JP2012127701A - Device and method for sound detection - Google Patents

Abstract

PROBLEM TO BE SOLVED: To detect a specific sound in a noise environment with ease and high accuracy.SOLUTION: A signal power calculation part and an inclination calculation part calculate a feature value indicating sound data characteristics of collected observation sounds in time series. Meanwhile, an expectation value of the feature value is calculated, as score parameters, from each of a plurality of learning data which is the kind of sound as detection object but different from one another, in time series. A score calculation part calculates a score evaluating sound data based on a difference between the feature value calculated from the sound data and the expectation value of the feature value calculated from the learning data. A generation section detecting part detects positions of maximum and minimum values of the score and detects a specific sound generation section in sound data of the observation sounds based on the positions of the maximum and minimum values.

Description

  The present invention relates to a sound detection apparatus and a sound detection method for detecting a specific type of sound.

  Conventionally, in security, detection of occurrence of an abnormal situation has been performed by paying attention to a specific sound. For example, when an abnormal sound such as a glass breaking sound is detected in the monitoring area, it can be determined that an abnormal situation has occurred. Moreover, although it is not determined that the sound is abnormal, it is necessary to determine whether or not an abnormal situation has occurred when a suspicious sound is detected. In order to automatically detect such a specific sound, it is necessary to detect the abnormal sound or the suspicious sound itself or the section where the abnormal sound or the suspicious sound is generated from all the observation sounds observed in the monitoring area. There is. In the following, unless otherwise specified, abnormal sounds and suspicious sounds are collectively referred to as suspicious sounds.

  Techniques for detecting specific sounds from sounds including environmental sounds have been proposed. For example, Patent Literature 1 discloses a technique for detecting a voice generation section using sound signal power. In Patent Literature 1, a predetermined sound generation section is detected by appropriately setting a threshold value for signal power. Further, for example, Patent Document 2 discloses a method of detecting a specific sound section using the number of zero crosses of a sound signal.

  Further, for example, as a technique used particularly for speech, a method for detecting a predetermined sound generation section by dividing a sound into a plurality of frequency bands and setting a threshold value for the signal power obtained in each band is known. It has been. This method is suitable for extracting a sound whose characteristic frequency band is known in advance, such as a human voice.

  Furthermore, in the detection of suspicious sound, a method of performing speech recognition processing on all sounds collected in a certain time without detecting the occurrence section is also conceivable. In this method, for example, the user designates the start and end points of voice recognition processing for the collected sound signal. The processing device performs voice recognition processing between the start and end points specified by the user of the transferred sound signal, and the user detects suspicious sound using the result of the voice recognition processing.

Japanese Patent No. 2521425 Japanese Patent No. 2944098

  By the way, when the technique of Patent Document 1 described above is used for detection of suspicious sound, if noise (for example, construction noise) is superimposed on the collected environmental noise and voice signal, the signal of suspicious sound The noise signal power is superimposed on the power. In this case, there is a problem that the detection of the suspicious sound may be missed or the section where the suspicious sound is generated may not be detected correctly. In addition, with respect to the technique of the above-mentioned Patent Document 2, the number of zero crosses is easily affected by noise, and when noise is superimposed on environmental sound, it is difficult to appropriately capture the characteristics of suspicious sound. there were.

  Furthermore, the method of dividing a sound into a plurality of frequency bands is effective for detecting a predetermined sound generation section in a noise environment if the signal power of the sound to be detected is concentrated in a specific frequency band. . However, the suspicious sound is substantially a real sound, and the signal power of the sound is not necessarily concentrated in a specific frequency band, and is not necessarily effective for detecting the suspicious sound section. was there.

  Furthermore, the method of performing voice recognition processing on all sounds collected in a certain time has a problem that abnormal sounds and suspicious sounds cannot be detected by the voice recognition processing itself. For example, in this method, a user observes a change in a parameter obtained as a result of the speech recognition process, and detects a specific sound such as a suspicious sound.

  In addition, when this method is applied to a monitoring device, it is necessary to perform speech recognition processing on all observation sounds being monitored over a long period of time, for example, 8 hours to 10 hours. From the aspect, there was a problem that was not realistic.

  On the other hand, it is also conceivable to reduce the calculation cost by installing one processing device (server) for a plurality of monitoring targets. However, even in this case, there is a problem that it is necessary to always transfer each of the observation sound data collected at a plurality of monitoring points to the server, which is not realistic in terms of communication cost.

  The present invention has been made in view of the above, and an object of the present invention is to detect a specific sound easily and with high accuracy even in a noisy environment.

  In order to solve the above-described problems and achieve the object, the present invention provides a feature value calculation means for calculating feature values indicating the characteristics of sound data along a time series, and is obtained in advance along the time series from learning data. Score calculating means for calculating a score for evaluating the sound data based on the difference between the expected value of the characteristic value and the characteristic value of the signal power time series of the sound data calculated by the feature value calculating means, and a maximum of the score And detecting means for detecting the position of the value and the position of the minimum value, and detecting a specific sound generation section in the sound data based on the position of the maximum value and the position of the minimum value.

