JP2008257110A - Object signal section estimation device, method, and program, and recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To precisely estimate an object signal section in such a situation that the arrival direction of an object signal cannot be precisely grasped in a noisy environment. <P>SOLUTION: Each of the signals observed by a plurality of sensors is segmented for each frame that is a prescribed time section, the segmented signal of each frame about each sensor is converted into a frequency domain, and a frequency domain signal for each time frequency bin is generated for each sensor. With the frequency domain signal corresponding to a reference sensor as a reference, each frequency domain signal corresponding to the sensor other than the reference sensor is normalized, and a normalized signal value is generated for each time frequency bin. An eccentricity index value indicating the eccentricity of the normalized signal value is calculated for each grid that is a prescribed time frequency section. With the eccentricity index value as an index, it is determined whether or not each grid corresponds to the object signal section. <P>COPYRIGHT: (C)2009,JPO&INPIT

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.

  In the target signal section estimation for the conventional cellular phone voice signal, the frequency spectrum of the signal, the energy of the entire band of the signal, the energy of each band after the band division, the number of zero crossings of the signal waveform, after noise suppression The target signal interval was estimated by using the frequency spectrum of the signal, the variance of the frequency spectrum, and their time differentiation as feature quantities (for example, pages 1 to 4 of Non-Patent Document 1, and Non-Patent Documents). 2 pages 40-43).

  In these target signal section estimation methods, an input signal is divided into frames of a certain fixed time length (for example, about 25 ms), the above-described acoustic feature amount is calculated for each frame, and the value is separately determined. If it exceeds the value, it is determined as a section where the target signal exists, and if not, it is determined as a non-target signal section. However, the acoustic feature amount as described above is easily affected by noise, and sufficient target signal section estimation accuracy cannot be obtained under environmental noise such as streets.

  As a method for accurately estimating a target signal section in such a noise environment, there is a method using not only the acoustic feature amount described above but also information on phase differences between a plurality of signals observed by a plurality of microphones. .

For example, if it is known that the target signal comes only from a certain direction, information on the phase difference between a plurality of signals is used to emphasize only the target signal coming from a certain direction. The accuracy of signal interval estimation can be improved (see, for example, Non-Patent Document 3). In addition, a method for determining a function value for an acoustic feature quantity such as the number of zero crossings based on reliability of the estimated arrival direction of the target signal (see, for example, Non-Patent Document 4) An estimation method (for example, see Non-Patent Document 5), a method for estimating a section in which the estimated arrival direction of the target signal is constant in time as a section where speech exists (for example, see Non-Patent Document 6), etc. is there.
ITU-T Recommendation G.729 Annex B., "A silence compression scheme for G. 729 optimized for terminals conforming to Recommendation V.70," 1996. ETSI standard document, "Speech Processing, Transmission and Quality Aspects (STQ); Distributed speech recognition; Advanced front-end feature extraction algorithm; Compression algorithms," ETSI ES 202 050 V1.1.5, 2007. 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, Taita Asano, Takashi Yoshimura, Yoichi Motomura, Hideki Aso, Isao Hara, Naoyuki Ichimura, Satoshi Ogata, Nobuhiko Kitakiwaki, “Evaluation of Spoken Interval 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.

  However, the conventional method has a problem in that the target signal section cannot be accurately estimated in a situation where the target signal arrival direction cannot be accurately known in a noisy environment.

  For example, in the method of Non-Patent Document 3, it is necessary to know the arrival direction of the target signal in advance, and when the arrival direction of the target signal is unknown, the target signal section cannot be estimated with high accuracy. The methods of Non-Patent Documents 4 to 6 are based on the premise that the arrival direction of the target signal can be accurately estimated. However, it is difficult to accurately estimate the signal arrival direction in an environment where a plurality of signals of all frequencies are simultaneously arriving from all directions (for example, a daily environment such as a street, a station, or an airport). In such a case, the target signal section cannot be accurately estimated by the methods of Non-Patent Documents 4 to 6.

  The present invention has been made in view of the above points, and accurately estimates a target signal section in a noise environment and in a situation where the arrival direction of the target signal cannot be accurately known. It aims at providing the technology that can be.

  In the present invention, in order to solve the above problems, a signal extraction unit that extracts each signal observed by a plurality of sensors for each frame that is a predetermined time interval, and each frame for each sensor that is extracted by the signal extraction unit A frequency domain conversion unit that converts each signal into the frequency domain and generates a frequency domain signal for each time frequency bin for each sensor, and a frequency domain signal corresponding to the reference sensor as a reference, and supports at least sensors other than the reference sensor Normalization unit that normalizes each frequency domain signal to generate a normalized signal value for each time frequency bin, and grid classification that classifies the normalized signal value of each time frequency bin for each grid that is a predetermined time frequency interval Each of the grids, the ubiquity index value calculation unit for calculating the ubiquity index value indicating the ubiquity of the normalized signal value for each grid, and the ubiquity index value as an index. Target signal interval estimation apparatus characterized by having, a determination section for determining whether or not corresponding to the target signal section is provided.

  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). In the present invention, this property is used to estimate a section with high uneven distribution as a target signal section in the noise environment. That is, the ubiquity index value indicating the ubiquity of the normalized signal value is calculated for each grid that is a predetermined time frequency interval, and the ubiquity index value is used as an index to determine whether each grid corresponds to the target signal interval. judge. This is the main feature unique to the present invention. 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.

  The target signal section estimation apparatus of the present invention preferably further includes a weight calculation unit that generates a weighting factor that is monotonically increased with respect to the absolute value of the amplitude of the frequency domain signal for each time frequency bin, and is unevenly distributed. The sex index value calculation unit weights the frequency of the normalized signal value of the time frequency bin corresponding to the weighting factor by the generated weighting factor, and calculates the ubiquitous index value using the weighted frequency.

  Normally, the frequency distribution of the environmental noise is uniform, while the target signal has a property (characteristic 2) that power is concentrated in a part of frequency bands. That is, the power of the normalized signal value of the time frequency bin corresponding to the target signal is the power of the normalized signal value of the frequency not included in the target signal or the normalized signal value of the time frequency bin corresponding to the environmental noise signal. It is significantly larger than power. In a preferred configuration of the present invention, an uneven distribution index value reflecting this property 2 is generated, and the target signal interval estimation accuracy is improved. That is, the ubiquity index value indicating the ubiquity of the normalized signal value is affected by the frequency of the value taken by the normalized signal value. That is, if the normalized signal value values are concentrated on a certain value and the frequency of the normalized signal value in the vicinity thereof increases, the generated ubiquitous index value is highly ubiquitous in the normalized signal value. It will be shown. From property 2, the power of the normalized signal value of the time frequency bin in which the target signal exists is significantly larger than the power of the other normalized signal values. Therefore, in a preferred configuration of the present invention, when calculating the ubiquitous index value, a large weight is given to the value of the normalized signal value of the time frequency bin in which the target signal exists, and the normalized signal of other time frequency bins A small weight is added to the value. 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 this case, more preferably, the weighting factor has a monotonically increasing value with respect to the absolute value of the amplitude of the frequency domain signal, and has a monotonically increasing relationship with the absolute value of the amplitude of the frequency domain signal. The value is normalized by the value obtained by adding the values for all frequency bins. Thereby, even in an environment where the power of the target signal varies from grid to grid, it is possible to calculate the uneven distribution index value while suppressing the influence of the variation. As a result, the estimation accuracy of the target signal section is improved.

