each microphone signal and then comparing the estimation results corresponding to the microphones to obtain a final target signal section estimation result. (Non-patent document 5). However, this method cannot sufficiently use the spatial information (information on the arrival direction of the target signal) by using a plurality of microphones.

On the other hand, sufficient target signal interval estimation accuracy using signal arrival time differences in environments where multiple acoustic signals simultaneously arrive in all directions and frequency bands (for example, daily environments such as streets, stations, and airports) As a technique for obtaining the above, there is a technique (Non-patent Document 6) that uses a degree (unevenness) in which a difference in arrival time of signals estimated in a certain time frequency interval is biased to a certain value.
Alvarez, A., Gomez, P., Nieto, V., Martinez, R., and Rodellar, V., "Application of a first-order differential microphone for efficient voice activity detection in a car platform", Proceedings of Interspeech, 2669-2672, 2005. Takamasa Tanaka, Yuka Tomita, Masato Nakayama, Takanobu Nishiura, “Examination of hands-free utterance detection based on weighted CSP method and speech features”, Proceedings of the Acoustical Society of Japan 2006 Spring Meeting, 1-P-3, pp 149-150, Mar. 2006. Kiyoshi Yamamoto, Tataka Asano, Takashi Yoshimura, Yoichi Motomura, Hideki Aso, Isao Hara, Naoyuki Ichimura, Satoshi Ogata, Nobuhiko Kitawaki, “Evaluation of Speech Section Detection and Separation System by Integration of Acoustic Information and Image Information,” Acoustical Society of Japan Proceedings of Autumn Research Presentation, 3-6-10, P121-122, 2003. Masayoshi Fujimoto, Yasuo Ariki, Shuji Doshita, “Person Recognition in News Video by Multimodal Interaction,” Journal of the Acoustical Society of Japan, Vol. 62, no. 3, P182-192, 2006. Akiko Araki, Masaki Fujimoto, Kentaro Ishizuka, Hiroshi Sawada, Shoji Makino, “Conference Speech Speaker Identification System Using Speech Interval Detection and Direction Information and Its Evaluation”, Spring Meeting of the Acoustical Society of Japan, pp. 1-4 2008. Juan E. Rubio, Kentaro Ishizuka, Hiroshi Sawada, Shoko Araki, Tomohiro Nakatani, and Masakiyo Fujimoto, "Two-Microphone Voice Activity Detection Based on the Homogeneity of the Direction of Arrival Estimates," Proceedings of the 32nd International Conference on Acoustics, Speech , and Signal Processing, Vol. 4, pp. 385-388, 2007.

  However, in the method of Non-Patent Document 6, a large amount of calculation is required to calculate the uneven distribution in all time frequency sections, and if there is directional noise, it is also detected as a target signal. There is a problem.

  The present invention has been made in view of the above points, and in a situation where it is under noisy environment and the arrival direction of the target signal cannot be accurately determined, the target signal section can be accurately obtained with a small amount of calculation. It is an object to provide a technique capable of estimating

  In the present invention, in order to solve the above-described problem, first, the signal extraction unit cuts out each signal observed by a plurality of sensors for each frame that is a predetermined time interval, and the frequency domain conversion unit generates a signal extraction unit. The signal of each frame cut out in step 1 is converted into the frequency domain, and a frequency domain signal for each time frequency bin is generated for each sensor. Further, the fundamental frequency estimation unit estimates the fundamental frequency of each frame signal cut out by the signal cutout unit, and the time frequency domain division unit includes a finite time frequency section including the fundamental frequency or each harmonic component thereof. 1 or more is specified for each frame, and the frequency domain signal of each time frequency bin belonging to each grid is extracted. Thereafter, the normalization unit normalizes at least each frequency domain signal corresponding to a sensor other than the reference sensor based on the frequency domain signal extracted by the time frequency domain division unit corresponding to the specific reference sensor included in the sensor. And a normalized signal value extracted by the time-frequency domain dividing unit corresponding to the direction of arrival of the signal observed by the sensor is generated for each time-frequency bin. Then, the ubiquity index value calculation unit obtains an ubiquitous value indicating the ubiquity of the normalized signal value for each grid, and uses the ubiquitous value for each grid to indicate the ubiquity of the normalized signal value for each frame. An index value is calculated.

  Here, the normalized signal value generated by the normalization unit of the present invention is a value corresponding to the direction of arrival of the signal. Normally, environmental noises arrive at the sensor from various directions, whereas the target signal has a property (characteristic 1) that arrives at the sensor only from a certain direction. Therefore, the normalized signal values of time frequency bins where the target signal does not exist are widely distributed (low unevenness), whereas the normalized signal values of the time frequency bin where the target signal exists are in the direction of arrival of the target signal. Distributed in the vicinity of the corresponding value (highly ubiquitous). Further, the fundamental frequency of the same target signal or its harmonic component (a frequency component that is an integral multiple of the fundamental frequency) is distributed narrowly in the time frequency domain, whereas the noise power is widely distributed in the time frequency domain (property 2). ). In the present invention, using these properties, an uneven value indicating the uneven distribution for each grid, which is a finite time frequency section including the fundamental frequency or its harmonic component, is obtained, and the uneven distribution value for each grid is used for normalization. An ubiquity index value indicating the ubiquity of the signal value for each frame is calculated. Thereby, the target signal section can be estimated with high accuracy. Also, in the present invention, since the normalized signal value is obtained only for each time frequency bin belonging to each grid and the uneven value is obtained only for each grid, the normalized signal value and the uneven value are obtained in all time frequency sections. Compared to the amount of calculation. In addition, when the uneven distribution of normalized signal values is used as an index, it is not necessary to know the arrival direction of the target signal accurately. Therefore, in the present invention, even when the accurate arrival direction of the target signal cannot be estimated, the target signal section can be estimated appropriately.

  Preferably, in the present invention, the uneven distribution index value calculation unit quantizes the normalized signal value, obtains the frequency of the quantized normalized signal value for each grid, and generates a histogram for each grid. Then, using the histogram for each grid, the ubiquity calculation unit that calculates the ubiquitous value indicating the deviation of the distribution of the histogram for each grid, the omnidirectional value of each grid corresponding to the same frame, and the average value And an average part calculated as an uneven distribution index value of the frame.

  In this way, by averaging the uneven distribution values of the grids corresponding to the same frame and calculating the average value as the uneven distribution index value of the frame, a noise component whose power and arrival direction are widely distributed in the time frequency domain Can be reduced, and the estimation accuracy of the target signal section can be improved.

  Preferably, in the present invention, the histogram generation unit weights the frequency of the quantized normalized signal value using a weighting factor, and generates a histogram for each grid using the weighted frequency. By setting this weighting factor as appropriate, an optimal target signal interval estimation method can be constructed for each environment.

  In the present invention, it is preferable that the ubiquitous calculation unit uses a histogram for each grid, and obtains a probability density function for each grid, which uses a value corresponding to each quantized normalized signal value as a random variable. A function generation unit; and a ubiquitous value calculation unit that obtains a function value monotonously increasing with respect to the entropy of the probability density function or a function value monotonously decreasing with respect to the entropy as an ubiquitous value. By obtaining the ubiquitous value in this way, a ubiquitous index value that takes a small value in a section where the target signal exists and takes a large value in a section where the target signal does not exist, or a large value in a section where the target signal exists. Therefore, it is possible to generate an uneven distribution index value that takes a small value in a section where there is no target signal.

  Preferably, in the present invention, the normalization unit uses the phase and / or amplitude of the frequency domain signal corresponding to the reference sensor as a reference, and at least the phase and / or each frequency domain signal corresponding to a sensor other than the reference sensor. The amplitude is normalized, and a normalized signal value that is the normalized value or a map thereof is generated. In this case, the normalized signal value is preferably a value obtained by normalizing frequency components and eliminating frequency dependence. When the frequency dependency of the normalized signal value is not excluded, the normalized signal value in the time frequency bin of the target signal is a value depending on the arrival direction and the frequency of the signal. On the other hand, when the frequency dependence of the normalized signal value is eliminated, the normalized signal value in the time frequency bin of the target signal is a value that depends only on the arrival direction of the signal. That is, even if the normalized signal values correspond to the same target signal, the normalized signal value from which the frequency dependence is eliminated is more unevenly distributed than the normalized signal value from which the frequency dependence is not eliminated. . As a result, it is possible to obtain an unevenness index value in which the unevenness of the normalized signal value caused by the target signal appears more clearly, and the accuracy of estimation of the target signal section performed using the unevenness index value as an index is improved.