  Further, according to the present invention, the feature value calculating means calculates the feature value indicating the feature of the sound data along the time series, and the score calculating means previously obtained from the learning data along the time series. A score calculation step for calculating a score for evaluating sound data based on the difference between the expected value of the feature value and the signal power time series feature value of the sound data calculated in the feature value calculation step, and a maximum value of the score And a detection step of detecting a specific sound generation section in the sound data based on the position of the maximum value and the position of the minimum value.

  According to the present invention, it is possible to easily and accurately detect a specific sound even in a noisy environment.

FIG. 1 is a block diagram schematically showing a configuration of an example of a sound detection apparatus applicable to the embodiment of the present invention. FIG. 2 is a functional block diagram of an example for explaining the function of the sound detection device in more detail. FIG. 3 is a schematic diagram for explaining the score parameter calculation method in more detail. FIG. 4 is a histogram showing an example of the distribution of the gradient of the signal power time series for each frame when the acoustic data of the glass impact sound is used as the learning data. FIG. 5 is a schematic diagram illustrating an example of the expected value μ and the variance value σ ² of each frame calculated based on the distribution of the slope of the signal power time series regarding each frame. FIG. 6 is a schematic diagram showing an example of values obtained for each frame i of the input acoustic data for the signal power y _LP (i), the slope y _GLP (i) of the signal power time series, and the score S (i). is there. FIG. 7 is a schematic diagram illustrating an example of the expected value μ _k and the variance value σ ² _k obtained for each frame k of the learning data. FIG. 8 is a graph in which the signal power, the slope of the signal power time series, and the score are plotted against the frame number. FIG. 9 is a schematic diagram illustrating an example of a suspicious sound generation section when the observed sound has little noise. FIG. 10 is a schematic diagram illustrating an example of a suspicious sound generation section when there is a lot of noise in the observed sound. FIG. 11 is a schematic diagram for explaining the delay of the calculated score. FIG. 12 is a schematic diagram for explaining correction of score delay. FIG. 13 is a flowchart of an example showing the detection processing of the suspicious sound generation section according to the present embodiment. FIG. 14 is a schematic diagram illustrating an example of signal power, signal power time series slope, and score when substantially random acoustic data is input to learning data based on a sound. FIG. 15 is a graph in which the signal power, the slope of the signal power time series, and the score are plotted with respect to the frame number when substantially random acoustic data is input with respect to the learning data based on the physical sound.

  Hereinafter, an embodiment of a sound detection device according to the present invention will be described in detail with reference to the accompanying drawings. In the embodiment of the present invention, the sound in the monitoring area is observed, and the generation period of the specific sound that is regarded as a suspicious sound or an abnormal sound is detected from the acoustic signal of the observed sound. Then, the detected sound signal of the specific sound generation section is cut out from the sound signal of the observation sound and output.

  The specific sound to be detected is a so-called sound that is different from the sound emitted by a person, and is similar to the environmental sound included in the observation sound. For this reason, in the present embodiment, generally, the expected value of the characteristic value is obtained in advance for the learning data based on the acoustic data of the same type of sound as the specific sound to be detected, and the obtained expected value and the time of the observation sound are determined. A specific sound generation interval is detected based on a score calculated using a difference from a feature value on the sequence.

  Here, the observation sound, environmental sound, suspicious sound and abnormal sound are defined. Observation sound refers to all sound collected in the surveillance area. An abnormal sound is a sound that should be output as an alarm, caused by intrusion behavior. A typical example of the abnormal sound is a breaking sound generated when glass or the like is broken. Suspicious sound is suspicious sound collected in the surveillance area, although it is not determined as abnormal sound. As an example of a suspicious sound, a striking sound can be considered. Environmental sound refers to sound other than suspicious sound and abnormal sound included in observation sound. Examples of environmental sounds include sounds resulting from natural phenomena such as wind, automobiles and trains.

  FIG. 1 schematically shows a configuration of an example of a sound detection apparatus applicable to the embodiment of the present invention. In FIG. 1, the sound detection device 100 includes an A / D conversion unit 11, a calculation unit 12, and a storage unit 13. For example, the observation sound collected by the microphone 10 in the monitoring area is converted into digital data by the A / D conversion unit 11 and supplied to the calculation unit 12 as input acoustic data 20.

  The arithmetic unit 12 includes, for example, a CPU (Central Processing Unit), a microprocessor, or a DSP (Digital Signal Processor). The storage unit 13 includes, for example, a semiconductor memory or an HDD (hard disk drive), and stores the input acoustic data 20 and calculates a score for evaluating the input acoustic data 20 created based on the learning data. The score parameters are stored in advance. The learning data is the same type as the suspicious sound or abnormal sound to be detected, and uses acoustic data of a plurality of different sounds. The storage unit 13 can also be used as a work area for the calculation unit 12.

