KR101530481B1 - Music classification device using autoregressive model and method thereof - Google Patents

Abstract

The present invention relates to a music classification device using an autoregressive model and a method thereof. The music classification device comprises: an input portion for receiving a sound source; a short-term feature extracting portion for extracting a short-term feature vector for a tone feature of the sound source inputted from the input portion; a long-term feature extracting portion for extracting a long-term feature vector by using the short-term feature vector; an AR modeling portion for extracting an LPC by using the extracted short-term feature vector, modeling the extracted LPC in the autoregressive model, producing a new feature vector having an increased degree, and converting the same into an LSP parameter; a feature selection portion for selecting a top feature vector having a high recognition rate among the short-term feature vector, the long-term feature vector, and the new feature vector; a model generation portion for producing a classification model of a music by using the selected feature vector; and a music classification portion for classifying a genre or a mode of a test music inputted based on the classification model. According to the present invention, through a selection process of the top feature vector after extracting the feature vector of the music, the music classification device using an autoregressive model enables to reduce time required to get a classification result and to get the high recognition rate by enabling to reduce the calculation amount required for a genre classification of inputted music data.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a music classification apparatus and a music classification method using an autoregressive model,

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a music classification apparatus and method using an autoregressive model, and more particularly, to a music classification apparatus and method using an autoregression model that can reduce a calculation amount and accurately distinguish genres or moods of inputted music.

Recently, a lot of pop music has been produced, and how to classify music data effectively by category, especially the classification of music data by genre, becomes an issue. In the past, music data was manually classified into categories by a person manually. However, in order to classify massive digital music data, algorithms for automatically classifying music by composer, singer, and genre are required.

The genres are ambiguous in terms of culture, singer, and market depending on the country or person. Automatic genre classification by music data analysis is applicable to various applications such as efficient data management and music recommendation. In addition, it is economically efficient because it is not classified by hand.

Genre classification by music data analysis has been progressed and developed by various methods such as feature vector extraction and classifier. To classify music genres or moods, music data must first be analyzed. There are various methods of analyzing music data such as Beat, Rhythm, Meter, Melody, and Chord.

After analyzing the music data, the genres are classified by various classifiers such as Gaussian Mixture Model (GMM), Hidden Marcov Model (HMM), Nearest Neighbor (NN), K- Nearest Neighbor (KNN) and Super Vector Machine (SVM).

Foote created a histogram of the music's 12th MFCC (Mel-Frequency Cepstral Coefficient) and classified the genres using NN as a classifier. Bagci classifies genres using MFCC and delta values of 13th order, Inter-Genre Similarity (IGS) model and GMM classifier.

Jiang has proposed Octave-based Spectral Contrast (Octave-based Spectral Contrast), another feature of MFCC, to improve the genre classification success rate, considering Octave of music. Using GMM as a classifier, we obtained 82% genre classification success rate for five genres of Jazz, Pop, Romantic, Baroque, and Rock.

Conventional music genre classification method or system recognizes genre by SVM using MFCC, Chroma, OSC and various feature vectors. However, the system is still under discussion for a music genre classification method or system that is expected to have a higher success rate due to its low recognition success rate.

The technology to be a background of the present invention is disclosed in Korean Patent Publication No. 2011-0013646 (published on Mar. 2, 2011).

An object of the present invention is to provide a music classification apparatus and a method thereof using an autoregressive model for reducing the amount of computation and correctly distinguishing genres or moods for input music.

According to an aspect of the present invention, there is provided a music classifying apparatus comprising: an input unit for receiving a sound source; A short-term feature extraction unit for extracting a short-term feature vector for a tone color feature of the input sound source; A long-term feature extraction unit for extracting a long-term feature vector using the short-term feature vector; (LPC) using the extracted short-term feature vector, modeling the extracted LPC with an autoregressive model to generate a new feature vector with an increased order, An AR modeling unit for converting the AR model into an AR model; A feature selecting unit for selecting the short-term feature vector, the long-term feature vector, and an upper feature vector having a higher recognition rate from among the new feature vectors; A model generation unit for generating a music classification model using the selected upper feature vector; And a music classifier for classifying the genre or mood of the test music inputted based on the classification model.

The short-term feature vector, the long-term feature vector, and the new feature vector may include a Mel-Frequency Cepstral Coefficient (MFCC) and a Decorrelated Filter Bank (DFB).

