JP2011248296A - Sound signal section extracting device and sound signal section extracting method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make possible precise extraction of a sound signal section, even if the input sound signal contains unsteady noise, without having to set a threshold for the power of input signals or process complex arithmetic operations.SOLUTION: Analog sound signals of oscillation generated by any sound, when input, are converted into digital sound signals by sampling and quantizing (a step 101); a two-dimensional picture is prepared by graphic conversion of amplitude values in a time series on the basis of the digital sound signals acquired at the step 101 to prepare binary picture data (a step 102); the binary picture data prepared at the step 102 are subjected to contraction processing to convert the value of a noted pixel from 1 to 0 if a predetermined number of pixels around which have one or more of 0 value, to extract a sound signal section that constitutes an agglomerated region cleared of any noise section present in the binary picture data (a step 103); and the start point and the end point of the agglomerated region in the binary picture data are detected in a time series to identify the sound signal section extracted at the step 103 (a step 106).

Description

  The present invention relates to a sound signal section extraction device and a sound signal section extraction method for extracting a signal section of a desired sound signal by removing a noise component contained in an input sound signal.

  Conventionally, various voice section detecting means are known for extracting a signal section of a desired sound signal by removing a noise component contained in an input sound signal. For example, in order to separate a signal section indicating a normal sound section such as a human speech voice or an operation sound generated from a machine from a noise section indicating a noise section other than the normal sound, the power of the input signal is separated. For example, a speech recognition device and a speech interval detection device that set a threshold value and compare the threshold value with the power of an input signal to distinguish a signal interval from a noise interval are known (for example, Patent Document 1, Patent). Reference 2).

  Also, in order to separate the signal section and the noise section, a zero-crossing method that sets a threshold value for comparison with the power of the input signal based on the number of times that the zero level is crossed within a certain time range (for example, patent document) 3), or an outline of a speech spectrum for improving the speech recognition rate by normalizing the input speech signal (see, for example, Patent Document 4), or capturing speech to detect a signal section. An autocorrelation (see, for example, Patent Document 5) used as a feature amount (periodic information of an input signal) is known.

Japanese Patent Publication No. 63-29754 JP 58-130395 A Japanese Patent Laid-Open No. 5-16596 Japanese Patent Publication No. 1-36959 JP 2007-328228 A

  However, in the speech recognition apparatus described in Patent Document 1 and the speech section detection apparatus described in Patent Document 2, physical characteristics such as the magnitude of noise do not vary greatly with time, so-called stationary noise, This is effective when the noise is relatively small compared to the signal interval, but the threshold may be exceeded for non-stationary noise where the noise level fluctuates irregularly in time series. There was a risk of misrecognition.

  In contrast, the speech detection device described in Patent Literature 3, the speech recognition device described in Patent Literature 4, and the signal processing device described in Patent Literature 5 are patented by using other physical indices. Although it is possible to improve the detection accuracy of the signal section from the speech recognition apparatus described in Document 1 and the speech section detection apparatus described in Patent Document 2, there is a problem that complicated arithmetic processing must be performed. It was. In this case, an arithmetic processing device suitable for the arithmetic processing is required.

  The present invention has been made to solve such a conventional problem, and does not set a threshold for the power of the input signal or perform complicated arithmetic processing even for an input sound signal including non-stationary noise. However, an object of the present invention is to provide a sound signal section extraction device and a sound signal section extraction method that can accurately extract a sound signal section.

  The sound signal section extraction device according to the first aspect of the present invention that achieves the above-mentioned object is detected by a sound information detection unit that detects vibration generated by sound and converts it into an analog sound signal, and a sound information detection unit When a sound analog sound signal is input, it is sampled and quantized and converted into a digital sound signal, and a two-dimensional graph of amplitude values in time series based on the digital sound signal acquired by the signal input unit. An image creation unit that creates an image to generate binary image data, and the binary image data created by the image creation unit has a value of the target pixel if the value of a predetermined number of peripheral pixels is one or more. An image processing unit having a contraction processing function for performing a contraction process for converting 1 to 0 and deleting a noise section existing in the binary image data to extract a sound signal section that is an agglomerated area And extracted by the image processor To identify a sound signal section, in which is composed of a sound signal section determining unit for detecting in time series the start and end points of agglomerated areas in the binary image data.

  According to a second aspect of the present invention, in the sound signal section extraction device according to the first aspect, the image processing unit sets the value of a predetermined number of peripheral pixels to 1 in a lump area in the binary image data. If there is one or more, the expansion processing function is performed to perform the expansion processing for converting the value of the target pixel from 0 to 1, and to restore a region partially deleted by the contraction processing performed by the contraction processing function.