  In the present invention, for example, the determination unit compares the ubiquity index values of each frame or their mapping with a predetermined threshold value, and determines whether each frame is a target signal section. Further, the determination unit has a division value that is a ratio of the ubiquity index value of the frame to be determined and the ubiquity index value of the frame in the non-target signal section, or the mapping of the division value is equal to or greater than a predetermined threshold value. The determination target frame may be determined to be the target signal section, or if the predetermined threshold is exceeded, the determination target frame may be determined to be the target signal section. Further, for example, the determination unit uses the pattern recognition using the relationship between the pre-learned frame ubiquitous index value and the determination result as to whether or not the frame is the target signal section. It may be determined whether or not the frame corresponding to the calculated uneven distribution index value is the target signal section.

  As described above, according to the present invention, it is possible to accurately estimate the target signal section with a small amount of calculation in a situation where the direction of arrival of the target signal cannot be accurately obtained even in a noisy environment. .

The best mode for carrying out the present invention will be described below with reference to the drawings.
FIG. 1 is a block diagram illustrating the overall configuration of a target signal section estimation device 10 of the present embodiment. FIG. 2 is a block diagram illustrating a detailed configuration of the uneven distribution index value calculation unit 16 of FIG. FIG. 3 is a block diagram illustrating a detailed configuration of the determination unit 17 of FIG.

<Configuration>
As illustrated in FIG. 1, the target signal interval estimation device 10 of the present embodiment includes a signal extraction unit 11, a frequency domain conversion unit 12, a fundamental frequency estimation unit 13, a time frequency domain division unit 14, a normalization unit 15, and an uneven distribution. A sex index value calculation unit 16, a determination unit 17, a control unit 18, and a storage unit 19 are provided, and signals observed by S (S ≧ 2) sensors 20-1 to S and sampled by the sampling unit 30 are obtained. It is a device that inputs and outputs the analysis result of the target signal section. As illustrated in FIG. 2A, the ubiquitous index value calculation unit 16 of this example includes a histogram generation unit 16a, a probability density function calculation unit 16b, an entropy calculation unit 16c, and an average unit 16d. As illustrated in FIG. 2B, the determination unit 17 of this example includes a relative value calculation unit 17a, a likelihood ratio calculation unit 17b, and a threshold determination unit 17c.

  The target signal section estimation device 10 is configured by causing a known computer including a CPU (central processing unit), a RAM (random access memory), a ROM (read only memory), and the like to execute a predetermined program, for example. It is what is done.

<Processing>
Next, the target signal section estimation method of this embodiment will be described.
In the target signal section estimation method of this embodiment, each signal observed by a plurality of sensors 20-1 to S (S ≧ 2) is subjected to time-frequency analysis, and a normalized signal value based on a specific reference sensor is obtained, Based on the uneven distribution of the normalized signal value in the grid that is a predetermined time frequency interval, the presence or absence of the target signal is detected and output. In this embodiment, a microphone is used as the plurality of sensors 20-1 to 20 -S, and each acoustic signal observed with them is used to detect and output the presence of a target signal such as a voice signal or a music signal. Illustrate. Further, although not specified below, the target signal section estimation device 10 executes each arithmetic processing based on the control of the control unit 18, and data obtained in the process of each arithmetic processing is sequentially stored in the storage unit 19, It is used for each subsequent calculation process.

  FIG. 4 is a flowchart for explaining the target signal section estimation method of the present embodiment. FIG. 5 is a flowchart for explaining details of step S7, and FIG. 6 is a flowchart for explaining details of step S8. Hereinafter, the target signal section estimation method of this embodiment will be described along these flowcharts.

First, each signal observed by S (S ≧ 2) sensors 20-1 to S is input to the sampling unit 30. These signals include environmental noise signals in addition to target signals such as audio signals and music signals. The sampling unit 30 samples each signal at a predetermined sampling frequency f _s (for example, 16,000 Hz), and thereby, the time domain signals x (1, t),. x (S, t) is extracted (step S1). Note that t represents the t-th sampling point.

  Signals x (1, t),..., X (S, t) in each time domain extracted by the sampling unit 30 are input to the signal extraction unit 11 of the target signal section estimation device 10. The signal cutout unit 11 cuts out each input signal x (1, t),... X (S, t) for each frame that is a predetermined time interval, and each of the sensors 20-1 to 20S. The signals x ′ (1, i, n),..., X ′ (S, i, n) of the frame i (i indicates a frame index) are extracted (step S2). Note that n represents the nth sample point in frame i. Specifically, for example, the signal cutout unit 11 applies a predetermined window function to each input signal x (1, t),..., X (S, t), for example, in the time axis direction. By multiplying while shifting (shifting) by 16 ms, for example, a signal x ′ (1, i, n),..., X ′ (S, i, n) having a time length of 32 ms is cut out. More specifically, for example, when the sampling frequency is 16,000 Hz, the signal extraction unit 11 performs, for example, for each input signal x (1, t),..., X (S, t), respectively. The Hanning window of Equation (1) is multiplied while moving (shifted) by 256 sample points (16,000 Hz × 16 ms), and a discrete signal of 512 sample points (16,000 Hz × 32 ms) is obtained for each sensor 20-1 to S. Cut out as a signal for one frame. Here, L represents the number of sample points (frame length: L = 512 in the above example) of the signal for one frame to be cut out.

信号切出部１１は、以上のように切り出した各センサ２０−１〜Ｓについての各フレームiの信号x’(1,i,n),...,x’(S,i,n)を出力し、これらは周波数領域変換部１２に入力される。

The signal cutout unit 11 outputs signals x ′ (1, i, n),..., X ′ (S, i, n) of each frame i for the sensors 20-1 to S cut out as described above. These are input to the frequency domain transform unit 12.

  The frequency domain converter 12 converts the signals x ′ (1, i, n),..., X ′ (S, i, n) of each frame i for the sensors 20-1 to S to the frequency domain. , Frequency domain signals (frequency domain spectrum) X (1, i, k),..., X (S, i, k) for each time frequency bin (i, k) are generated for each sensor 20-1 to S. (Step S3). When performing this transformation by the discrete Fourier transform, the frequency domain transformation unit 12 uses frequency domain signals X (1, i, k),..., X (S, i, k) as shown in the following equation (2). Is calculated.

ここで、ｊは虚数単位を示し、ｓ（s∈{1,...,S}）は各センサ２０−１〜Ｓの番号を示す。また、ｋ（k=0,...,M-1）は周波数インデックスであり、サンプリング周波数ｆ_ｓをＭ等分した離散点を表す。Ｍはフレーム長Ｌ以上の自然数であり、例えば、M=512とする。周波数領域変換部１２は、以上のような変換によって得られた周波数領域信号（周波数スペクトル）X(1,i,k),...,X(S,i,k)を出力する。

Here, j indicates an imaginary unit, and s (sε {1,..., S}) indicates the number of each sensor 20-1 to S. K (k = 0,..., M−1) is a frequency index, which represents a discrete point obtained by dividing the sampling frequency f _s into M equal parts. M is a natural number greater than or equal to the frame length L, for example, M = 512. The frequency domain transform unit 12 outputs frequency domain signals (frequency spectrum) X (1, i, k),..., X (S, i, k) obtained by the above transformation.

Further, the signals x ′ (1, i, n),..., X ′ (S, i, n) of each frame i for the sensors 20-1 to S cut out by the signal cutout unit 11 are: It is also input to the fundamental frequency estimation unit 13. The fundamental frequency estimator 13 uses the time domain signals x ′ (1, i, n),..., X ′ (S, i, n) of each frame i and uses the fundamental for each sensor s and frame i. The frequencies F ₀ (1, i),..., F ₀ (S, i) are estimated (step S4). For this estimation, for example, the following autocorrelation method (see, for example, “Quatieri, TF,“ Discre-time Speech Signal Processing principles and practice, ”Prentice-Hall, 2002; pp. 504-505”) is used. In this case, the fundamental frequency estimation unit 13 first obtains the coefficient c (s, i, n) of the autocorrelation function for each sensor s and frame i for n = 1,.

なお、周波数領域変換部１２が離散フーリエ変換によって周波数領域信号X(1,i,k),...,X(S,i,k)を算出する場合、基本周波数推定部１３は、周波数領域変換部１２の出力である各周波数領域信号X(1,i,k),...,X(S,i,k)の絶対値を２乗して逆フーリエ変換し、各自己相関関数の係数c(s,i,n)を求めることもできる。

When the frequency domain transform unit 12 calculates the frequency domain signals X (1, i, k),..., X (S, i, k) by discrete Fourier transform, the fundamental frequency estimation unit 13 The absolute value of each frequency domain signal X (1, i, k),..., X (S, i, k) that is the output of the transform unit 12 is squared and inverse Fourier transformed, and each autocorrelation function The coefficient c (s, i, n) can also be obtained.

