KR100978913B1 - A query by humming system using plural matching algorithm based on svm - Google Patents

Abstract

PURPOSE: A sound source search system and a method combining matching algorithms based on a plurality of SVM searching humming are provided to combine DTW(Dynamic Time Warping), EMD(Earth Mover's Distance) and matching score obtained from LS(Linear Scaling) method thereby providing exact sound source search result. CONSTITUTION: An audio signal analyzer(120) analyzes audio signal and extracts feature value of the audio signal. An audio signal matching searcher](140) uses feature value of corrected audio signal and applies the first and the second matching algorithm and produces the first and the second output value. The audio signal matching searcher uses the first and the second output value as an input and produces similarity measured value and searches and extracts sound source having high similarity based on the produced similarity measured value from sound source data base.

Description

A sound source retrieval system and method combining SMB-based multiple matching algorithms {A ＱＵEＲＹ ＢＹ HＵＭＭＩＮ ＳＹＳＴＥＭ ＵＳＩＮＧ PLC

The present invention relates to a sound source retrieval system and method incorporating a plurality of SVM-based matching algorithms, and more particularly, to receive and filter and normalize an audio signal such as a humming, and to various pitch characteristics of a normalized audio signal. The present invention relates to a system and method for retrieving a sound source from a database by matching matching scores calculated by applying a plurality of different matching algorithms through an SVM.

Conventional sound source search method takes a symbol melody sequence by taking the relative value of each note in the input humming melody, and similarity with the feature value of the sound source stored in the database using dynamic programming (DP) matching Measure and determine the success of the match.

In this method, the melody similarity calculator calculates the similarity between the input symbol melody sequence and the feature value of each audio in the metadata DB. In this case, the similarity may be a method such as Euclidean distance or Sum of absolute difference, which measures the distance between two vectors using DP matching. After that, the distance-based classifier sorts the calculated distance information in the order of the smallest distance value, extracts the corresponding metadata information from the metadata DB, and delivers the metadata information to the client through the Internet as search result information.

In another conventional sound source searching method, a breakpoint is set using a frequency spectrum energy distribution of an input melody, and the input melody is converted into clear sounds based on the excitation point. The confidence level is determined by the indicator value of the Spectrum Energy Distribution and the energy level of the input melody. The matching process is performed in such a way as to minimize the error associated with inserting and deleting the notes and the error related to the confidence level.

Important features used in this method are instrumentation and beat. For instrumentation measurements, we used a spectrum-based distance measurement method called EMD (Earth Mover's Distance), and a beat based rythmic based measure method. The similarity is measured by weighting the measured values from these two methods.

In another method, the similarity is measured by extracting a feature through a pitch tracker and a feature converter in humming and matching it with a template of a melody database.

In this method, the distance is first measured using the EMD algorithm, and then the threshold is determined by ascending order of distance values. The remainders are removed except for candidates smaller than the predetermined threshold. The candidates are then matched using the DTW (Dynamic Time Warping) algorithm to select the one with the smallest distance from the database.

Existing related methods use the simpler method of Euclidean distance or absolute difference, or combine matching scores (similarity results) from two or more measurement methods into weighted sums according to specific rules. It was used by combining a method or a method of reducing candidate groups to be searched by applying a matching algorithm sequentially.

However, when a simple Euclidean distance or an absolute difference value is used, a large error may occur when the chord of the input hum is different or noise is included. In addition, when using two or more matching methods, weighting (reliability) is combined to each matching method according to a specific rule, so that weights are empirically determined based on the accuracy of each known method rather than an obvious theoretical basis. There is a problem that it is not adaptive depending on the matching environment. In the case of reducing candidates to be searched by applying different matching methods hierarchically or stepwise (sequential), the matching result is calculated based on a single threshold, so that the correct sound source is incorrect at an early stage even in the case of humming with low noise. Exclusion may occur.