  The calculation unit 12 calculates a feature value of the input acoustic data 20 on a time series, evaluates the calculated feature value on the time series using a score parameter stored in the storage unit 13, and performs detection Detect suspicious and abnormal sound occurrence sections. When detecting the suspicious sound or abnormal sound generation section to be detected from the input acoustic data 20, the calculation unit 12 cuts out the detected suspicious sound or abnormal sound generation section from the input acoustic data 20, and detects the suspicious sound generation section. Are output as output acoustic data 21. The output acoustic data 21 is transmitted to a monitoring server or the like via a communication network, for example.

  In the present embodiment, the time-series gradient of the signal power of the acoustic data is used as the characteristic value of the acoustic data. Note that this is not limited to this example, and the feature value may be another value as long as it is a value indicating the feature of the acoustic data. For example, the acoustic data signal power itself or the number of zero crosses in a predetermined section may be used as the feature value.

  FIG. 2 is a functional block diagram of an example for explaining the functions of the sound detection device 100 in more detail. In FIG. 2, the same reference numerals are given to portions common to those in FIG. 1 described above, and detailed description thereof is omitted. In the present embodiment, the above-described suspicious sound and abnormal sound are both detected. Therefore, in the following, suspicious sound and abnormal sound are collectively described as suspicious sound unless otherwise specified.

  In FIG. 2, the signal power calculation unit 101, the slope calculation unit 102, the score calculation unit 103, and the generation interval detection unit 104 are included in the calculation unit 12. The score parameter 111 is created in advance based on the learning data and stored in the storage unit 13. The storage unit 110 is, for example, an area in the storage unit 13 described above, and temporarily stores the input acoustic data 20 obtained by converting the observation sound into digital data by the A / D conversion unit 11. The data used for detecting the suspicious sound generation section, such as the score to be corrected and the delay time correction amount, is stored.

  The signal power calculation unit 101 calculates the signal power of the input acoustic data 20 supplied from the A / D conversion unit 11 on a time series. More specifically, the signal power calculation unit 101 calculates the signal power in units of a predetermined number of samples that are continuous on the time series of the input acoustic data 20. Here, the unit for calculating the signal power is called a frame, and the number of samples included in the frame is the frame width.

The signal power y _LP (i) of the i-th frame of the input acoustic data 20 is, for example, given by the following equation (1), where the frame width is W and the n-th waveform data (sample value) in the frame is a value x (n). Is calculated by

In the frame for calculating the signal power y _LP (i), samples to be used are shifted for every predetermined frame interval D with respect to the head of the frame so that some samples overlap the previous frame. To set. As an example, assuming that the frame width W = 160 samples and the mth frame m is composed of the first sample to the 160th sample, the m + 1st frame (m + 1) is the 81st sample to the 240th sample. The (m + 2) th frame (m + 2) is composed of the 161st to 320th samples. In this case, the frame interval D = 80 samples. In this example, the length of the overlapping portion of the frame is the frame width W / 2, but this is not limited to this example.

The inclination detection unit 102 calculates a time-series inclination (referred to as a signal power time-series inclination) y _GLP (i) of the signal power y _LP (i) for each frame calculated by the signal power calculation unit 101. To do. The gradient y _GLP (i) of the signal power time series of the frame i is, for example, the frame i, the frame (i-4), the frame (i-3), and the frame (i-1) for which the signal power has already been calculated. It can be calculated by the following equation (2) using the signal power of 4 frames out of the frame (i-4) that goes back 4 frames from the target frame i.

In Equation (2), the gradient y _GLP (i) is calculated using the signal power of four frames from the previous four frames with respect to the target frame i, but this is not limited to this example. In other words, the number of frames going back from the target frame i to use data is determined according to the configuration of the equation for _obtaining the slope y _GLP (i) of the signal power time series. Also, the value of each constant is not limited to the value used in this example.

The score calculation unit 103 uses the signal power time-series gradient y _GLP (i) calculated for the input acoustic data 20 by the gradient calculation unit 102 and the learning parameter to create score parameters stored in the storage unit 13 in advance. 111, a score for evaluating the input acoustic data 20 is calculated. The score parameter 111 includes an expected value for a suspicious sound to be detected, which is created based on learning data prepared in advance. Then, the suspicious sound generation interval is detected based on the time-series change of the calculated score.

  The score parameter 111 is created as follows. First, a plurality of acoustic data having the same type as the suspicious sound to be detected and having different sounds are prepared as learning data. For example, when the suspicious sound to be detected is a glass breaking sound, each acoustic data obtained by collecting the breaking sound when the glass is broken under various conditions such as different sizes, thicknesses, and materials is used as learning data. .