Next, the fundamental frequency estimation unit 13 performs a constant search range of n for each sensor s and frame i, for example, 32 ≦ n ≦ 320 (a frequency range from 50 Hz to 500 Hz when the sampling frequency f _{s is} 16,000 Hz). N) where the coefficient c (s, i, n) of the autocorrelation function is maximized. The resulting n corresponds to the period length of the most dominant periodic component in the search range of the input signal x ′ (1, i, n), ..., x ′ (S, i, n). , If the input signals x '(1, i, n), ..., x' (S, i, n) are each a single perfect periodic signal (eg sine wave), it corresponds to the period length To do. The fundamental frequency estimator 13 divides the sampling frequency f _s by n obtained for each sensor s and frame i, so that the fundamental frequency F ₀ (1, i),. , F ₀ (S, i) are generated and output. Note that parallel processing, SIFT algorithm, cepstrum analysis, etc. may be used as the fundamental frequency estimation method (for example, “Sadaaki Furui,“ Digital Speech Processing ”, Tokai University Press, ISBN4-486-00896-0”). reference).

Next, each time frequency domain dividing unit 14 is supplied with each fundamental frequency F ₀ (1, i),..., F ₀ (S, i) and frequency domain signal X (1, i, k),. , X (S, i, k) are input. The time-frequency domain dividing unit 14 inputs each input fundamental frequency F ₀ (1, i),..., F ₀ (S, i) or each harmonic component (a frequency component that is an integral multiple of the fundamental frequency). A grid that is a finite time frequency interval including one or more is specified for each sensor s and frame i, and the frequency domain signal XGRID ₁ (1, i, k), ..., of each time frequency bin belonging to each grid. XGRID _G (S, i, k) is extracted and output (step S5). Each grid of the frame i of the sensor s is a fixed time frequency interval in the vicinity of the fundamental frequency F ₀ (s, i) or each harmonic component thereof, for example, at the fundamental frequency F ₀ (s, i). A time frequency interval in the predetermined time frequency range from the nearest time frequency bin, and a time frequency interval in the predetermined time frequency range from the time frequency bin closest to each harmonic component of the fundamental frequency F ₀ (s, i) It is. For example, if the frequency bin closest to the fundamental frequency F ₀ (s, i) and the frequency bin closest to each harmonic component of the fundamental frequency F ₀ (s, i) are expressed as k ′ (f = f _s · k ′) / M is the closest to the fundamental frequency F ₀ (s, i) or its harmonic component), the frequency domain signal XGRID _g (s, i, k) of each time frequency bin belonging to each grid of frame i of sensor s (G = 1, ..., G) can be expressed as follows.
XGRID _g (s, i, k) = {X (s, i + P, k '+ Q)} ... (4)

ここで、Aは時間方向の幅を示し、Bは周波数方向の幅を示し、｛・｝は・を要素とする集合を意味する。A、Bには例えばA=9,B=5（時間幅160ms、周波数幅156.25Hz）を用いる。また、Ｇは基本周波数及びその各倍音成分の合算数を示す。Ｇは例えば定数である。

Here, A indicates the width in the time direction, B indicates the width in the frequency direction, and {·} means a set whose element is. For A and B, for example, A = 9, B = 5 (time width 160 ms, frequency width 156.25 Hz) is used. G represents the total number of fundamental frequencies and their harmonic components. G is a constant, for example.

[Preferred grid width setting method]
As described above, in the present invention, an uneven value indicating the uneven distribution of each normalized signal value for each grid is obtained, and the uneven distribution value indicating the uneven distribution of each normalized signal value for each frame is obtained using the uneven distribution value for each grid. The sex index value is calculated, and it is determined whether or not the frame is the target signal section. Here, if the time frequency section of the grid is too wide, the uneven distribution of the normalized signal value in the grid is flattened, and it is difficult to determine whether the signal is the target signal section or the non-target signal section from the uneven distribution. It becomes. On the other hand, if the time frequency section of the grid is too narrow, the number of samples is small and the unevenness of the normalized signal value in all grids becomes high, and it is determined whether it is the target signal section or the non-target signal section from the uneven distribution. It becomes difficult to judge. Therefore, it is necessary to set the grid width within a range where such a problem does not occur. A preferable grid width setting method will be described below.

<< About A in Formula (5) >>
When the signal is an audio signal, A corresponding to a time length of 50 to 300 ms in which the steadiness of the audio signal can be assumed is determined. That is, if the width of the frame shift is SF ms, an integer value between 50 / SF and 300 / SF may be A. Also, if the speaker's speech rate SR syllables / sec (the number of syllables spoken per second) is known in advance, an integer value in the vicinity of (1000 / SR) / SF (for example, the closest) may be set to A. (For example, if SR = 7 syllables / sec and SF = 16 ms, (1000 / SR) / SF = (1000/7) /16=8.93, so A = 9). If the target signal is a music signal, it is desirable to use a value for obtaining A in the same way from the rhythm of music (corresponding to the SR of sound).

<< About B in Formula (5) >>
Basically, a width obtained from the main lobe width of the window function w (n) may be used. For example, the discrete Fourier transform value of the window function w (n) is W (k), and the maximum value satisfying 20 log ₁₀ (W (k) / W (0))>-60dB in the range of 1 <k <M / 2 And cf · 2 + 1 (for example, the closest integer value) is B. This value varies depending on the sampling frequency f _s , the analysis frame length L, and the total number M of frequency bins of the discrete Fourier transform (for example, if the sampling frequency is 8 kHz, the width of the window function is 256 sampling points, and M = 256, cf = 2 and B = 5).

Further, when the fundamental frequency estimated by the fundamental frequency estimation unit 13 is F ₀ (s, i) Hz, for example, to prevent harmonic components of two or more audio signals from entering one grid, for example, If B = 2 · F ₀ (s, i) / (f _s / M) +1 and this is larger than the width obtained from the main lobe width, the value obtained from the main lobe width should be adopted. It is good. For example, if the sampling frequency is 8 kHz, the window function width is 256 sampling points, and M = 256, then F = ₀ (s, i) = 50 Hz, so B = 2 · 50 · (8000/256) + 1 = 4.2 For example, B = 4. On the other hand, if F ₀ (s, i) = 200 Hz, B = 2 · 200 · (8000/256) + 1 = 13.8, but this is larger than the value B = 5 obtained from the above main lobe width. = 5 is adopted. This is because the arrival direction of the audio signal is unevenly distributed only within the main lobe width. The same applies to the case where the target signal is a music signal (end of description of “preferred grid width setting method”).

Next, the frequency domain signals XGRID ₁ (1, i, k),..., XGRID _G (S, i, k) of each time frequency bin belonging to each grid output from the time frequency domain dividing unit 14 are normalized. Is input to the conversion unit 15. The normalization unit 15 uses the frequency domain signal XGRID ₁ (s _B , i, k),... Extracted by the time frequency domain division unit 14 corresponding to a specific reference sensor s _B ∈ {1,. .., XGRID _G (s _B , i, k) as a reference, each frequency domain signal XGRID extracted by the time frequency domain dividing unit 14 corresponding to at least a sensor s (≠ s _B ) other than the reference sensor s _{B 1} (s, i, k), ..., XGRID _G (s, i, k) is normalized and the normalized signal value ZGRID ₁ (i, k), corresponding to the direction of arrival of the signal observed by the sensor ..., ZGRID _G (i, k) is generated for each time frequency bin (i, k) (step S5). Each such normalized signal value ZGRID ₁ (i, k), ..., ZGRID _G (i, k) is the arrival of the target signal in the time frequency bin (i, k) where the target signal exists. It becomes biased to a value corresponding to the direction. An example of the normalized signal value ZGRID _g (i, k) (g = 1,..., G) generated by the normalization unit 15 is shown below.

[Example of normalized signal value ZGRID _g (i, k)]
In this embodiment, as an example of the normalized signal value ZGRID _g (i, k), S = 2, the frequency domain signal XGRID _g (1, i, k) corresponding to the reference sensor 20-1, and the other sensor 20- 2 is estimated from the frequency domain signal XGRID _g (2, i, k) corresponding to 2, and the estimated signal arrival direction is set as a normalized signal value ZGRID _g (i, k) (normalized signal value). Example 1 of ZGRID _g (i, k). In this example, the normalization unit 15 sets the signal arrival direction θ _g (i, k) calculated by the following equations (7) and (8) as the normalized signal value ZGRID _g (i, k). Ν represents the speed of sound (about 340 m / sec), d represents the distance between sensors (m), f represents the discrete frequency f = f _s · k / M corresponding to the frequency bin k, arg (·) Indicates the phase (deflection angle). Further, τ _g (i, k) indicates a signal arrival time difference from the signal source to each of the sensors 20-1, 2 and θ _g (i, k) indicates a signal arrival direction estimated value. Further, the signal arrival direction θ _g (i, k) calculated by the equation (8) passes through the midpoint of the line connecting the sensors 20-1 and 20-2, and is an angle with the direction orthogonal to the line as 0radian. (Radian). The normalized signal value ZGRID _g (i, k) calculated in this way is a value in which the frequency component f is normalized and the frequency dependency is eliminated.

