JP3474131B2 - High-speed signal search method and apparatus and recording medium thereof - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a high-speed signal search method, apparatus, and program for causing a computer to execute a high-speed signal search method for searching a bank of signals similar to a signal registered in advance from a signal sequence. The present invention relates to a recorded computer-readable recording medium, which can be used for acoustic signal detection, for example. That is, a signal detection that can detect the time when a specific commercial is broadcast from the broadcast audio signal and automatically record it, or detect a specific theme song and start or stop video recording. Related to technology.

Further, the present invention relates to a technique that enables automatically monitoring the time when a clapping sound is emitted from a broadcast, the time when a laughing voice is emitted, or the like, and it is possible to search for a specific scene. Is also relevant. Furthermore, the present invention can be applied to detection of general signals such as video signals as well as audio signals.

[0003]

2. Description of the Related Art Conventionally, there is a matched filter as a method of detecting a region in which a target signal (hereinafter referred to as a reference signal) exists from a signal sequence. The matched filter calculates the correlation value between the waveform of the input signal in the region of interest (hereinafter referred to as the window of interest) and the waveform of the reference signal while moving the window of interest, and the correlation value is calculated. This is a method of determining that "the reference signal exists within the window of interest" when the value exceeds a certain value.

[0004]

However, in this method, it is necessary to calculate the correlation value while finely moving the position of the window of interest with respect to the entire area of the input signal.
There is a problem that the amount of calculation becomes huge and the calculation speed becomes slow. In order to solve this problem, we have already searched for a location of an acoustic signal similar to the acoustic signal registered in advance in “High-speed signal search method, device and its recording medium” (Japanese Patent Application No. 11-130630). We proposed a signal detection method and a method with improved detection accuracy.

However, these methods have a problem in that the collation is always performed assuming the worst change in the similarity, so that the collation may be performed unnecessarily for the actual signal. The present invention has been made in view of such circumstances, and a high-speed signal search method, apparatus, and recording medium therefor capable of performing signal search at higher speed by performing signal detection processing with a small number of collation calculations. The purpose is to provide. In particular, the above-mentioned “high-speed signal search method, device and recording medium thereof” (Japanese Patent Application No. 11-130630).
No.) in the skip width calculation process, the window of interest is skipped with a larger width than the embodiment in this prior application, and a backtrack necessity determination process is newly provided to detect the missed detection. The main point of the invention is that the suppression enables the higher-speed signal detection while ensuring the same accuracy.

[0006]

In order to achieve the above-mentioned object, the invention according to claim 1 of the present invention is
In a high-speed signal search method for finding a portion similar to a reference signal registered in advance, a reference feature amount calculation process for deriving a feature amount sequence from the reference signal, and a feature amount sequence from the input signal, for the feature amount sequence The similarity between the input feature quantity calculation process of setting the attention window, the feature quantity series derived in the reference feature quantity calculation process, and the feature quantity series in the attention window set in the input feature quantity calculation process is calculated. Based on the similarity calculation process and the similarity calculated in the similarity calculation process, the skip width of the attention window is calculated, and the skip width calculation process of moving the attention window by the skip width, and the skip width. At the position of the window of interest moved in the calculation process, based on the similarity obtained in the similarity calculation process,
A backtracking necessity determining step of determining whether or not the moved window of interest needs to be returned in the opposite direction of time, and the similarity calculating step, the skip width calculating step, and the backtracking necessity determining step are performed. By repeating the process, for a plurality of locations of the input signal, calculating the degree of similarity with the reference signal and comparing the degree of similarity and a target similarity set in advance, the reference signal is It is characterized in that it is determined whether or not the input signal exists at the location.

According to a second aspect of the present invention, in the high-speed signal search method according to the first aspect, the skip width calculation process predicts future similarity based on a probability density distribution of similarity. The skip width is determined based on the prediction.

According to a third aspect of the present invention, in the high-speed signal search method according to the first or second aspect, the reference feature amount calculation process and the input feature amount calculation process are performed on the obtained feature amount series. On the other hand, a histogram is created, and the similarity is calculated based on the histogram in the similarity calculation process.

The invention according to claim 4 is the high-speed signal search method according to any one of claims 1 to 3, wherein the window of interest has the same time length as that of the reference signal. .