The short-term feature vector includes a statistical feature for a tone color extracted using a texture window technique, and the statistical feature may include an average value and a variance value.

The long-term feature vector includes Feature-based Modulation Spectral Flatness Measures (FMSFM) and Feature-based Modulation Spectral Measurements (FMSCM) extracted using a Feature-based Modulation Spectrum (FMS). The FMSFM and the FMSCM Can be defined by the following equation.

The average of the FMS and the FMS can be defined by the following equation.

Here, T is the number of total texture windows, X _t (k, p) is the kth element of the short-term feature of the pth frame in the tth texture window, P _i is the total number of frames And M is the magnitude of the Modulation Fourier-transform.

The long-term feature vector includes a statistical feature, which may include an average value and a variance value.

Wherein the music classification unit extracts a feature vector corresponding to the upper feature vector for the test music when the test music to be classified is input and compares the feature vector with the classification model based on the extracted feature vector, The genre or mood of the test music can be classified.

The feature selecting unit selects the upper feature vector using a SVM (Support Vector Machine) ranker, and the music classifier selects a genre or mood of the test music using a one-against-one SVM (Support Vector Machine) Can be classified.

According to another embodiment of the present invention, there is provided a music classification method using a music classification apparatus, comprising: extracting a short-term feature vector for a tone color feature of an input sound source; Extracting a long-term feature vector using the short-term feature vector; Extracting a linear prediction coefficient (LPC) using the extracted short-term feature vector; Modeling the extracted LPC with an autoregressive model to generate a new feature vector with an increased order and transforming the new feature vector into LSP parameters; Selecting an upper feature vector having a higher recognition rate from the Short-term feature vector, the Long-term feature vector, and the new feature vector; And generating a music classification model using the selected upper feature vector; And classifying the genre or mood of the inputted test music based on the classification model.

According to the present invention, by selecting a feature vector having a high recognition rate from the feature vectors of music, it is possible to reduce the amount of calculation required for the genre classification of the inputted music, so that it is possible to shorten the time required to obtain the classification result, .

1 is a configuration diagram of a music classification apparatus according to an embodiment of the present invention.
2 is a flowchart of a music classification method according to an embodiment of the present invention.
3 is a waveform diagram of an input music signal.
4 is a diagram for explaining the Analysis Window.
5 is a view for explaining the Texture Window.
6 is a feature extraction block diagram using the MFCC algorithm according to the embodiment of the present invention.
7 is a feature extraction block diagram using the DFB algorithm according to the embodiment of the present invention.
8 is a feature extraction block diagram using an OSC algorithm according to an embodiment of the present invention.
9 is a diagram for explaining an LSP conversion process according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. In order to clearly illustrate the present invention, parts not related to the description are omitted, and similar parts are denoted by like reference characters throughout the specification.

First, the music genre recognizing apparatus of the present invention will be described.

1 is a configuration diagram of a music classification apparatus according to an embodiment of the present invention.

1, the music classification apparatus 100 includes an input unit 110, a short-term feature extraction unit 120, a long-term feature extraction unit 130, an AR modeling unit 140, (150), a model generating unit (160), and a music classifying unit (170)).

The input unit 110 receives voice or music signals on a frame basis. As an embodiment of the present invention, the input unit 110 may directly receive a human voice through a microphone or receive music data stored from a music database.

The short-term feature extraction unit 120 extracts a timbre feature or a statistical feature corresponding to the short-term feature of the voice or music signal input to the input unit 110. [

The long-term feature extraction unit 130 extracts the long-term feature using the timbre feature extracted from the short-term feature extraction unit 120. Also, the long-term feature extraction unit 130 extracts statistical features.

The AR modeling unit 140 extracts a linear prediction coefficient LPC using the timbre feature extracted from the short-term feature extraction unit 120 and outputs the extracted linear prediction coefficient LPC to an autoregressive model (AR model) to generate a new feature vector with an increased order. Then, the AR modeling unit 140 converts the new feature vector into LSP parameters.

The feature selecting unit 150 selects an upper feature having a higher recognition rate among the feature vectors extracted from the short-term feature unit 120, the long-term feature unit 130, and the AR modeling unit 140 using an algorithm called an SVM ranker Select a vector. That is, the feature selector 150 measures the recognition rate for the Short-term feature vector, the Long-term feature vector, and the LSP-transformed new feature vector, and selects an upper feature vector whose recognition rate is higher than the reference value. Since only the upper feature vector selected from the whole feature vectors is used for music genre classification, the selection unit 150 can reduce the amount of calculation required for feature extraction and improve the recognition rate.