  In the present invention, the normalization unit uses, for example, 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 amplitude of each frequency domain signal corresponding to a sensor other than the reference sensor. And a normalized signal value that is the normalization 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, preferably, the normalization unit generates two or more types of normalized signal values for each time frequency bin, and the uneven distribution index value calculation unit includes two or more types of normalized signal values belonging to each grid. For each grid, the determination unit weights the two or more uneven distribution index values for each grid, and uses the weighted uneven distribution index value as an index. It is determined whether each grid corresponds to the target signal section.

  As described above, by using two or more kinds of normalized signal values for each time frequency bin, an increase in the uneven distribution of the normalized signal values due to the presence of the target signal can be captured more accurately. Further, two or more uneven distribution index values are calculated for each grid, weighted to them, and whether or not each grid corresponds to a target signal section is determined using the uneven distribution index value after the weighting as an index. As a result, it is possible to determine whether to correspond to the target signal interval by placing importance on the higher reliability among the two or more uneven distribution index values.

  In the present invention, preferably, the normalization unit generates two or more types of normalized signal values for each time frequency bin, and the uneven distribution index value calculation unit includes two or more types of normalized signal values belonging to each grid. For each grid, and the determination unit uses the ubiquity index value indicating the vector ubiquity as an index, and whether each grid corresponds to the target signal interval. Determine whether. Thereby, the increase in the uneven distribution of the normalized signal value due to the presence of the target signal can be captured more accurately.

  In addition, the determination unit of the present invention compares, for example, the ubiquity index value of each grid or a map thereof and a predetermined threshold value to determine whether each grid corresponds to a target signal section. .

  In addition, the determination unit of the present invention substitutes, for example, the uneven distribution index value of the determination target grid into a predetermined function, and has a first value that is monotonically increasing with respect to the probability indicating the likelihood of the target signal section of the grid. And a first value calculation unit that calculates a non-target signal section grid subordination index value is substituted into a predetermined function, and a second that is monotonically increasing with respect to the probability indicating the target signal section likelihood of the grid. A second value calculation unit that calculates a value, and a division value that is a ratio between the first value and the second value or a map of the division value is equal to or greater than a predetermined threshold value, the determination target grid is a target signal interval Or a threshold determination unit that determines that the determination target grid corresponds to the target signal section when the predetermined threshold is exceeded.

  In addition, the determination unit of the present invention performs, for example, an uneven distribution index value by pattern recognition using a relationship between a pre-learned grid uneven distribution index value and a determination result of whether or not the grid is a target signal section. It is determined whether or not the grid corresponding to the uneven distribution index value calculated by the calculation unit corresponds to the target signal section.

  As described above, according to the present invention, it is possible to accurately estimate the target signal section in a situation where the target signal arrival direction cannot be accurately known even under 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. 2A is a block diagram illustrating a detailed configuration of the uneven distribution index value calculation unit 16 of FIG. FIG. 2B is a block diagram illustrating a detailed configuration of the determination unit 17 in FIG.

<Configuration>
As illustrated in FIG. 1, the target signal section estimation device 10 according to the present embodiment includes a signal extraction unit 11, a frequency domain conversion unit 12, a normalization unit 13, a weight calculation unit 14, a grid classification unit 15, and an uneven distribution index value. A calculation unit 16, a determination unit 17, a control unit 18, and a storage unit 19 are provided, and a signal observed by S (S ≧ 2) sensors 20-1 to S and sampled by the sampling unit 30 is input. This device outputs the analysis result of the target signal section. Further, as illustrated in FIG. 2A, the ubiquitous index value calculation unit 16 in this example includes a histogram generation unit 16a, a probability density function calculation unit 16b, and an entropy calculation unit 16c. Further, as illustrated in FIG. 2B, the determination unit 17 in this example includes a first value calculation unit 17a, a second value calculation unit 17b, a relative value calculation unit 17c, an average likelihood ratio calculation unit 17d, and a threshold value. And a determination unit 17e.

  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 interval estimation method of the present embodiment, each signal observed by the plurality of sensors 20-1 to 20-S is subjected to time frequency analysis, a normalized signal value with respect to a specific reference sensor is obtained, and a predetermined time frequency interval is determined. The presence / absence of a target signal is detected and output based on the uneven distribution of normalized signal values in the grid. 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. 3 is a flowchart for explaining the target signal section estimation method of the present embodiment. Hereinafter, the target signal section estimation method of this embodiment will be described along this flowchart.

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, 8.000 Hz), whereby 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 8,000 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 128 sample points (8,000 Hz × 16 ms), and a discrete signal of 256 sample points (8,000 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 = 256 in the above example) of the signal for one frame to be cut out.

  In FIG. 7A, such a window function is multiplied while being shifted to the waveform of the time domain signal x (1, t) extracted by the sampling unit 30, and the signal x ′ (1, i, n) of each frame is multiplied. It is a figure which illustrates the process of cutting out). In this figure, the process of cutting out the signal x ′ (1, i, n) of each frame from the time domain signal x (1, t) corresponding to the sensor 20-1 is shown. Similarly for 20-2 to S, signals x ′ (2, i, n),..., X ′ (S, i, n) of each frame are cut out.

  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 the frequency domain signals X (1, i, k),..., X (S, i, k) as shown in the following equation (2). Is calculated.

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, and f = f _s · k / M (k = 0, ..., M-1) where f is a discrete frequency. Satisfy the relationship. Note that f _s is the sampling frequency as described above, and M is a natural number greater than or equal to the frame length L.

  FIG. 7B shows an example in which the signal x ′ (1, i, n) of each frame corresponding to the sensor 20-1 is converted into the frequency domain signal X (1, i, k) by the discrete Fourier transform. FIG. The same conversion is performed for the other sensors 20-2 to 20-S.

  8A illustrates a time domain acoustic signal observed by environmental noise, and FIG. 8B illustrates a frequency obtained by converting the time domain acoustic signal illustrated in FIG. 8A into a frequency domain. It is a figure expressing a spectrum (frequency domain signal). Here, the acoustic signal illustrated in FIG. 8A passes through the center of a line segment connecting two microphones (sensors 20-1, 2 / S = 2) arranged at intervals of 4 cm under street noise. The acoustic signal observed with one microphone (sensor 20-1) in an environment where the target signal arrives from a position 40 to 60 cm away in the direction orthogonal to the line segment is shown. Further, in FIG. 8B, the horizontal axis is the frame index i, the vertical axis is the frequency index k, and the magnitude of the power of the frequency domain signal X (1, i, k) is expressed by color shading. In addition, it shows that the power of the frequency domain signal X (1, i, k) is larger as the color is lighter (whiter), and the frequency of the partial frequency domain signal X (1, i, k) is higher as the color is darker (darker). Indicates that the power is small. FIG. 8B shows that the power of the frequency domain signal X (1, i, k) is large in the target signal section, and the power is concentrated at some frequencies (white stripe pattern).

  The frequency domain transform unit 12 outputs the frequency domain signals X (1, i, k),..., X (S, i, k) obtained by the above transformation.

