JP3537727B2 - Signal detection method, signal search method and recognition method, and recording medium - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a signal detection method for searching for a signal similar to a specific signal from a signal sequence, and is widely applied to general signal detection such as sound signal detection and video signal detection. The present invention relates to a signal detection technology that can be used for music and video use management on the Internet. The present invention also relates to a signal searching method and a signal recognizing method executed by a voice or video search device and a recognition device.

[0002]

2. Description of the Related Art In a conventional technique for detecting a predetermined signal from a signal sequence, an input signal to be processed is collated with a reference signal registered in advance. For example, in the high-speed signal search method disclosed in Japanese Patent Application No. 11-130630, an input signal is compared with a pre-registered reference signal using respective feature amount sequences, and an input signal having a signal similar to the reference signal is detected. I'm trying to find the position.

In speech recognition, a speech to be recognized is specified in advance, and the speech signal is compared with a supplied speech signal. For this reason, it is necessary to previously hold voice signals of various voices to be recognized, and the current voice recognition device registers these voice signals in a built-in dictionary in advance at the time of manufacture.

[0004]

By the way, in the conventional signal detection technique described above, the entire reference signal registered in advance is used for comparison with the input signal. That is, the matching is performed on the assumption that the entire reference signal is included in the input signal. For this reason, when only a part of the reference signal is included in the input signal, there is a problem that it is often impossible to detect this. For example, when checking the use of music on the Internet,
Since it is also assumed that only an arbitrary part of the music is used, it is difficult to register the reference signal by defining the part to be used in advance. Is difficult to detect.

As a method of solving this problem, a reference signal is divided into a plurality of sections in advance by hand or the like, and the signals in those sections are sequentially compared with the input signal as reference signals, thereby obtaining the original signal. It is conceivable to detect a signal in each section that is a part of the reference signal. However, how to determine the section for dividing the reference signal cannot be uniquely determined,
Each time, the appropriate section must be snooped. For this reason, depending on the method of division, detection leakage may occur, or useless collation calculation may be required, and it is difficult to expect efficient signal detection.

On the other hand, speech recognition requires a speech recognition device to register speech to be recognized in advance, whereas general television broadcasts such as news programs use current affairs terms and personal names. For example, voices of words that are difficult to register in advance during device manufacturing frequently appear. For this reason, in speech recognition, there is a problem that the speech of such words hinders recognition.

The present invention has been made to solve such a conventional problem, and provides a signal detection technique capable of efficiently detecting a signal partially including a signal specified in advance. It is aimed at.

Further, the present invention makes it possible to automatically learn a signal such as a target voice by voice search or voice recognition or the like by using such a signal detection technique, so that various signals can be easily searched or retrieved. It aims to provide a technology that can be recognized.

[0009]

According to the first aspect of the present invention,
A reference feature value calculating step of deriving a feature value sequence from a pre-registered reference signal, an input feature value calculating process of deriving a feature value sequence from an input signal, and a part of the feature value sequence derived in the reference feature value calculating process. , An attention window setting step of setting an attention window for each of a part of the feature amount series derived in the input feature amount calculation step, and the reference feature amount calculation step and the input feature amount calculation step. Of each feature quantity series,
A similarity calculation step of calculating a similarity for a feature amount within each attention window set in the attention window setting step, and a skip width of the attention window based on the similarity calculated in the similarity calculation step And a skip width calculating step of determining a setting position of each of the attention windows when the attention window setting step is performed next, and wherein the attention window setting step, the similarity calculation step, and the skip width calculation step The attention window in which the similarity between the reference signal and the input signal is calculated by repeating the process and comparing the similarity calculated in the similarity calculation process with a preset target similarity. Is determined whether or not each set point is similar, and a portion in the input signal that is similar to a partial section of the reference signal is searched for.

According to a second aspect of the present invention, in the signal detection method according to the first aspect, the similarity calculating step includes creating a histogram for the feature amount sequence and calculating the similarity based on the histogram. It is characterized by.

According to a third aspect of the present invention, in the signal detection method according to the first or second aspect, the skip width calculating step is based on the similarity calculated in the similarity calculating step.
The skip width is calculated from an upper limit value of the similarity when the respective attention windows for the reference signal and the input signal are moved near the current position.