In order to solve the above problems, the problem of the existing methods is solved by matching the match values calculated in the DTW, EMD and LS matching algorithms to the SVM in the same step.

In addition, in the QbH system, by matching matching values obtained from DTW, EMD, and LS matching methods in the same step, than using a single matching method or using a combination of weighted SUMs, It is an object of the present invention to provide accurate sound source search results.

A sound source retrieval system incorporating a plurality of matching algorithms of the present invention for achieving the above object includes an audio signal input unit for receiving an audio signal, an audio signal analysis unit for analyzing the audio signal and extracting features of the audio signal; The audio signal correction unit corrects the audio signal in consideration of the characteristics of the audio signal, calculates a first output value by applying a first matching algorithm by inputting the corrected characteristics of the audio signal, and calculates the characteristics of the corrected audio signal. As a input, a second output value is calculated by applying a second matching algorithm, a similarity measurement value is calculated by combining the first output value and the second output value, and a sound source with high similarity is calculated from the sound source database using the calculated similarity measurement value. It may include an audio signal matching search unit for searching and extracting.

Also, in the above-described configuration, the audio signal corrector may further include a zero remover that removes a sound having a zero pitch from the analyzed audio signal.

In addition, in the above-described configuration, the audio signal corrector may include a filter that performs predetermined filtering on the analyzed audio signal in units of predetermined filter lengths.

In addition, in the above-described configuration, the filter unit may include a median filter unit that performs median filtering on the analyzed audio signal by a predetermined filter length unit, and performs an average filtering on the audio signal to which the median filtering is applied by a predetermined filter length unit. An average filter unit may be included.

In addition, in the above-described configuration, the audio signal corrector may further include a scaling unit that scales the pitch value of the filtered audio signal using the maximum and minimum values of the pitch.

Also in the above-described configuration, the feature of the corrected audio signal includes pitch values of the corrected audio signal,

The audio signal matching search unit calculates a first output value by applying a DTW matching algorithm by inputting a feature of the corrected audio signal, and calculates a second output value by applying an EMD matching algorithm by inputting the feature of the corrected audio signal. The third output value may be calculated by applying an LS matching algorithm using the corrected audio signal as an input, and the similarity measurement value may be calculated by combining the first output value, the second output value, and the third output value.

In addition, in the above-described configuration, the audio signal matching search unit may receive at least two or more outputs calculated by applying matching algorithms and combine them using an SVM (Support Vector Machine) to generate a similarity measurement value as an output.

In addition, a sound source search method combining a plurality of matching algorithms of the present invention for achieving the above object, the step of receiving an audio signal, analyzing the audio signal to extract the features of the audio signal, and in the analyzed audio signal Correcting the audio signal in consideration of the extracted feature; calculating a first output value by applying a first matching algorithm by inputting the corrected feature of the audio signal; and applying a second feature of the corrected audio signal as an input. Calculating a second output value by applying a matching algorithm; searching for and extracting a sound source having a high similarity to the input audio signal from a sound source database using a similarity measure calculated by combining the first output value and the second output value; It may include.

Also, in the above-described configuration, the step of correcting the audio signal may include removing a zero pitch sound from the analyzed audio signal, filtering an audio signal from which zero pitch is removed, and using the maximum and minimum values of the pitch. And scaling a pitch value of the filtered audio signal.

In the above-described configuration, the filtering may include performing median filtering of the audio signal from which zero sound is removed in a predetermined filter length unit, and performing average filtering of the audio signal to which the median filtering is applied in a predetermined filter length unit. It may include the step of performing.

In addition, in the above configuration, the output values calculated by applying matching algorithms may be combined using a SVM (Support Vector Machine).
Preferably, the first matching algorithm and the second matching algorithm are any one of a DTW matching algorithm, an EMD matching algorithm, and an LS matching algorithm.

According to the present invention, when searching a sound source using the QbH system, a more accurate sound source search is possible by combining three matching scores calculated using different algorithms such as DTW, EMD, and LS matching algorithms.