For each of the learning data, the signal power is obtained for each frame as described above, and the slope of the signal power time series is calculated using the obtained signal power. Then, the expected value μ _k and the variance value σ ² _k of the slope of the signal power time series of the corresponding frame k of the plurality of learning data are calculated. The calculated expected value μ _k and variance value σ ² _k are stored in the storage unit 13 as the score parameter 111 of the frame k.

  The calculation method of the score parameter 111 will be described in more detail with reference to FIG. First, a frame is set for learning data. The frame width W and the frame interval D are the same as the frames set for the input acoustic data 20 by the signal power calculation unit 101 described above. Note that the slope of the signal power time series is calculated using the signal power value from a frame that is four frames back from the target frame i according to the above-described equation (2).

  In the following description, the first frame (the first frame (when the frame necessary for calculating the slope of the signal power time series of the frame is traced from the frame including the start position of the suspicious sound generation section) is referred to as the first frame ( Frame # 1).

  A frame (frame # 5 in the example of FIG. 3) related to the rise time of the waveform of the learning data, that is, the start position of the suspicious sound generation interval is set. Then, frames are sequentially set at a frame interval D from the frame # 5 toward the direction of attenuation of the waveform of the learning data (end position direction of the suspicious sound generation interval) (frames # 6 to # 8). In addition, when calculating the slope of the signal power time series for a frame related to the start position of the suspicious sound generation section, in this example, since the frame is required from a position that is four frames back from the frame, these frames are also Set. In the example of FIG. 3, frames # 4 to # 1 are set in a direction that goes back in time series from frame # 5. Note that the frame (frame # 5) related to the start position of the suspicious sound generation section is preferably set so that the approximate center of the frame is the start position.

  Similarly, each frame is set with respect to each of a plurality of learning data having different sounds on the basis of the rise time of the waveform.

  In this example, the number of frames used to calculate the slope of the signal power time series is 4 frames, and the total number of frames used to calculate the score parameter 111 from the learning data is 8 frames. For example, more frames may be used. In addition, the frame that aligns the center with respect to the rise time of the waveform of the learning data is the fifth frame, but this is not limited to this example, and is matched to the total number of frames used to calculate the score parameter 111. Another frame may be used. Furthermore, there may be a case where there is no data before the waveform rise time in the learning data. In this case, the frame is set on the assumption that data of “0” (silence data) exists.

Next, signal power is calculated for each frame set for each learning data as described above, and the slope of the signal power time series is calculated for each learning data. Then, the expected value μ _k and variance value σ ² _k of the slope of the signal power time series in the corresponding frame k of each learning data are calculated.

  As an example, a histogram relating to the slope of the signal power time series is created for each of the frames # 5 to # 8 based on the learning data with the frame # 5 including the rising portion of the waveform as a base point. FIG. 4 shows an example of the distribution (histogram) of the gradient of the signal power time series for each of the frames # 5 to # 8 when the acoustic data of the glass impact sound is used as the learning data. 4A shows an example of frame # 5, FIG. 4B shows an example of frame # 6, FIG. 4C shows an example of frame # 7, and FIG. 4D shows an example of frame # 8. 4 (a), 4 (b), 4 (c), and 4 (d), the horizontal axis represents the slope of the signal power time series, and the vertical axis represents the frequency.

Based on these histograms of FIGS. 4A to 4D, the expected value μ and the variance value σ ² can be obtained for each of the frames # 5 to # 8. Note that the calculation of the expected value μ _k and the variance value σ ² _k can be performed using a known method, and thus the description thereof is omitted here. FIG. 5 shows an example of the expected value μ and variance value σ ² of each of frame # 5 to frame # 8 calculated based on the histograms of FIGS. 4 (a) to 4 (d). Expected value μ _k and variance value σ ² _k are calculated for each frame k. The calculated expected value μ _k and variance value σ ² _k of each frame k are stored in the storage unit 13 as the score parameter 111.

The score calculation unit 103 uses the expected value μ _k and the variance value σ ² _k of each frame k thus calculated and stored as the score parameter 111 in the storage unit 13, and the input sound exemplified by the following equation (3) A score calculation formula for calculating the score S (i) in the frame i of the data 20 is obtained. The frame i in the input acoustic data 20 can be evaluated by the score S (i) calculated by this score calculation formula.

In equation (3), the value “8” included in the gradient y _GLP (i + k−8) and the value “8” indicating the end of the sum are the total frames used to calculate the score parameter 111 from the learning data. Is a number. The value “5” indicating the start of the sum is a frame number counted from the first frame used for calculating the score parameter 111 from the learning data of the frame number including the start position of the suspicious sound generation section. These values are determined according to the configuration of an equation for _obtaining the slope y _GLP (i) of the signal power time series. Further, in the expression (3), in order to set the maximum score value to “0”, a negative sign is assigned to the entire right side.