The invention according to claim 5 is a high-speed signal search device for searching a portion similar to a reference signal registered in advance from an input signal, and a reference feature amount calculating means for deriving a feature amount sequence from the reference signal. An input feature quantity calculating means for deriving a feature quantity series from the input signal and setting a window of interest for the feature quantity series; a feature quantity series derived by the reference feature quantity calculating means; and the input feature quantity. Similarity calculation means for calculating the similarity with the feature amount series in the window of interest set by the calculation means, based on the similarity calculated by the similarity calculation means, to calculate the skip width of the window of interest, A skip width calculating means for moving the window of interest by the skip width,
At the position of the window of interest moved by the skip width calculation means, it is determined whether or not the moved window of interest needs to be returned in the opposite direction of time based on the similarity calculated by the similarity calculation means. And a similarity calculation means for calculating the similarity with the reference signal at a plurality of points of the input signal, and the similarity and the target similarity set in advance. It is characterized in that whether or not the reference signal is present at the relevant portion of the input signal is determined by comparing with the frequency.

According to a sixth aspect of the present invention, in the high-speed signal search device according to the fifth aspect, the reference feature quantity calculating means and the input feature quantity calculating means respectively operate on the obtained feature quantity series. It is characterized in that a histogram is created, and the similarity calculation means calculates the similarity based on the histogram.

According to a seventh aspect of the present invention, in the high-speed signal search device according to the fifth or sixth aspect,
The window of interest has the same time length as the reference signal.

The invention according to claim 8 is a high-speed signal search method for searching for a portion similar to a reference signal registered in advance from an input signal, and a reference feature quantity calculation process for deriving a feature quantity sequence from the reference signal, An input feature quantity calculation process of deriving a feature quantity sequence from the input signal and setting a window of interest for the feature quantity sequence, a feature quantity sequence derived in the reference feature quantity calculation process, and the input feature quantity calculation process In the similarity calculation process of calculating the similarity with the feature amount series in the window of interest set in step 1, based on the similarity calculated in the similarity calculation process, the skip width of the window of interest is calculated, and the skip A skip width calculation process of moving the window of interest by the width,
At the position of the window of interest moved in the skip width calculation process, based on the similarity obtained in the similarity calculation process,
A backtracking necessity determining step of determining whether or not the moved attention window needs to be returned in the opposite direction of time, and processing by the similarity calculating step, skip width calculating step, and backtracking necessity determining step By repeating the calculation of the similarity with the reference signal for a plurality of locations of the input signal and comparing the similarity with a preset target similarity. A gist of a computer-readable recording medium recording a program for causing a computer to execute a high-speed signal search method, which determines whether or not the signal is present at the location.

According to the first to fourth aspects of the invention, in the high-speed signal search method for searching the input signal for a part similar to the reference signal registered in advance, the window of interest set in the input signal at the time of signal search is moved. In this case, a backtrack necessity judgment process is newly provided to judge whether or not it is necessary to move the window of interest significantly in the skip width calculation process and return the moved window of interest to the opposite direction of time. Since the omission of detection is suppressed, the signal detection process can be performed with a smaller number of collation calculations, and the signal can be searched at a higher speed.

According to the invention described in claims 5 to 7, in the high-speed signal search device for searching the input signal for a portion similar to the reference signal registered in advance, the window of interest set in the input signal at the time of signal search is used. When moving, the skip width calculation means is used to significantly move the window of interest and a backtrack necessity determination means is newly provided to determine whether or not the moved window of interest needs to be returned in the opposite direction of time. Since the omission of detection is suppressed, the signal detection processing can be performed with a smaller number of collation calculations, and the signal can be searched at a higher speed.

Further, according to the invention of claim 8, in a high-speed signal search method for searching a portion similar to a reference signal registered in advance from an input signal, a reference feature quantity calculation process for deriving a feature quantity sequence from the reference signal. An input feature quantity calculation process for deriving a feature quantity sequence from the input signal and setting a window of interest for the feature quantity series; a feature quantity sequence derived in the reference feature quantity calculation process; and the input feature quantity. Based on the similarity calculation process of calculating the similarity with the feature amount series in the window of interest set in the calculation process and the similarity calculated in the similarity calculation process, the skip width of the window of interest is calculated, At the skip width calculation step of moving the window of interest by the skip width, and at the window of interest moved in the skip width calculation step, the movement is performed based on the similarity obtained in the similarity calculation step. A backtracking necessity determining process for determining whether or not the attention window needs to be returned in the opposite direction of time, and repeating the processes by the similarity calculating process, the skip width calculating process, and the backtracking necessity determining process. , Calculating the degree of similarity with the reference signal for a plurality of points of the input signal, and comparing the degree of similarity with a preset target degree of similarity, the reference signal is Since a program for causing a computer to execute a high-speed signal search method characterized by determining whether or not it exists at a location is recorded on a computer-readable recording medium, the program is read into a computer system. By doing so, the signal detection processing can be performed with a smaller number of collation calculations than before, and the signal can be processed faster. It is possible to search for.