また、前述の式（７）で算出された信号到達時間差τ_g(i,k)を正規化信号値ZGRID_g(i,k)としてもよい（正規化信号値ZGRID_g(i,k)の例２）。なお、このように算出された正規化信号値ZGRID_g(i,k)も周波数成分ｆが正規化され、周波数依存性が排除された値となる。

The signal arrival time difference tau _g (i, k) calculated in formula (7) described above the normalized signal value zgrid _g (i, k) may be (normalized signal value zgrid _g (i, k) Example 2). Note that the normalized signal value ZGRID _g (i, k) calculated in this way is also a value in which the frequency component f is normalized and the frequency dependency is eliminated.

The phase difference arg (XGRID _g (2, i of the frequency domain signal _{XGRID g (2, i, k} ) a frequency domain signal XGRID _g with respect to the phase of the (1, i, k), k) / XGRID g (1, i , k)) may be set as the normalized signal value zGRID _g (i, k) (eg normalized signal value _{zGRID g (i, k) 3} ), the frequency domain signal _{XGRID g (2, i, k} ) phase and frequency domain signal _{XGRID g (1, i, k} ) the difference between the phase of _{arg (XGRID g (2, i} , k)) - arg (XGRID g (1, i, k)) normalized signal value ZGRID _g (i, k) may be used (Example 4 of normalized signal value ZGRID _g (i, k)). Further, the ratio of the amplitude of the frequency domain signal XGRID _g (1, i, k) to the amplitude of the frequency domain signal XGRID _g (2, i, k) | XGRID _g (2, i, k) | / | XGRID _g (1, i, k) | (may be a i, k) (normalized signal value zGRID _g (i, k) a normalized signal value zgrid _g example 5), the frequency domain signal _{XGRID g (1, i, k} ) Of power of frequency domain signal XGRID _g (2, i, k) to power of | XGRID _g (2, i, k) | ² / | XGRID _g (1, i, k) | ² is normalized signal value ZGRID _g (i, k) may also be used (Example 6 of normalized ZGRID _g (i, k)). In either case, only in the time frequency bin (i, k) where the target signal exists, a value biased to a value corresponding to the direction of arrival of the target signal is taken, so that the normalized signal value ZGRID _g (i, k) Whether or not the target signal exists can be determined using the uneven distribution as an index.

Moreover, although the case where the number of sensors was two was illustrated above, when the number of sensors is three or more, for example, as shown below, the arrival azimuth angle estimated value θ _g (i, k) of the target signal and the elevation angle estimation value phi _g (i, k) and seeking, these two values time-frequency bins (i, k) the normalized signal values for zGRID _g (i, k) may be (normalized signal value zgrid _g (i, Example 7) of k).

First, a coordinate vector in the space of each sensor 20-s (s = 1,..., S) is set to d _s = [x coordinate, y coordinate, z coordinate]. Further, a J (J∈ (1,..., S)) th sensor 20-J is set as a reference sensor, and a distance vector D between the reference sensor 20-J and each sensor 20-s is expressed by the following equation (9). Set as follows. [•] ^T indicates transposition of a vector.

D = [d ₁ -d _J , d ₂ -d _J , ..., d _S -d _J ] ^T ... (9)
Further, a signal arrival time difference τ _g (s, i, k) between the reference sensor 20-J and each sensor 20-s is obtained by the following equation (10), and a signal arrival time difference vector τ _g ′ () using them as elements. i, k) is obtained as in the following equation (11).

τ_g'(i,k)=[τ_g(1,i,k),τ_g(2,i,k),...,τ_g(S,i,k)]^T ...(11)
上述の式（９）〜（１１）には以下の式（１２）の関係が成り立ち、以下の式（１２）から目的信号の到来方位角推定値θ_g(i,k)と仰角推定値φ_g(i,k)とを求める。なお、式（１２）におけるD^-1はムーア・ペンローズ型一般化逆行列などの一般化逆行列である。また、目的信号の到来方位角とはｘ−ｙ平面上の目的信号の到来方向を意味し、目的信号の仰角とはｘ−ｚ平面上の目的信号の到来方向を意味する。また、ｙ軸線方向が0radianである。

τ _g '(i, k) = [τ _g (1, i, k), τ _g (2, i, k), ..., τ _g (S, i, k)] ^T ... (11 )
The relationship of the following equation (12) is established in the above equations (9) to (11). From the following equation (12), the arrival azimuth angle estimated value θ _g (i, k) of the target signal and the elevation angle estimated value φ _{Find g} (i, k). In Equation (12), D ⁻¹ is a generalized inverse matrix such as a Moore-Penrose type generalized inverse matrix. Also, the arrival azimuth angle of the target signal means the arrival direction of the target signal on the xy plane, and the elevation angle of the target signal means the arrival direction of the target signal on the xz plane. The y-axis direction is 0radian.

ν ・ D ⁻¹・ τ _g ′ (i, k) = (cosθ _g (i, k) cosφ _g (i, k), sinθ _g (i, k) sinφ _g (i, k), sinφ _g (i , k)] ^T ... (12)
Further, the normalized signal value ZGRID _g (i, k) illustrated in Examples 1 to 7 of the above-described normalized signal value Z (i, k) is combined, and two or more normalized values are obtained for each time frequency bin (i, k). The normalized signal value ZGRID _g (i, k) may be calculated (Example 8 of the normalized signal value ZGRID _g (i, k)). For example, the ratio of the phase difference arg (XGRID _g (2, i, k) / XGRID _g (1, i, k)) to the amplitude | XGRID _g (2, i, k) | / | XGRID _g (1, i, k k) | may be a normalized signal value ZGRID _g (i, k) of the time frequency bin (i, k). Also, for example, S = 3, and the ratio of the phase difference arg (XGRID _g (2, i, k) / XGRID _g (1, i, k)) and amplitude | XGRID _g (3, i, k) | / | XGRID The pair with _g (1, i, k) | may be the normalized signal value ZGRID _g (i, k) of the time frequency bin (i, k). Moreover, (End description of the example of the normalized signal values _{ZGRID g (i, k)]} ) mapping the normalized signal value ZGRID _g (i, k) of the value generated as described above may be.

As described above, in step S6, the normalization unit 15 generates and outputs the normalized signal values ZGRID ₁ (i, k),..., ZGRID _G (i, k) as described above.

Each normalized signal value ZGRID ₁ (i, k),..., ZGRID _G (i, k) output from the normalizing unit 15 is input to the uneven distribution index value calculating unit 16. The ubiquitous index value calculation unit 16 has an ubiquitous value H ₁ (i, k) indicating the ubiquity of each normalized signal value ZGRID ₁ (i, k),..., ZGRID _G (i, k) for each grid. , ..., H _G (i, k), and using the uneven value H ₁ (i, k), ..., H _G (i, k) for each grid, the normalized signal value frame i The ubiquity index value H (i) indicating the ubiquity for each is calculated (step S7). Details of step S7 will be exemplified below.

[Example of Step S7]
In this example, first, the histogram generation unit 16a (FIG. 2) of the ubiquitous index value calculation unit 16 inputs the respective normalized signal values ZGRID ₁ (i, k), ..., ZGRID _G (i, k) is quantized into C values Z (c) (c = 1,..., C), respectively, and the frequency bin ₁ (i, k, c), ..., bin _G (i, k, c) (c = 1, ..., C) is obtained for each grid, and a histogram for each grid is generated (step S71). For example, when the normalized signal value ZGRID _g (i, k) is the signal arrival direction θ _g (i, k) and C = 32, each normalized signal value ZGRID _g (i, k) is It is quantized into such C normalized signal values Z (c).

Z (1) (-π / 2 ≦ ZGRID _g (i, k) <-7π / 16)
Z (2) (-7π / 16 ≦ ZGRID _g (i, k) <-3π / 16)
...
Z (C) (7π / 16 <ZGRID _g (i, k) <π / 2)
When the signal arrival time difference τ _g (i, k) calculated by the above equation (7) is the normalized signal value ZGRID _g (i, k), the histogram generation unit 16a, for example, | τ _g The normalized signal value ZGRID _g (i, k) is quantized to C in units of (i, k) | ≦ (d / ν) × α (α is a positive constant).