Next, frequency domain signals X (1, i, k),..., X (S, i, k) are input to the normalization unit 13, and the normalization unit 13 receives the reference sensor s _B ε {1, ..., S} with reference to the frequency domain signal X (s _B , i, k), and each frequency domain signal X (1 (1) corresponding to at least a sensor s (≠ s _B ) other than the reference sensor s _B , i, k),..., X (S, i, k) are normalized to generate a normalized signal value Z (i, k) for each time frequency bin (i, k) (step S4). Note that such a normalized signal value Z (i, k) is biased to a value corresponding to the arrival direction of the target signal in the time frequency bin (i, k) where the target signal exists. An example of the normalized signal value Z (i, k) generated by the normalizing unit 13 is shown below.

[Example of normalized signal value Z (i, k)]
In this embodiment, as an example of the normalized signal value Z (i, k), S = 2 is set, and the frequency domain signal X (1, i, k) corresponding to the reference sensor 20-1 and the other sensor 20-2 are set. The signal arrival direction is estimated from the corresponding frequency domain signal X (2, i, k), and the signal arrival direction estimate is defined as a normalized signal value Z (i, k) (normalized signal value Z (i, k) Example 1) of k). In this example, the normalization unit 13 calculates a normalized signal value Z (i, k) by the following equations (3) and (4). 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 index k, and arg (·) Indicates the phase (deflection angle). Further, τ (i, k) indicates a signal arrival time difference from the signal source to each of the sensors 20-1, 2 and θ (i, k) indicates a signal arrival direction estimated value. Further, the signal arrival direction θ (i, k) calculated by the equation (4) passes through the midpoint of the line segment connecting the sensors 20-1 and 20-2, and is an angle (0radian) where the direction orthogonal to the line segment is 0radian. radian). The normalized signal value Z (i, k) calculated in this way is a value in which the frequency component f is normalized and the frequency dependency is eliminated.

  FIG. 9A shows the frequency domain signal X (1, i, k) of FIG. 8B in the range of k = 1,..., M / 2. FIG. 9B is a diagram showing the signal arrival direction estimation value θ (i, k) calculated according to equations (3) and (4). In FIG. 9B, the horizontal axis is the frame index i, the vertical axis is the frequency index k, and the signal arrival direction estimation value is expressed by color shading. Note that the estimated signal arrival direction θ (i, k) in FIG. 9B passes through the midpoint of the line segment connecting the sensors 20-1 and 20 within the range of ± π / 2 radian and is orthogonal to the line segment. The value is expressed as an absolute value. In addition, in FIG. 9B, the darker the color (the darker the color), the closer the estimated signal arrival direction is to 0radian, and the lighter the color (the brighter the color), the farther the estimated signal arrival direction is from 0radian. Is shown.

  From the comparison between FIGS. 9A and 9B, the signal arrival direction estimation value in FIG. 9B is obtained in the time frequency bin area where the target signal exists (the white striped pattern area shown in FIG. 9A). θ (i, k) is uniformly expressed in black, and it can be seen that the signal arrival direction estimation value θ (i, k) is biased in a specific direction. That is, in the time frequency bin where the target signal exists and the power of the frequency domain signal X (1, i, k) is biased, the signal arrival direction estimation value θ (i, k) is also biased in a specific direction. I understand. On the other hand, in the domain of the time frequency bin where the target signal does not exist and the power of the frequency domain signal X (1, i, k) is uniform, the signal arrival direction estimation value θ (i, k) is also biased. Absent. When such a signal arrival direction estimation value θ (i, k) is a normalized signal value Z (i, k), the target signal exists using the uneven distribution of the normalized signal value Z (i, k) as an index. It can be determined whether or not.

  The same can be said for the case where the signal arrival time difference τ (i, k) calculated by the above equation (3) is the normalized signal value Z (i, k) (normalized signal value Z (i, k) Example 2) of k). Note that the normalized signal value Z (i, k) calculated in this way is also a value in which the frequency component f is normalized and the frequency dependency is eliminated.

Also, the phase difference arg (X (2, i, k) / X (1, i, k)) of the frequency domain signal X (1, i, k) with respect to the phase of the frequency domain signal X (1, i, k) May be used as the normalized signal value Z (i, k) (Example 3 of the normalized signal value Z (i, k)), and the phase of the frequency domain signal X (1, i, k) and the frequency domain signal X ( The difference arg (X (2, i, k))-arg (X (1, i, k)) from the phase of 1, i, k) may be used as the normalized signal value Z (i, k) (normal) Example 4) of the generalized signal value Z (i, k). Further, the ratio of the amplitude of the frequency domain signal X (1, i, k) to the amplitude of the frequency domain signal X (1, i, k) | X (2, i, k) | / | X (1, i, k) ) | May be a normalized signal value Z (i, k) (Example 5 of normalized signal value Z (i, k)), or a frequency domain signal corresponding to the power of the frequency domain signal X (1, i, k). The power ratio of X (1, i, k) | X (2, i, k) | ² / | X (1, i, k) | ² may be used as the normalized signal value Z (i, k) ( Example 6 of normalized signal value Z (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 Z (i, k) is unevenly distributed. Whether or not the target signal exists can be determined using the sex 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, the arrival azimuth angle estimated value θ (i, k) and the elevation angle estimated value of the target signal are as follows. φ (i, k) is obtained, and these two values may be used as the normalized signal value Z (i, k) for the time frequency bin (i, k) (an example of the normalized signal value Z (i, k)) 7).

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 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 (5). Set as follows. [•] ^T indicates transposition of a vector.
D = [d ₁ -d _J , d ₂ -d _J , ..., d _S -d _J ] ^T ... (5)

Further, a signal arrival time difference τ (s, i, k) between the reference sensor 20-J and each sensor 20-s is obtained by the following equation (6), and a signal arrival time difference vector τ ′ (i, k) using these as elements. k) is obtained by the following equation (7).

The relationship of the following equation (8) holds in the above equations (5) to (7). From the following equation (8), the arrival azimuth angle estimated value θ (i, k) of the target signal and the elevation angle estimated value φ ( i, k). In Equation (8), 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 0 radian.
ν ・ D ⁻¹・ τ '(i, k) = [cosθ (i, k) cosφ (i, k), sinθ (i, k) sinφ (i, k), sinφ (i, k)] ^T
... (8)

  In addition, the normalized signal value Z (i, k) exemplified in Examples 1 to 7 of the above-described normalized signal value Z (i, k) is combined, and two or more normalizations are performed for each time frequency bin (i, k). The signal value Z (i, k) may be calculated (Example 8 of normalized signal value Z (i, k)). For example, the phase difference arg (X (2, i, k) / X (1, i, k)) and the amplitude ratio | X (2, i, k) | / | X (1, i, k) | May be the normalized signal value Z (i, k) of the time frequency bin (i, k). For example, S = 3, and the phase difference arg (X (2, i, k) / X (1, i, k)) and the phase difference arg (X (3, i, k) / X (1, i, The pair with k)) may be the normalized signal value Z (i, k) of the time frequency bin (i, k). Further, the mapping of the values generated as described above may be used as the normalized signal value Z (i, k) (end of description of [Example of normalized signal value Z (i, k)]).

  As described above, in step S4, the normalization unit 13 generates and outputs the normalized signal value Z (i, k) as described above.

