KR101356039B1 - Blind source separation method using harmonic frequency dependency and de-mixing system therefor - Google Patents

Abstract

The blind signal separation method according to the present invention is a method of separating signals in a frequency domain, and the signals from two or more sound sources are mixed and received only between frequency bins included in the same harmonic frequency group among a plurality of harmonic frequency groups. Demixing the received signal assuming a dependency.

Description

Blind Signal Separation Method Using Dependency Between Harmonic Frequency and Demixing System for It {BLIND SOURCE SEPARATION METHOD USING HARMONIC FREQUENCY DEPENDENCY AND DE-MIXING SYSTEM THEREFOR}

The present invention relates to a method for separating tacit signals using dependencies between harmonic frequencies and a demixing system therefor, and more particularly, to signals through independent vector analysis based on dependencies between harmonic frequencies in speech and / or music signals. It relates to a blind signal separation algorithm for separating the.

Independent Component Analysis (ICA) is a Blind Source Separation (BSS) algorithm that uses statistical independence between output signals. Frequency Domain Independent Component Analysis (FDICA) has been used for the convolutive BSS algorithm because it allows the convolution of the convolutional signal in the time domain in the frequency domain to an instantaneous mixture Because it can be modeled. This modeling can simplify the separation problem. FDCIA successfully separates the signal components of each frequency channel. However, the random permutation problem of separated frequency components between frequency bins occurs.

Independent Vector Analysis (IVA), a multivariate extension of independent component analysis, solves the permutation uncertainty problem by utilizing dependencies between frequency components. Studies based on conventional basic IVA models (KIM, T., ATTIAS, HT, LEE, S.-Y., and LEE, T.-W .: 'Blind source separation exploiting higher-order frequency dependencies', IEEE In Trans.Audio Speech Lang.Process., 2007, 15, (1), pp. 70-79), the sound source signal is represented as a frequency component vector, and independent vector analysis is a priori a combined probability density function that is radially symmetric in full band. (PDF: Probability Density function) can be used to model the sound source signal to isolate the signal. Here, the probability density function assumes a dependency between frequency components. More specifically, in the basic IVA model as described above, there is a dependent relationship between the first assumption that there is a probabilistic independent relationship between each signal before mixing and the different frequency components of each signal. A second assumption was used.

Subsequently, an independent vector analysis method is presented that represents an improved frequency dependency model. This is a subband local group IVA model (LEE, I., JANG, G.-J., and LEE, T.-W .: 'Independent vector analysis using densities represented by chain-like overlapped cliques in graphical models for separation of convolutedly mixed signals', Electronics Letters, 2009, 45, (13), pp. 710-711). This method maintains the first hypothesis of the basic IVA model but assumes a dependent relationship between near frequency components even between different frequency components of each signal, and corrects the second hypothesis that there are no dependencies between distant frequency components. Indicates.

In recent years, there is a growing need for technology that can effectively separate signals from sound sources such as voice or music, which we usually encounter.

Korean Laid-Open Publication No. 10-2008-0019879 (2008.03.05)

Disclosure of Invention The present invention has been made to solve the problems of the prior art, and an object of the present invention is to provide a technique capable of effectively separating signals through independent vector analysis for sound source signals having a harmonic frequency structure, such as voice and / or music signals. It is done.

The technical objects to be achieved by the present invention are not limited to the above-mentioned technical problems, and other technical subjects which are not mentioned can be clearly understood by those skilled in the art from the description of the present invention .

In accordance with an aspect of the present invention, there is provided a method of separating a signal in a frequency domain, the method comprising: receiving a mixture of signals from two or more sound sources; And demixing the received signal assuming a dependency only between frequency bins included in the same harmonic frequency group among a plurality of harmonic frequency groups.

Demixing the received sound source signal according to the present invention may be performed through independent vector analysis.

Demixing system for blind signal separation according to the present invention is a demixing system for separating the signal in the frequency domain, the signal receiving unit for receiving a mixture of signals from two or more sound sources; And a demixing filter for demixing the received signal assuming a dependency only between frequency bins included in the same harmonic frequency group among a plurality of harmonic frequency groups.

According to the present invention, it is possible to provide an algorithm and a method for effectively separating signals through independent vector analysis on a sound source signal having a harmonic frequency structure, such as a voice and / or music signal. In addition, according to the present invention, it is possible to prevent the frequency bin permutation problem at the time of blind signal separation.