Then, the histogram generator 16a determines which normalized signal value Z (c) corresponds to the normalized signal value ZGRID _g (i, k) for each time frequency bin (i, k), and the frequency thereof. Is counted for each grid, and a histogram for each grid is generated. At this time, the histogram generation unit 16a weights the frequency of the quantized normalized signal value Z (c) using a certain weighting coefficient, and generates a histogram for each grid using the weighted frequency. Also good. For example, the histogram generation unit 16a may weight the frequency with the weighting factor W (i, k) of the corresponding time frequency bin (i, k) when counting the frequency. More specifically, for example, if the value obtained by quantizing the normalized signal value ZGRID _g (i, k) of the time frequency bin (1, 2) is Z (5), the frequency for Z (5) is Count W (1,2). That is, the frequency bin _g (i, k, c) (c = 1, ..., C) of the quantized normalized signal value Z (c) with respect to the normalized signal value ZGRID _g (i, k) is For example, you may count like the following formula | equation (13).

bin _g (i, k, c) = ΣW (i + P, k + Q) if ZGRID _g (i + P, k + Q) ∈Z (c) ... (13)

<< Example of weighting factor W (i, k) >>
An example of the weighting factor W (i, k) is shown below. As an example of the weighting factor W (i, k), for example, the frequency domain signals X (1, i, k),..., X (S, i, k) is summed and normalized by the sum of the power of the frequency domain signals X (1, i, k), ..., X (S, i, k) for all sensors and all frequencies. (Example 1 of weighting factor W (i, k)).

また、重み係数W(i,k)として、例えば、以下の式（１５）のように、全センサについての周波数領域信号X(1,i,k),...,X(S,i,k)の振幅の絶対値を合算し、それを全センサ・全周波数についての周波数領域信号X(1,i,k),...,X(S,i,k)の振幅の絶対値の総和で正規化した値を用いてもよい（重み係数W(i,k)の例２）。

Further, as the weighting factor W (i, k), for example, the frequency domain signals X (1, i, k),..., X (S, i, The absolute value of the amplitude of k) is summed, and the sum of the absolute values of the amplitudes of the frequency domain signals X (1, i, k), ..., X (S, i, k) for all sensors and all frequencies. A value normalized by the sum may be used (Example 2 of weighting factor W (i, k)).

また、式（１４）（１５）のような正規化を行わないで重み係数W(i,k)を求めてもよい（重み係数W(i,k)の例３）。この場合であっても、雑音環境によっては十分に目的信号区間推定が可能な場合もある。例えば、以下の式（１１）（１２）のように重み係数W(i,k)を求めてもよい。

Further, the weighting factor W (i, k) may be obtained without performing normalization as in the equations (14) and (15) (Example 3 of the weighting factor W (i, k)). Even in this case, the target signal section may be sufficiently estimated depending on the noise environment. For example, the weighting coefficient W (i, k) may be obtained as in the following equations (11) and (12).

また、全センサについての周波数領域信号X(1,i,k),...,X(S,i,k)の振幅の絶対値やパワーを合算するのではなく、一部のセンサについての周波数領域信号X(1,i,k),...,X(S,i,k)の振幅の絶対値やパワーを合算したり、以下の式（１８）（１９）のように１個のセンサ２０−Ｊの周波数領域信号X(J,i,k)の振幅の絶対値やパワーを重み係数W(i,k)としたりしてもよい（重み係数W(i,k)の例４）。なおこの場合には、できるだけ信号源に近い（出来れば最も近い）センサ２０−Jの周波数領域信号X(J,i,k)を用いることが望ましい。信号源に近いセンサ２０−Jほど、遅延や畳み込みの影響が少なく、適切な重み係数W(i,k)を算出できるからである。

Also, instead of adding the absolute values and powers of the amplitudes of the frequency domain signals X (1, i, k), ..., X (S, i, k) for all sensors, The absolute values and powers of the amplitudes of the frequency domain signals X (1, i, k),..., X (S, i, k) are added together, or one as shown in the following equations (18) and (19) The absolute value or power of the amplitude of the frequency domain signal X (J, i, k) of the sensor 20-J may be used as the weighting factor W (i, k) (an example of the weighting factor W (i, k)) 4). In this case, it is desirable to use the frequency domain signal X (J, i, k) of the sensor 20-J as close to the signal source as possible (closest if possible). This is because the sensor 20-J closer to the signal source has less influence of delay and convolution, and an appropriate weighting factor W (i, k) can be calculated.

W (i, k) = | X (J, i, k) | ... (18)
W (i, k) = | X (J, i, k) | ² ... (19)
The weight coefficient W (i, k) may be a fixed value such as 1. In addition, the weighting factor W (i, k) is set to a fixed value such as 1 according to the noise environment and the state of the target signal, and the weighting factor W as in Examples 1 to 4 of the weighting factor W (i, k). A configuration in which (i, k) is sequentially calculated and a switchable control may be employed (end of explanation of << example of weighting factor W (i, k) >>).

FIG. 11 shows the histogram generated as described above, with the horizontal axis representing quantized normalized signal value (signal arrival direction) Z (c) and the vertical axis representing normalized frequency bin _g (i , k, c). Here, FIG. 11A is a histogram created for a grid including a time frequency bin in which the target signal exists, and FIG. 11B shows a time in which only the noise signal exists without the target signal. It is an example of the histogram produced about the grid containing a frequency bin. In these examples, the weighting factor W (i, k) is 1.

  As can be seen from the comparison between FIGS. 11 (a) and 11 (b), the grid histogram (FIG. 11 (a)) including the time-frequency bin where the target signal exists has a normalized signal value Z (c) of a specific value. The histogram of the grid (FIG. 11 (b)) including the time frequency bin where the target signal does not exist and only the noise signal exists while the distribution is uneven (highly uneven) has a widely distributed shape. I understand that

The histogram generation unit 16a outputs bin _g (i, k, c) (c = 1,..., C) for specifying the histogram for each grid generated as described above, and bin _g (i, k, c) is input to the probability density function calculator 16b.

Probability density function calculation section 16b uses the bin _g (i, k, c), regarded as the following equation (20) a histogram probability density function P _g as (i, k, c), the quantized A probability density function P _g (i, k, c) having values c = 1,..., C corresponding to the respective normalized signal values as random variables is calculated and output (step S72).

出力されたグリッド毎の各確率密度関数P_g(i,k,c)は、エントロピー計算部１６ｃに入力され、エントロピー計算部１６ｃは、以下の式（２１）のようにグリッド毎のエントロピーH_g(i,H)を求め、各グリッドの偏在値として出力する。

The output probability density functions P _g (i, k, c) for each grid are input to the entropy calculation unit 16c, and the entropy calculation unit 16c performs entropy H _{g for} each grid as shown in the following equation (21). (i, H) is obtained and output as an uneven value of each grid.

このように算出したエントロピーH_g(i,k)は、正規化信号値Z(c)のヒストグラムが特定の値に偏った分布をみせる場合には低い値となり、幅広く分布する場合には高い値となり、ヒストグラムの分布の偏りを示す。すなわち、図１１（ａ）のように、目的信号が存在する時間周波数ビンを含むグリッドのヒストグラムは、正規化信号値Z(c)が特定の値に偏るため、エントロピーH_g(i,k)は小さくなる。なお、このエントロピーの大小を反転させるため、エントロピー計算部１６ｃがさらに以下の計算を行い、その演算結果を各グリッドの偏在値として出力としてもよい。

The entropy H _g (i, k) calculated in this way is low when the histogram of the normalized signal value Z (c) shows a distribution biased to a specific value, and high when it is widely distributed. Thus, the bias of the histogram distribution is shown. That is, as shown in FIG. 11 (a), the histogram of the grid including the time frequency bin where the target signal is present has the entropy H _g (i, k) because the normalized signal value Z (c) is biased to a specific value. Becomes smaller. In order to invert the magnitude of the entropy, the entropy calculation unit 16c may further perform the following calculation and output the calculation result as an unevenly distributed value of each grid.

この演算結果は、エントロピーH_g(i,k)の大小を反転させたものであり、目的信号が存在するグリッドで大きい値をとり、それ以外のグリッドで小さい値をとり、ヒストグラムの分布の偏りを示す。以下では、エントロピー計算部１６ｃから出力される各グリッドの偏在値を式（２１）の演算結果も含めてH_g(i,k)と表現する。

The result of the calculation is the inversion of the entropy H _g (i, k), which takes a large value in the grid where the target signal exists, takes a small value in the other grids, and biases the histogram distribution. Indicates. Hereinafter, the uneven distribution value of each grid output from the entropy calculation unit 16c is expressed as H _g (i, k) including the calculation result of Expression (21).

The uneven distribution value H _g (i, k) of each grid output from the entropy calculation unit 16c is input to the averaging unit 16d. The averaging unit 16d averages the uneven distribution value H _g (i, k) of each grid corresponding to the same frame i, and calculates the average value as the uneven distribution index value H (i) of the frame i (step S74). ). That is, the averaging unit 16d adds the uneven distribution values H _g (i, k) of the grids corresponding to the same frame i with respect to g = 1,. Sex index value H (i) is calculated.