The model generation unit 160 generates each music classification model based on the upper feature vector selected by the selection unit 150. [ Specific genres can be classified into Classical, Country, Disco, Hiphop, Jazz, Rock, Blues, Reggae, Pop, Metal and so on.

The music classification unit 170 classifies the genre or mood of the test music using the recognizer by using the upper vector extracted from the test music based on the music classification model generated by the model generation unit 160. [

Hereinafter, a music genre recognition method using the music classification apparatus 100 according to an embodiment of the present invention will be described in more detail.

2 is a flowchart of a music classification method according to an embodiment of the present invention.

The method for recognizing a music genre according to an embodiment of the present invention includes a learning step for generating a classification model as a whole and a testing step for classifying the genre or mood of the inputted test music through the generated model. As shown in FIG. 2, the learning step includes steps S210 to S250, and the testing step includes steps S260 to S300.

In order to finally classify music genres or moods, it is necessary to recognize the genre or mood of the music. In order to recognize the genre or mood of the music, a music classification model must be accurately generated.

In order to recognize the music genre, the music classification apparatus 100 according to the embodiment of the present invention first extracts a feature vector for each genre of music. The feature vectors include Short-term and Long-term feature vectors, LSP And a new feature vector for the transformed Short-term.

3 is a waveform diagram of an input music signal. FIG. 4 is a view for explaining an Analysis Window, and FIG. 5 is a view for explaining Texture Windows.

As shown in FIG. 3, when a music signal is input through the input unit 110, the short-term feature extraction unit 120 extracts a short-term feature vector for each frame through the window shown in FIG. 4 (S210). The short-term feature extraction unit 120 extracts a statistical feature vector using the texture window technique of FIG. 5 based on the short-term feature vector extracted from each frame (Analysis Window). The statistical features include at least one of Mean or Variance.

Hereinafter, the step S210 of extracting the timbre feature corresponding to the short-term feature vector will be described in more detail.

Short-term feature extraction according to one embodiment of the present invention is performed by using MFCC (Mel-Frequency Cepstral Coefficient), mel-scale band pass filter, and high pass Using a filter

DFB (Decorrelated Filter Bank), the characteristics of the music signal by octave band, and the Octave-based Spectral Contrast (OSC) used for music recognition.

Hereinafter, the MFCC algorithm will be described with reference to FIG.

6 is a feature extraction block diagram using the MFCC algorithm according to the embodiment of the present invention.

First, the short-term feature extraction unit 120 performs Fast Fourier Transform (FFT) after applying a Hamming window to each frame in the time domain. Then, after scaling the spectrum with a Mel-scale band-pass filter with B number of bands, the weighting sum of each spectrum WS (b) _{b = 1, ... , And B} are obtained. Then apply log to this value (1 ≤ b ≤ B). Finally, the short-term feature extraction unit 120 extracts a K-dimensional MFCC feature vector by applying Discrete Cosine Transform (DCT) to each weighted sum.

The feature that the short-term feature extraction unit 120 extracts using the MFCC is defined by Equation 1 below.

Next, the DFB (Decorrelated Filter Bank) will be described with reference to FIG.

7 is a feature extraction block diagram using the DFB algorithm according to the embodiment of the present invention.

를 통과하게 한다. 최종적으로Short-term 특징 추출부(120)는 아래의 수학식 2와 같이 D차원의 DFB 특징 벡터를 추출한다.The short-term feature extraction unit 120 generates a log (WS (b)) (1? B? B) by adding scaled values for each band as in MFCC through a DFB algorithm and applying log. As shown in FIG. 7, in the last step, an FIR High-pass-filter

. Finally, the short-term feature extraction unit 120 extracts the D-dimensional DFB feature vector as shown in Equation (2) below.

Next, an octave-based spectral contrast (OSC) will be described with reference to FIG.

8 is a feature extraction block diagram using an OSC algorithm according to an embodiment of the present invention.

Unlike MFCC or DFB, OSC extracts features based on octave rather than auditory model. The short-term feature extraction unit 120 extracts a feature vector considering the peaks and valleys of the spectrum of each band using the OSC algorithm. In most music, strong peaks are associated with the harmonic part, and stronger valleys are associated with the non-harmonic part. Therefore, OSC can represent the harmonic and non-harmonic components of music by considering the peak and valley values of spectrum per band.