1 illustrates an environment of a demixing system in which a blind signal separation algorithm may be executed in accordance with an embodiment of the present invention.
2 is a harmonic of a wide frequency group of the basic IVA model, (b) a subband local frequency group of the subband local IVA model, and (c) a harmonic frequency group IVA model according to an embodiment of the present invention. Represents a frequency group.
Figure 3 shows a simulation environment for testing the performance of the blind signal separation algorithm according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a detailed description of preferred embodiments of the present invention will be given with reference to the accompanying drawings. The shape and the size of the elements in the drawings may be exaggerated for clarity of explanation and the same reference numerals are used for the same elements and the same elements are denoted by the same quote symbols as possible even if they are displayed on different drawings Should be. In the following description, well-known functions or constructions are not described in detail to avoid unnecessarily obscuring the subject matter of the present invention.

The signal mainly dealt with by human is voice or music signal. Looking at the spectrum of the voice signal and the music signal, it can be seen that a harmonic relationship exists between their frequency components. That is, the voice signal and the music signal may be referred to as having a strong harmonic structure.

In an embodiment of the present invention, in addition to the voice signal and the music signal, for a sound source signal having harmonic frequency components, a technique capable of separating signals by using dependencies between the harmonic frequency components is proposed. That is, an embodiment of the present invention proposes an improved frequency dependent model for independent vector analysis using dependency relations between harmonic frequency components. Compared with the conventional frequency dependent model, the frequency dependent model according to the embodiment of the present invention is very effective for separating sound signals having strong harmonic structures such as voice and music signals.

1 illustrates an environment of a demixing system in which a blind signal separation algorithm may be executed in accordance with an embodiment of the present invention. As shown in FIG. 1, it can be considered that a sound source signal is mixed from two or more sources 10 and 12 and received by one or more signal receiving units 20 and 22. 1 illustrates an indoor environment. Therefore, the signals from the sources 10 and 12 not only reach the signal receiving units 20 and 22 through the direct paths D11, D12, D21 and D22, , R22). The sound source signal thus received may be input to the demixing system 30. [ The de-mixing performed through the demixing system 30 may separate the received sound source signals. Hereinafter, the demixing system 30 may be referred to as a concept including the signal receivers 20 and 22.

At this time, the state in which there is no information about the sound source signal or the mixed environment is referred to as the "implicit" state. That is, an embodiment of the present invention provides an algorithm for separating the received signal in the tacit state.

In the following, independent vector analysis based on dependencies within a harmonic frequency clique according to an embodiment of the present invention will be described.

First, the mixed signal received by the demixing system 30 is represented in the frequency domain through a Short-Time Fourier Transform (STFT). The convolutional rock signal separation algorithm in the frequency domain begins by approximating a bin-wise instantaneous mixing model. For each frequency bin, the model can be formulated as follows:

_{수학식 (1)}

_{Equation (1)}

Here, the subscript k indicates the frequency bin index. y _k denotes a signal that is demixed by the demixing system 30 and is separated, x _k denotes a signal received by the signal receiving units 20 and 22 and input to the demixing system 30, and s _k represents a signal from the sound sources 10 and 12.

Although FIG. 1 illustrates two sound sources and two microphones, it is assumed that there are N sound source signals and N microphones. Each of the NXN matrices A _k and W _k represents an instant mixing and de-mixing matrix. That is, A _k represents a transfer function for the path of the sound source signal, and W _k represents a transfer function of the demixing filter. Therefore, it should be possible to obtain the signal s _k from the sound source by multiplying the received signal x _k by the demixing matrix W _k . K represents the number of frequency bins and t represents the time index of the frame.

In the embodiment of the present invention, the separation signal y may be represented using a multivariate probability density function proposed by Equation (2) below.

_{수학식 (2)}

_{Equation (2)}

Here, y _ik represents the i-th element of y _k = [y _lk , ..., y _Nk ] ^T. H is the total number of harmonic groups, and σ _hk is the standard deviation of the absolute value of the separated signal of the k-th frequency bin group belonging to the harmonic group h, for example, can be set to one. Ch represents a frequency bin group belonging to the harmonic group h. The fundamental frequency of H is represented by Fh and defined as follows.

_{수학식 (3)}

_{Equation (3)}

In an embodiment of the present invention, F1 may be set to 55 Hz, for example. Also, the total number of harmonic populations can be illustrated as 49. The fundamental frequencies of these harmonic groups represent note frequencies from A1 (55 Hz) to A5 (880 Hz). This frequency range may cover the entire frequency of the human speech signal. Under the condition that 1≤h≤H-1, the following equation may be established.

_{수학식 (4)}

_{Equation (4)}

Where f _k denotes the frequency of the k-th bin and Ch denotes the h-th local frequency clique. Population C _k contains frequency bins of the first eight multiples of F _h (ie, M = 8). The bandwidth of the m th multiple frequency of F _h (ie, mF _h ) is 2δmF _h , and two consecutive harmonic groups can overlap each other. At this time, it is illustrated that 50% of consecutive harmonic groups overlap. The population C _H = {1, K, K } includes all frequency bins. This can prevent permutation of frequency bins with frequencies less than 55 Hz and improve the learning speed of the demixing filter. Figure 2 (c) shows a harmonic frequency group of the harmonic frequency group IVA model according to an embodiment of the present invention. It can be seen from FIG. 2C that 49 harmonic frequency groups are illustrated on the vertical axis, and 8 harmonic frequency bins are included in the same harmonic frequency group. In this case, the number of harmonic frequency groups, the number of harmonic frequency bins included in the harmonic frequency group, and the degree of overlap of consecutive harmonic groups may be variously changed according to an embodiment of the present invention.