[0017]

DESCRIPTION OF THE PREFERRED EMBODIMENTS Next, embodiments of the present invention will be described in detail with reference to the drawings. A high-speed signal search device according to an embodiment of the present invention is a high-speed signal search method for searching a portion similar to a reference signal registered in advance from an input signal, and calculating a reference feature quantity from the reference signal. A process, an input feature amount calculation process of deriving a feature amount sequence from the input signal, and setting a window of interest for the feature amount sequence, and a feature amount sequence derived in the reference feature amount calculation process,
Based on the similarity calculation process of calculating the similarity with the feature amount series in the window of interest set in the input feature amount calculation process and the similarity calculated in the similarity calculation process, the skip width of the window of interest And a skip width calculation step of moving the target window by the skip width, and a target window position moved in the skip width calculation step, based on the similarity obtained in the similarity calculation step, A backtracking necessity determining step of determining whether or not the moved window of interest needs to be returned in the opposite direction of time, and processing by the similarity calculating step, the skip width calculating step, and the backtracking necessity determining step. By repeating the above, for a plurality of locations of the input signal, the similarity with the reference signal is calculated, and the similarity is compared with a target similarity set in advance to obtain the reference. No. is a device for implementing the high-speed signal search method for determining whether present in the portion of the input signal.

FIG. 1 shows the configuration of a high speed signal searching apparatus according to an embodiment of the present invention. The high-speed signal search device according to the embodiment of the present invention targets an acoustic signal as an input signal. In FIG. 1, the high-speed signal search device 1 according to the present embodiment requires a reference feature quantity calculation means 10, an input feature quantity calculation means 12, a similarity degree calculation means 14, a skip width calculation means 16, and a backtrack. And a sex determining means 18.

The high-speed signal search device shown in FIG. 1 is specifically composed of a computer system including a CPU (central processing unit) and its peripheral circuits. The computer system functions as each unit shown in FIG. 1 by being controlled by a control program recorded in a predetermined recording medium (magnetic disk, semiconductor memory, etc.). The control program can be distributed via a communication line.

High-speed signal search device 1 according to the present embodiment
Is a reference signal input from the input terminal 100, that is, an acoustic signal to be searched as a sample and an input signal input from the input terminal 102, that is, an acoustic signal to be searched, and the similarity with the reference signal is preset. Data indicating a portion in the input signal that is equal to or greater than the set value (referred to as a search threshold) θ is output from the output terminal 104.

The reference feature amount calculation means 10 derives a feature amount sequence from the reference signal. Further, the input feature quantity calculation means 12 derives a feature quantity series from the input signal and sets a window of interest for the feature quantity series. The similarity calculation means 14 calculates the similarity between the feature quantity series derived in the reference feature quantity calculation process and the feature quantity series in the window of interest set in the input feature quantity calculation process.

The skip width calculating means 16 calculates the skip width of the window of interest based on the similarity calculated by the similarity calculating means 14 and moves the window of interest by the skip width. The backtrack necessity determining unit 18 determines, based on the similarity output by the similarity calculating unit 14 at the position of the window of interest moved by the skip width calculating unit 16,
Determine if the window of interest needs to be moved backwards in time.

Next, the operation of the high-speed signal search apparatus according to the embodiment of the present invention shown in FIG. 1 will be specifically described with reference to FIGS. 2 and 3. FIG. 2 is a flow chart showing the operation of the high speed signal search device according to the present embodiment.
FIG. 4 is an explanatory diagram showing the principle of overskipping of the high-speed signal search device according to the present embodiment. In the following,
For simplification of explanation, a case where the window of interest has the same time length as that of the reference signal will be described. "High-speed signal search method, device and recording medium thereof" (Japanese Patent Application No. 11-1306)
The same can be applied to the case of temporally dividing the window of interest as described in No. 30).