As shown in FIG. 8, the OSC uses an octave-scale band-pass filter instead of the Mel-scale band-pass filter, unlike MFCC and DFB. do. In the present invention, eight octave-scale band-pass filters are used as shown in Table 1 below.

Band Frequency (Hz) One 0-100 2 100 to 200 3 200 to 400 4 400 to 800 5 800 ~ 1600 6 1600 to 3200 7 3200 ~ 8000 8 8000-22050

When the length of one frame and the FFT point are equal to N, the FFT spectrum for each frame is {x ₁ , x ₂ , ... , x _N }. If the number of FET points corresponding to the i-th band is K _{i ,} the spectrum of the i-th band is {x _{i , 1} , x _{i , 2} , ... , x _{i , Ki} }. To find peaks and valleys, first sort the spectra for each band in descending order. The spectrum of the ith band in descending order {x? _{, 1} , x? _{, 2} , ... , x? _{, Ki} }, we can extract Peak and Valley using the following equation.

In this case, α is a constant for the range of the peripheral value, and as a result of testing from 0.02 to 0.2, α has no significant effect on the performance, so it can be set to 0.02. The difference between the peak value and the valley value obtained by Equations (3) and (4) is calculated and the spectral contrast (SC _i ) is obtained as shown in Equation (5).

The OSC uses Spectral Contrast and Valley {SC ₁ , SC ₂ , ... , SC _I , V ₁ , V ₂ , ... , V _I } are used as feature vectors. In the present invention, 16-order OSC feature vectors are used because eight bands are used. For the i-th band, OSC uses Spectral Contrast and Valley as feature vectors such as {SC _i , V _i }.

Short-term feature extraction unit 120 extracts the mean (μ _t (k)) and the variance (σ _t ² (k)) of the feature vector extracted from the Analysis Window from the following equation 6 and equation 7 of .

In the above equation, X _t (k, p) is the kth element of the timbre feature of the p th frame in the t th Texture Window, and P is the total number of frames contained in the Texture Window.

The statistical feature vectors μ _t (k) and σ _t ² (k) obtained in the t-th texture window are averaged for the entire texture window.

When the short-term feature extraction unit 120 extracts the short-term feature, as shown in Equations (8) and (9), the long-term feature extraction unit 130 extracts the short- -term feature is extracted (S220). Hereinafter, a spectro-temporal feature using a modulation spectrum, which is one of the long-term features, will be described.

According to the embodiment of the present invention, the long-term feature extraction unit 130 extracts a feature vector using a modulation spectrum, such as Modulation Spectral Flatness Measures (MSFM), Modulation Spectral Crest Measures (MSCM) : Modulation spectral crest measurement).

In addition, the long-term feature extraction unit 130 extracts feature-based modulation spectral flatness measures (FMSFM) and feature-based modulation spectral crest measures (FMSCM) based on the MSFM / MSCM Based modulation spectral crest measurement) can be extracted.

MSFM / MSCM uses a modulation spectrum of an octave band sum (OBS), whereas the FMSFM / FMSCM according to an embodiment of the present invention uses a Feature-based Modulation Spectrum (FMS) . The long-term feature extraction unit 130 extracts a modulation spectrum for each Feature Dimension using the FMS, thereby knowing how each Feature Dimension changes with time.

Step S220 will be described in more detail. First, the long-term feature extraction unit 130 extracts an FMS as shown in Equation (10).

Where X _t (k, p) is the k th element of the timbre feature of the p th frame in the t th texture window, P _t is the total number of frames in the t th texture window, M is the modulated Fourier transform (Modulation Fourier-transform). The long-term feature extraction unit 130 extracts an average FMS for the entire texture window as follows.

Where T is the total number of texture windows.

Next, the long-term feature extraction unit 130 extracts the FMSFM / FMSCM using the results of Equation (10) and Equation (11) as follows.

A small value (for example, 0) and a large value (for example, 1 or 1 or more) of the FMSFM calculated in the equation (12) represent the average and the flatness of the average FMS, respectively. FMSCM shows a tendency to contradict with FMSFM. If the kth FMSFM has a very small value, the kth modulation frequency in the input music has a repetitive pattern.

The long-term feature extraction unit 130 uses the Octave Band Sum (OBS), which is the energy in the Octave Band, to extract the MSFM / MSCM.