  That is, the equation (3) is obtained by calculating the square of the difference between the slope of the signal power time series of the frame i to be score-calculated in the input acoustic data 20 and the expected value from the start position of the suspicious sound generation interval in the learning data. The score of frame i is calculated based on the sum obtained by sequentially shifting the frames. The variance value normalizes the numerator value. In equation (3), the square of the difference between the slope of the signal power time series and the expected value is used, but this is not limited to this example, and the absolute value of the difference may be used, for example.

The score calculation unit 103 sequentially applies the gradient y _GLP (i) of the signal power time series calculated for each frame i of the input acoustic data 20 by the gradient calculation unit 102 to the equation (3), and calculates the value of each frame i. Score S (i) is calculated.

The generation section detection unit 104 calculates the score time-series slope GS (i) of the frame i from the score S (i) of the frame i of the input acoustic data 20 calculated by the score calculation unit 103. In this example, the slope GS (i) of the score time series is 4 from the frame (i-4) that is four frames backward from the target frame i, as in the calculation of the slope of the signal power time series described above. Using the score S (i-4), the score S (i-3), the score S (i-1), and the score S (i), for example, the following equation (4) is used.

  Note that, here, four scores are used to calculate the slope GS (i) of the score time series, but this is not limited to this example. In addition, the score time series slope may be calculated by the score calculation unit 103.

The occurrence section detection unit 104 determines whether the score S (i) is a maximum value or a minimum value based on the calculated slope GS (i) of the score time series. That is, when the slope GS (i) of the score time series satisfies the following condition (A), the score S (i) takes a maximum value.
GS (i-1)> 0 and GS (i) ≦ 0 (A)

Similarly, when the slope GS (i) of the score time series satisfies the following condition (B), the score S (i) takes a minimum value.
GS (i−1) <0 and GS (i) ≧ 0 (B)

  When the score S (i) is a maximum value, the occurrence section detection unit 104 determines whether or not the score S (i) exceeds a threshold, and if it exceeds, the score S (i) is set to the score S (i). It is determined that the corresponding frame i includes the rising position of the waveform. The threshold value is obtained in advance by an experimental method or the like and stored in the storage unit 13. Here, in the learning data, when the approximate center of a frame (for example, frame # 5) related to the waveform rising position is set to be the start position, the position of the approximate center of the frame i is the waveform rising position. It is said.

  On the other hand, for the score S (i) that first takes the minimum value after detecting the frame including the rising position of the waveform, the generation section detection unit 104 has the waveform (i) corresponding to the score S (i) as the waveform. It is determined that the falling position is included. Also in this case, in the learning data, when the approximate center of the frame related to the rising position of the waveform (for example, frame # 5) is set to be the rising position, the position of the approximate center of the frame i is the waveform rising position. The position is lowered.

  When the rising position and the falling position of the waveform are detected, the generation section detection unit 104 sets the detected rising position of the waveform as the start position of the suspicious sound generation section, and sets the falling position as the end position of the suspicious sound generation section. And Thereby, a suspicious sound generation area is detected.

Here, the meaning of the score calculation formula shown in Formula (3) will be described. The slope y _GLP (i) of the signal power time series used in the equation (3), which is a score calculation formula, is calculated by the above equation (2), and the signal power y _LP (i) used in the equation (2) is , Calculated by the above-described equation (1). FIG. 6 shows an example of values obtained for the frame i of the input acoustic data 20 for the signal power y _LP (i), the slope y _GLP (i) of the signal power time series, and the score S (i) thus obtained. .

The expected value μ and the variance value σ ² of the feature value based on the learning data are respectively a frame including a rising edge of a sound of a sound (suspicious sound) in the learning data and a predetermined number of frames (four frames in the above example) from the frame. Is a value calculated for. FIG. 7 shows an example of the expected value μ _k and the variance value σ ² _k obtained for each frame k of the learning data.

  FIG. 8 shows a graph in which the signal power, the slope of the signal power time series, and the score shown in FIG. 6 are plotted against the frame number. In this graph, the delay associated with score calculation described later is not corrected. In the example of FIG. 8, the score plot is delayed by 3 frames relative to the signal power plot.

  The score based on the difference between the expected value μ and the characteristic value of the input sound data 20 at the position where the similarity between the sound in the learning data and the sound included in the input sound data 20 is high, that is, the rising position of the sound, is the maximum. Take the value (position of frame # 9 in the score plot). Therefore, with reference to the Σ portion of Equation (3), the score has a maximum value in that the sum of the predetermined number of frames takes the largest value, and that frame is set as a frame including the start position of the sound generation interval.

The signal power of the input sound data 20 is attenuated after the start position of the sound generation interval (frame numbers # 7 to # 10 in the signal power plot). Accordingly, the slope of the signal power time series of the input acoustic data 20 takes a negative value (frames # 8 and # 9 in the plot of the slope of the signal power time series). Therefore, the value of the square of “y _GLP (i + k−8) −μ _k ” in the equation (3) is large, and the score S (i) in which the sum of the square value of 4 frames is added with a negative sign is It becomes a small value (frames # 12 and # 13 in the score plot). When the value of the score S (i) is the smallest, the score S (i) takes a minimum value (frame # 13 in the score plot), and the frame that takes this minimum value includes the end position of the sound generation interval. It can be regarded as a frame.