  Further, the frequency domain signals X (1, i, k),..., X (S, i, k) output from the frequency domain converter 12 (step S3) are also input to the weight calculator 14. . The weight calculator 14 is a weighting coefficient W (i, k) that is monotonically increasing with respect to the absolute value of the amplitude of the frequency domain signal X (1, i, k),. k) is generated for each time frequency bin (i, k) (step S5). Preferably, the weighting factor is a monotonically increasing value with respect to the absolute value of the amplitude of the frequency domain signal X (1, i, k), ..., X (S, i, k). A value that is monotonically increasing with respect to the absolute value of the amplitude of the frequency domain signal X (1, i, k), ..., X (S, i, k) is summed up for all frequency bins. It is a normalized value. The reason is as described above. Hereinafter, an example of the weighting factor W (i, k) will be described.

[Example of weighting factor W (i, k)]
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)).

  FIG. 10A shows the frequency domain signal X (1, i, k) of FIG. 8B in the range of k = 1,..., M / 2. FIG. 10B shows the weighting factor W (i, k) calculated according to the equation (9). In FIG. 10 (b), the horizontal axis is the frame index i, the vertical axis is the frequency index k, and the weighting coefficient W (i, k) is expressed in shades of color. In FIG. 10B, the darker the color (the darker the color), the smaller the value of the weighting factor W (i, k), and the lighter the color (the whiter) the weighting factor W (i, k). The value is large.

  From the comparison between FIGS. 10A and 10B, the weighting factor W (i, k) is also a large value in the time frequency bin where the target signal exists and the normalized signal value Z (i, k) is biased. It can be seen that the weighting factor W (i, k) is also small in the time frequency bin where the target signal does not exist and the normalized signal value Z (i, k) is a uniform value. That is, the weighting factor W (i, k) generated in this way is used to emphasize the uneven distribution of the normalized signal value Z (i, k) in the time frequency bin (i, k) where the target signal exists. It can be used as information. Details of this will be described later in step S9. Also, the weighting factor W (i, k) in equation (9) is the power of the frequency domain signal X (1, i, k), ..., X (S, i, k) for all sensors and all frequencies. It is the relative value of the power sum of the frequency domain signals X (1, i, k),..., X (S, i, k) for all sensors with respect to the sum. Therefore, for example, the power of the frequency domain signals X (1, i, k),..., X (S, i, k) for each frame index i so that the background noise power varies at each time. Even when the frequency fluctuates uniformly for all frequencies, the weighting factor W (i, k) can be set appropriately. For example, even when the power of the background noise is extremely large, the weighting factor W (i, k) of the time frequency bin (i, k) where the target signal exists and the time frequency bin (i, k) where the target signal does not exist The ratio with the weighting factor W (i, k) can be made sufficiently large.

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)).

Further, the weighting factor W (i, k) may be obtained without performing normalization as in equations (9) and (10) (Example 3 of weighting factor W (i, k)). This is because the amount of calculation can be reduced in this case, and the target signal section can 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).

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 value and power of the amplitudes of the frequency domain signals X (1, i, k),. 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) | ... (13)
W (i, k) = | X (J, i, k) | ² ... (14)

  The weight coefficient W (i, k) may be a fixed value such as 1, and the weight calculator 14 and its processing may be omitted. 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 switching control is possible (end of description of [example of weighting factor W (i, k)]).

  The weight calculation unit 14 outputs the weighting coefficient W (i, k) for each time frequency bin (i, k) generated as described above.

Thereafter, the normalized signal value Z (i, k) output from the normalization unit 13 and the weighting coefficient W (i, k) output from the weight calculation unit 14 are input to the grid classification unit 15. The grid classification unit 15 classifies the normalized signal value Z (i, k) of each time frequency bin (i, k) for each grid that is a predetermined time frequency section, and centers the time frequency bin (i, k). A set of normalized signal values Z (i, k) of time frequency bins included in the grid is output as GRID _z (i, k) (step S6). Further, the grid classification unit 15 classifies the weighting factor W (i, k) of each time frequency bin (i, k) for each grid that is a predetermined time frequency section, and centers the time frequency bin (i, k). A set of weighting factors W (i, k) of time frequency bins included in the grid is output as GRID _W (i, k) (step S7).

The classification of the grid is performed according to the following formulas (15) to (18), for example. In addition, {•} means a set having • as an element.
GRID _z (i, k) = {Z (i + P, k + Q)} ... (15)
GRID _W (i, k) = {W (i + P, k + Q)} ... (16)

  FIG. 11A is a diagram showing a frequency domain signal, a normalized signal value, and a weighting factor in a grid including a time frequency bin in which a target signal exists. FIG. 11B is a diagram showing a frequency domain signal, a normalized signal value, and a weighting factor in a grid composed of time frequency bins in which no target signal exists and only a noise signal exists. In FIGS. 11A and 11B, the horizontal axis is the frame index i, the vertical axis is the frequency index k, and the magnitude of each value of the frequency domain signal, the normalized signal value, and the weighting coefficient is represented by the color shading. This is expressed (similar to FIG. 8B, FIG. 9B, and FIG. 10B). In FIGS. 11A and 11B, the signal arrival direction estimation value of Expression (4) is used as the normalized signal value, and the weighting coefficient of Expression (9) is used.

  As can be seen from FIG. 11A, in the region of the time frequency bin where the target signal exists, the normalized signal value is biased to a specific value (specific signal arrival direction estimated value), and the value of the weighting factor becomes large. On the other hand, as can be seen from FIG. 11B, in the region of the time frequency bin where the target signal does not exist, the specific value of the normalized signal value and the weighting factor are widely and uniformly distributed.

  As described above, the present invention refers to the uneven distribution of the normalized signal value Z (i, k) in units of grids and determines whether or not the grid is the target signal section. Here, if the time frequency interval of the grid in the normalized signal value Z (i, k) is too wide (for example, the time frequency interval including a plurality of white stripe patterns in FIG. 8B), normalization in the grid is performed. The uneven distribution of the normalized signal value Z (i, k) is flattened, and it is difficult to determine whether the signal is a target signal section or a non-target signal section from the uneven distribution. Conversely, if the time frequency interval of the grid in the normalized signal value Z (i, k) is too narrow (for example, a time frequency interval consisting of a few time frequency bins), normalization in all grids due to the small number of samples. The ubiquity of the converted signal value Z (i, k) becomes high, and it is difficult to determine whether it is the target signal section or the non-target signal section from the ubiquity.

  Therefore, it is necessary to set the grid width of the normalized signal value Z (i, k) within a range in which such a problem does not occur. A preferable grid width setting method will be described below.

[About A in Formula (17)]
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 uttered per second) is known in advance, an integer value in the vicinity of (1000 / SR) / SF (for example, the nearest) may be 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 (17)]
Preferably, a width obtained from the main lobe width of the window function w (n) is basically 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).

However, if the fundamental frequency F0 Hz of the audio signal is known in advance, for example, B = 2 · F0 / (f _s / M so that harmonic components of two or more audio signals do not enter one grid. ) Determined by +1. When this is larger than the width obtained from the main lobe width, the value obtained from the main lobe width is adopted. For example, when the sampling frequency is 8 kHz, the width of the window function is 256 sampling points, and M = 256, if F0 = 50 Hz, B = 2 · 50 · (8000/256) + 1 = 4.2. To do. On the other hand, if F0 = 200 Hz, B = 2 · 200 · (8000/256) + 1 = 13.8, but B = 5 is adopted because it is larger than the value B = 5 obtained from the main lobe width. This is because the arrival direction of the audio signal is unevenly distributed only within the main lobe width. These are the same when the target signal is a music signal.