Here, entropy is used as an index indicating the bias of the histogram and is used as the ubiquitous index value H (i), but other indexes indicating the ubiquity of the normalized signal value ZGRID _g (i, k) are present. It may be the sex index value H (i). Examples of other uneven distribution index values H (i) are shown below.

<Modification of uneven distribution index value H (i)>
For example, the ubiquitous index value calculation unit 16 of FIG. 7 may be used instead of the ubiquitous index value calculation unit 16 of FIG. 2 (modified example 1 of the ubiquitous index value H (i)). In this example, the variance is used as the uneven distribution index value H (i). In this case, first, the normalized signal value ZGRID _g (i, k) is input to the average value calculation unit 16e of the uneven distribution index value calculation unit 16. The average value calculation unit 16e weights each normalized signal value ZGRID _g (i, k) with a weighting factor W (i, k) for each time frequency bin (i, k) as shown in the following equation (22). Then, the weighted average value E _g (i, k) is obtained and output for each grid. Note that μ is the number of elements of the normalized signal value ZGRID _g (i, k) for each grid.

偏在性指標値算出部１６の分散計算部１６ｆには、平均値Ｅ_g(i,k)と、各正規化信号値ZGRID_g(i,k)とが入力され、以下の式（２３）のように分散H_g(i,k)を計算し、それを各グリッドの偏在値H_g(i,k)として出力する。

An average value E _g (i, k) and each normalized signal value ZGRID _g (i, k) are input to the variance calculation unit 16f of the uneven distribution index value calculation unit 16, and the following equation (23) is obtained. Thus, the variance H _g (i, k) is calculated and output as the uneven distribution value H _g (i, k) of each grid.

分散計算部１６ｆから出力された各グリッドの偏在値H_g(i,k)は、平均部１６ｄに入力される。平均部１６ｄは、同一のフレームiに対応する各グリッドの偏在値H_g(i,k)を平均し、その平均値を当該フレームiの偏在性指標値H(i)として算出する。

The uneven distribution value H _g (i, k) of each grid output from the variance calculation unit 16f is input to the averaging unit 16d. The averaging unit 16d averages the uneven distribution value H _g (i, k) of each grid corresponding to the same frame i, and calculates the average value as the uneven distribution index value H (i) of the frame i.

  Further, the ubiquitous index value calculator 16 shown in FIG. 8 may be used instead of the ubiquitous index value calculator 16 shown in FIG. 2 (Modification 2 of the ubiquitous index value H (i)). In this example, kurtosis is used as the uneven distribution index value H (i).

In this case, first, the normalized signal value ZGRID _g (i, k) is input to the average value calculation unit 16e of the uneven distribution index value calculation unit 16. The average value calculation unit 16e weights each normalized signal value ZGRID _g (i, k) with a weighting coefficient W (i, k) for each time frequency bin (i, k) as shown in Expression (22). The average value E _g (i, k) after weighting is obtained and output. In addition, the mean value E _g (i, k) and each normalized signal value ZGRID _g (i, k) are input to the variance calculation unit 16g of the uneven distribution index value calculation unit 16, and the equation (23) Similarly, the variance σ _g (i, k) is calculated and output.

Further, the variance σ _g (i, k), the average value E _g (i, k), and each normalized signal value ZGRID _g (i, k) are input to the kurtosis calculation unit 16h, and the kurtosis calculation unit For 16h, for example, the kurtosis H _g (i, k) is obtained and output by the following equation (24).

尖度計算部１６ｈから出力された各グリッドの偏在値H_g(i,k)は、平均部１６ｄに入力される。平均部１６ｄは、同一のフレームiに対応する各グリッドの偏在値H_g(i,k)を平均し、その平均値を当該フレームiの偏在性指標値H(i)として算出する。

The uneven distribution value H _g (i, k) of each grid output from the kurtosis calculation unit 16h is input to the averaging unit 16d. The averaging unit 16d averages the uneven distribution value H _g (i, k) of each grid corresponding to the same frame i, and calculates the average value as the uneven distribution index value H (i) of the frame i.

Also, the statistic indicating the uneven distribution of other normalized signal values ZGRID _g (i, k) such as standard deviation is defined as the uneven distribution value H _g (i, k) of each grid, which is averaged for each frame i. The index value H (i) may be used.

Furthermore, when two or more kinds of normalized signal values ZGRID _g (i, k) (for example, phase difference and amplitude ratio) are generated for each time frequency bin (i, k), the two or more kinds of normalization signal values are generated. Two or more uneven distribution index values H (i) each indicating the uneven distribution of the normalized signal value ZGRID _g (i, k) may be calculated, or the two or more types of normalized signal values ZGRID _g (i, k) may be calculated. ) May be calculated as the ubiquity index value H (i) indicating the ubiquity of the vector, but when calculating two or more ubiquity index values H (i) and one type of ubiquity index value The content of processing in the determination unit 17 to be described later is different from the case of calculating H (i) (<< variation of uneven distribution index value H (i) >> [end of description of step S7]).

  As described above, the normal ubiquity index value H (i) output from the ubiquitous index value calculation unit 16 is input to the determination unit 17, and the determination unit 17 uses the ubiquity index value H (i) as an index. It is determined whether or not each frame is a target signal section (step S8).

  The determination unit 17 of the present embodiment is configured such that the division value that is the ratio of the ubiquity index value of the frame to be determined and the ubiquity index value of the frame in the non-target signal section or a mapping of the division value is equal to or greater than a predetermined threshold value If it is, it is determined that the determination target frame is the target signal section, or if the predetermined threshold is exceeded, it is determined that the determination target frame is the target signal section.

[Details of Step S8]
In the example shown in FIGS. 3 and 6, first, the relative value calculation unit 17 a (FIG. 3) of the determination unit 17 calculates the ubiquitous index value calculated by the equation (21) and output from the ubiquitous index value calculation unit 16. Among them, the ubiquity index value H ′ (i) of the frame to be determined is set as the ubiquity index value λ of the frame in the non-target signal section where it is estimated that the target signal does not exist, and a division value γ that is a ratio thereof (i) is calculated and output as follows (step S81). An example of a non-target signal section in which no target signal is estimated, for example, an initial section such as i = 1,.

γ (i) = H '(i) / λ ... (25)
Next, the division value γ (i) is input to the likelihood ratio calculation unit 17b, and the likelihood ratio calculation unit 17b calculates and outputs the likelihood ratio Λ (i) according to the following equation (26) (Step S1). S82). In addition, the logarithm of Formula (26) is a natural logarithm. The likelihood ratio calculation formula is, for example, Shon, J, Kim, N.-S., and Sung, W., “A Statistical Model-based Voice Activity Detection,” IEEE Signal Processing Letters, Vol. , No. 1, pp.1-3, 1999.

次に、閾値判定部１７ｃに尤度比Λ(i)が入力され、閾値判定部１７ｃは尤度比Λ(i)と所定の閾値thとを比較し、尤度比Λ(i)に対応するフレームiが目的信号区間であるか否か、すなわち、フレームiが目的信号区間であるか否かを判定し、その判定結果を出力する（ステップＳ８３）。具体的には、閾値判定部１７ｃは、例えば、尤度比Λ(i)が所定の閾値thより大きい場合（「閾値th以上の場合」としてもよい）、目的信号がフレームiに含まれるとして１を出力し（ステップＳ８４）、尤度比Λ(i)が所定の閾値thより小さい場合（「閾値th以下の場合」としてもよい）、目的信号がフレームiに含まれないとして０を出力する（ステップＳ８５）。なお、閾値thは、尤度比Λ(i)の時間長平均（複数のフレームｉに対する平均）や分散などの統計量を用いて設定されてもよいし、th=0.2などの固定値を事前に設定しておいてもよい。分散などの統計量を用いて閾値thを設定する場合の一例としては、目的信号が存在しないと推定されるフレームを判定対象のフレームとして尤度比Λ(i)を求め、それらの平均値から所定のマージンを設けた値を閾値thとする方法がある。

Next, the likelihood ratio Λ (i) is input to the threshold determination unit 17c, and the threshold determination unit 17c compares the likelihood ratio Λ (i) with a predetermined threshold th and corresponds to the likelihood ratio Λ (i). It is determined whether or not the frame i to be processed is the target signal section, that is, whether or not the frame i is the target signal section, and the determination result is output (step S83). Specifically, the threshold determination unit 17c determines that the target signal is included in the frame i, for example, when the likelihood ratio Λ (i) is greater than the predetermined threshold th (may be “when the threshold is greater than or equal to the threshold th”). 1 is output (step S84), and if the likelihood ratio Λ (i) is smaller than the predetermined threshold th (may be “below the threshold th”), 0 is output because the target signal is not included in the frame i. (Step S85). Note that the threshold th may be set using a statistic such as a time length average (average for a plurality of frames i) or variance of the likelihood ratio Λ (i), or a fixed value such as th = 0.2 is set in advance. It may be set to. As an example of setting the threshold th using a statistic such as variance, the likelihood ratio Λ (i) is obtained with a frame estimated to be absent of the target signal as a frame to be determined, and the average value thereof is calculated. There is a method in which a value provided with a predetermined margin is used as a threshold th.