  According to a third aspect of the present invention, in the sound signal section extraction device according to the second aspect, the image processing unit extracts the agglomerated region of the sound signal section in the binary image data by morphological operation processing. .

  The sound signal section extraction method according to the fourth aspect of the present invention includes a first step of converting a digital sound signal by sampling and quantizing when an analog sound signal of vibration generated by sound is input. , A second step of creating a two-dimensional image in which amplitude values are graphed in time series based on the digital sound signal acquired in the first step to obtain binary image data, and a binary generated in the second step If the image data has a predetermined number of peripheral pixel values of 0 or more, a contraction process for converting the value of the target pixel from 1 to 0 is performed, and noise existing in the binary image data A third step of extracting a sound signal section that becomes a clustered area by deleting the section, and a clustered area in the binary image data in order to specify the sound signal section extracted in the third step The start and end points of Leaving ones and a fourth step of detecting a column.

  According to a fifth aspect of the present invention, in the sound signal section extraction method according to the fourth aspect, the third step is to set a predetermined number of peripheral pixel values in the agglomerated region in the binary image data. When one or more 1s are included, expansion processing for converting the value of the target pixel from 0 to 1 is performed, and a region partially deleted by the contraction processing is restored.

  The sound signal section extraction device according to the first aspect and the sound signal section extraction method according to the fourth aspect have physical characteristics such as signal power and spectrum information as in the conventional sound section detection means. Rather than removing the noise component contained in the input sound signal, the sound signal section is extracted based on visual information that allows a person to visually grasp the presence of the sound signal section from the signal waveform of the sound signal. The present invention provides a means for extracting a sound signal section that has not existed before. Specifically, the horizontal axis (X axis) is time and the vertical axis (Y axis) is zero at the center in order to make the signal waveform of the sound signal visual information that humans can visually grasp. Expressed in a graph with positive and negative amplitude values in the direction. By graphing the signal waveform of the sound signal in this way, a person has an amplitude value on the vertical axis that is large up and down, and a clustered area that appears continuously on the horizontal axis is a sound signal section. It will be determined visually.

  The sound signal segment extraction device according to the first aspect and the sound signal segment extraction method according to the fourth aspect apply image recognition by human vision, and an analog sound signal of the sound detected by the sound information detection unit Is converted into a digital sound signal by the signal input unit and sent to the image creating unit, and the image creating unit creates binary image data based on the digital sound signal. Since this binary image data is an image of a digital sound signal including a noise component, if this binary image data is subjected to contraction processing by the contraction processing function of the image processing unit, it exists in the binary image data. It is possible to extract a sound signal section that is an agglomerated area by deleting existing noise sections. Note that even if the noise component of the sound signal is non-stationary noise, the amplitude value is low, or even if the amplitude value is high, it becomes a long and narrow area with a narrow width on the time axis. When binary image data has a predetermined number of peripheral pixel values of 0 or more, a contraction process is performed to convert the value of the target pixel from 1 to 0, thereby removing the noise component and generating a sound signal. Sections can be agglomerated. Then, the threshold value for the power of the input signal is set even for the input sound signal including non-stationary noise by detecting the start point and the end point of the clustered region in the binary image data in time series by the sound signal section determination unit. Therefore, it is possible to extract the sound signal section without performing complicated calculation processing.

  The sound signal section extraction device according to the second aspect and the sound signal section extraction method according to the fifth aspect are the expansion processing function of the image processing unit, and agglomeration in the binary image data by the contraction processing by the contraction processing function. If a predetermined number of peripheral pixel values is 1 or more in the predetermined area, a partially deleted area is restored by performing dilation processing for converting the value of the target pixel from 0 to 1 Therefore, the extraction accuracy of the sound signal section is improved.

  Further, in the sound signal section extraction device according to the third aspect, the image processing unit extracts the agglomerated region of the sound signal section in the binary image data by morphological operation processing, so that the contraction processing by the contraction processing function And expansion processing by the expansion processing function can be used in combination.

  According to the sound signal section extraction device and the sound signal section extraction method of the present invention, even for an input sound signal including non-stationary noise, a threshold value for the power of the input signal is not set or complicated arithmetic processing is not performed. The sound signal section can be accurately extracted.