  9 and 10 show examples of the suspicious sound generation interval detected as described above. FIG. 9 shows an example when the observation sound has a little noise (environmental sound), and FIG. 10 shows an example when the observation sound has a lot of noise. In FIG. 9 and FIG. 10, the suspicious sound generation interval is detected using the same score parameter 111 and the threshold value for detecting the suspicious sound generation interval.

  9 and 10, the upper graph shows the input sound data 20, and the lower graph shows the score for the input sound data 20 and the suspicious sound generation interval detected based on the score. The suspicious sound generation section indicates a suspicious sound generation section having a high level value. In FIGS. 9 and 10, in the score graph, a delay related to score calculation described later is corrected.

  In FIG. 9, the graph of the input sound data 20 on the upper side shows that a suspicious sound is generated near the time “3000”, and this suspicious sound is rapidly attenuated at a time “200”. On the other hand, in the lower graph, according to the condition (A) described above, the score has a large maximum value near the time “3000” and a slightly large maximum value near the time “5500”. In this example, it is assumed that the maximum value of the score near time “3000” exceeds the threshold, and the maximum value of the score near time “5500” does not exceed the threshold. Further, the score takes a local minimum value near the time “3300” after the local maximum value near the time “3000” exceeding the threshold in accordance with the condition (B) described above. Therefore, it is possible to determine that the vicinity of time “3000” to time “3300” is the suspicious sound generation interval.

  In addition, it can be seen that the same result as in the case where the observation sound in FIG. Thereby, it turns out that a suspicious sound generation area can be easily detected by using the sound detection device of the present embodiment even in a noisy environment.

  By the way, as already described, when calculating the score, frames before and after the target frame of the input acoustic data 20 are used. Therefore, as shown in FIG. 11, the rising position and the falling position of the waveform due to the suspicious sound in the input acoustic data 20 obtained based on the calculated score are the rising position and the falling position of the waveform in the actual input acoustic data 20. Has a delay with respect to position. Therefore, in order to cut out the suspicious sound generation section from the input acoustic data 20, it is necessary to correct this delay.

  The delay correction amount for correcting the delay depends on the sampling frequency, the frame width W, and the frame interval D of the input acoustic data 20. That is, in the above example in which 8 frames of frame # 1 to frame # 8 are used for score calculation, and frame # 5 is associated with the rising position of the waveform in the learning data, as illustrated in FIG. Therefore, it takes a time of one frame width W to calculate the signal power, and the signal power of each frame is calculated for each frame interval D. Further, calculation of the slope of the signal power time series requires 5 frames later, that is, 1 frame width W + 4 frame intervals D. Furthermore, since 4 frames are used to calculate the score, a 4-frame interval D is required. Therefore, in order to calculate the score of frame i, 1 frame width W + 7 frame interval D = 9 frame interval D is required.

  As a more specific example, in the example where the sampling frequency of the input acoustic data 20 is 16 kHz (kilohertz), the number of samples of one frame width W is 160 samples, and the number of samples is 80 samples of one frame interval D, the delay correction amount is 80 Sample × 9 = 720 samples. This delay correction amount is 720 samples × (1/16000) = 0.045 sec (45 milliseconds) in terms of time.

  The generation section detection unit 104 sets the value obtained by subtracting this delay correction amount from the time of the start position and end position of the detected suspicious sound generation section as the time of the start position and end position of the corrected suspicious sound generation section. . Then, the corrected suspicious sound generation section data is cut out from the input acoustic data 20 stored in the storage unit 13 and output as output acoustic data 21.

  FIG. 13 is a flowchart of an example showing the detection processing of the suspicious sound generation section according to the present embodiment. Each process according to this flowchart is executed by, for example, a CPU (not shown) included in the calculation unit 12 according to a program stored in the storage unit 13 in advance. The program includes, for example, modules that respectively implement a signal power calculation unit 101, an inclination calculation unit 102, a score calculation unit 103, and a generation interval detection unit 104. When the program is executed by the CPU, each program is stored in a main memory (not shown). Expand and run the module.

  Not only this but the signal power calculation part 101, the inclination calculation part 102, the score calculation part 103, and the generation | occurrence | production area detection part 104 which are contained in the calculating part 12 are each comprised by separate hardware, and each part cooperates in a flowchart. Each process may be executed.

  In FIG. 13, in step S <b> 100, an analog audio signal according to the collected observation sound is output from the microphone 10. The analog audio signal is converted into a digital audio signal by the A / D conversion unit 11 and supplied to the signal power calculation unit 101 as input acoustic data 20. The input acoustic data 20 is also supplied to and stored in the storage unit 110.