Further, the grid classification (step S7) of the weighting factor W (i, k) is not necessarily required, and a method that does not execute step S7 may be used.
The grid classification unit 15 outputs the sets GRID _z (i, k) and GRID _W (i, k) generated as described above, and these are input to the ubiquitous index value calculation unit 16.

The ubiquitous index value calculation unit 16 uses the set GRID _z (i, k) and GRID _W (i, k), and the ubiquitous index value H () indicating the ubiquity of the normalized signal value Z (i, k). i, k) is calculated for each grid (step S8). As an example, in this embodiment, a histogram of the normalized signal value Z (i, k) is generated for each grid while weighting the frequency by the weighting factor W (i, k), and the generated histogram is converted into a probability density function. And the entropy is defined as an uneven distribution index value H (i, k).

First, the histogram generation unit 16a (FIG. 2A) of the ubiquitous index value calculation unit 16 calculates each normalized signal value Z (i, k) that is an element of the input set GRID _z (i, k). Quantize to C values Z (c) (c = 1,..., C). Then, the frequency for each quantized normalized signal value Z (c) is weighted by the weighting factor W (i, k) that is an element of GRID _W (i, k) for each time frequency bin (i, k). Count and generate a histogram. For example, when the normalized signal value Z (i, k) is the signal arrival direction θ (i, k) and C = 32, each normalized signal value Z (i, k) is represented by the following C Quantized to the normalized signal value Z (c).
Z (1) (-π / 2 ≦ Z (i, k) <-7π / 16)
Z (2) (-7π / 16 ≦ Z (i, k) <-3π / 16)
...
Z (C) (7π / 16 <Z (i, k) <π / 2)

  When the signal arrival time difference τ (i, k) calculated by the above equation (3) is the normalized signal value Z (i, k), for example, | τ (i, k) | ≦ (d The normalized signal value Z (i, k) is quantized to C in units of / ν) × α (α is a positive constant).

Then, for each time frequency bin (i, k), it is determined which normalized signal value Z (i, k) corresponds to which normalized signal value Z (c), and the frequency is counted. In the counting, the frequency is weighted by the weighting factor W (i, k) of the corresponding time frequency bin (i, k). For example, if the value obtained by quantizing the time frequency bin (1, 2) is Z (5), W (1, 2) is counted as the frequency for Z (5). That is, for the normalized signal value Z (i, k) belonging to GRID _z (i, k), the frequency bin (i, k, c) of the quantized normalized signal value Z (c) (c = 1, ..., C) are counted as in the following equation (19).
bin (i, k, c) = ΣW (i, k) if Z (i, k) ∈Z (c) ... (19)

  FIG. 12 shows the histogram generated in this manner, with the abscissa indicating the quantized normalized signal value (signal arrival direction) Z (c) and the ordinate indicating the normalized frequency bin (i, k). , c). Here, FIG. 12 (a) is a histogram created for a grid (FIG. 11 (a)) including time frequency bins where the target signal exists, and FIG. 12 (b) shows that the target signal does not exist. It is an illustration of the histogram created about the grid (FIG.11 (b)) containing the time frequency bin in which only a noise signal exists.

  As can be seen from the comparison of FIGS. 12 (a) and 12 (b), the histogram (FIG. 12 (a)) of the grid including the time frequency bin where the target signal is present has the normalized signal value Z (c) set to a specific value. The histogram of the grid (FIG. 12 (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 generator 16a outputs bin (i, k, c) (c = 1, ..., C) for specifying the histogram generated as described above, and bin (i, k, c) is It is input to the probability density function calculator 16b.
The probability density function calculation unit 16b uses bin (i, k, c), regards the histogram as a probability density function P (i, k, c) as shown in the following equation (20), and sets the probability density function P (i , k, c) is calculated and output.

The probability density function P (i, k, c) is input to the entropy calculation unit 16c, and the entropy calculation unit 16c obtains entropy as shown in the following equation (21), and obtains the ubiquitous index value H (i, Output as k).

  The entropy H (i, k) calculated in this way is a low value when the histogram of the normalized signal value Z (c) shows a distribution biased to a specific value, and a high value when the histogram is widely distributed. Become. That is, as shown in FIG. 12A, in the histogram of the grid including the time-frequency bin where the target signal exists, the normalized signal value Z (c) is biased to a specific value, so the entropy H (i, k) is Get smaller.

  FIG. 13A is a graph illustrating the entropy H (i, k) thus obtained. In FIG. 13A, the horizontal axis is the frame index i, the vertical axis is the frequency index k, and the size of the entropy H (i, k) is expressed by the color shading. In addition, it shows that entropy H (i, k) is so large that a color is light (it becomes white), and entropy H (i, k) is so small that a color is dark (it becomes black).

  As can be seen by comparing FIG. 13 (a) and FIG. 10 (a), the entropy H (i, k) value is small and only noise exists in the grid including the time frequency bin where the target signal exists. In the time frequency bin to be performed, the value of entropy H (i, k) becomes large. Therefore, if this entropy H (i, k) is the ubiquitous index value H (i, k), the target signal interval can be estimated using the size of the ubiquitous index value H (i, k) as an index.

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

[Modified example of uneven distribution index value H (i, k)]
For example, the ubiquitous index value calculation unit 16 shown in FIG. 4A may be used instead of the ubiquitous index value calculation unit 16 shown in FIG. 2A (transformation of the ubiquitous index value H (i, k). Example 1). In this example, the variance is used as the uneven distribution index value H (i, k). In this case, first, each normalized signal value Z (i, k), which is an element of GRID _z (i, k), and GRID _W (i, k) are input to the average value calculation unit 16d of the uneven distribution index value calculation unit 16. ) Is input as a weighting factor W (i, k). The average value calculation unit 16d weights each normalized signal value Z (i, k) with a weighting coefficient W (i, k) for each time frequency bin (i, k) as shown in the following equation (22). The average value E (i, k) after weighting is obtained and output. Note that μ is the number of elements of GRID _W (i, k).

The variance calculation unit 16e of the uneven distribution index value calculation unit 16 includes an average value E (i, k), each normalized signal value Z (i, k) that is an element of GRID _z (i, k), and a GRID. _A weighting factor W (i, k), which is an element of W (i, k), is input, and a variance H (i, k) is calculated as in the following equation (23). Output as (i, k).

  The variance H (i, k) in Expression (23) is a small value for a grid including a time frequency bin where the target signal exists, and is large for a grid including a time frequency bin where only noise exists. Therefore, if the variance H (i, k) is the ubiquitous index value H (i, k), the target signal interval can be estimated using the size of the ubiquitous index value H (i, k) as an index.

  Further, the ubiquitous index value calculation unit 16 of FIG. 4B may be used instead of the ubiquitous index value calculation unit 16 of FIG. 2A (transformation of the ubiquitous index value H (i, k)). Example 2). In this example, kurtosis is used as the ubiquitous index value H (i, k).