It is a block diagram of a system configuration showing a preferred embodiment in the sound signal section extraction device of the present invention. It is a flowchart which shows the preferable example of embodiment in the sound signal area extraction method of this invention. It is a figure for demonstrating the sound signal area extraction method of this invention, (A) is the amplitude pattern image of an input sound signal, (B) is the image which performed the contraction process with respect to the amplitude pattern image of (A), ( C) is an image obtained by performing an expansion process on the image of (B). It is explanatory drawing which shows the relationship between the attention pixel and peripheral pixel in the shrinkage | contraction process and expansion process of an image. It is explanatory drawing which shows the relationship between the amplitude pattern image of FIG. 3, and the image which performed the expansion process.

  The best mode for carrying out the sound signal section extraction device and the sound signal section extraction method of the present invention will be described below with reference to the drawings.

  As shown in FIG. 1, the sound signal section extraction device of the present invention includes a sound information detection unit 11 that detects vibration generated by sound and converts it into an analog sound signal, and an A / D conversion unit (not shown). When an analog sound signal of the sound detected by the detection unit 11 is input, a signal input unit 12 that converts the digital sound signal by sampling and quantization, and an amplitude value based on the digital sound signal acquired by the signal input unit 12 An image creation unit 13 that creates a time-series graphed two-dimensional image to obtain binary image data, and an image processing unit 14 that has a shrinkage processing function for performing shrinkage processing on the binary image data created by the image creation unit 13. And a sound signal section determination section 15 for detecting the start and end points of the clustered area in the binary image data in time series in order to specify the sound signal section extracted by the image processing section 14. .

  The sound information detection unit 11 corresponds to a microphone or a vibration sensor that can detect vibration generated by sound and convert it into an analog sound signal. Since sound is generated by vibration of air in the microphone, an analog waveform corresponding to the vibration of the air can be acquired by the signal input unit 12 by converting the vibration of the air into an electric signal. The vibration sensor captures vibration quantitatively by displacement, speed, or acceleration, and obtains an analog waveform corresponding to the vibration by the signal input unit 12 by converting the measured physical quantity into an electric signal. Can do.

  In order to store the signal information of the digital sound signal, the signal input unit 12 stores, for example, the first column with the sample point number, the second column with the amplitude value corresponding to the first column, and N rows of the total number of samples ( N: 0, 1, 2,..., N−1). Further, by storing the value of the sampling frequency F, the time of an arbitrary sample point n (0 ≦ n ≦ N−1) can be calculated by n × (1 / F). Therefore, the first column in the signal information matrix stores time information.

The image creating unit 13 creates an image in which the sample points in the first column in the signal information matrix are expressed on the X axis as a time axis, and the amplitude values in the second column are expressed on the Y axis as a signal power. . For example, n _{a to} n _b (a, b: 0 ≦ a <b, a <b ≦ N−1) between arbitrary sample points is set to one pixel in the X-axis direction, and between n _{a to} n _b The number of pixels corresponding to the average value, median value, and maximum value of a certain amplitude value is expressed on the Y axis. The Y axis represents a positive amplitude value above the center of the Y axis and a negative amplitude value below the center of the Y axis.

This image is a binary image in which the black portion expressing the amplitude value is 1 and the white portion other than that is 0. In this case, the image size S _{X in} the X-axis direction is N / (n _b −n _a ). The image size Sy in the Y-axis direction is a value obtained by adding the maximum value Ap and the minimum value Am of the amplitude value when the original value of the amplitude value is used as it is, but may be an arbitrary size. In order to set the image size to an arbitrary size, an equation corresponding to image enlargement / reduction processing, Sy = (Ap + Am) × C, is preferable. In this case, the average or median amplitude value in between any sample point n _a ~n _b, the number of pixels of value component that corresponds to the maximum value even in the C doubled to represent the Y-axis. Here, C is an enlargement process if C> 1, and a reduction process if 0 <C <1.
Although the amplitude value is expressed linearly here, it is not limited to this and may be expressed nonlinearly.

  The contraction processing function of the image processing unit 14 performs contraction processing for converting the value of the target pixel from 1 to 0 when the value of the predetermined number of peripheral pixels is one or more. This image contraction processing is effective for smoothing a convex shape with a narrow pixel width extending outward from the contour of the target object in the binary image, and for removing so-called sesame salt noise in the background of the target object. If one or more 0 is present in the peripheral pixels of the target pixel in the process, the target pixel is set to zero. Note that the pixels x and y have values satisfying 0 ≦ x <Sx−1 and 0 ≦ y <Sy−1 so that the region of the sound signal section is not largely deleted by the contraction process. The range of surrounding pixels is arbitrary, but if the range is too large, there is a possibility that the region where the signal section exists is deleted. Therefore, it is best to use 8 pixels adjacent to the target pixel as the minimum component. is there. That is, the center of 3 × 3 pixels is set as the target pixel, and when the target pixel is 1, if any one of the eight surrounding pixels is 0, the target pixel is set to 0. When this contraction processing is repeatedly performed on the entire binary image data created by the image creation unit 13, the value 1 region is gradually reduced, and eventually the region with a small area disappears. However, only the area with the originally large area remains on the image.