Next, the AR modeling unit 140 extracts the linear prediction coefficients LPC using the timbre feature extracted from the short-term feature extraction unit 120, and outputs the extracted linear prediction coefficients LPC to the self- A new feature vector for the short-term whose degree is increased is modeled by a regression model (AR model) (S230). Then, the AR modeling unit 140 converts the new feature vector into LSP parameters.

Hereinafter, the process of generating the feature vector for the LSP-transformed Short-term in step S230 will be described in more detail with reference to FIG.

9 is a diagram for explaining an LSP conversion process according to an embodiment of the present invention.

The AR modeling unit 140 extracts a linear prediction coefficient (LPC) of a time axis using a Levinson algorithm, which is similar to a modulation spectrum but is not a Fourier transform.

When a linear prediction coefficient (LPC) is modeled through an autoregressive model (AR model) as shown in FIG. 9, a new feature vector having an order higher than that of the existing short-term feature vector is generated.

For example, when the inputted short-term feature vector is 13th, 130-degree new feature vectors are generated, which are increased 10-fold by using the autoregressive model (AR model).

However, since the linear prediction coefficient (LPC) has a large dynamic range, a recognition error may occur. Therefore, the AR modeling unit 140 converts the new feature vectors generated by the autoregressive model (AR model) Line Spectrum Pairs Parameter) to generate each genre / mood model.

The LSP parameter uses the linear prediction coefficient (LPC) only because the dynamic range is relatively smaller than the linear prediction coefficient (LPC), and the poles and zeros of the spectrum can be precisely estimated by the relative distance of adjacent LSPs. And has excellent performance compared to the case.

In this way, the short-term feature extraction unit 120, the long-term feature extraction unit 130, and the AR modeling unit 140 respectively generate feature vectors, and the feature vectors corresponding to all features extracted from the actual system are Table 2 shows examples.

Feature vectors Dimension Texture window Mean MFCC 13 DFB 13 OSC 16 Variance MFCC 13 DFB 13 OSC 16 Feature-based modulation spectrum FMSFM MFCC 13 DFB 13 OSC 16 FMSCM MFCC 13 DFB 13 OSC 16 Feature-based autoregressive model LSP MFCC 130 DFB 130 OSC 160 Total 588

As shown in Table 2, the short-term feature extraction unit 120, the long-term feature extraction unit 130, and the AR modeling unit 140 can generate 588 feature vectors for the input music . That is, as described in step S210, the short-term feature extraction unit 120 obtains thirteen feature vectors for the mean of MFCC for each of the thirteen orders among the music signals input through the texture window technique , 13 feature vectors are obtained with respect to the average value of DFB for each of the 13 orders.

As described in step S220, the long-term feature extraction unit 130 obtains 13 FMSFM-calculated feature vectors for each of 13 orders using the feature-based modulation spectrum, The FMSCM computed 13 feature vectors of the DFB are obtained.

Similarly, as described in step S220, the AR modeling unit 140 obtains 130 LSP-transformed feature vectors of the MFCC for each of the 130 orders whose orders are increased by a factor of 10 using the feature-based autoregressive model, 130 feature vectors of the DFB are obtained for each of the orders.

Next, the feature selecting unit 150 selects an upper feature vector from the extracted short-term or long-term feature vectors (S240). Here, the upper feature vector means a feature vector whose recognition rate is higher than the reference value.

As described above, according to the embodiment of the present invention, by selecting a feature vector corresponding to the priority among all the feature vectors unlike the prior art, the recognition rate can be improved and the amount of calculation can be reduced. Taking Table 2 as an example, the feature selecting unit 150 selects 160 higher feature vectors having excellent recognition rates from 588 feature vectors obtained. Therefore, in the test step, only the upper feature vector selected by the feature selecting unit 150 among all the extracted features is used for genre recognition.

The feature selector 150 according to the embodiment of the present invention can select an upper feature vector corresponding to a priority order by using a Best-first search and a method using a Support Vector Machine (SVM) ranker. That is, the feature selector 150 can select a feature vector by evaluating the value of each feature using an SVM recognizer called an SVM-ranker, and then rank the feature. The technique of ranking through the SVM-ranker is easy The detailed description will be omitted.