According to a fourth aspect of the present invention, in the signal detection method according to any one of the first to third aspects, the skip width calculating step includes determining any one of the positions of the window of interest satisfying the calculated skip width. Is determined as a setting position of each of the attention windows when the attention window setting process is performed next.

According to a fifth aspect of the present invention, there is provided a method of registering a signal to be searched and searching for the registered signal from a supplied processing target signal. In the method, a signal that can be a search target is learned or accumulated by using the signal detection method described in the section, and is used as a signal to be searched.

According to a sixth aspect of the present invention, there is provided a method of registering a signal to be recognized and recognizing the registered signal in a supplied signal to be processed.
A signal that can be recognized is learned or accumulated by using the signal detection method according to any one of the above items, and is used as a signal to be recognized.

According to a seventh aspect of the present invention, there is provided a signal detection method according to any one of the first to fourth aspects, a signal search method according to the fifth aspect, or a signal recognition method according to the sixth aspect. ,
A computer-readable recording medium that records a program to be executed by the computer when the program is executed by the computer.

In particular, the present invention differs from the method according to the earlier application (Japanese Patent Application No. 11-130630) in that, in the attention window setting process, a new reference signal is partially added. The gist is that an attention window is set such that only the attention is paid, and a function of calculating the attention window setting position for both the input signal and the reference signal is provided in the skip width calculation process. With such a unique configuration, even when only a part of the signal provided as the reference signal is included in the input signal, it is possible to detect a part of the signal.

[0017]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS <Basic Embodiment> An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a partial signal detection device to which a signal detection method according to one embodiment of the present invention is applied. The present partial signal detection device is intended for an audio signal, and a signal detection process mainly for an audio signal will be described below. However, the audio signal is only an example of the signal to be processed, and the same configuration as that of the partial signal detection device can be applied to various general signals such as various time-series signals (audio signal of an audio signal and video signal of a moving image). And the signal detection processing can be similarly performed.

This partial signal detecting apparatus includes a reference feature value calculating means 1, an input feature value calculating means 2, and a notice window setting means 3
And a similarity calculating means 4 and a skip width calculating means 5. Then, the previously registered reference signal and the sequentially supplied input signal are input, and the similarity with a certain part of the reference signal is set to a predetermined value (hereinafter, this is referred to as a “search threshold” and represented by θ). A portion in the input signal that exceeds the portion and the portion in the reference signal are output as signal detection results. Here, the reference signal is an acoustic signal including, as a part thereof, a signal to be detected (a signal such as a sound to be searched) as a sample, and is registered and provided in a predetermined storage unit or the like in advance. On the other hand, the input signal is an audio signal such as a sound to be searched, and an audio signal generated by various means such as a television broadcast, the Internet, and a sound reproducer can be used.

The reference feature value calculating means 1 derives a feature value sequence from the reference signal, and supplies the feature value sequence to the attention window setting means 3. The input feature value calculating means 2 derives a feature value sequence from the input signal and supplies the feature value sequence to the attention window setting means 3. Various features can be used as the feature amount here, but in the present embodiment, a spectrum feature of an acoustic signal is used as an example. In the case where a video signal is to be used, a video feature such as a color feature is used, and the sequence is extracted by the reference feature value calculating means 1 and the input feature value calculating means 2, respectively. If we do the same thing,
It is also possible to search for a video signal by the partial signal detection device.

The attention window setting means 3 includes a part of the feature value sequence of the reference signal derived by the reference feature value calculation means 1 and a part of the feature value sequence of the input signal derived by the input feature value calculation means 2. For each of these, a predetermined attention window is set (hereinafter, these attention windows are referred to as a “reference attention window” and an “input attention window”, respectively).
That. ). The window of interest here is set based on the head position of each window of interest supplied from the skip width calculation means 5,
The attention window setting unit 3 supplies the similarity calculation unit 4 with the feature amount sequence of the reference signal within the set reference attention window and the feature amount sequence of the input signal within the input attention window.