In order to demix the received signal to the demixing system 30, it is necessary to calculate a transfer function W for demixing in the demixing filter. For example, the parameters of the transfer function W may be obtained from the filter parameter calculator in the following manner.

In order to perform independent vector analysis while maximizing the characteristics of the dependency model between frequency bins, the following log-likelihood function can be used as a cost function.

_{수학식 (5)}

_{Equation (5)}

In order to obtain an optimized separation signal, the following natural gradient learning rule may be applied.

_{수학식 (6)}

_{Equation (6)}

Where I represents the unit matrix of NXN. Φ (y _k ) denotes a column vector of NX 1, where the i th element is defined as follows.

_{수학식 (7)}

_{Equation (7)}

Here, S _k is an index group of groups including the k th frequency bin.

In an embodiment of the present invention, the filter parameter calculator (not shown) included in the demixing system 30 may obtain a transfer function W for demixing using a cost function expressed as Equation (5). Can be. Thereafter, the signal received by the demixing system 30 is demixed in the demixing filter using the calculated transfer function W. FIG.

In this case, the filter parameter calculator may receive an output from the demixing filter, and repeatedly obtains the filter parameter according to the learning rule expressed by Equation (6) and supplies the filter parameter to the demixing filter. Thus, the demixing filter can operate adaptively. That is, the transfer function W may be adaptively obtained by repeatedly performing a calculation according to the learning rule of Equation (6). Thereafter, it is determined whether or not the transfer function W converges. If not, the flow returns to the previous step to calculate the transfer function W again and perform demixing.

The separated signal y may be obtained using the transfer function W obtained in this adaptive manner, which may then be transformed to be expressed in the time domain as needed.

Simulation results ( Simulation Results ):

Various 2 × 2 BSS experiments were performed to evaluate the performance of the proposed algorithm according to an embodiment of the present invention. In this experiment, 8-s-length real speech signals from the TIMIT database, monophonic music signals at 8 kHz sampling rate, and imaging (see ALLEN, JB, and BERKLEY, DA: Room impulse responses generated by 'Image method for efficiently simulating small-room acoustics', J. Acoust.Soc.Am., 1979, 65, (4), pp. 943-950) were used. .

Figure 3 shows a simulation environment for testing the performance of the blind signal separation algorithm according to an embodiment of the present invention. Two microphones (1, 2) and two source signals (A to K) were placed in a cube room measuring 7m x 5m x 2.75m. The reverberation time was set to 100 ms and the reflection coefficients of the walls, floor and ceiling were set to 0.57. In this experiment a 2048-point Fast Fourier Transform (FFT), a 2048-tab hanning window, and a shift size of 512 samples were used.

In order to compare the performance of the proposed model with the performance of other models according to an embodiment of the present invention, in this experiment, the basic IVA (see: KIM, T., ATTIAS, HT, LEE, S.-Y., and LEE, T.-W .: 'Blind source separation exploiting higher-order frequency dependencies', IEEE Trans.Audio Speech Lang.Process., 2007, 15, (1), pp. 70-79) and subband local population IVA (See LEE, I., JANG, G.-J., and LEE, T.-W .: 'Independent vector analysis using densities represented by chain-like overlapped cliques in graphical models for separation of convolutedly mixed signals', Electronics Letters, 2009, 45, (13), pp. 710-711) the same experiment was carried out. The subband local group consists of seven 256-bin shifted groups that are 128-bin shifted. 2 (a) and 2 (b) show the wideband frequency band of the basic IVA model and the subband local frequency band of the subband local IVA model, respectively.

The separation performance of the blind signal separation algorithm may be measured in terms of signal-to-interference ratio (SIR) defined by Equation (8) below.

_{수학식 (8)}

_{Equation (8)}

와 같이 정의되는 전역 임펄스 응답을 나타내며, 여기서 a_ijk 및 w_ijk는 각각 A _k와 W _k의 (i, j)번째 성분을 나타낸다. Where q (i) represents the separated sound source index at which the i-th sound source appears, and h _{iq (j)}

And refers to the global impulse response is defined as, where a and w _{ijk ijk} represents the (i, j) th element of A _k and W _k, respectively.