  Note that the method of determining the target signal section using the uneven distribution index value H (i) as an index is not limited to this. As described above, the size of the uneven distribution index value H (i) is a value that varies depending on whether or not each frame i is the target signal section. Any method may be used as long as it evaluates the size of the ubiquitous index value H (i) and associates the evaluation result with the determination result of whether or not each frame i is the target signal section. . Below, the modification of the target signal area determination method is shown.

[Modification of target signal section judgment method]
For example, the determination unit 17 in FIG. 9 may be used instead of the determination unit 17 in FIG. 3 (Modification 1 of the target signal section determination method). In the case of this modification, the first value calculation unit 17a has the above-described determination of the ubiquity index value H ′ (i) of the frame and the ubiquity index of the frame in the non-target signal section estimated that the target signal does not exist. A value λ is input, and a division value γ (i), which is a ratio of the values λ, is calculated and output as in the above equation (25). Next, the division value γ (i) is input to the threshold value determination unit 17d, and the threshold value determination unit 17d compares the division value γ (i) with the threshold value th for each frame i, and the division value γ (i) is the threshold value. If it is greater than th (“may be greater than or equal to the threshold th”), the frame i corresponding to the division value γ (i) corresponds to the target signal interval, otherwise it corresponds to the division value γ (i). It is determined that the frame i is a non-target signal section, and the determination result (1 or 0) is output. Also, instead of using the division value γ (i), a subtraction value obtained by subtracting the ubiquitous index value λ from the ubiquitous index value H ′ (i) is used, and the threshold processing similar to the above is performed on the subtracted value. Thus, it may be estimated whether or not it is the target signal section.

  Further, for example, the determination unit 17 in FIG. 10A may be used instead of the determination unit 17 in FIG. 3 (Modification 2 of the target signal section determination method). In the case of this modification, the ubiquitous index value H (i) calculated by the equation (21) and output from the ubiquitous index value calculating unit 16 is input to the threshold determining unit 17i of the determining unit 17, and the threshold determining unit 17i compares the division value γ (i) with the threshold value th for each frame i, and if the division value γ (i) is larger than the threshold value th (may be “when it is equal to or greater than the threshold value th”), the division value γ It is determined that the frame i corresponding to (i) corresponds to the target signal section, otherwise the frame i corresponding to the division value γ (i) corresponds to the non-target signal section, and the determination result (1 or 0) Is output. The threshold th is dynamically set based on, for example, a statistic such as an average value for each frame i of the uneven distribution index value H (i) input by the threshold calculation unit 17h. Further, the threshold th may be a fixed value.

  Note that the target signal section may be determined as described above using an uneven distribution index value H (i) other than that described above. The threshold determination in this case is based on the characteristic of the uneven distribution index value H (i). That is, when using the ubiquitous index value H (i), which increases in value as the ubiquity increases, the ubiquitous index value H (i) or its mapping exceeds a predetermined threshold (or )) Is determined to be the target signal section, and if the ubiquitous index value H (i) or its mapping is less than a predetermined threshold (or “in the following case”), it is determined not to be the target signal section. On the other hand, when using the ubiquitous index value H (i) whose value increases as the ubiquity is low, the ubiquitous index value H (i) or its mapping exceeds a predetermined threshold (or ]), It is determined that it is not the target signal section, and is determined to be the target signal section when the ubiquitous index value H (i) or the mapping thereof is less than a predetermined threshold (or “in the following case”).

Further, time-frequency bins (i, k) normalized signal values of two or more for each ZGRID _g (i, k) are generated, two or more of the normalized signal values belonging to each grid ZGRID _g (i, k) Even when the ubiquity index value H (i) indicating the ubiquity of the vector having the element is calculated for each frame i, the determination unit 17 determines whether or not the target signal section is the same as described above. Can be determined.

On the other hand, time-frequency bins (i, k) normalized signal values of two or more for each ZGRID _g (i, k) are generated, two or more of the normalized signal values belonging to each grid ZGRID _g (i, k) When two or more uneven distribution index values H (i) each indicating the uneven distribution of each frame i are calculated for each frame i, for example, the determination unit 17 sets two or more uneven distribution index values H (i) for each frame i. Are weighted, and the unevenness index value after the weighting is used as an index to determine whether or not each frame i is the target signal section. Specifically, for example, it is determined whether or not the frame i is the target signal section, based on whether or not the weighted sum of two or more uneven distribution index values H (i) exceeds a predetermined threshold.

  Further, as described above, instead of determining whether or not the target signal section is a comparison by comparing the ubiquity index value H (i) or its mapping with a predetermined threshold value, the pre-learned frame ubiquity The frame corresponding to the ubiquitous index value calculated by the ubiquitous index value calculation unit by the pattern recognition using the relationship between the sex index value and the determination result of whether or not the frame is the target signal section is the target signal section It may be determined whether or not it corresponds to. In this case, for example, as in the determination unit 17 in FIG. 10B, the parameter learning unit 17h causes the frame acoustic feature amount (such as an uneven distribution index value or γ (i)) and the frame to be the target signal section. A learning sample consisting of a pair with the determination result is input and pattern recognition learning is performed by the parameter learning unit 17h to obtain a model parameter. Then, this parameter and the ubiquity index value H (i) to be determined are input to the pattern recognition unit 17i, and the frame i corresponding to the ubiquity index value H (i) by the pattern recognition belongs to the target signal section. It is determined whether or not. Examples of pattern recognition techniques include known support vector machines (Koji Tsuda, “What is a support vector machine”, Journal of the Institute of Electronics, Information and Communication Engineers, 2000: 460-466 pages), and hidden Markov models (Kenji Kitaken). , Satoshi Nakamura, Masaaki Nagata, “Spoken Language Processing”, Mori Publishing Co., Ltd., 1996: 57-90).

  The determination unit 17 does not output the determination result as to whether or not it is the target signal section, but the above-described likelihood ratio Λ (i) itself, or Λ (i) / (1 + Λ (i)) may be output.

<Experimental result>
The experimental result for showing the effect of this form is shown. In this experiment, two microphones are used as sensors, an acoustic signal in which an audio signal and a noise signal are mixed is observed, the acoustic signal is analyzed by the signal interval estimation method of this embodiment, and an audio signal interval is detected. Indicates. In this experiment, the signal arrival direction estimation value is used as the normalized signal value Z (i, k), and the ubiquity index value H (i) is calculated by the equation (21) and is calculated from the ubiquity index value calculation unit 16. The target signal interval was estimated using the output ubiquitous index value.

  The acoustic signal data used was a signal that recorded the utterances that students gave their research presentations using posters in a university laboratory, and was sampled discretely at a sampling frequency of 16 kHz and a quantization bit rate of 16 bits. Using. Two microphones are used for recording, and the two microphones are arranged on the same line at intervals of 4 cm. FIG. 12A shows an acoustic signal recorded. In FIG. 12A, the horizontal axis is time, and the vertical axis is the amplitude of the acoustic signal. The sound signal contains sound (directional noise) that opens and closes the door of the laboratory where the poster is being presented at the beginning. The signal section estimation method according to the present embodiment is applied to this acoustic signal by setting the time length of one frame to 32 ms (512 sample points) and moving the frame start point every 16 ms (256 sample points). FIG. 12B is a graph illustrating the uneven distribution index value H (i) (acoustic feature amount) estimated in each frame. In FIG. 12A, the horizontal axis is time, and the vertical axis is the amplitude of the uneven distribution index value H (i). Moreover, the uneven distribution index value H (i) obtained by the method described in Non-Patent Document 6 is shown in FIG.