The similarity calculation means 4 includes, in the feature amount sequences supplied from the reference feature amount calculation means 1 and the input feature amount calculation means 2, the features in each attention window set by the attention window setting means 3. Calculate similarity for quantities. Then, by determining whether the calculated similarity is greater than the search threshold θ, it is determined whether or not the feature amount sequence in the currently set reference attention window exists at the position of the currently set input attention window. It is determined, and if it exists, the position of the reference window of interest and the position of the input window of interest are output as signal detection results. Further, the similarity calculating means 4 supplies the calculated similarity to the skip width calculating means 5.

The skip width calculating means 5 calculates the skip width of each window of interest from the upper limit of the similarity when each window of interest is moved near the current position. That is, based on the similarity from the similarity calculating means 4, a range in which the similarity when each window of interest is moved does not exceed the search threshold θ is calculated, and a width in which each window of interest can be skipped is calculated. Based on this, the head position of each window of interest to be compared (calculation of similarity value, comparison with a search threshold, etc.) is determined and supplied to the window setting means 3. As a result, the setting positions of the reference window of interest and the input window of interest in the next search stage are determined, and a part of the feature amount sequence of the reference signal and the input signal within the window of interest is respectively transmitted from the window of interest setting means 3 The processing is supplied to the similarity calculating means 4 and the processing in each means is repeated in the same manner as described above.

In this manner, the processing by the attention window setting means 3 to the skip width calculating means 5 is repeated, and the similarity between some parts of the input signal and some parts of the reference signal is calculated and calculated. By comparing the similarity with a preset target similarity, it is determined whether or not the relevant part of the reference signal and the relevant part of the input signal are similar. Thereby, a part similar to a partial section of the reference signal registered in advance is searched for from the input signal.

<Specific Embodiment> Next, the processing in the above-described reference feature value calculating means 1 to skip width calculating means 5 will be described more specifically, and its processing principle will be clarified (the processing principle will be described later). A specific example for this is shown in FIG.

The reference feature value calculating means 1 first reads a given reference signal. Next, feature extraction is performed on the read reference signal. In the present embodiment, since a spectral feature is used as the feature, the feature extraction here can be performed by, for example, a band-pass filter.

As a specific example of such feature extraction, for example, when it is desired to search for a specific audio signal of about 15 seconds from an audio signal provided by television broadcasting or the Internet, specific setting of feature extraction Can be obtained as follows. That is, seven band-pass filters are used, and their center frequencies are set at equal intervals on a logarithmic axis. Then, 60 to the reference signal
An analysis window with a time length of about milliseconds is set, and the analysis window is set to 1
The average value of the square of the output of each band-pass filter in the analysis window is calculated while moving by 0 millisecond, and the obtained seven average values are grouped as a 7-dimensional feature vector. In this case, one feature vector is obtained every 10 milliseconds of moving the analysis window. In this way, the reference feature value calculating means 1 sequentially obtains, in time series, feature vectors having components of the respective frequency bands of the reference signal as elements, and these are used as the feature value series of the reference signal. Will be output to

It should be noted that the present partial signal detecting apparatus can be used for searching for an image by using the image feature as described above. In this case, the image of one frame of the video is divided into four equal parts in the horizontal direction and three equal parts in the vertical direction and divided into a total of 12 regions, and the RGB values in each divided region are used as features to obtain a total of 36-dimensional feature vectors. Good results can be obtained by using the represented video features.

On the other hand, the input feature quantity calculating means 2 first reads a sound signal to be searched as an input signal. Next, feature extraction is performed on the read input signal. The feature extraction here is performed by performing the same operation on the input signal as performed in the reference feature amount calculating means 1.

That is, in the case of the above specific example, an analysis window having a time length of about 60 milliseconds is set for an input signal by using seven band-pass filters whose center frequencies are set in the same manner as described above. The average value of the square of the output of each bandpass filter in the analysis window is calculated while moving the analysis window by 10 milliseconds, and the obtained seven average values are grouped into a seven-dimensional feature vector. . In this case, one feature vector is obtained every 10 milliseconds. In this way, the input feature value calculating means 2 sequentially obtains, in time series, feature vectors of spectral features having components of the respective frequency bands of the input signal as elements. The result is output to the setting means 3.

The attention window setting means 3 first reads the series of feature vectors output from the reference feature value calculating means 1 and the input feature value calculating means 2, respectively. Subsequently, an attention window is set for each of the series of the feature vectors of the reference signal and the input signal. In the present embodiment, the length of the window of interest set here is D. The window length of interest D is the length of the feature vector sequence on the time axis. For convenience, the window length D is given by the number of feature vectors corresponding to the length (included in the time of the length). The number of feature vectors corresponding to a certain time length including only a part of the entire signal is defined as a window length of interest D.