In the reference feature quantity calculating means 10, first,
Read the given reference signal. Next, feature extraction is performed on the read reference signal (step 201). In this embodiment, since the spectral feature is used as the feature, the feature extraction can be performed by a bandpass filter, for example. A specific example of feature amount extraction will be described below.
For example, when it is desired to search for a sound signal of about 15 seconds from a broadcast signal of a television, a radio or the like, a good result can be obtained by extracting the feature amount with the device shown in FIG. That is, seven band pass filters are used, their center frequencies are set at equal intervals on the logarithmic axis, and an analysis window having a time length of about 60 milliseconds is set for the output waveform of each band pass filter. The average value of the square of the output waveform within the analysis window is calculated. The seven average values thus obtained are combined into a seven-dimensional feature vector. The feature vector is sequentially calculated while moving the analysis window by 10 milliseconds.

Next, a histogram of the feature vector is created from the time series of the feature vector (step 202).
Each of these histograms is created by dividing a range of values (minimum value to maximum value) that each element of the feature vector can take into a plurality of bins (sections). For example, if the range of values that each element can take is divided into three sections (assuming sections A, B, and C) and the number of elements in each feature vector is 7, each of the 7 elements is Belongs to one of the sections A, B, and C.

Therefore, considering one feature vector, as a combination of the elements of the feature vector,
Indicates that all seven elements belong to section A (A, A,
From (A, A, A, A, A) to (C, C, C, C, C, C, C), which indicates that all seven elements belong to section C,
A total of 3 to 7 different combinations are possible. From the above, the total number of bins in the histogram of the feature vector, that is, the number of sections arranged on the horizontal axis of the histogram is 3 to the 7th power. Therefore, when the horizontal axis of the histogram is set in this manner, each feature vector is classified into any one of the 3 to 7 7 sections.

In the input feature quantity calculation means 12, first,
An acoustic signal which is an input signal is read from the input terminal 102. Next, feature extraction is performed on the read input acoustic signal (step 203). The feature extraction is performed by the same operation as that performed by the reference feature amount calculation means 10. Then, a window of interest is set for the series of feature vectors of the input signal (step 204). First, a window of interest having the same time length as the reference signal given to the reference feature amount calculation means 10 is set. At the beginning of detection, the window of interest is set at the beginning of the input signal on the time axis, but in the process of detection, as will be described later, the window of interest is sequentially shifted in the time axis direction for detection. The amount of shift in the time axis direction is given by the skip width calculation means 16.

Next, a histogram of feature vectors is created from the time series of feature vectors in the window of interest (step 205). The histogram is created by dividing the value of each element of the feature vector into several bins (intervals). The classification is performed by the same method as that used by the reference feature amount calculation means 10. The similarity calculation means 14 first reads the histogram of the feature vector output from the reference feature quantity calculation means 10 and the input feature quantity calculation means 12.

The histogram output from the reference feature amount calculation means 10 is HR, and the histogram output from the input feature amount calculation means 2 is HI. However, R and I respectively indicate that the histogram is created from the feature vector of the reference signal and the feature vector of the input signal. Subsequently, the similarity calculation means 14 calculates the histogram of the reference signal,
The similarity SRI with the histogram of the input signal is calculated (step 206). The similarity SRI between the reference signal histogram HR and the input signal histogram HI is defined as follows.

[Equation 1] Here, L is the total number of bins in the histogram (3 to the 7th power in the above example), D is the total frequency of the histogram (total number of feature vectors derived from one reference signal), hRL, h
Il, represents the number (frequency) of the feature vectors included in the l-th bin of HR and HI, respectively.

The similarity calculating means 14 outputs the value of the similarity SRI to the skip width calculating means 16. In addition to this, if the value of the similarity SRI exceeds a preset value θ (when the window of interest is divided in the time axis direction, the value of the similarity is previously set in all time divisions). If it is found that the value exceeds the set value (step 207), it means that the reference signal was present in the input signal. Therefore, the similarity calculation means 14 determines that the reference signal The number and the current position of the window of interest for the input signal are output from the output terminal 104 (step 21).
2).

On the other hand, in the skip width calculating means 16, first, the similarity SR output from the similarity calculating means 14 is outputted.
I is read, and the skippable width w is calculated based on this similarity SRI (step 208). According to the signal detection method already proposed by us, the skippable width w is obtained by the following equation.

