JP2001027895A - Signal separation and apparatus therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To synthesize signals of original signal sources respectively independently or to an arbitrary number of signals from mixed signals from plural signal sources by continuously analyzing the mixed signals with lapse of time, separating the same by a single signal source and reconstituting the signal of the desired signal source from the results of the separation. SOLUTION: When a frequency component appears in a proximity frequency band, a harmonic structure decision section forms a harmonic structure and decides this structure as a candidata for the frequency component constituting the single signal source. Next, the manipulation to change frequency resolution is repeated by a frequency resolution changing section to detect the harmonic component group of integer times in a central frequency and is determined as the frequency component of the candidate for constituting the signal from the single signal source. Next, the rise and fall of the previously obtained frequency component are compared by noticing a Bregman rule and the frequency component which is the same in either of the rise or fall time is extracted. The frequency component of the single signal source is identified and is reconstituted, by which the separation of the single signal source is made possible.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for separating signals, and more specifically, for example, to a mixed sound of a voice signal and environmental noise in a living environment and a simultaneous performance sound of a plurality of musical instruments. The present invention relates to a method and an apparatus for separating a signal of a desired sound source from a mixed signal of acoustic signals from a plurality of sound sources.

[0002]

2. Description of the Related Art Conventionally, a comb filter or the like has been used for separating an acoustic signal, but it has been difficult to separate a signal other than a sound signal or a sound signal in a specific frequency band.

In the acoustic signal separation method described in Japanese Patent Application Laid-Open No. 10-228296, it is possible to separate a mixed signal into a plurality of sound sources. However, since a template matching method is basically adopted, a target signal It is necessary to store all the waveforms that may be included in the template as a template. Therefore, when the properties of the signal or the sound source included in the target signal are unknown, the signal cannot be accurately separated and lacks flexibility.

Japanese Patent Application Laid-Open No. 05-127668 discloses a method for automatically transcribing MIDI codes and the like from audio signals by wavelet transform. However, this publication discloses that a signal source is not a pure tone of a single frequency component, but a complex signal in which a plurality of frequency components are superimposed, and a mixed signal having a complicated configuration in which a plurality of signal sources exist. On the other hand, means for accurately obtaining a MIDI code is not disclosed.

[0005]

However, in a normal living environment or the like, sound signals or sound signals from a plurality of unspecified sound sources are mixed. In such a situation, a means for separating a signal from a desired sound source without specifying a frequency band, a signal waveform, or the like is desired.

SUMMARY OF THE INVENTION It is an object of the present invention to satisfy such a demand and to provide a signal separating method and apparatus for separating a signal of a desired signal source from a mixed signal of signals from a plurality of signal sources.

[0007]

SUMMARY OF THE INVENTION A signal separation method according to the present invention is a signal separation method in which an observation signal separates a mixed signal of signals from a plurality of signal sources. And a reconstructing step of reconstructing a signal of a desired signal source from a result of the separation in the separating step.

A signal separating apparatus according to the present invention is a signal separating apparatus for separating a mixed signal of observation signals from signals from a plurality of signal sources. It is characterized by having analysis separation means for separating each source, and reconstruction means for reconstructing a signal of a desired signal source from the separation result of the separation means.

[0009]

Embodiments of the present invention will be described below in detail with reference to the drawings.

FIG. 1 is a diagram showing a schematic configuration and a flow chart of an embodiment of the present invention. For audio signals, Breg
There are four heuristic rules of man. That is, (1) rules relating to common rise / fall, (2) rules relating to asymptotic changes, (3) rules relating to harmonic relationships, and (4) rules relating to changes occurring in one acoustic event. The last rule (4) means that a change occurring in one acoustic event has a similar effect on each component constituting the sound.

In this embodiment, a desired acoustic signal is separated from a mixed signal composed of a plurality of acoustic signals by using these rules or rules as physical constraints.