Next, the model generating unit 160 generates a music classification model using the selected upper feature vector (S240). Specific genres can be classified into Classical, Country, Disco, Hiphop, Jazz, Rock, Blues, Reggae, Pop, Metal and so on. To describe a specific method, the model generation unit 160 performs modeling for each genre of music using the extracted top feature vectors. That is, the model generation unit 160 groups the feature values of the upper feature vectors corresponding to each music genre and generates a model for each genre.

When the classification model is generated in this manner, a test step for classifying the genre or mood of the inputted test music is performed.

First, when a test music signal to be classified is input (S260), the music classifying unit 170 classifies a short-term feature vector, a long-term feature vector, New feature vectors generated by the AR model are sequentially extracted (S270, S280, S290). That is, in the case of Table 2, the music classifier 170 does not need to extract the feature vectors for all 588, but extracts only 160 feature vectors corresponding to the priorities.

For example, if it is assumed that the feature vectors of music signals corresponding to the first, fourth, sixth, ninth, and thirteenth orders among the thirteen orders are included in the priority order with respect to the mean value of MFCC obtained through the texture window technique , The music classification unit 170 extracts the feature vectors corresponding to the five test music signals.

Here, a step of extracting a short-term feature vector, a step of extracting a long-term feature vector, and a step of extracting a new feature vector generated by the AR model (S270, S280, S290) -term feature vector, a long-term feature vector, and a new feature vector, which are substantially the same as those of S210, S220 and S230 described above, and redundant explanations are omitted.

The short-term feature extraction unit 120 extracts a short-term feature vector, a long-term feature vector, and a new feature vector (S270, S280, and S290) The long-term feature extraction unit 130, and the AR modeling unit 140 may directly perform the test process.

Next, the music classifying unit 170 classifies the genre or mood of the inputted test music based on the classification model generated by the model generating unit 160 using the feature vector corresponding to the extracted upper feature vector (S300 ). In particular, according to an embodiment of the present invention, the music classifier 170 can classify music genres or moods using a one-against-one SVM.

As described above, according to the embodiment of the present invention, since only the upper feature vector is extracted in the test step and used to classify the music genre or the mood, the conventional technique of classifying music genre or mood by extracting all the feature vectors of the input test music The computation amount can be significantly reduced.

As one experimental example according to an embodiment of the present invention, a GTZAN database and an MIREX mood clustering method are used for music genre classification performance evaluation.

That is, performance evaluation was performed using GTZAN database for genre recognition and MIREX mood clustering method for mood recognition.

GTZAN includes 10 genres including Classical, Country, Disco, Hiphop, Jazz, Rock, Blues, Reggae, Pop and Metal. It has 100 songs per genre and 30 seconds per song in 16bit, 22050Hz, mono and AU file formats. Consists of.

Classifier Accuracy (%) genre Mood Order All feature set 82.4 62.2 588 + Feature selection 85.0 69.7 160

As shown in the accuracy results shown in Table 3, the upper 160 feature vectors are selected, compared with the case where all the 588 feature vectors are extracted to recognize the genre or the mood, and the feature vectors selected according to the embodiment of the present invention The recognition of the genre or the mood is highly accurate when the genre or the mood is recognized.

As described above, according to the method and apparatus for classifying music genres according to the embodiment of the present invention, it is possible to reduce the amount of calculation required for classifying genre of input music data through the process of selecting an upper feature vector after extracting features of music, Can be shortened. Even if the genre or mood of the music data inputted through the selection of the upper feature vector is classified, a high recognition rate can be obtained as compared with the conventional method using the entire feature vector.

In extracting feature vectors, LSP parameters based on the timbre feature are extracted and used as feature vectors by using an autoregressive model (AR model) in addition to the timbre feature, so that the performance of genre and mood recognition system Can be improved.

The present invention has been described above with reference to the embodiments. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. Therefore, the scope of the present invention is not limited to the above-described embodiments, but should be construed to include various embodiments within the scope of the claims and equivalents thereof.