In addition, matching matching values (matching scores) of DTW, EMD and LS algorithms are combined in the same step, thereby preventing information loss and obtaining a more accurate nonlinear classifier.

In addition, in the case of using the SVM, combining three matching matches can provide more accurate results than using one matching match or combining the matching matches hierarchically (sequentially). .

The foregoing and further aspects of the present invention will become more apparent through the preferred embodiments described with reference to the accompanying drawings. Hereinafter will be described in detail to enable those skilled in the art to easily understand and reproduce through these embodiments of the present invention.

1 is a block diagram illustrating a sound source search system according to an embodiment of the present invention.

As shown in FIG. 1, the sound source search system 100 of the present invention includes an audio signal input unit 110, an audio signal analyzer 120, an audio signal corrector 130, and an audio signal match search unit 140. It is made to include.

The audio signal input unit 110 may receive a part of a humming, song, or sound source as an input. The input audio signal may be stored in a storage unit, and the audio signal input unit may be configured to directly receive the audio signal or to receive the audio signal in the form of a sound source file recognizable by the sound source search system.

The audio signal analyzer 120 extracts feature values by analyzing the audio signal input through the audio signal inputter 110. The extracted feature values include pitch information, instrumentation information, tempo information, or beat information of the input signal.

The audio signal corrector 130 calculates an average value of the frequency band and the frequency of the input audio signal. The average value of the pitch values is also calculated. In general, in the case of humming, there is a difference in note, pitch, frequency, etc. with the sound source to be searched, and thus the search performance is reduced when searching for similar sound sources. After calculating the average values of the frequencies, the frequency information of the input audio signal is corrected similarly to the frequency information of the sound source (Mean shifting). In addition, the audio signal correction unit 130 performs various corrections to improve the performance of the matching sound source search even after the mean shifting, and details thereof will be described later in detail.

The audio signal matching search unit 140 inputs a plurality of matching search units for generating a matching score by performing at least two algorithms and the matching scores calculated by the plurality of matching search units, and finally receives them through the SVM (Support Vector Machine). Calculate the similarity measure.

The matching search unit may generate a matching score by using a dynamic time warping (DTW) matching algorithm, an early moveover distance (EMD) matching algorithm, an LS matching algorithm, or the like.

2 is a block diagram illustrating an audio signal correction unit according to an embodiment of the present invention.

As shown in FIG. 2, the audio signal corrector 200 includes a zero remover 210, an average corrector 220, a median filter unit 230, an average filter unit 240, and a scaling unit 250. It is made to include.

The zero remover 210 removes a signal having a pitch value of 0, that is, a signal corresponding to silence, from the input audio signal.

The average corrector 220 performs a similarly corrected process to some extent because the input audio signal may have a higher or lower pitch than the target sound source.

The median filter 230 performs median filtering to remove noise included in the input audio signal. Median filtering is used to remove randomly generated peak noise.

The average filter unit 240 performs average filtering to correct and smooth the shaking of the sound with respect to the median filtering audio signal.

The scaling unit 250 performs Min-Max scaling so that the ranges of pitch values coincide with each other in order to solve the difference in the pitch range of the input audio signal and the comparison target sound file. To this end, the scaling unit 250 calculates the minimum and maximum values of the pitches of the input audio signal and the comparison target sound source signal and normalizes them so that the ranges between the minimum and maximum values match.

Hereinafter, a function performed by the audio signal corrector 200 will be described in detail.

3 illustrates a change in frequency of an input audio signal and a comparison target sound source. As shown in Fig. 3, the frequency average value of the input audio signal and the data of the sound source file do not coincide. Therefore, the average corrector 220 corrects the frequency average value of the audio signal to match the frequency average value of the sound source to be compared. More preferably, it is possible to correct the frequency average of the input audio signal and the comparison target sound source to match all zeros.

4 illustrates extracted pitch information of a comparison target sound source and an input audio signal.