That is, the reference window of interest for the feature vector sequence of the reference signal is composed of D feature vectors (mth to (m + D-1) th feature vectors after the mth feature vector from the top of the reference signal feature vector sequence. Vector). The input window of interest for the feature vector sequence of the input signal is the D feature vectors (the nth to (n + D-1) th feature vectors) after the nth feature vector from the beginning of the feature vector sequence of the input signal. Will be included. Here, m for designating the head position of each window of interest
And the value of n are supplied from the skip width calculating means 5 (details will be described later). Such a reference window of interest and an input window of interest are set by the window-of-interest setting means 3, and the feature vectors in the reference windows of the reference signal and the input signal are output to the similarity calculating means 4.

The similarity calculation means 4 first reads the feature vectors in the attention window for both the reference signal and the input signal, which are output from the attention window setting means 3. next,
A histogram of the feature vector is created from the feature vectors in the window of interest.

This histogram is created by dividing the value of each element of the feature vector into several bins (sections). For example, if each element is divided into three sections, and if the number of elements of each feature vector is 7, there are three combinations of sections for each of the seven elements. That is, the number of sections arranged on the horizontal axis of the histogram is 3 to the seventh 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 7 sections.

The similarity calculating means 4 counts the number of feature vectors classified into each section, thereby obtaining the feature vector in the reference attention window received from the attention window setting means 3,
A histogram for the reference signal and a histogram for the input signal are created from the feature vectors in the input window of interest. Here, the histogram for the reference signal is H _R , and the histogram for the input signal is H _I.
Here, R and I indicate that the histogram is created from the feature vector of the reference signal and the feature vector of the input signal, respectively.

[0035] Then the similarity calculating unit 4 calculates a histogram H _R of the reference signal a similarity between the histogram H _I of the input signal. As the similarity in this case is susceptible to various definitions, wherein is set to be used overlapping ratio of the histogram as an example, defines a similarity S _RI of the histogram H _R and the histogram H _I as follows .

[0036]

(Equation 1)

Where L is the total number of histogram bins
(3 to the 7th power in the above example), and h _Rl, H_IlIs it
Each histogram H_R, H_IFeatures included in the l-th bin of
Represents the number (frequency) of sign vectors. D is a histogram
Is the total frequency of, derived from the reference or input signal
Of the feature vectors that are included in one window of interest,
The window length D of interest corresponds to the total number.
You.

The similarity calculating means 4 calculates the value of the similarity _SRI according to the above equation (1) and outputs it to the skip width calculating means 5. In addition, the similarity calculating means 4 compares the calculated value of the similarity _SRI with a preset search threshold θ. At this time, if the similarity _SRI exceeds the search threshold θ, it means that the reference signal and the input signal are well similar at the positions of the currently set reference window of interest and input window of interest. . Therefore, when the similarity _SRI exceeds the search threshold value θ, the current position of the reference window of interest with respect to the reference signal and the current position of the input window of interest with respect to the input signal are output as signal detection results (each Outputs the time or time from the beginning of the signal.)

The skip width calculating means 5 first reads the similarity _SRI output from the similarity calculating means 4. Next, the skip width w of the reference attention window for the reference signal
The skip width w _I of the input window of interest for _R and the input signal is calculated.

Here, the total skippable width w of w = w _R + w _I is considered. Now, since the attention window length is given by the number of feature vectors corresponding to the time length, the skip widths w _R and w _I
Assuming that the unit for the total skippable width w is also expressed as the number of feature vectors, the total skippable width w is obtained by the following equation.

[0041]

(Equation 2)

In Equation 2, floor (•) represents rounding down.
D is the window length of interest, that is, the length of the matching section,
θ is the above-described search threshold set in advance.