[Equation 2] However, the unit of w is the number of feature vectors, and
r (•) represents rounding down. Expression (2) is currently SRI
If <θ, it means that the SRI never exceeds θ even if the attention window is shifted by w−1. This is because when the window of interest is shifted, all the feature vectors that go out of the window of interest do not contribute to the overlapping of the histograms,
It can be easily understood by considering the case where all the feature vectors that come into the window of interest contribute to the overlap of the histograms (when the similarity reaches the set value earliest).
That is, the upper limit of SRI when the window of interest is shifted by w−1 is θ.

On the other hand, if SRI ≧ θ, w = 1 (full search) to find a local peak of similarity. On the other hand, the main object of the present invention is to reduce the number of times of matching that gives a degree of similarity equal to or less than the search threshold θ (that is, to reduce useless matching calculation), as compared with the method already proposed by us. is there. The principle of reducing the number of collations will be described below. In FIG. 3, it is assumed that the input signal is compared with the reference signal at time t0 to obtain the similarity S0. From the equation (1), the absolute value of the rate of change of the similarity does not exceed 1 / D, so the upper limit of the similarity is the slope 1 / D shown in FIG.
Is represented by a straight line M. Therefore, the time tθ = t0 + w at the intersection of the straight line M and the straight line S = θ is the skip position in the signal detection method already proposed by us.

On the other hand, in the overskipping performed to move the time window in the high speed signal searching apparatus according to this embodiment, the skipping is performed until time tp (> tθ). A straight line N having a slope of −1 / D passing through the point (tθ, θ) has the highest degree of similarity and reaches the search threshold from time t0.
It represents the lower limit of the degree of similarity when the degree of similarity decreases most suddenly thereafter. If the similarity S _{REAL at} time tp is less than or equal to the straight line N, it can be seen that the similarity never exceeded the search threshold θ between time t0 and time tp. Therefore, even if the search is continued as it is, no omission occurs in the search.

On the other hand, if the similarity S _REAL is above the straight line N, there is a possibility that the similarity exceeds the search threshold value θ between the time t0 and the time tp. Backtrack (ie, go backwards through the window of interest) to see if it has been exceeded. When backtracking occurs frequently, the number of times of matching increases rather than the method already proposed by us, so it is important to properly set the overskipping position.

The overskip rate α is

[Equation 3] And the backtrack risk rate is p, the expected value R of the matching frequency ratio (ratio of the number of matching times when overskipping is performed to that when overskipping is not performed, provided that the similarity is less than or equal to a threshold value) is

[Equation 4] Given in. Here, u is the average number of collations required per backtrack.

From these equations,

[Equation 5] and

[Equation 6] Is derived (S 0 <θ, S _P <θ). However, S _P is the value of the straight line N at tp.

Since R ^-1 represents the effect of reducing the number of collations, it should be maximized. For example, the normal distribution of similarity N
Assuming that (m = 0.3, σ ² = 0.05 ² ), S _P = m + c
If σ (c is a constant) and u = 1, θ = 0.8, So = m,

[Equation 7] The maximum value is about 1.76 near c = 2 (matching frequency ratio 0.57). That is, it can be expected that the number of useless collations is reduced by 40% or more. At this time, exactly the same precision as the known method is guaranteed.

In a section where the degree of similarity is relatively small (steady section), the similarity distribution can be estimated from the accumulation result of the degrees of similarity. In a section where the degree of similarity is relatively large (similar section), the skippable width is small, so there is little replacement of feature vectors within the time window due to skipping, and this must be taken into account in the estimation of the similarity distribution. For example, a method of linearly predicting a change in similarity can be considered. Based on the above principle, the skip width calculating means 16 calculates the skip width of the window of interest as αw (step 208), and the skip width α.
The data of w is output to the input feature amount calculation means 12.

When the backtrack necessity determining means 18 determines that there is no need for backtracking (step 209), the input feature quantity calculating means 12 shifts the target window by αw to determine the input signal target window. The position is determined (step 211), the similarity calculation means 14 determines the presence or absence of the reference signal (step 212), the process returns to step 205, and the processing already described is repeated. Further, when the target window is shifted by αw and the target window exceeds the end of the input signal, it means that all the input signals have been searched, and the signal detection processing is terminated.

In the backtrack necessity determining means 18,
First, the similarity SRI output from the skip width calculation means 16 is read. The similarity SRI is S in FIG.
Equivalent to _REAL . The backtrack necessity determining means 18 first receives the skip width output from the skip width calculating means 16 (the one output in the previous collation).
As a result, the data at time tp in FIG. 3 is obtained. Next, the backtrack necessity determination means 18 calculates the value of the straight line N (FIG. 3) at the time tp.