As shown in FIG. 4B, the input audio signal generates a number of peak points within the pitch contour. The main cause of this is the tremor of voices generated from audio inputs such as hum and the noise generated from the surroundings or input terminals. Since the performance of the matching search is degraded due to these noises, the audio signal corrector 200 corrects the input audio signal through various filtering. As shown in (b) of FIG. 4, it can be seen that the input audio signal has been removed from the pitch having zero pitch. In addition, it can be seen that the input audio signal is lower than the average sound height of the sound source file to be compared, which can be corrected by the average correction unit as described above.

5 is a diagram illustrating median filtering performed by a median filter unit.

The median filter portion has a filter size of a predetermined size. In general, the filter size is preferably set in radix, and FIG. 5 shows a median filter having a filter size of 5. FIG. For convenience, the filter size is assumed to be 5 and will be described.

The median filter unit rearranges the pitches of the first to fifth notes of the input audio signal in ascending order. Next, replace the third centered value (4 in Figure 5) with the third pitch value. That is, the peak value generated by noise or the like is corrected by correcting the median value of five notes to four pitches. Next, the pitch values are arranged in ascending order for the pitches of the second to sixth notes of the input audio signal, and the median value is corrected to the pitch values of the fourth note. Next, filter the third to seventh notes and correct the median to the fifth pitch value. This process is repeated sequentially, the pitch values of the last five notes are arranged in sequence and then repeated until filtering is performed. When the median filtering is completed, all pitch values are corrected except for the first two notes and the last two notes.

6 is a view for explaining average filtering performed by the average filter unit.

Average filtering is similar to the median filtering process as a whole, but differs in that the replaced value is not an intermediate value but an average value of all pitch values included in the filter length. As shown in FIG. 6, when the filter size is 5, 8.6, which is an average value of five pitch values, is calculated, and 8, which is lower than the decimal point, may be corrected to the third negative pitch value. If the mean value is not an integer, rounding may be performed or it may be designed to correct a non-integer value.

When the average filtering is performed first and the median filtering is performed, it is confirmed that the median filtering is not properly performed. Therefore, it is more preferable to perform median filtering to relax peak values and then perform average filtering.

7 is a diagram illustrating Min-Max scaling performed in the scaling unit.

As described above, the height of the sound may not match with the input audio signal and the comparison target sound source. In this case, the difference between the minimum value and the maximum value also does not coincide. That is, when the input audio signal has a pitch value lower than the comparison target sound source by the humming input, a difference between the minimum value and the maximum value occurs smaller than the difference between the maximum value and the minimum value of the pitch of the comparison target sound source. For example, if the pitch value of the sound source to be compared is 63, 65, or 69, but the pitch value of the humming input is 60, 62, 64, the maximum value of the pitch values of the sound source and the input audio signal is 5 and the minimum value is both. It can be scaled to -5 and the remaining pitch values scaled from -5 to 5 in proportion to the ratio.

Next, the function of the audio signal matching search unit will be described in detail.

8 is a block diagram illustrating an audio signal matching search unit according to an embodiment of the present invention.

The audio signal matching search unit 800 performs at least two different matching algorithms, preferably DTW matching search using the pitch information of the input audio signal 810, EMD matching using the pitch information of the input audio signal. The LS matching search 830 using the search 820 and the pitch information of the input audio signal is calculated. The matching algorithm applied may include a recursive alignment (RA) matching algorithm.

Since the DTW matching algorithm uses the DP matching algorithm, a dynamic programming (DP) matching algorithm will be described. DP matching is generally the same between the input music information (R = [ r _o , r ₁ ... r _NR-1 ]) and the stored sound source (Q = [ q _o , q ₁ ... q _NQ-1 ]). Even if it is music, it is an algorithm that calculates the similarity of both sound sources considering the difference in length.

9 is a diagram for explaining a dynamic matching matching algorithm.