Equation 2 indicates that if S _RI <θ at present, the similarity S _RI will never exceed the search threshold θ even if the reference window and the input window are shifted by w-1 feature vectors. Means This is because, when the two attention windows are shifted from each other, all the feature vectors that go out of each attention window do not contribute to the overlap of the histograms, and the feature vectors that enter into each attention window. Can be easily understood by considering the case in which all the factors contribute to the overlap of the histograms (the case where the similarity reaches the search threshold first). That is, since the similarity in such cases is increased most, assuming a case, the upper limit of the similarity S _RI when shifted by the feature vector w-1 pieces of combined reference target window and the input focus window search It becomes the threshold value θ. From this, when S _RI <θ, the skip width for w feature vectors for which there is a possibility that the similarity S _RI exceeds the search threshold θ is set as the total skippable width. If S _RI ≧ θ at present, w = 1 (full search) to find a local peak of similarity.

As described above, first, the feature vector number m (corresponding to the time position in the reference signal) from the head of the reference signal and the feature vector number n (corresponding to the time position in the input signal) from the head of the input signal ) Represents the relationship between the skip widths w _R and w _I on the mn plane with the coordinate axes as the range of possible values of the skip widths w _R and w _I under the constraint of the total skippable width w. , Centering on the current window position of interest (m, n),
A square having a diagonal length of 2w is formed. Then, inside the (not including Henjo) area of the square, the value of the similarity S _RI that is guaranteed not exceed the search threshold theta, the next target window position set outside this region do it. This is the processing principle of signal search in the present partial signal detection device. Based on this, the position of the window of interest is sequentially set, and the similarity _SRI is calculated.

It is to be noted that since the evaluation has been made that the similarity _SRI does not exceed the search threshold θ for the area formed inside the above-mentioned square, this area is hereinafter referred to as “evaluated area”. Call. On the other hand, the other area, that is, the area for which it is not clear whether the value of the similarity _SRI does not exceed the search threshold value θ up to the present time is referred to as an “unevaluated area”. Assuming that the value obtained by subtracting the window length of interest D from the total number of feature vectors of the reference signal and the input signal is M and N, the search is completed by filling up the M × N search space in the evaluated area. Where the values M, N
Corresponds to the time obtained by subtracting the time length of the window of interest from the time lengths of the reference signal and the input signal, respectively, and M ≧ 0 and N ≧ 0.

FIG. 2 shows a comparison between the processing mode of the present partial signal detection apparatus based on the above principle and the processing mode of the method according to the prior application (Japanese Patent Application No. 11-130630) mentioned above as a conventional technique. FIG. In this figure,
The upper part shows the processing form by the method of the prior application, and the lower part shows the processing form by the present partial signal detection device. In each processing form, a portion corresponding to a point to be actually checked on the mn plane is shaded. (Hereinafter, such a point will be referred to as an “active point.” Note that the illustrated mn plane is at an intermediate stage of the search.) . However, a point on the mn plane is represented by a small square area in a matrix since m and n are feature vector numbers (integers), and coordinates (m,
The point corresponding to n) (small square area) corresponds to the case where the matching is performed using the head position of the reference window of interest as the m-th feature vector and the head position of the input window of interest as the n-th feature vector.

In the processing according to the prior application method in the upper part, the signal search is repeated while the position of the window of interest with respect to the reference signal is gradually changed. In other words, the prior application does not disclose that a window of interest is set only for a part of the reference signal and that its skip width (the head position of the next comparison) is specified. In order to detect a partial signal in the reference signal, it is necessary to repeat the signal search while gradually changing the position of the attention window including the entire reference signal (the reference signal itself before being supplied is divided). It is also conceivable to repeat the signal search in such a case, but that is not appropriate as a comparison because it is necessary to repeat all processes including the first feature value calculation.) For this reason, the head position of the reference signal attention window corresponding to the reference attention window in the partial signal detection device is, as shown by the active point on the m-axis in the figure, from the top of the feature vector sequence of the reference signal. The positions are sequentially set to one feature vector at a time (therefore, the reference signal is searched at all the feature vector positions after all). Then, the matching is performed as shown in the n-axis direction in the figure while calculating the skip width of the input signal attention window for each position of the reference signal attention window. still,
Although not shown, the full search corresponds to a processing mode in which all points are active points, and the number of times of matching is M × N.