In FIG. 3, the backtrack necessity determining means 18 has no possibility that the similarity exceeds the search threshold value θ from time t0 to time tp if S _REAL is equal to or less than the straight line N. It is determined that backtracking (reversing the window of interest) is not necessary (step 209). S _REAL
Is above the straight line N, the similarity may have exceeded the search threshold θ between the time to and the time tp, so it is determined that backtracking is necessary (step 20).
9).

In this case, the backtrack necessity determining means 18 calculates the backtrack width and sends the backtrack width to the input feature quantity calculating means 12 via the skip width calculating means 16 (step 210). The optimum backtrack width is the time at which the most time interval (this is called an unevaluated interval) in which the similarity value may have exceeded the search threshold θ can be excluded by performing matching at that backtrack position. is there. This time can also be estimated by predicting the similarity change, but since the unevaluated section is usually short, extra backtrack matching can be performed even by a simple method such as setting the backtrack position at the center of the unevaluated section. It is considered that it is unlikely to be necessary (u = 1 in most cases).

If the unevaluated section disappears by the collation at the backtrack position, the backtrack is terminated and the forward search is restarted. If there is an unevaluated section, it is determined that there is a need for backtracking until there is no unevaluated section, and the same collation is repeated for each unevaluated section. In a series of backtrack matching, if the similarity exceeds the search threshold, a full search (no skip) is performed before and after the search threshold, as in the conventional method.

Next, an operation experiment example of the high-speed signal search apparatus according to this embodiment will be shown. In order to confirm the effect of the present invention,
6 hours of sound signal as input signal, randomly selected 8
A search is performed for each reference signal (15 seconds), and what is the ratio of the number of times the similarity is equal to or less than the search threshold as compared with the case where the present invention is not applied (in the present specification). This is referred to as the matching frequency ratio). The search parameters are: sampling frequency = 11.025 kHz, number of bandpass filters = 7, analysis window length of frequency analysis = 60 ms, analysis window shift = 10 ms, number of bins in each feature dimension =
3, no time window division, search threshold θ = 0.8.

In the present experiment, the similarity of less than 0.5 was regarded as the steady section, and the others were regarded as the similar section. Sp in the steady section
= M + 2σ, α was set for each collation from equation (5). In the similar section, α is set using the extrapolation line of the similarity and the estimated value of the variance. As a result of this experiment, the average number of times of collation was 2503 times in the conventional method and 1603 times in the proposed method. That is, the matching frequency ratio was 0.64. This is shown in FIG. Further, FIG. 6 shows a search result for an input signal for 3 minutes. The vertical line in FIG. 6 indicates the time when the matching is actually performed. With the high-speed signal search device according to the embodiment of the present invention, it can be confirmed that the number of times of matching has decreased.

As described above, according to the high-speed signal search device according to the embodiment of the present invention, the window of interest is moved to a greater extent than in the prior art, and means for preventing the omission of detection (backtrack necessity). By providing a determination means), a search can be performed while eliminating unnecessary collation calculation, and signal detection processing can be performed with a smaller number of collation calculation times than in the conventional case, thereby enabling faster signal detection. There is an effect.

In the high-speed signal search method for searching the input signal for a portion similar to the reference signal registered in advance,
A reference feature quantity calculation step of deriving a feature quantity sequence from the reference signal; an input feature quantity calculation step of deriving a feature quantity series from the input signal and setting a window of interest for the feature quantity series; and the reference feature quantity calculation. A similarity calculation process for calculating the similarity between the feature series derived in the process and the feature series in the window of interest set in the input feature calculation process, and the similarity calculated in the similarity calculation process Based on, the skip width of the attention window is calculated, and the skip width calculation step of moving the attention window by the skip width, and the attention window position moved in the skip width calculation step, in the similarity calculation step. And a backtracking necessity determining step of determining whether or not the moved window of interest needs to be returned in the opposite direction of time based on the obtained degree of similarity, the similarity calculating step, the skip width calculating step And a backtracking necessity determining process are repeated to calculate the similarity with the reference signal at a plurality of points of the input signal, and the similarity is compared with a preset target similarity. Due to
The reference signal is recorded on a computer-readable recording medium for recording a program for causing a computer to execute a high-speed signal search method, which determines whether or not the reference signal is present at the location of the input signal. You may make it implement | achieve the function of a high-speed signal search device by making a computer system read and execute the program recorded on the recording medium.