  Here, since the amplitude value of the noise section is low, the portion where the value on the image is 1 is a small area, and is therefore deleted as a result of the contraction process. In the noise section, the amplitude value may increase with a short time length. However, since the portion where the value on the image is 1 is a narrow and narrow area in the X-axis direction, this area is also the result. Are deleted by the contraction process. On the other hand, in the signal section, since the area of the value 1 on the image is large, it is not deleted even if the contraction process is performed. This contraction process is repeated at least k times (k ≧ 1) until a region where a target sound signal section exists can be extracted or a region where a noise section exists is deleted.

In this image shrinking process, depending on the shape of the image, a part of the area where the sound signal section exists may be deleted. Therefore, the image processing unit 14 may be provided with an expansion processing function for performing the expansion process on the image. . The expansion processing function of the image processing unit 14 performs expansion processing for converting the value of the pixel of interest from 0 to 1 when one or more of the predetermined number of peripheral pixel values is one. This image expansion process is an effective process for smoothing a concave shape with a narrow pixel width extending inside a target object in a binary image and removing so-called sesame salt noise inside the target object. This process is repeated at least k times (k ≧ 1) for the entire binary image data created by the image creation unit 13 until a partially deleted region appears. Since the amplitude pattern area where no sound information exists is deleted by the contraction process, the original area does not appear even if the expansion process is performed. That is, the region of value 1 remaining in the image after the expansion processing is a region where a sound signal section exists. As described above, in the dilation processing function of the image processing unit 14, there is one or more 1 in the values of the predetermined number of peripheral pixels in the agglomerated region in the binary image data by the constriction processing by the contraction processing function Since the partially deleted region can be restored by performing expansion processing for converting the value of the target pixel from 0 to 1, the extraction accuracy of the sound signal section is improved.
Such shrinkage processing and expansion processing of the image processing unit 14 are preferably morphological calculation processing. This is because the morphological calculation process can be used in combination with the shrinking process and the expanding process.

  The sound signal section determination unit 15 extracts the agglomerated area in the binary image data serving as the sound signal section by the signal input unit 12, the image creation unit 13, and the image processing unit 14. The sound signal section can be identified efficiently and accurately only by detecting the starting point and the ending point of the sound in time series.

  Hereinafter, the sound signal section extraction method by the sound signal section extraction apparatus configured as described above will be described with reference to FIGS. 1, 2, 3, 4, and 5. Note that the image processing unit 14 has a contraction processing function and an expansion processing function.

  An analog sound signal of the sound detected by the sound information detection unit 11 is converted into a digital sound signal by the signal input unit 12 and sent to the image creation unit 102 (step 101), and the image creation unit 102 performs binary processing based on the digital sound signal. Image data is created (step 102). As shown in FIG. 3A, the binary image data is a binary image in which the black portion expressing the amplitude value is 1 and the white portion that is the other background is 0. Since this binary image data is an image of a digital sound signal including a noise component, if this binary image data is subjected to a contraction process by the contraction processing function of the image processing unit 14 (step 103), the binary image It is possible to extract a sound signal section that is an agglomerated area by deleting a noise section existing in the data (step 104). As shown in FIG. 4, the contraction process uses a central pixel e of 3 × 3 pixels as a target pixel, and when the target pixel e is 1, the surrounding eight pixels a, b, c, d, f, g, If at least one of h and i is 0, the target pixel is set to 0. Steps 103 and 104 are repeatedly executed until the area where the noise section exists is deleted. The sound signal section that is an agglomerated area by deleting the noise section is an image as shown in FIG.

  Further, the image processing unit 14 is an expansion processing function, and by performing expansion processing on the agglomerated region in the binary image data by contraction processing by the contraction processing function, it is possible to restore a partially deleted region ( Step 105). As shown in FIG. 4, when the target pixel e is 0, this expansion processing is performed if any one of the surrounding eight pixels a, b, c, d, f, g, h, i is 1 Let the pixel be 1. Step 105 is repeatedly executed until an area partially deleted by the shrinking process appears. The sound signal section that is an agglomerated region where the partially deleted region appears is an image as shown in FIG.