On the other hand, in the processing by the present partial signal detection device at the lower stage, the evaluation area is sequentially obtained as described above using the reference attention window and the input attention window for only a part of the reference signal, and the other unevaluated areas are obtained. The position of each window of interest is set in the area. The dotted line in the figure indicates the boundary between the evaluated area and the unevaluated area in this case.
A point that does not require collation in both axial directions is eliminated, and a minimum required active point is obtained more efficiently than the prior application method as shown in the figure (hereinafter, such an active area and an unevaluated area are compared with each other). The boundary will be referred to as the "active edge." Although the number of active points is small as described above, since the attention window including only a part of the feature vector sequence of the reference signal is set for each of the feature vector sequences of the reference signal and the input signal, the reference signal A portion in the input signal that includes only a part of the signal is properly detected, and leakage of the detection can be prevented.

The processing in the present partial signal detection apparatus determines an active point to be collated next from active points based on the similarity _SRI sequentially calculated by the similarity calculating means 4, and assigns each active window a target window. This is realized by repeating the setting of an active point. In this embodiment, the following algorithm is used in the processing of the skip width calculating means 5 in determining the active point in this case.

1. It is assumed that the initial active point P ₀ = (0, 0) (the head positions of the reference window of interest and the input window of interest are respectively set at the head of the feature vector series of the reference signal and the input signal). 2. The one with the smallest n-coordinate value is selected from the existing active points, the reference position of the reference attention window and the head position of the input attention window are set for that point, the matching is performed, and the already evaluated active area is updated and the newly created active edge is determined. Ask. 3. The above 2. And the existing active edge or a straight line m = 0, n = 0, m = M−1 or n = N−
The intersection with 1 is a new active point. 4. The existing active points in the updated evaluation area are deleted. Also, the existing active edge is updated. 5. If there are no unerased active points that have not yet been verified, the process ends. 2. If there is already, Return to

In this algorithm, 2. In this step, the matching is performed from the active point where the value of n is small. When the active points to be compared next are determined in this way, only the upper two sides of the square forming a new active edge intersect with the existing active edges based on the matching result. It is good to check the intersection only for the side (this is also shown in FIG. 2).

The skip width calculating means 5 obtains the start position (m, n) of the reference window of interest and the input window of interest to be set next on the basis of the above principle, and obtains the start position (m, n).
n) is output to the attention window setting means 3. Thereby, the attention window setting means 3 shifts the reference attention window for the reference signal and the input attention window for the input signal to the head position (m, n) output from the skip width calculation means 5, respectively. 3. The processing in the similarity calculation means 4 and the skip width calculation means 5 is repeated in the same manner as described above. If the value of n exceeds the end of the feature vector sequence of the input signal, it means that all the input signals have been searched, and this is detected by the skip width calculation means 5 or the attention window setting means 3 and The signal detection processing by the partial signal detection device ends.

<Experimental Example> Next, an operational experimental example of the present partial signal detecting apparatus will be described. In order to confirm the effect of the signal detection method executed by this partial signal detection device, the same acoustic signal for 6 hours of recording a television broadcast is used as both an input signal and a reference signal, and iterative matching by full search is performed. Experiments were performed to compare the number of times of matching by three methods, namely, searching, iterative matching searching by the method of the above-mentioned prior application, and searching by the present partial signal detecting device. However, both the cases where the calculated similarity was a value equal to or less than the search threshold and the case where the calculated similarity was a value exceeding the search threshold were counted.

In the present partial signal detection device, the method of generating feature vectors and the algorithm of classification thereof can be changed to various forms according to the purpose.
As described in “Specific embodiment”, the reference feature amount calculating unit 1
A 7-channel bandpass filter bank is used for creating feature vectors in the input feature amount calculating means 2, and a feature classifying method based on a combination of quantization for each dimension is used for classifying feature vectors in the similarity calculating means 4. Was. The length of the collation section (interest window length D) is equivalent to 15 seconds, and the search threshold θ is 0.8.

In this experiment, a feature vector is generated every 10 milliseconds for each signal, and the minimum step width of movement of a window of interest (time window) is set to 10 feature vectors in all methods. Therefore, it means that the attention window is moved at intervals of a minimum of 0.1 second. Note that it is unknown that the input signal and the reference signal are the same, and the entire mn plane is set as a search target.