In this case, since the program for realizing the function of the high-speed signal search device is recorded in the computer-readable recording medium, the program recorded in this recording medium is read into the computer system and executed. The signal detection process can be performed with a smaller number of collation calculations than in the past, and the signal can be searched at a higher speed.

The "computer system" referred to here includes an OS and hardware such as peripheral devices. The "computer-readable recording medium" means a floppy disk, a magneto-optical disk, an RO.
It means a storage medium such as M and CD-ROM that is distributed as a portable medium, and a storage device such as a hard disk built in a computer system.

[0050]

According to the invention described in claims 1 to 4,
In the high-speed signal search method that searches for a portion similar to the reference signal registered in advance from the input signal, when moving the window of interest set for the input signal during signal search, the window of interest is moved significantly during the skip width calculation process. , A backtrack necessity judgment process is newly provided to judge whether or not the moved window of interest needs to be returned in the opposite direction of time, so that omission of signal detection is suppressed. Signal detection processing can be performed, and signals can be searched at higher speed.

According to the invention described in claims 5 to 7, in the high-speed signal search device for searching the input signal for a portion similar to the reference signal registered in advance, the window of interest set in the input signal at the time of signal search is set. When moving, the skip width calculation means is used to significantly move the window of interest and a backtrack necessity determination means is newly provided to determine whether or not the moved window of interest needs to be returned in the opposite direction of time. Since the omission of detection is suppressed, the signal detection processing can be performed with a smaller number of collation calculations, and the signal can be searched at a higher speed.

Further, according to the invention of claim 8, in a high-speed signal search method for searching a portion similar to a reference signal registered in advance from an input signal, a reference feature quantity calculation process for deriving a feature quantity sequence from the reference signal. An input feature quantity calculation process for deriving a feature quantity sequence from the input signal and setting a window of interest for the feature quantity series; a feature quantity sequence derived in the reference feature quantity calculation process; and the input feature quantity. Based on the similarity calculation process of calculating the similarity with the feature amount series in the window of interest set in the calculation process and the similarity calculated in the similarity calculation process, the skip width of the window of interest is calculated, At the skip width calculation step of moving the window of interest by the skip width, and at the window of interest moved in the skip width calculation step, the movement is performed based on the similarity obtained in the similarity calculation step. A backtracking necessity determining process for determining whether or not the attention window needs to be returned in the opposite direction of time, and repeating the processes by the similarity calculating process, the skip width calculating process, and the backtracking necessity determining process. , Calculating the degree of similarity with the reference signal for a plurality of points of the input signal, and comparing the degree of similarity with a preset target degree of similarity, the reference signal is Since a program for causing a computer to execute a high-speed signal search method characterized by determining whether or not it exists at a location is recorded on a computer-readable recording medium, the program is read into a computer system. By doing so, the signal detection processing can be performed with a smaller number of collation calculations than before, and the signal can be processed faster. It is possible to search for.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a high-speed signal search device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing the operation of the high-speed signal search device according to the embodiment of the present invention shown in FIG.

FIG. 3 is an explanatory diagram showing the principle of overskipping in the high-speed signal search device according to the embodiment of the present invention.

FIG. 4 is a block diagram showing a configuration example for extracting spectral features in the high-speed signal search device according to the embodiment of the present invention.

FIG. 5 is an explanatory diagram showing the measurement result of the number of times of matching at the time of signal search by the high-speed signal search device according to the embodiment of the present invention, in comparison with the conventional example.

FIG. 6 is an explanatory diagram showing an experimental result of performing a signal search targeting an acoustic signal by the high-speed signal searching device according to the embodiment of the present invention, in comparison with a conventional example.

[Explanation of symbols]

1 High-speed signal search device 10 Reference feature amount calculation means 12 Input feature quantity calculation means 14 Similarity calculation means 16 Skip width calculation means 18 Backtrack necessity judgment means 100, 102 input terminals 104 output terminals

─────────────────────────────────────────────────── ─── Continuation of front page (51) Int.Cl. ⁷ Identification code FI G10L 3/00 561H 571A (56) References JP 2000-312343 (JP, A) JP 4-198999 (JP, A) Kunio Kashiwano, Hiroshi Murase, High-speed search of multiple acoustic signals-Reducing the number of collations in OR search of reference signals-, Proceedings of the Acoustical Society of Japan, March 10, 1999, p. 447-448 Kunio Kashino, Hiroshi Murase, Acceleration of time-series active search method by overskipping, Proceedings of Acoustical Society of Japan, September 29, 1999, p. 445-446 (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 15/10 H04N 15/00 JISST file (JOIS)