100: music classification device, 110: input part,
120: Short-term feature extraction unit, 130: Long-term feature extraction unit,
140: AR modeling unit, 150: feature selecting unit,
160: model generation unit, 170: music classification unit

Claims (17)

An input unit for receiving a sound source;
A short-term feature extraction unit for extracting a short-term feature vector for a tone color feature of the input sound source;
A long-term feature extraction unit for extracting a long-term feature vector using the short-term feature vector;
Term feature vector (LPC) using the extracted short-term feature vector, and modeling the extracted LPC with an autoregressive model to obtain a new feature An AR modeling unit for generating a vector and converting the vector into an LSP parameter;
A feature selecting unit for selecting the short-term feature vector, the long-term feature vector, and an upper feature vector having a higher recognition rate from among the new feature vectors;
A model generation unit for generating a music classification model using the selected upper feature vector; And
And a music classifier for classifying a genre or a mood of the test music inputted based on the classification model,
The short-term feature vector, the long-term feature vector,
A Mel-Frequency Cepstral Coefficient (MFCC) and a Decorrelated Filter Bank (DFB)
The short-term feature vector includes a statistical feature for a tone color extracted using a texture window technique,
Wherein the statistical characteristic includes an average value and a variance value. delete delete The method according to claim 1,
The long-term feature vector may be expressed as:
Based modulation spectral flatness measures (FMSFM) and feature-based modulation spectral crest measures (FMSCM) extracted using Feature-based Modulation Spectrum (FMS)
Wherein the FMSFM and the FMSCM are defined by the following equations.

5. The method of claim 4,
The average of the FMS and the FMS,
A music classification apparatus defined by the following equation:

Here, T is the number of total texture windows, X _t (k, p) is the kth element of the short-term feature of the pth frame in the tth texture window, P _i is the total number of frames And M is the magnitude of the Modulation Fourier-transform. The method according to claim 1,
The long-term feature vector includes a statistical feature,
Wherein the statistical characteristic includes an average value and a variance value. The method according to claim 1,
Wherein the music classification unit comprises:
Extracting a feature vector corresponding to the upper feature vector with respect to the test music and comparing the extracted feature vector with the classification model based on the extracted feature vector, Or mood. The method according to claim 1,
Wherein the feature selecting unit comprises:
Selects an upper feature vector using a SVM (Support Vector Machine) ranker,
Wherein the music classification unit comprises:
A music classification apparatus for classifying a genre or a mood of the test music using a one-against-one SVM (Support Vector Machine). A music classification method using a music classification apparatus,
Extracting a short-term feature vector for a tone color feature of the input sound source;
Extracting a long-term feature vector using the short-term feature vector;
Extracting a linear prediction coefficient (LPC) using the extracted short-term feature vector;
Modeling the extracted LPC with an autoregressive model to generate a new feature vector for the short-term feature vector with an increased order and transforming the new feature vector into LSP parameters;
Selecting an upper feature vector having a higher recognition rate from the Short-term feature vector, the Long-term feature vector, and the new feature vector; And
Generating a music classification model using the selected upper feature vector;
Classifying the genre or mood of the test music inputted based on the classification model,
The short-term feature vector, the long-term feature vector,
A Mel-Frequency Cepstral Coefficient (MFCC) and a Decorrelated Filter Bank (DFB)
The short-term feature vector includes a statistical feature for a tone color extracted using a texture window technique,
Wherein the statistical feature includes an average value and a variance value. delete delete 10. The method of claim 9,
The long-term feature vector may be expressed as:
Based modulation spectral flatness measures (FMSFM) and feature-based modulation spectral crest measures (FMSCM) extracted using Feature-based Modulation Spectrum (FMS)
Wherein the FMSFM and the FMSCM are defined by the following equation.

13. The method of claim 12,
The average of the FMS and the FMS,
A music classification method defined by the following equation:

Here, T is the number of total texture windows, X _t (k, p) is the kth element of the short-term feature of the pth frame in the tth texture window, P _i is the total number of frames And M is the magnitude of the Modulation Fourier-transform. 10. The method of claim 9,
The long-term feature vector includes a statistical feature,
Wherein the statistical feature includes an average value and a variance value. 10. The method of claim 9,
Wherein classifying the genre or mood of the input music based on the classification model comprises:
Extracting a feature vector corresponding to the upper feature vector with respect to the test music when the test music to be classified is input; And
And classifying the genre or mood of the test music by comparing the extracted feature vector with the classification model. 10. The method of claim 9,
In the step of selecting a feature whose recognition rate is higher than a reference value,
SVM (Support Vector Machine) ranker,
In the step of classifying the genre or the mood of the inputted music based on the classification model,
A music classification method for classifying a genre or a mood of the test music using a one-against-one SVM (Support Vector Machine). 10. The method of claim 9,
The step of extracting a linear prediction coefficient (LPC) using the extracted short-term feature vector comprises:
And extracting the linear prediction coefficient (LPC) using a Levinson algorithm.

Images