In this case, first, each normalized signal value Z (i, k), which is an element of GRID _z (i, k), and GRID _W (i, k) are input to the average value calculation unit 16d of the uneven distribution index value calculation unit 16. ) Is input as a weighting factor W (i, k). The average value calculation unit 16d weights each normalized signal value Z (i, k) with a weighting coefficient W (i, k) for each time frequency bin (i, k) as shown in the equation (22). The later average value E (i, k) is obtained and output. In addition, the variance calculation unit 16e of the uneven distribution index value calculation unit 16 includes an average value E (i, k) and each normalized signal value Z (i, k) which is an element of GRID _z (i, k). , GRID _W (i, k), which is an element of the weight coefficient W (i, k), is input, and the variance σ (i, k) is calculated and output in the same manner as in equation (23).

Further, the kurtosis calculation unit 16f has a variance σ (i, k), an average value E (i, k), and each normalized signal value Z (i, k) that is an element of GRID _z (i, k) , GRID _W (i, k) and the weighting factor W (i, k) are input, and the kurtosis calculation unit 16f obtains the kurtosis H (i, k) by the following equation (24), for example. This is output as an uneven distribution index value H (i, k).

  Further, a statistic indicating the uneven distribution of other normalized signal values Z (i, k) such as standard deviation may be used as the uneven distribution index value H (i, k).

  Further, when two or more kinds of normalized signal values Z (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 are performed. Two or more unevenness index values H (i, k) each indicating the uneven distribution of the signal value Z (i, k) may be calculated for each grid, or the two or more kinds of normalized signal values Z (i , k) may be calculated for each grid, and the ubiquitous index value H (i, k) may be calculated for each grid. The processing contents in the determination unit 17 to be described later are different between the case and the case of calculating one kind of uneven distribution index value H (i, k) ([Modified example of uneven distribution index value H (i, k)]. End of description).

  As described above, the normalized signal value Z (i, k) output from the ubiquitous index value calculation unit 16 is input to the determination unit 17, and the determination unit 17 determines the ubiquity index value H (i, k). It is determined whether each grid corresponds to the target signal section as an index (step S9).

  In this embodiment, the ubiquitous index value H (i, k) is input, and the ubiquitous index value H (i, k) of the determination target grid is substituted into a predetermined function to indicate the likelihood of the target signal section of the grid. The first value H ′ (i, k) that is monotonically increasing with respect to the probability is calculated, and the uneven distribution index value H (i, k) of the grid in the non-target signal section is substituted into the predetermined function, A second value λ (k) that is monotonically increasing with respect to the probability indicating the likelihood of the target signal section of the grid is calculated. If the division value that is the ratio between the first value and the second value or the mapping of the division value is equal to or greater than a predetermined threshold value, it is determined that the determination target grid corresponds to the target signal interval, or the predetermined value When the threshold is exceeded, it is determined that the determination target grid corresponds to the target signal section.

  As an example of this, the likelihood of the target signal existing section for each frequency is calculated as a ratio (likelihood ratio) to the likelihood of nonexistent section (likelihood ratio), and the average likelihood ratio over the entire frequency band and a predetermined threshold value A method of comparing the magnitude relation and determining whether or not it is the target signal section can be exemplified. This method will be described below.

First, the ubiquitous index value H (i, k), which is the aforementioned entropy, is input to the first value calculation unit 17a, and the first value H ′ (i, k) is calculated and output by the following equation (25). To do. The first value H ′ (i, k) is a value obtained by reversing the size of the ubiquitous index value H (i, k), which is entropy. The first value H ′ (i, k) takes a large value in the grid where the target signal exists, Takes a small value in a grid that does not exist. That is, the first value H ′ (i, k) is a value that is in a monotonically increasing relationship with the probability indicating the likelihood of the target signal section of the determination target grid.
H '(i, k) = (1- H (i, k)) / log ₂ (C) ... (25)

  Also, the second value calculation unit 17b receives the ubiquitous index value H (i, k), which is the above-described entropy, corresponding to a grid in which the target signal does not exist (or is predicted to not exist). The second value λ (k) is calculated and output by (24). In addition, as a grid where it is predicted that the target signal does not exist, for example, a grid composed of frames with the first frame index i = 1,.

Next, the first value H ′ (i, k) and the second value λ (k) are input to the relative value calculation unit 17c, and the relative value calculation unit 17c obtains the first value by the following equation (26). A division value γ (i, k), which is a ratio with the second value, is calculated and output.
γ (i, k) = H '(i, k) / λ (k) ... (26)

Next, the division value γ (i, k) is input to the average likelihood ratio calculation unit 17d, and the average likelihood ratio calculation unit 17d calculates the average likelihood ratio Λ (i) according to the following equation (27). Output. In addition, the logarithm of Formula (26) is a natural logarithm. The calculation formula for the average likelihood ratio is, for example, Shon, J, Kim, N.-S., and Sung, W., “A Statistical Model-based Voice Activity Detection,” IEEE Signal Processing Letters, Vol. 6, No. 1, pp.1-3, 1999.

  FIG. 13B is a graph illustrating the average likelihood ratio Λ (i) calculated in this way. In FIG. 13B, the horizontal axis is the frame index i, and the vertical axis is the average likelihood ratio Λ (i). As can be seen from comparison with FIG. 10B, the average likelihood ratio Λ (i) takes a large value in the section where the target signal exists.

  Next, the average likelihood ratio Λ (i) is input to the threshold determination unit 17e, and the threshold determination unit 17e compares the average likelihood ratio Λ (i) with a predetermined threshold th, and the average likelihood ratio Λ (i ) Is a target signal section, that is, whether the frame corresponding to the frame index i is a target signal section, and the determination result is output. Specifically, the threshold determination unit 17e, for example, if the average likelihood ratio Λ (i) is larger than a predetermined threshold th (may be “when it is equal to or larger than the threshold th”), the target signal is a frame corresponding to the frame index i. Is output as 1 and the average likelihood ratio Λ (i) is smaller than a predetermined threshold th (may be “below the threshold th”), the target signal is not included in the frame for the frame index i. Is output as 0. Note that the threshold th may be set using a statistic such as a time length average (average for a plurality of frame indexes i) or variance of the average likelihood ratio Λ (i), or a fixed value such as th = 1.0. May be set in advance.

  Note that the method of determining the target signal section using the uneven index value H (i, k) as an index is not limited to this. As described above, the size of the uneven distribution index value H (i, k) is a value that varies depending on whether or not each grid is the target signal section. Any method can be used as long as it evaluates the size of the ubiquitous index value H (i, k) and associates the evaluation result with the determination result of whether each grid is the target signal section or not. Good. 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. 5 may be used instead of the determination unit 17 in FIG. 2B (Modification 1 of the target signal section determination method). In the case of this modification, the ubiquitous index value H (i, k), which is the aforementioned entropy, is input to the first value calculation unit 17a, and the first value H ′ (i, k) is obtained by the above-described equation (24). Calculate and output. Also, the second value calculation unit 17b receives the ubiquitous index value H (i, k), which is the above-described entropy, corresponding to a grid in which the target signal does not exist (or is predicted to not exist). The second value λ (k) is calculated and output by (24). Next, the first value H ′ (i, k) and the second value λ (k) are input to the relative value calculation unit 17c, and the relative value calculation unit 17c obtains the first value by the above-described equation (25). A division value γ (i, k), which is a ratio with the second value, is calculated and output. Next, the division value γ (i, k) is input to the threshold value determination unit 17g, and the threshold value determination unit 17g compares the division value γ (i, k) with the threshold value th for each (i, k), and performs division. If the value γ (i, k) is greater than the threshold th (“may be greater than or equal to the threshold th”), the grid corresponding to the division value γ (i, k) corresponds to the target signal interval, and For example, it is determined that the grid corresponding to the division value γ (i, k) corresponds to the non-target signal section, and the determination result (1 or 0) is output.