The signal power calculation unit 101 sets a frame i for the input sound data 20 that has been input, and calculates the signal power y _LP (i) of the set frame i according to the above-described equation (1) (step S101). ). The value of the calculated signal power y _LP (i) is temporarily stored in the storage unit 110, for example. In the next step S102, the slope calculation unit 102 retrieves the predetermined number of signal power values already calculated from the storage unit 110, and calculates the slope y _GLP (i) of the signal power time series according to the above-described equation (2). To do. The calculated signal power time series gradient y _GLP (i) is held in the storage unit 110.

Next, in step S103, the score calculation unit 103 extracts from the storage unit 110 the value of the already calculated slope y _GLP (i) of the signal power time series and the score parameter 111 calculated in advance based on the learning data. Then, the score S (i) of the frame i is calculated according to the above equation (3). The calculated score S (i) is held in the storage unit 110. In the next step S104, the occurrence section detection unit 104 extracts the already calculated score value from the storage unit 110, and calculates the score time-series slope GS (i) according to the above-described equation (4).

  In the next step S105, the occurrence section detection unit 104 refers to the condition (A) described above, and determines whether or not the calculated slope GS (i) of the score time series is a maximum value. If it is determined that the score S (i) is a maximum value, the process proceeds to step S106, and it is determined whether or not the score S (i) that is the maximum value exceeds a predetermined threshold value. If it is determined that it does not exceed, the process returns to step S100.

  On the other hand, if it is determined in step S106 that the score S (i) exceeds the threshold value, the process proceeds to step S107, and it is assumed that the rising of the waveform of the suspicious sound is detected at the approximate center of the frame i. Then, the process returns to step S100.

  If it is determined in step S105 described above that the score S (i) is not the maximum value, the process proceeds to step S108. In step S108, the occurrence section detection unit 104 refers to the condition (B) described above, and determines whether or not the score S (i) is a minimum value. If it is determined that the value is not the minimum value, the process returns to step S100.

  On the other hand, if it is determined in step S108 that the score S (i) is a minimum value, the process proceeds to step S109. In step S109, the occurrence section detection unit 104 determines whether or not the local minimum value is a local minimum value detected for the first time after the local maximum value is detected in step S106 described above. If it is determined that it is not the minimum value detected for the first time after the detection of the maximum value, the process returns to step S100.

  In step S108, when the generation section detection unit 104 determines that the score S (i) is the minimum value detected for the first time after the maximum value is detected in step S106, the process shifts to step S110, and the frame i It is assumed that the falling edge of the suspicious sound waveform was detected in the approximate center of In step S110 and the above-described step S107, the rising and falling edges of the suspicious sound waveform are detected.

  In the next step S111, the generation section detection unit 104 sets the sampling frequency, the frame width W, and the frame interval D of the input acoustic data 20 and the number of frames used when calculating the slope of the signal power time series for the learning data. The rise position and fall position of the waveform of the suspicious sound are corrected using the delay correction amount calculated based on the delay correction amount. The corrected falling position and falling position are set as the start position and end position of the suspicious sound generation section, and the suspicious sound generation section is detected (step S112).

  As described above, according to the present embodiment, time-series data of feature values is extracted from input acoustic data based on collected observation sounds, and calculated using the extracted feature values and learning data in advance. The score is obtained by comparing the expected value of the feature value, and the suspicious sound generation interval is detected based on the change of the score over time. Therefore, it is possible to easily detect a suspicious sound generation section even in a noisy environment.

  In the present embodiment, a threshold for detecting suspicious sound is applied to the score calculated from the feature value of the input acoustic data. This score takes a substantially constant value even in the observation sound in a noisy environment. Therefore, there is no need to change the threshold according to the environment of the monitoring area. At the same time, in the present embodiment, the score, which is a value that is not easily affected by noise, is used for detection of suspicious sound, so that robust detection against noise is possible, and the noise level and noise level change are possible. This makes it possible to stably detect the suspicious sound generation section even in an environment where the sound is generated.

  Here, it will be described that the detection method of the suspicious sound generation section according to the present embodiment is robust against a noise environment.

  In the present embodiment, the score calculated by the equation (3) has a property of taking a large value when the similarity between the learning data and the input acoustic data 20 is high. On the other hand, since the similarity between noise (noise acoustic data) and learning data is low, the score takes a substantially constant value and does not change significantly. Therefore, by setting a threshold value for the score and determining whether the score exceeds the threshold value, it is more reliable than a conventional sound detection method that sets a threshold value for signal power, for example. It is possible to detect the occurrence interval.

  That is, when acoustic data based on suspicious sound is input, the waveform of the suspicious sound generation interval in the input acoustic data has a high similarity to the waveform based on the learning data, so the score greatly changes at the beginning of the suspicious sound generation interval. . In this embodiment, this score is compared with a threshold value, and when the score exceeds the threshold value, it is determined that a suspicious sound occurrence section has been detected.