Dynamic matching matching (DP matching) is a method for measuring similarity between two patterns having different lengths as shown in FIG. 9, and a method of measuring pattern similarity for each path by an insertion and removal method. to be. When matching using this method, there are a total of five constraints as shown in FIG. First, since the matching between the start point and the end point of the input pattern and the reference pattern is performed as the end point constraint, accurate speech section detection is required. The second is the monotonically increasing constraint that the optimal path should always be monotonically increased. The third condition is local constrained constraints that restrict the path to reach a node on the grid to prevent it from contracting or expanding in time. The fourth is a global path constraint, which restricts the allowable area over the entire interval between the input speech pattern and the reference pattern having different lengths. Finally, as the slope weighting condition, when calculating the cost of the local path, all of them should be given different weights according to the slope without giving the same weight to prevent them from changing irrationally with time.

 In general, since the input music information such as the humming may include information on an arbitrary portion of the stored sound source data, the input music information R values are slid and compared as shown in FIG. 10.

The following equation is an example of an equation for calculating the difference measurement between two pieces of information.

[Equation 1]

In Equation 1, r _i (m) is comparison target sound source information stored in the sound source database, and q _j (m-ps) is input audio signal information. At this time, as a ps q _j (m) sliding going to be measured different from the level between the Equation 1 two pieces of information (r _i (m) and q _j (m)) by. Details of the DP matching algorithm are well known, and thus detailed descriptions thereof will be omitted.

As mentioned above, the matching method using DP is often used in consideration of the length difference generated between the input hum and the stored sound source. To use DTW, five constraints are essential, as described above, the first of which must be observed.

 The first constraint of the DTW is that the start and end points of the two waveforms to be compared must be the same.

11 is a diagram representing an example of the first condition. Assuming that (a) is an input signal, (b) is a signal in which (a) is doubled, and (c) is a signal in which another signal is added after (a). Let's assume (a) as input data and (b) and (c) as stored data. Visually, since (c) contains the same signal as (a), one might think that the matching results of (a) and (c) will be better than the matching results of (a) and (b). . However, in case of the matching of (a) and (c), the first constraint of DTW is ignored, so the difference is not good. The advantage of DTW, that insertion and deletion occurs freely during matching, is established under the same condition of starting point and ending point, which can be confirmed by matching with (b). In fact, when (a) and (b) are matched, the start point and the end point coincide, and during insertion and deletion, which are the characteristics of the DTW, it is possible to check the difference near zero through insertion. Therefore, when an input audio signal such as a humming data is input, this section detection process is absolutely necessary to determine exactly where the start point and the end point of the sound source are input.

 The fourth of the five constraints of the DTW is the global path constraint. This is a condition that limits the allowable area over the entire range when matching the input data with the sound source data. When a person hums or sings a sound, the speed difference with the sound source usually does not occur so much. That is, since the humming waveform rarely contracts or expands significantly compared to the sound source, it is not necessary to allow the search of all regions. For example, if it is assumed that the length of the input data is 5 and the length of the sound source data is 5 as shown in FIG. 12, a total of 25 search areas appear. At this time, since there is no extreme case where a person humming five times faster or slower than a sound source, it is not necessary to calculate all the regions. Therefore, if the path search is restricted to some extent as shown in the red line of FIG. 12, and the matching is allowed only in the region, the processing time can be shortened and the matching accuracy can be increased.

In the present invention, the information such as the humming may include information on any portion of the stored sound source data, so that the comparison is performed while the input data moves in the same manner as in FIG. 13. In this case, the size of the region to be matched in the sound source data is preferably selected based on the length of the input data.

Next, the EMD algorithm will be described. The EMD algorithm is a special solution to the transportation problem, one of the linear programming methods, and is effectively used in various fields. Fig. 14 shows an example of a transport problem. The EMD algorithm finds the most optimized method by calculating the transport distance and the supply amount expressed by. Not only can it be applied to the whole data, but also partial region matching is possible. In addition, since the calculation amount is small and easy to implement in the true-metric method, it is generally used to distinguish and match a multimodal distribution.