  For example, instead of the determination unit 17 in FIG. 2B, the determination unit 17 in FIG. 6A may be used (modified example 2 of the target signal section determination method). In the case of this modification, the ubiquitous index value H (i, k), which is the aforementioned entropy, is input to the threshold determination unit 17i of the determination unit 17, and the threshold determination unit 17i performs the division value γ for each (i, k). (i, k) is compared with the threshold value th, and when the divided value γ (i, k) is larger than the threshold value th (may be “when it is equal to or greater than the threshold value th”), the divided value γ (i, k) It is determined that the corresponding grid corresponds to the target signal section, otherwise the grid corresponding to the division value γ (i, k) corresponds to the non-target signal section, and the determination result (1 or 0) is output. The threshold th is dynamically set based on a statistic such as an average value of the uneven distribution index value H (i, k) 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 the ubiquitous index value H (i, k) other than entropy. The threshold determination in this case is based on the characteristics of the uneven distribution index value H (i, k). That is, when using the ubiquitous index value H (i, k) that increases in value as the ubiquity increases, the ubiquitous index value H (i, k) or its mapping exceeds a predetermined threshold (or "If above") it is determined that it is the target signal section, and if the ubiquitous index value H (i, k) or its mapping is less than a predetermined threshold (or "if below"), it is not the target signal section judge. On the other hand, when using the ubiquitous index value H (i, k) that increases in value as the ubiquity is low, the ubiquitous index value H (i, k) or its mapping exceeds a predetermined threshold (or If it is determined that it is not the target signal section in the above-mentioned case, and the ubiquitous index value H (i, k) or its mapping is less than a predetermined threshold (or “in the following case”), it is the target signal section judge.

  Also, two or more types of normalized signal values Z (i, k) are generated for each time frequency bin (i, k), and two or more types of normalized signal values Z (i, k) belonging to each grid are used as elements. Even if the ubiquity index value H (i, k) indicating the ubiquity of the vector is calculated for each grid, the determination unit 17 determines whether or not the target signal section is the same as described above. Judgment can be made.

  On the other hand, two or more kinds of normalized signal values Z (i, k) are generated for each time frequency bin (i, k), and two or more kinds of normalized signal values Z (i, k) belonging to each grid are unevenly distributed. When two or more ubiquitous index values H (i, k) each indicating the sex are calculated for each grid, the determination unit 17 may, for example, have two or more ubiquitous index values H (i, k) for each grid. Is weighted, and the uneven distribution index value after the weighting is used as an index to determine whether each grid corresponds to the target signal section. Specifically, for example, it is determined whether or not it is a target signal section depending on whether or not the weighted sum of two or more uneven distribution index values H (i, k) 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, k) or its mapping with a predetermined threshold value, a pre-learned grid is used. The grid corresponding to the ubiquitous index value calculated by the ubiquitous index value calculation unit is obtained by pattern recognition using the relationship between the ubiquitous index value and the determination result of whether or not the grid is the target signal section. You may determine whether it corresponds to a signal area. In this case, for example, as in the determination unit 17 in FIG. 6B, the parameter learning unit 17h includes a set of the uneven distribution index value of the grid and the determination result as to whether or not the grid is the target signal section. A learning sample is input, pattern recognition learning is performed by the parameter learning unit 17h, and model parameters are obtained. Then, this parameter and the ubiquity index value H (i, k) to be determined are input to the pattern recognition unit 17i, and the grid corresponding to the ubiquity index value H (i, k) is obtained by pattern recognition. It is determined whether it is a thing. Pattern recognition technology includes 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 Kita, Nakamura). Tetsu, Masaaki Nagata, “Spoken Language Processing”, Mori Publishing Co., Ltd., 1996: 57-90).

<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), the above entropy is used as the ubiquitous index value H (i, k), and the average likelihood ratio Λ (i) The target signal interval was estimated by comparing with the threshold. In this experiment, the time length of one frame is set to 32 ms (256 sample points), the start point of the frame is moved (shifted) every 16 ms (128 sample points), and the average likelihood ratio Λ (i) is set to each frame. Asked. Further, the average likelihood ratio Λ (i) obtained in this way was compared with a fixed threshold th = 1.08, and the target signal section was estimated.

  The signal data used is an acoustic signal obtained by adding noise noise recorded on the street to a voice (target signal) uttered by one woman with a signal-to-noise ratio of 10 dB, and having a sampling frequency of 8 kHz, quantum Discrete-sampled with 16 bits. In addition, recording of the target signal that is voice is performed by two microphones arranged at intervals of 4 cm, and the generation position of the voice that is the target signal at this time is the midpoint of the line connecting the two microphones. It is a position 40 to 60 cm away in a direction perpendicular to the line segment. Street noise and noise were also recorded by the same microphone.

  FIG. 14A is a graph showing an audio signal before adding noise and noise, and FIG. 14B is a graph showing an audio signal added with noise and noise as described above. In FIGS. 14A and 14B, the horizontal axis represents discrete real time, and the vertical axis represents signal amplitude. FIG. 14C is a graph showing the average likelihood ratio Λ (i). In FIG. 14C, the horizontal axis is the frame index i, and the vertical axis is the average likelihood ratio Λ (i). FIG. 14D is a graph showing a signal section determination result using the average likelihood ratio Λ (i). The horizontal axis of FIG.14 (d) is the frame index i, and a vertical axis | shaft is the determination result of (0) whether it is the target signal area (1).

  From these figures, it can be seen that the method of the present embodiment is effective for estimating the target signal section in a noisy environment.