As shown in the figure, it is understood that the ubiquitous index value H (i) output by the target signal section estimation method according to the present embodiment shows a high value in the voice signal existence section B and a small value in the other sections. . Further, when FIGS. 12B and 12C are compared, in the section A in which only the door opening / closing sound included in the data of FIG. While the value of (i) remains small (FIG. 12 (b)), the method described in Non-Patent Document 6 shows a high value similar to that of the audio signal interval (FIG. 12 (c)). From this, it can be seen that the ubiquitous index value H (i) obtained by the present embodiment is robust against directional noise such as door opening / closing sound that does not have harmonics. In addition, in this embodiment, only the ubiquity of the arrival time difference in the vicinity of the fundamental frequency and its harmonic component is used, so that the ubiquity of the arrival time difference is calculated in the entire time frequency band than the method described in Non-Patent Document 6. The uneven distribution index value H (i) can be calculated at high speed. In this experiment, the method of this embodiment was able to calculate the acoustic features with a calculation time of 9% of the method described in Non-Patent Document 6. The worst calculation amount is
(Sampling frequency / lower frequency limit of search range of fundamental frequency estimation) / discrete Fourier transform score. In the case of the sampling frequency (16 kHz) and the discrete Fourier transform score (512 points) exemplified in this embodiment and the lower limit frequency (50 Hz) of the search range of the fundamental frequency estimation, (16000/50) / 512 = about 31.25% Just the amount of calculation. Actually, since the estimated fundamental frequency value is distributed in the range of 50 to 500 Hz, the calculation can be performed in a time shorter than the worst calculation amount as shown by the above experiment.

  From the above, it can be seen that this embodiment makes it possible to detect the target acoustic signal section at high speed without being affected by the directional noise. In addition, compared with the method described in Non-Patent Document 6, it is necessary to estimate the fundamental frequency of the observation signal in this embodiment, but the fundamental frequency estimation method described in this embodiment can be executed at high speed. It does not affect the overall computational complexity.

  The present invention is not limited to the embodiment described above. For example, in step S4 of the present embodiment, the fundamental frequency estimation unit 13 estimates fundamental frequencies corresponding to all sensors, and uses them for processing corresponding to each subsequent sensor. However, in step S4, the fundamental frequency estimation unit 13 may estimate only the fundamental frequencies corresponding to some sensors (for example, one sensor) and use it for processing corresponding to all subsequent sensors. .

  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. Moreover, the structure which includes the sampling part 30 may be sufficient as the signal area estimation apparatus 10, and the structure which carries out the distributed process of the function of the signal area estimation apparatus 10 with a some computer may be sufficient. Needless to say, other modifications are possible without departing from the spirit of the present invention.

  Further, when the above-described configuration is realized by a computer, processing contents of functions that each device should have are described by a program. The processing functions are realized on the computer by executing the program 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. Furthermore, 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 device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the 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 in this embodiment includes information that is used for processing by an electronic computer and that conforms 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 this embodiment, the apparatus is configured by executing a predetermined program on a computer. However, at least a part of the processing contents may be realized by hardware.

  The fields of application of the present invention include, for example, encoding of target signals, suppression of noise signals, dereverberation, automatic speech recognition, etc., in an environment where target signals such as voice signals and music signals are observed together with noise signals. The acoustic signal processing field can be exemplified. Of course, the present invention may be applied to signal processing other than acoustic signals.

FIG. 1 is a block diagram illustrating the overall configuration of the target signal section estimation device of the present embodiment. FIG. 2 is a block diagram illustrating a detailed configuration of the uneven distribution index value calculation unit 16 of FIG. FIG. 3 is a block diagram illustrating a detailed configuration of the determination unit 17 of FIG. FIG. 4 is a flowchart for explaining the target signal section estimation method of the present embodiment. FIG. 5 is a flowchart for explaining details of step S7. FIG. 6 is a flowchart for explaining details of step S8. FIG. 7 is a block diagram illustrating a configuration example of the uneven distribution index value calculation unit. FIG. 8 is a block diagram illustrating a configuration example of the uneven distribution index value calculation unit. FIG. 9 is a block diagram illustrating a configuration example of the determination unit. 10A and 10B are block diagrams illustrating a configuration example of the determination unit. FIGS. 11A and 11B are graphs showing histograms obtained by the processing of this embodiment. FIG. 12A is a graph showing acoustic signals recorded in the experiment. FIG. 12B is a graph showing the ubiquitous index value H (i) (acoustic feature amount) estimated in each frame in the experiment. FIG. 12C is a graph showing the uneven distribution index value H (i) obtained by the method described in Non-Patent Document 6.

Explanation of symbols

10 Signal section estimation device

Claims (10)

A target signal section estimation device for estimating a target signal section,
A signal extraction unit that extracts each signal observed by a plurality of sensors for each frame that is a predetermined time interval;
A frequency domain conversion unit that converts the signal of each frame extracted by the signal extraction unit into a frequency domain, and generates a frequency domain signal for each time frequency bin for each sensor;
A fundamental frequency estimation unit that estimates the fundamental frequency of each frame signal cut out by the signal cutout unit;
Time-frequency domain division that identifies one or more grids that are finite time-frequency sections each including the fundamental frequency or each harmonic component thereof for each frame, and extracts the frequency-domain signal of each time-frequency bin belonging to each grid And
Based on the frequency domain signal extracted by the time frequency domain dividing unit corresponding to a specific reference sensor included in the sensor, it is extracted by at least the time frequency domain dividing unit corresponding to the sensor other than the reference sensor. Normalizing each frequency domain signal, and generating a normalized signal value corresponding to the direction of arrival of the signal observed by the sensor for each time frequency bin,
The ubiquitous value for calculating the ubiquity index value indicating the ubiquity of the normalized signal value for each frame by using the ubiquitous value for each frame by obtaining the ubiquitous value indicating the ubiquity of the normalized signal value for each grid. A sex index value calculator,
A target signal section estimation device comprising: The target signal section estimation device according to claim 1,
A determination unit that determines whether each frame corresponds to the target signal section using the uneven distribution index value as an index;
A target signal section estimation device characterized by the above. The target signal section estimation device according to claim 1 or 2,
The uneven distribution index value calculation unit
A histogram generating unit that quantizes the normalized signal value, calculates a frequency of the quantized normalized signal value for each grid, and generates a histogram for each grid;
Using the histogram for each grid, an uneven distribution calculating unit that calculates an uneven value indicating the distribution of the histogram for each grid,
An average part that averages the uneven value of each grid corresponding to the same frame, and calculates the average value as the uneven index value of the frame,
A target signal section estimation device comprising: The target signal section estimation apparatus according to claim 3, wherein
The uneven distribution calculation unit
A probability density function generating unit for obtaining a probability density function for each grid using a histogram for each grid and using a value corresponding to each quantized normalized signal value as a random variable;
A function value that monotonically increases with respect to the entropy of the probability density function, or a function value that monotonously decreases with respect to the entropy, and an uneven value calculation unit that calculates the uneven value,
A target signal section estimation device comprising: The target signal section estimation device according to any one of claims 2 to 4,
The determination unit is
Comparing the ubiquitous index value of each frame or a map thereof and a predetermined threshold value to determine whether each frame is the target signal section;
A target signal section estimation device characterized by the above. The target signal section estimation device according to any one of claims 2 to 4,
The determination unit is
The division value that is the ratio of the ubiquitous index value of the frame to be determined and the ubiquitous index value of the frame of the non-target signal section or the mapping of the divided value is greater than or equal to a predetermined threshold value. A threshold determination unit that determines that the determination target frame is the target signal interval or determines that the determination target frame is the target signal interval when the predetermined threshold is exceeded.
A target signal section estimation device characterized by the above. The target signal section estimation device according to any one of claims 2 to 4,
The determination unit is
The uneven distribution calculated by the uneven distribution index value calculation unit by pattern recognition using a relationship between the pre-learned frame unevenness index value and a determination result of whether or not the frame is the target signal section. Determining whether the frame corresponding to the sex index value is the target signal interval;
A target signal section estimation device characterized by the above. A target signal section estimation method of a target signal section estimation device for estimating a target signal section,
A step in which the signal extraction unit cuts out each signal observed by a plurality of sensors for each frame that is a predetermined time interval;
A frequency domain transforming unit transforming the signal of each frame cut out by the signal cutout unit into a frequency domain, and generating a frequency domain signal for each time frequency bin for each sensor;
A step of estimating a fundamental frequency of a signal of each frame cut out by the signal cutout unit,
The time-frequency domain dividing unit specifies one or more grids that are finite time-frequency sections each including the fundamental frequency or each harmonic component thereof for each frame, and the frequency-domain signal of each time-frequency bin belonging to each grid Extracting the
The time frequency domain corresponding to at least the sensor other than the reference sensor based on the frequency domain signal extracted by the time frequency domain dividing unit corresponding to the specific reference sensor included in the sensor. Normalizing each frequency domain signal extracted by the dividing unit and generating a normalized signal value corresponding to the direction of arrival of the signal observed by the sensor for each time frequency bin;
An unevenness index value calculation unit obtains an uneven value indicating the unevenness of the normalized signal value for each grid, and indicates the unevenness of the normalized signal value for each frame by using the uneven value for each grid. Calculating an uneven distribution index value;
A target signal section estimation method comprising:   8. A target signal section estimation program for causing a computer to function as the target signal section estimation apparatus according to claim 1.   A computer-readable recording medium storing the target signal section estimation program according to claim 9.

Images