In the EMD algorithm, there are suppliers and consumers. For example, suppose that a supplier accurately delivers the minimum amount of soil a consumer needs to fill a hole. Finally, the EMD algorithm measures the amount of work carried out by digging up the amount of soil needed. If the EMD algorithm measures two points in a finite discrete distribution, they are represented by x and y in the equation below.

[Equation 2]

In the above formula, w and u are the amount of weight given to x and y values. Let s be the sum of the weights of x and s are the sum of the weights of y. In the case of Figure 38, the sum of the weights is all 1. Also, x has 2 values and y has 3 values. The area of the circle surrounding the point is proportional to the weight, and the result of the EMD algorithm between the two values with the same weight is proportional to the amount of work each is doing. do.

 The quantity transferred from xi to yj is expressed in fij and is called flow. dij represents the distance between xi and yj. By using this, the total transport volume at each location can be calculated using the following equation.

&Quot; (3) "

Substituting the example of FIG. 15 into the above equation, the total transport volume of 0.23 * 155.7 + 0.51 * 252.3 + 0.26 * 316.30 = 246.7 can be obtained. However, Figure 38 is not an optimized result of the EMD algorithm, and in the actual EMD algorithm, x and y rarely have the same weight.

16 shows a case in which x and y each have a different sum of weights. In the case where the actual EMD algorithm is used, the above cases are mostly used. The total transport in this case is 0.23 * 155.7 + 0.25 * 252.3 + 0.26 * 198.2 = 150.4. The EMD algorithm divides the least amount of work by the value moved from the heaviest position to the lightest position. In the case of a perfect match, the optimized total transport is divided by the total supply and normalized. This is expressed as follows.

&Quot; (4) "

As a result, when the result of the EMD algorithm is obtained using the smallest supply amount = 0.74, the value of EMD (x, y) = 150.4 / 0.74 = 203.3 is obtained.

There are several advantages to applying the EMD algorithm with the above-mentioned calculation methods to the QbSH system. Firstly, the EMD algorithm uses a top-down approach. In the case of string matching using the bottom-up method, time and money are consumed while inserting and deleting, and an error occurs at the beginning, end, and time delay of the part to be executed. Has characteristics that are robust to this problem. Second, the processing time is faster because of the smaller amount of calculation compared to sliding matching or DTW. For example, when an input data needs to be matched with a large amount of sound source data, a complicated and time-consuming algorithm takes a long time to process and cannot be used in real time, but an EMD algorithm provides a real-time program. Can be used for Third, data can be grouped or divided into parts. In this case, the EMD algorithm can determine whether to use the data globally or locally because it can determine how detailed the data is analyzed. In addition, most bottom-up-based note-based cases do not preserve both pitch and tempo information, and some loss occurs. When using the EMD algorithm as a global domain, tempo and duration information for a long time is preserved. It can be used to view and measure pitch and tempo information globally.

Next, the LS algorithm will be described.

LS (linear scaling) algorithm is known as the simplest method for melody recognition as it compares the pitch vector of the input data with the sound source stored in the database while decreasing or increasing the linearly with some time. When the matching is attempted through the LS algorithm, the input data varies in length from 0.5 times to 2.0 times the original length. FIG. 17 shows the comparison of the input data by decreasing or increasing the input data by 0.5, 0.75, 1, 1.25 and 1.5 times using LS algorithm, respectively.

 As shown in FIG. 17, when the input data is increased by 1.25 times, the most similar to the sound source can be seen, and the Euclidean distance at this time is the smallest.

Next, a function of calculating a final similarity measure by applying an SVM with inputs of a plurality of matching scores calculated by applying different mapping algorithms will be described.

18 shows an SVM classifier in three-dimensional space.