Claims (8)

(57) [Claims] 1. A high-speed signal search method for finding a portion similar to a reference signal registered in advance from an input signal, wherein a reference feature amount calculation step of deriving a feature amount sequence from the reference signal, and a feature amount sequence from the input signal. An input feature quantity calculation process of deriving and setting a window of interest for the feature quantity sequence, a feature quantity sequence derived in the reference feature quantity calculation process, and a feature in the attention window set in the input feature quantity calculation process Based on the similarity calculation process of calculating the similarity with the quantity series and the similarity calculated in the similarity calculation process, the skip width of the window of interest is calculated, and the window of interest is moved by the skip width. Skip width calculation process, at the window position of interest moved in the skip width calculation process, based on the similarity obtained in the similarity calculation process,
A backtracking necessity determining step of determining whether or not the moved window of interest needs to be returned in the opposite direction of time, and the similarity calculating step, the skip width calculating step, and the backtracking necessity determining step. By repeating the process, for a plurality of locations of the input signal, calculating the degree of similarity with the reference signal and comparing the degree of similarity and a target similarity set in advance, the reference signal is A high-speed signal search method characterized by determining whether or not the input signal exists at the location. 2. The skip width calculation process predicts a future similarity based on a probability density distribution of similarities, and determines a skip width based on the prediction. High-speed signal search method. 3. The reference feature quantity calculation process and the input feature quantity calculation process create a histogram for each of the obtained feature quantity series, and calculate the similarity based on the histogram in the similarity calculation process. The high-speed signal search method according to claim 1 or 2, characterized in that. 4. The high-speed signal search method according to claim 1, wherein the window of interest has the same time length as that of the reference signal. 5. A high-speed signal search apparatus for searching a portion similar to a reference signal registered in advance from an input signal, wherein a reference characteristic amount calculation means for deriving a characteristic amount sequence from the reference signal, and a characteristic amount from the input signal. An input feature quantity calculating means for deriving a sequence and setting a window of interest for the feature quantity series; a feature quantity series derived by the reference feature quantity calculating means; and a window of interest set by the input feature quantity calculating means The similarity calculation means for calculating the similarity with the feature amount series, based on the similarity calculated by the similarity calculation means, the skip width of the attention window is calculated, and only the skip width is applied to the attention window. At the skip width calculating means for moving, and the position of the window of interest moved by the skip width calculating means, the moved window of interest is moved in the opposite direction of time based on the similarity calculated by the similarity calculating means. Backtracking necessity determining means for determining whether or not there is a need to restore, and the similarity calculating means calculates the similarity with the reference signal for a plurality of points of the input signal, By comparing the degree with a target similarity set in advance,
A high-speed signal search device, characterized in that it determines whether or not the reference signal is present at the location of the input signal. 6. The reference feature quantity calculating means and the input feature quantity calculating means create a histogram for each of the obtained feature quantity series, and the similarity calculating means calculates the similarity based on the histogram. The high-speed signal search device according to claim 5, wherein 7. The high-speed signal search device according to claim 5, wherein the window of interest has the same time length as that of the reference signal. 8. A high-speed signal search method for finding a portion similar to a reference signal registered in advance from an input signal, wherein a reference feature amount calculation step of deriving a feature amount sequence from the reference signal, and a feature amount sequence from the input signal An input feature quantity calculation process of deriving and setting a window of interest for the feature quantity sequence, a feature quantity sequence derived in the reference feature quantity calculation process, and a feature in the attention window set in the input feature quantity calculation process Based on the similarity calculation process of calculating the similarity with the quantity series and the similarity calculated in the similarity calculation process, the skip width of the window of interest is calculated, and the window of interest is moved by the skip width. Skip width calculation process, at the window position of interest moved in the skip width calculation process, based on the similarity obtained in the similarity calculation process,
A backtracking necessity determining step of determining whether or not the moved window of interest needs to be returned in the opposite direction of time, and the similarity calculating step, the skip width calculating step, and the backtracking necessity determining step. By repeating the process, for a plurality of locations of the input signal, calculating the degree of similarity with the reference signal and comparing the degree of similarity and a target similarity set in advance, the reference signal is A computer-readable recording medium having recorded therein a program for causing a computer to execute a high-speed signal search method characterized by determining whether or not the input signal is present at the location.