FIG. 3 shows the results of this experiment. In the search by the partial signal detection device of (3), the number of times of matching is greatly reduced as compared with the iterative matching search by the full search of (1) and the iterative matching search by the prior application method of (2). ,
It is noted that the number is reduced to 1/10 or less as compared with the method (2). The effect of reducing the number of times of collation by the present partial signal detection device becomes more conspicuous as the value of the window of interest D is larger, and becomes more conspicuous as the minimum step width of the window of interest is smaller. Alternatively, the effect of reducing the number of times of collation can be enhanced by reducing the minimum step width of the movement of the window of interest.) According to this partial signal detection device having a small number of times of matching, the calculation cost of search can be reduced by the small number of times of matching, and signal detection can be performed efficiently.

As described above, the present partial signal detection apparatus newly sets a window of interest for a reference signal,
It has a function of setting an attention window for both the input signal and the reference signal so as to improve detection efficiency while preventing detection leakage. Therefore, according to the present partial signal detection device, even when it is desired to detect a part of the reference signal included in the input signal, more efficient partial signal detection can be performed as compared with the conventional method. There is an advantage that you can.

<Application Example> The above-described signal detection method executed by the partial signal detection device is also used for checking the use of music on the Internet for the purpose of properly managing the use of the music. be able to. That is, if the audio signal of the target music to be used for the use check is registered in advance as the reference signal, by searching for an audio signal file on the Internet based on the signal,
It is possible to detect an audio signal file on the Internet that includes a part of the target music. Furthermore,
The above-described signal detection method can be similarly applied to detection of a video signal for the purpose of managing video use on the Internet, and can also be applied to detection of a general signal or file.

The signal detection method described above can also be used for learning and storing a target signal in a signal search method and a signal recognition method, and can also be used for, for example, a speech recognition device. In the current speech recognition device, it is necessary to register speech to be recognized in a dictionary in advance, but in speech recognition in general television broadcasting such as a news program, it is necessary to manufacture devices such as current affairs terms and personal names. Occasionally, speech signals that are difficult to register in advance frequently appear, and the speech of such words hinders recognition. Therefore, a signal that repeatedly appears in an audio signal is extracted by the above-described signal detection method (the above-described search processing is performed on an audio signal such as a television broadcast), and a repeatedly appearing partial signal that can be a recognition target is extracted. Detected) and register it in the dictionary to make it a new speech signal to be recognized. As a result, the recognition target signal is automatically and appropriately learned, and a speech signal that is difficult to register at the time of manufacture, such as current affairs related terms, can be newly accumulated as needed. Voice recognition can be easily performed. Furthermore,
Similarly, in the voice search device, a pattern that appears automatically and repeatedly by using the above-described signal detection method is extracted in advance, and a search target signal is automatically registered and appropriately learned or accumulated. , It is possible to improve the search speed.

The processing according to the present signal detection method may be performed by recording a program defining the procedure on a computer-readable recording medium, and reading and executing the program recorded on the recording medium by a computer system. Good. Here, the computer system includes an OS and hardware such as peripheral devices as necessary, and in a system using a WWW system, also includes a homepage providing environment (or display environment). Computer-readable recording media include a floppy (registered trademark) disk, a magneto-optical disk, a ROM, and a CD-R.
Any storage device such as a portable medium such as an OM or a hard disk built in a computer system may be used. Further, a communication line for transmitting a program through a network such as the Internet or a communication line such as a telephone line. In addition, the program is held for a certain period of time, such as a program that dynamically holds a program for a short time (transmission medium or transmission wave) and a volatile memory in a computer system serving as a server or a client in that case. Or the like. Further, the above-mentioned program may define a part of the above-mentioned various processes, and may be a so-called difference file (difference file) that realizes the signal detection method in combination with a program already recorded in the computer system. Program).

[0061]

As described above, according to the present invention, an attention window is set for each of a part of the feature amount sequence of the reference signal and a part of the feature amount sequence of the input signal. In addition to calculating the similarity for the feature amount within and comparing it with the target similarity set in advance, the skip width of the attention window is calculated based on the calculated similarity, and the next setting of each attention window is performed. Since the position is determined, signal detection is performed for each of the reference signal and the input signal while sequentially setting a window of interest for detecting only a part of the reference signal. This makes it possible to search for a partial signal in the input signal that is similar only to a partial section of the reference signal, and to efficiently detect a portion in the input signal that partially includes the signal specified in advance. Is obtained.