As shown in FIG. 18, when the matching scores to which the DTW, EMD, and LS matching algorithms are applied are input, they are represented as vectors in a three-dimensional space. In FIG. 13, blue vectors are vectors based on matching values extracted when the same sounds are compared, and red vectors are vectors based on matching values when different sounds are compared. The yellow part shows the classifier obtained through training.

SVM goes through training to get the correct classifier. A matching value vector is obtained using matching matching values obtained by applying the above-described different matching algorithms to the input audio signal. At this time, the matched value vector obtained by comparing the same sound and the matched value vector obtained by comparing the different sounds is obtained. In the case of equality comparison, since the difference between the original sound source will be small, the measured distance value will be small, and thus, the vector distribution close to the origin will be obtained in FIG. Relatively far away.

The matched value vectors as shown in FIG. 18 obtained by comparing the same and different sounds may be classified into an optimal classifier that distinguishes two distributions as shown in FIG. 19 through an SVM training process. That is, in the training process, the matching values by comparing the same tones may be assigned an output value of −1, and the matching values by comparing different tones may be specified as an output value of 1 to obtain an optimal classifier. In the ideal case, the distribution formed by the SVM output values using the matched values has two distributions as shown in FIG. 19, and distinguishes equal and different sound comparisons by using a threshold that divides the two distributions with a minimum error. do. That is, the classifier may be obtained by setting a threshold value from the two distributions of the similarity measurement values obtained through the training process to the point where the same sound comparison error and the other sound comparison error become minimum at the same time.

That is, when such a threshold value is determined in the training data, if the SVM output value for the test data is less than or equal to this threshold, it is determined by equality comparison (that is, the sound source currently matching the input audio signal is determined to be the same sound), and the SVM output value is determined. If it is above this threshold, it is determined by comparison of other sounds (i.e., the sound source currently matching the input audio signal is determined to be different sound). In other words, the smaller the similarity measured value, the higher the similarity of the sound source.

When the query data through the user's humming is input, the matching with the sound source databases is performed through the above three matching methods (DTW, EMD, LS). Through this, three matching values are obtained and inputted into the pre-learned SVM to obtain one output value, and it is determined whether or not the sound source compared with the humming is the same song according to the predetermined threshold in the distribution as shown in Fig. 2. do.

That is, the similarity measurement value is obtained by the SVM to determine the matching rank for all sound sources in the order of the smallest similarity measurement value, and the matching sound source can be searched by extracting the sound source having the smallest similarity measurement value.

So far I looked at the center of the preferred embodiment for the present invention.

Those skilled in the art will appreciate that the present invention can be implemented in a modified form without departing from the essential features of the present invention. Therefore, the disclosed embodiments should be considered in descriptive sense only and not for purposes of limitation. The scope of the present invention is shown in the claims rather than the foregoing description, and all differences within the scope will be construed as being included in the present invention.

1 is a block diagram illustrating a sound source search system according to an embodiment of the present invention.

2 is a block diagram illustrating an audio signal correction unit according to an embodiment of the present invention.

3 illustrates a change in frequency of an input audio signal and a comparison target sound source.

4 illustrates extracted pitch information of a comparison target sound source and an input audio signal.

5 is a diagram illustrating median filtering performed by a median filter unit.

6 is a view for explaining average filtering performed by the average filter unit.

7 is a diagram illustrating Min-Max scaling performed in the scaling unit.

8 is a block diagram illustrating an audio signal matching search unit according to an embodiment of the present invention.

9 and 10 are diagrams for describing a dynamic match matching algorithm.

11 to 13 are diagrams for explaining a DTW matching algorithm.

14 shows an example of a transport problem for explaining the EMD matching algorithm.

15 shows an example of an Equal-weighted distribution EMD algorithm.

16 shows an example of an Unequal-weight distributed EMD algorithm.

17 illustrates an example of using an LS algorithm.

18 and 19 illustrate the SVM according to the present invention.