  On the other hand, when acoustic data due to noise is input, the waveform due to noise has a small similarity to the waveform due to learning data, so the change in score is extremely small. Therefore, the possibility that the score exceeds the threshold value is small, and the occurrence of erroneous detection is suppressed.

FIG. 14 shows signal power and signal power when substantially random acoustic data (acoustic data due to noise) is input when the learning data is acoustic data obtained by collecting a physical sound (suspicious sound to be detected). FIG. 15 shows a graph in which the values of the items illustrated in FIG. 14 are plotted against the frame numbers. The expected value μ _k and the variance value σ ² of the learning data are the same as the values shown in FIG.

  In the graph illustrated in FIG. 15, delay correction associated with score calculation described later is not performed, and the score plot is delayed by three frames with respect to the signal power plot.

  As illustrated in FIG. 15, the signal power of acoustic data due to noise changes relatively large. Therefore, when the threshold value 200 is set for the signal power as in the prior art, frame numbers # 3, # 16, # 18, and # 19 are erroneously detected. On the other hand, when acoustic data whose waveform is significantly different from the learning data is input, the score does not change greatly. Therefore, when a threshold value is set for a score according to the present embodiment, the possibility that the score exceeds the threshold value is low, and a specific sound generation interval can be accurately detected even in a noisy environment.

  In the present embodiment, the suspicious sound generation section is detected by extracting the characteristics of the suspicious sound. For this reason, not only voice but also various kinds of sound can be detected in the suspicious sound generation section.

  Furthermore, by applying the present embodiment, it is possible to accurately detect a suspicious sound generation interval. Thereby, the improvement in the precision in the sound recognition process with respect to the acoustic data of the detected suspicious sound generation area can be expected. In addition, by detecting the suspicious sound generation section in advance for the sound recognition process, the system may perform the recognition process only on the acoustic data of the detected suspicious sound generation section, and the entire sound recognition system It is possible to reduce the calculation cost and the communication cost for performing acoustic data communication.

  The sound detection device according to the present embodiment is provided in a security device that outputs an alarm when a suspicious person is detected in the monitoring area, or the output from the sound detection device according to the present embodiment is input to the security device. can do. Thereby, since the suspicious sound in a monitoring area | region can be detected easily and with high precision, it becomes possible to prevent the false alarm by a security device.

DESCRIPTION OF SYMBOLS 10 Microphone 11 A / D conversion part 12 Calculation part 13 Storage part 20 Input acoustic data 21 Output acoustic data 100 Sound detection apparatus 101 Signal power calculation part 102 Inclination calculation part 103 Score calculation part 104 Generation | occurrence | production area detection part 111 Score parameter

Claims (8)

A feature value calculating means for calculating a feature value indicating a feature of the sound data along a time series;
The sound data is evaluated based on the difference between the expected value of the characteristic value obtained in advance from the learning data along the time series and the characteristic value of the signal power time series of the sound data calculated by the characteristic value calculating means. Score calculating means for calculating a score to be performed;
And detecting means for detecting a position of a maximum value and a position of a minimum value of the score, and detecting a specific sound generation section in the sound data based on the position of the maximum value and the position of the minimum value. Sound detection device. The detection means includes
When the maximum value exceeds a threshold, the position of the maximum value is determined as the start position of the specific sound generation section,
The sound detection device according to claim 1, wherein the position of the minimum value that first appears after the start position is determined as the end position of the specific sound generation section. The sound detection device according to claim 1, wherein the expected value of the feature value is obtained by using a plurality of sound data of the same type and different sounds as the learning data. The expected value of the characteristic value of the signal power time series is obtained for each of a plurality of predetermined ranges arranged in a time series order with some overlapping learning data,
The feature value calculating means includes
The feature value is calculated for a predetermined range arranged in chronological order with a part of the sound data overlapping,
The score calculation means includes
The score is calculated based on a sum of differences between feature values of a plurality of predetermined ranges of the sound data and the expected values of the plurality of predetermined ranges obtained for the learning data. The sound detection device according to any one of claims 3 to 4. The score calculation means includes
The sound detection apparatus according to claim 4, wherein the score is calculated by normalizing the difference using a variance value of characteristic values of signal power time series obtained in advance from learning data. The sound detection apparatus according to claim 1, wherein the feature value is a slope of a signal power time series of sound data. The delay correction means for correcting a delay generated when the score calculation means calculates the score for the specific sound generation section detected by the detection means. Item 7. The sound detection device according to any one of items 6 to 6. A feature value calculating means for calculating a feature value indicating the feature of the sound data along a time series;
The score calculation means is based on the difference between the expected value of the feature value obtained in advance along the time series from the learning data and the feature value of the signal power time series of the sound data calculated in the feature value calculating step. A score calculating step for calculating a score for evaluating the sound data;
Detecting a position of a maximum value and a position of a minimum value of the score, and detecting a specific sound generation section in the sound data based on the position of the maximum value and the position of the minimum value, Sound detection method.

Images