Further, the present invention uses such a signal detection method to learn or accumulate a signal that can be an object by a method of searching or recognizing a signal. Therefore, the learning or accumulation is automatically performed. An effect is obtained that various signals and the like can be easily searched or recognized.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a partial signal detection device according to an embodiment of the present invention.

FIG. 2 is a diagram showing a processing mode by the partial signal detection device in comparison with a processing mode by a conventional method.

FIG. 3 is a diagram showing an experimental result of a search process for an acoustic signal by the partial signal detection device and the like.

[Explanation of symbols]

1. Reference feature value calculation means 2 Input feature calculation means 3 Attention window setting means 4 Similarity calculation means 5 Skip width calculation means

Continuation of the front page (56) References JP-A-2001-92486 (JP, A) JP-A-2000-312343 (JP, A) JP-A-9-68994 (JP, A) Kunio Kashino, Gavin Smith, Hiroshi Murase, “Histogram A Fast Search Method for Acoustic Signals Using Features-Time-Series Active Search Method-", IEICE Transactions on Digital Science, September 1999, Vol. J82-D-II, No. 9, p. 1365-1373 Hiroshi Murase, V.A. V. Vinod, "High-speed object search using local color information-active search method-", IEICE Transactions D-II, September 1998, Vol. J81-D-II, No. 9, p. 2035-2042 Kunio Kashino, Takayuki Kurozumi, Hiroshi Murase, "High-speed AND / OR Search of Sound and Video Based on Time-Series Active Search", NTT R & D, February 2002, Vol. 50, No. 11, p. 895-901 (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 11/02 G10L 15/04 G10L 15/10 JICST file (JOIS)

Claims (7)

(57) [Claims] 1. a reference feature value calculating step for deriving a feature value sequence from a reference signal registered in advance; an input feature value calculating process for deriving a feature value sequence from an input signal; and a feature value derived in the reference feature value calculating process An attention window setting step of setting an attention window for each of a part of the series and a part of the feature amount series derived in the input feature amount calculation step; the reference feature amount calculation step and the input feature amount calculation A similarity calculating step of calculating a similarity for a feature amount in each of the attention windows set in the attention window setting step among the respective feature amount sequences derived in the process; and calculating in the similarity calculation step. Calculating a skip width of the window of interest based on the calculated similarity, and a skip width calculation step of determining a setting position of each window of interest when the window of interest setting step is performed next. , Said The reference signal and the input signal are compared with the similarity calculation process and the skip width calculation process by comparing the similarity calculated in the similarity calculation process with a previously set target similarity. Determining whether or not each set point of the window of interest for which the similarity is calculated is similar, and searching for a part in the input signal similar to a partial section of the reference signal. Detection method. 2. The signal detection method according to claim 1, wherein in the similarity calculating step, a histogram is created for the feature amount series, and the similarity is calculated based on the histogram. 3. The method according to claim 2, wherein the skip width calculating step includes calculating a similarity when the respective attention windows for the reference signal and the input signal are moved near a current position based on the similarity calculated in the similarity calculating step. The signal detection method according to claim 1, wherein the skip width is calculated from an upper limit value of degrees. 4. The step of calculating the skip width includes the step of calculating any one of the positions of the window of interest satisfying the calculated skip width.
4. The method according to claim 1, wherein a setting position of each attention window is set when the attention window setting process is performed next.
The signal detection method according to any one of the above items. 5. Registering a signal to be searched for,
A method of searching for the registered signal from a supplied processing target signal, wherein the signal which can be a search target is learned or accumulated by using the signal detection method according to any one of claims 1 to 4. A signal search method, wherein the signal is used as a signal to be searched. 6. A signal to be recognized is registered,
A method of recognizing the registered signal in the supplied processing target signal, wherein the signal that can be recognized is learned or accumulated by using the signal detection method according to any one of claims 1 to 4. A method for recognizing a signal, wherein the signal is used as a signal to be recognized. 7. A method for detecting a signal according to claim 1, a method for retrieving a signal according to claim 5, or a method for recognizing a signal according to claim 6, using a computer. A computer-readable recording medium on which a program to be executed by the computer is recorded.