  In this way, after performing the n-time contraction process on the image, the opening process for performing the n-time expansion process is performed because an image obtained by binarizing the signal waveform of the input sound signal has so-called sesame salt noise. Although it does not occur, unevenness due to the magnitude of the amplitude value occurs. Impulse noise, which is a place where the amplitude value is large in a pulse shape, has a convex shape with a narrow pixel width in the X-axis direction. The area becomes narrow. The contraction process can remove these noises, and as a result, it is possible to leave only the sound signal section that becomes a clustered area on the image. However, since the width of the sound signal section, which is an agglomerated area extracted by the contraction process, decreases in the X-axis direction by the number of contractions, the original X-axis width of the area is reduced. In order to return, it is necessary to perform expansion processing for the number of times of contraction processing.

  As shown in FIG. 5, the image of the region of the sound signal section obtained in this way detects the start point A and the end point B of the agglomerated area in the binary image data in a time series as shown in FIG. (Step 106). By performing image processing in this way, even an input sound signal including non-stationary noise can be a sound that becomes an agglomerated region without setting a threshold value for the power of the input signal or performing complicated arithmetic processing. The signal interval can be extracted efficiently and accurately.

Such a sound signal section extraction device and sound signal section extraction method of the present invention can improve the extraction accuracy of a sound signal section by using it together with a conventional speech section detection means.
Further, the object of the present invention can also be achieved by reading and executing by a computer a recording medium that records software program codes for realizing the functions of the sound signal section extracting device and the sound signal section extracting method of the present invention. it can.

  Although the present invention has been described with the specific embodiments shown in the drawings, the present invention is not limited to the embodiments shown in the drawings, and is known so far as long as the effects of the present invention are achieved. It goes without saying that any configuration can be adopted.

  Such a sound signal section extraction apparatus and sound signal section extraction method of the present invention may be misrecognized by noise when analyzing the operation sound of a machine, for example, in order to detect an abnormal sign of the machine at an early stage. Can be prevented.

DESCRIPTION OF SYMBOLS 1 ... Sound signal area extraction device 11 ... Sound information detection part 12 ... Signal input part 13 ... Image preparation part 14 ... Image processing part 15 ... Sound signal area determination part

Claims (5)

A sound information detector that detects vibration generated by sound and converts it into an analog sound signal;
When the analog sound signal of the sound detected by the sound information detection unit is input, a signal input unit that converts to a digital sound signal by sampling and quantizing,
An image creation unit that creates a binary image data by creating a two-dimensional image in which amplitude values are graphed in time series based on the digital sound signal acquired by the signal input unit;
The binary image data created by the image creation unit is subjected to a contraction process for converting the value of the target pixel from 1 to 0 when the predetermined number of peripheral pixel values is one or more, and An image processing unit having a contraction processing function for extracting a sound signal section that is an agglomerated area by deleting a noise section existing in binary image data;
In order to specify the sound signal section extracted by the image processing unit, the sound signal section determination unit configured to detect the start point and the end point of the clustered area in the binary image data in time series. A sound signal section extraction device characterized by comprising:   The image processing unit converts the value of the pixel of interest from 0 to 1 when the lumped area in the binary image data has one or more of the predetermined number of neighboring pixel values. The sound signal section extraction device according to claim 1, further comprising an expansion processing function for performing expansion processing and restoring a region partially deleted by the contraction processing performed by the contraction processing function.   The sound signal section extraction device according to claim 2, wherein the image processing section extracts the agglomerated region of the sound signal section in the binary image data by morphological operation processing. When an analog sound signal of vibration generated by sound is input, a first step of converting to a digital sound signal by sampling and quantizing;
A second step of generating a two-dimensional image in which amplitude values are graphed in time series based on the digital sound signal acquired in the first step, and obtaining binary image data;
The binary image data created in the second step is subjected to a contraction process for converting the value of the target pixel from 1 to 0 when the predetermined number of peripheral pixel values is one or more, A third step of extracting a sound signal section that is a clustered area by deleting a noise section existing in the binary image data;
A fourth step of detecting, in time series, a start point and an end point of the agglomerated region in the binary image data in order to specify the sound signal section extracted in the third step. A characteristic sound signal segment extraction method.   The third step converts the value of the target pixel from 0 to 1 when the lumped area in the binary image data has one or more of the predetermined number of neighboring pixel values. The sound signal section extraction method according to claim 4, further comprising: performing an expansion process to restore a region partially deleted by the contraction process.

Images