The experimental results of separating the two voice signals and the experimental results of separating the voice signal and the music signal are shown in Tables 1 and 2, respectively. Table 1 shows that two voice signals were used in the experiment, and Table 2 shows that one voice signal and one music signal were used in the experiment.

Experiment number One 2 3 4 5 6 7 Source locations A, I B, G E, G I, K C, D E, F I, J Input SIR 0.7 0.0 0.0 -0.4 -0.0 0.0 -0.4 Original IVA 16.1 14.2 14.5 8.7 10.2 14.5 14.1 Sub-band local clique IVA 19.9 17.3 17.3 8.3 11.9 18.3 15.0 Harmonic clique IVA 21.3 18.6 18.1 10.1 13.4 19.2 15.8

Experiment number One 2 3 4 5 6 7 Source locations A, I B, G E, G I, K C, D E, F I, J Input SIR 0.2 0.4 0.3 -0.5 -0.1 0.1 -0.4 Original IVA 10.2 8.9 9.7 6.8 7.0 9.2 10.0 Sub-band local clique IVA 14.2 13.2 10.8 9.2 8.9 12.6 12.4 Harmonic clique IVA 15.4 14.7 14.4 9.6 10.2 14.1 13.0

As can be seen from Table 1 and Table 2, when the blind signal separation algorithm according to the embodiment of the present invention, denoted Harmonic clique IVA, is used in terms of SIR, original IVA and subband local population IVA (Sub-). consistently better performance than band local clique IVA).

As described above, in the embodiment of the present invention, for a sound source signal having a harmonic structure such as voice and music signals, a BSS algorithm considering such a strong harmonic structure is proposed. Harmonic groups a priori assign a dependency between frequency bins for multiples of the fundamental frequencies of the sound source. According to the blind signal separation algorithm proposed according to the embodiment of the present invention from the above simulation results, it is possible to more effectively achieve separation of voice and music signals than the conventional BSS algorithm.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art. will be. Therefore, it should be understood that the above-described embodiments are to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than the foregoing description, It is intended that all changes and modifications derived from the equivalent concept be included within the scope of the present invention.

10, 12: source
20, 22: Signal receiving section
30: De-mixing system

Claims (12)

In a method for separating a signal in the frequency domain,
Mixing and receiving signals from two or more sound sources; And
Demixing the received signal assuming a dependency only between frequency bins included in the same harmonic frequency group among a plurality of harmonic frequency groups;
The blind signal separation method comprising a. The method of claim 1,
Wherein the demixing comprises:
A blind signal separation method characterized in that it is carried out through an independent vector analysis. The method of claim 1,
Wherein the demixing comprises:
Cost function

Calculate the transfer function (W) using

Repeatedly calculating the transfer function according to the transfer function; And
And demixing the received signal using the transfer function,
Where L is the log likelihood function

A blind signal separation method characterized in that. Claim 4 has been abandoned due to the setting registration fee. 4. The method according to any one of claims 1 to 3,
And a signal from the sound source comprises at least one of a voice signal and a music signal. 4. The method according to any one of claims 1 to 3,
The method of claim 2, wherein consecutive harmonic frequency groups of the plurality of harmonic frequency groups overlap each other. 4. The method according to any one of claims 1 to 3,
The plurality of harmonic frequency groups includes 49 harmonic frequency groups, and each of the plurality of harmonic frequency groups includes frequency bins consisting of the first eight multiples of the corresponding fundamental frequency Fh. Way. In a demixing system for separating signals in the frequency domain,
A signal receiving unit which receives signals from two or more sound sources mixed together; And
A demixing filter demixing the received signal assuming a dependency only between frequency bins included in the same harmonic frequency group among a plurality of harmonic frequency groups;
Demixing system for blind signal separation comprising a. Claim 8 was abandoned when the registration fee was paid. The method of claim 7, wherein
The blind signal separation is a demixing system for blind signal separation, characterized in that performed through independent vector analysis. Claim 9 has been abandoned due to the setting registration fee. The method of claim 7, wherein
Cost function

Calculate the transfer function (W) using

Further comprising a filter parameter calculator for repeatedly calculating the transfer function according to,
Where L is the log likelihood function

Demixing system for blind signal separation, characterized in that. Claim 10 has been abandoned due to the setting registration fee. 10. The method according to any one of claims 7 to 9,
The signal from the sound source may comprise any one or more of a speech signal and a music signal. Claim 11 was abandoned when the registration fee was paid. 10. The method according to any one of claims 7 to 9,
And a continuous harmonic frequency group among the plurality of harmonic frequency groups overlaps each other. Claim 12 is abandoned in setting registration fee. 10. The method according to any one of claims 7 to 9,
The plurality of harmonic frequency groups includes 49 harmonic frequency groups, and each of the plurality of harmonic frequency groups includes frequency bins consisting of the first eight multiples of the corresponding fundamental frequency Fh. Demixing system for

Images