Claims (15)

An audio signal input unit for receiving an audio signal; An audio signal analyzer configured to analyze the audio signal and extract feature values of the audio signal; An audio signal correction unit configured to correct an audio signal by inputting a feature value of the audio signal; A first output value is calculated by applying a first matching algorithm as a feature value of the corrected audio signal, and a second matching algorithm different from the first matching algorithm is obtained by inputting the feature value of the corrected audio signal. Calculates a second output value by using the first output value and the second output value as an input, calculates a similarity measurement value, and uses the calculated similarity measurement value to search for and extract from a sound source database from a sound source with high similarity A sound source search system incorporating a plurality of matching algorithms, characterized in that it comprises a signal matching search unit. The method of claim 1, wherein the audio signal corrector, And a zero removal unit for performing a correction to remove a sound having a pitch of zero from the analyzed audio signal. The method of claim 1, wherein the audio signal corrector, And a filter unit for filtering the analyzed audio signal in units of predetermined filter lengths. The method of claim 3, wherein the filter unit, A median filter unit for performing median filtering on the analyzed audio signal in units of predetermined filter lengths; And And a plurality of matching algorithms for performing average filtering on a predetermined filter length unit for the audio signal to which the median filtering is applied. The method of claim 2, wherein the audio signal correction unit, And a scaling unit for scaling the pitch value of the filtered audio signal using the maximum and minimum values of the pitch. The method of claim 1, The characteristic value of the corrected audio signal includes pitch values of the corrected audio signal, The audio signal matching search unit, The first output value is calculated by applying the DTW matching algorithm by using the feature value of the corrected audio signal, and the second output value is calculated by applying the EMD matching algorithm by using the feature value of the corrected audio signal as input. A third output value is calculated by applying an LS matching algorithm using the feature value of the audio signal, and the similarity measurement value is calculated by inputting a first output value, a second output value, and a third output value. Sound source retrieval system combining number matching algorithm. The method of claim 1, wherein the audio signal matching search unit, And a plurality of matching algorithms as outputs by receiving at least two outputs calculated by applying the matching algorithms and combining them using a support vector machine (SVM). The method of claim 7, wherein the SVM, And a classifier obtained based on data obtained through a training process, wherein the classifier determines whether the sound is the same according to the similarity measure. The method of claim 8, wherein the classifier, A sound source retrieval system incorporating a plurality of matching algorithms, wherein the equality comparison error and the different tone comparison error are obtained from a region where the similarity comparison error and the different tone comparison error are minimum from the distribution of the result of the similarity measurement obtained through the training process. Receiving an audio signal; Analyzing the audio signal to extract feature values of the audio signal; Correcting the audio signal by inputting a feature value extracted from the analyzed audio signal; Calculating a first output value by applying a first matching algorithm as a feature value of the corrected audio signal, and calculating a second output value by applying a second matching algorithm as a feature value of the corrected audio signal as an input. ; And And searching for and extracting a sound source having the highest similarity from a sound source database by using the similarity measure calculated by combining the first output value and the second output value. . The method of claim 10, wherein correcting the audio signal comprises: Removing a pitch of zero pitch from the analyzed audio signal; Filtering the audio signal from which zero tones have been removed; And And scaling a pitch value of the filtered audio signal using the maximum and minimum values of the pitch. The method of claim 11, wherein the filtering comprises: Performing median filtering on the audio signal from which the zero sound is removed in predetermined filter length units; And And performing a mean filtering on a predetermined filter length unit for the audio signal to which the median filtering is applied. The method of claim 10, And output values calculated by applying the matching algorithms are combined using a support vector machine (SVM). A computer-readable medium having recorded thereon a program for executing the method of claim 10. The method according to any one of claims 1 to 9, The first matching algorithm and the second matching algorithm may be any one of a DTW matching algorithm, an EMD matching algorithm, or an LS matching algorithm, and the first matching algorithm and the second matching algorithm may be different matching algorithms. Sound source retrieval system combining number matching algorithm.

Images