  The present invention is not limited to the embodiment described above. For example, the various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. 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. The computer-readable recording medium may be any medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, or a semiconductor memory. Specifically, for example, the magnetic recording device may be a hard disk device or a flexible Discs, magnetic tapes, etc. as optical disks, DVD (Digital Versatile Disc), DVD-RAM (Random Access Memory), CD-ROM (Compact Disc Read Only Memory), CD-R (Recordable) / RW (ReWritable), etc. As the magneto-optical recording medium, MO (Magneto-Optical disc) or the like can be used, and as the semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory) or the like can 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 own 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. 2A is a block diagram illustrating a detailed configuration of the uneven distribution index value calculation unit of FIG. FIG. 2B is a block diagram illustrating a detailed configuration of the determination unit in FIG. FIG. 3 is a flowchart for explaining the target signal section estimation method of the present embodiment. 4A and 4B are block diagrams showing a modification of the uneven distribution index value calculation unit. FIG. 5 is a block diagram illustrating a modification of the determination unit. FIGS. 6A and 6B are block diagrams showing a modification of the determination unit. FIG. 7A is a diagram illustrating a process of cutting out the signal of each frame by multiplying the window function while shifting it to the waveform of the signal in the time domain extracted by the sampling unit. FIG. 7B is a diagram illustrating an example in which a signal of each frame corresponding to the sensor is converted into a frequency domain signal by discrete Fourier transform. FIG. 8A is a diagram illustrating a time-domain acoustic signal observed with environmental noise. FIG. 8B is a diagram expressing a frequency spectrum (frequency domain signal) obtained by converting the time domain acoustic signal illustrated in FIG. 8A into the frequency domain. FIG. 9A shows the frequency domain signal of FIG. 8B. FIG. 9B is a diagram illustrating the signal arrival direction estimated value calculated according to the equations (3) and (4). FIG. 10A shows the frequency domain signal of FIG. 8B. FIG. 10B is a diagram illustrating the weighting coefficient calculated according to Equation (9). FIG. 11A is a diagram showing a frequency domain signal, a normalized signal value, and a weighting factor in a grid including a time frequency bin in which a target signal exists. FIG. 11B is a diagram showing frequency domain signals, normalized signal values, and weighting coefficients in a grid composed of time frequency bins in which there is no target signal and only noise signals are present. FIG. 12 shows an example in which the generated histogram is displayed with the normalized signal value (signal arrival direction) on the horizontal axis and the frequency after weighting on the vertical axis. Here, FIG. 12 (a) is a histogram created for a grid (FIG. 11 (a)) including time frequency bins where the target signal exists, and FIG. 12 (b) shows that the target signal does not exist. It is an illustration of the histogram created about the grid (FIG.11 (b)) containing the time frequency bin in which only a noise signal exists. FIG. 13A is a graph illustrating the obtained entropy. FIG. 13B is a graph illustrating the calculated average likelihood ratio. FIG. 14A is a graph showing an audio signal before adding noise and noise, and FIG. 14B is a graph showing an audio signal added with noise and noise as described above.

Explanation of symbols

10 Signal section estimation device

Claims (14)

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 for each sensor 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 normalization unit that normalizes each frequency domain signal corresponding to the sensor other than the reference sensor and generates a normalized signal value for each time frequency bin, with the frequency domain signal corresponding to the reference sensor as a reference,
A grid classifying unit that classifies the normalized signal value of each time frequency bin for each grid that is a predetermined time frequency interval;
An ubiquitous index value calculation unit that calculates an ubiquitous index value indicating the ubiquity of the normalized signal value for each grid;
A determination unit that determines whether or not each grid corresponds to the target signal section using the uneven distribution index value as an index,
A target signal section estimation device comprising: The target signal section estimation device according to claim 1,
A weight calculation unit that generates a weighting factor that is monotonically increasing with respect to the absolute value of the amplitude of the frequency domain signal for each time frequency bin;
The uneven distribution index value calculation unit
Weighting the frequency of the normalized signal value of the time frequency bin corresponding to the weighting factor by the weighting factor, and calculating the unevenness index value using the weighted frequency,
A target signal section estimation device characterized by the above. The target signal section estimation device according to claim 2,
The weighting factor is
The value that is monotonically increasing with respect to the absolute value of the amplitude of the frequency domain signal, and the value that is monotonically increasing with respect to the absolute value of the amplitude of the frequency domain signal are summed for all frequency bins. , Which is the normalized value,
A target signal section estimation device characterized by the above. The target signal section estimation device according to claim 1,
The normalization part
Based on the phase and / or amplitude of the frequency domain signal corresponding to the reference sensor as a reference, at least the phase and / or amplitude of each frequency domain signal corresponding to the sensor other than the reference sensor is normalized, and the normalized value or its Generating the normalized signal value which is a map;
A target signal section estimation device characterized by the above. The target signal section estimation device according to claim 4,
The normalized signal value is
The frequency component is normalized and the frequency dependence is eliminated.
A target signal section estimation device characterized by the above. The target signal section estimation device according to any one of claims 1, 4 and 5,
The normalization part
Generate two or more kinds of normalized signal values for each time frequency bin,
The uneven distribution index value calculation unit
Calculating two or more uneven distribution index values respectively indicating the uneven distribution of two or more kinds of normalized signal values belonging to each grid, for each grid;
The determination unit is
Weighting two or more uneven distribution index values for each grid, and determining whether each grid corresponds to the target signal section, using the uneven distribution index value after the weight as an index,
A target signal section estimation device characterized by the above. The target signal section estimation device according to any one of claims 1, 4 and 5,
The normalization part
Generate two or more kinds of normalized signal values for each time frequency bin,
The uneven distribution index value calculation unit
Calculating the ubiquitous index value indicating the ubiquity of a vector having two or more kinds of normalized signal values belonging to each grid as elements;
The determination unit is
Using the ubiquitous index value indicating the ubiquity of the vector as an index, it is determined whether each grid corresponds to 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 1, 6, and 7,
The determination unit is
Comparing the ubiquitous index value of each grid or their mapping with a predetermined threshold value to determine whether each grid corresponds to the target signal interval;
A target signal section estimation device characterized by the above. The target signal section estimation device according to any one of claims 1, 6, and 7,
The determination unit is
A first value calculation unit that substitutes the ubiquitous index value of the determination target grid into a predetermined function and calculates a first value that is monotonically increased with respect to a probability that indicates the target signal interval of the grid;
A second value calculation for substituting the ubiquitous index value of the grid in the non-target signal section into the predetermined function and calculating a second value that is monotonically increasing with respect to the probability indicating the likelihood of the target signal section of the grid. And
If the division value that is the ratio between the first value and the second value or the mapping of the division value is greater than or equal to a predetermined threshold value, it is determined that the determination target grid corresponds to the target signal interval, A threshold determination unit that determines that the determination target grid corresponds to the target signal section 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 1, 6, and 7,
The determination unit is
The uneven distribution calculated by the uneven distribution index value calculation unit by pattern recognition using the relationship between the pre-learned grid uneven distribution index value and the determination result of whether or not the grid is the target signal section. Determining whether the grid corresponding to the sex index value corresponds to the target signal interval;
A target signal section estimation device characterized by the above. A target signal section estimation method for estimating a target signal section,
A signal extraction process of extracting each signal observed by a plurality of sensors for each frame that is a predetermined time interval;
A frequency domain conversion process for converting each frame signal for each sensor extracted in the signal extraction process into a frequency domain, and generating a frequency domain signal for each time frequency bin for each sensor;
A normalization process for normalizing each frequency domain signal corresponding to the sensor other than the reference sensor and generating a normalized signal value for each time frequency bin, with the frequency domain signal corresponding to the reference sensor as a reference,
A grid classification process for classifying the normalized signal value of each time frequency bin for each grid that is a predetermined time frequency interval;
An uneven distribution index value calculation process for calculating an uneven distribution index value indicating the uneven distribution of the normalized signal value for each grid;
A determination process for determining whether or not each grid corresponds to the target signal section using the uneven distribution index value as an index,
A target signal section estimation method comprising: It is the target signal area estimation method according to claim 11,
A weight calculation process for generating a weighting factor that is monotonically increasing with respect to the absolute value of the amplitude of the frequency domain signal for each time frequency bin;
The uneven distribution index value calculation process is as follows:
Weighting the frequency of the normalized signal value of the time frequency bin corresponding to the weighting factor by the weighting factor, and calculating the uneven distribution index value using the weighted frequency.
A target signal interval estimation method characterized by the above.   11. 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 13.