Paying attention to Bregman's law (3), it utilizes the fact that the frequency components of many instrument sounds excluding some instruments such as percussion instruments form a harmonic structure. Decompose into frequency components of the input signal with a predetermined frequency resolution, and extract the fundamental frequency component and the frequency component whose center frequency is in a harmonic relationship,
The change of the frequency resolution is sequentially repeated until the harmonic structure is detected.

In this embodiment, it is assumed that, at the start of the analysis, the fundamental frequency and the frequency of the harmonic component contained in the observation signal (signal to be separated) are unknown. For example, referring to the audible range of a person, the audible range of a person is generally about 20-2.
Since the frequency is set to be 0000 Hz, the frequency resolution to be initially set is set to the lower frequency side of the audible range, and the lowest pass frequency is set to 2
Set to 0Hz, frequency 40 which is an integral multiple of the lowest passing frequency
Hz, 60Hz, 80Hz, ..., 20,000Hz
Is formed, and the observation signal is filtered by setting the pass bandwidth to 20 Hz which is equivalent to the fundamental frequency.

If a frequency component appears in an adjacent frequency band, the harmonic structure determination unit determines a candidate for a frequency component forming a harmonic structure and constituting a single signal source. Even if the first frequency resolution is determined by the initial frequency resolution setting function without knowing the nature of the observation signal, it is difficult to capture the harmonic structure well. Therefore, in the present embodiment, the frequency resolution changing unit repeats, for example, an operation of sequentially shifting the lowest pass frequency. In the second operation, when the lowest pass frequency is changed to nHz on the high frequency side, the lowest pass frequency becomes (2
0 + n) Hz, and the pass frequency on the high frequency side is (40 + 2).
n) Hz, (60 + 3n) Hz, (80 + 4n) Hz,
.., (20,000 + 1000n) Hz. By repeating this frequency resolution changing operation, a harmonic component group whose center frequency is an integral multiple is eventually detected, which becomes a candidate frequency component constituting a signal from a single signal source. By this means for sequentially changing the frequency resolution and repeating and repeating the frequency decomposition, it is possible to obtain a candidate for a frequency component of a harmonic structure constituting a signal from an unknown signal source.

As described above, if the frequency resolution is simply shifted, the optimum frequency resolution may not be detected by skipping the optimum frequency resolution when the setting of the shift width is too wide. Alternatively, the frequency component may not be detected because the pass bandwidth of the filter does not match the bandwidth of the analysis signal. However, the frequency resolution changing unit and the analysis condition changing unit change the analysis conditions such as the frequency resolution or, if a filter is used, the pass bandwidth of the filter until the harmonic structure is detected, and the harmonic structure is changed. The operation of changing the analysis conditions is repeated until a detection is made. Thus, it is possible to prevent a false determination that there is no harmonic structure.

In order to expedite the search for the harmonic structure, for example, without limiting the details of the analysis condition changing means, it is sufficient to refer to information on the magnitude of the signal of the decomposed frequency component. It is appropriate to use some convergence algorithm, such as Newton's method or steepest descent method. In the case of using the Newton method, a function that defines the gradient of the shift width of the frequency resolution and the change in the magnitude of the signal of the detected frequency component is defined, and the solution is obtained. The optimum frequency resolution can be obtained.

In the flow shown in FIG. 1, "analysis continuation judging section 1: whether or not the analysis end condition has been reached" means that there is no truly harmonic structure, or that there are practically negligible frequency components. Is provided to avoid falling into an infinite loop. A predetermined analysis end condition is determined in advance, and the process is forcibly ended.

Focusing on Bregman's rule (1), the rise and fall of the previously obtained frequency components are compared, and frequency components having at least one of the same rise and fall times are extracted. Focusing not only on the harmonic structure but also on the rise and fall times, the frequency components of different signal sources, but as a result of the fact that the frequency bands are close for some reason,
This is because the components that are regarded as the frequency components of the same signal source are discriminated and removed only by the above-described means, and the frequency components of different signal sources are prevented from being erroneously reconstructed. As a result, signal source separation accuracy is improved.

The reason why it is possible to select whether to pay attention to both the rise time and the fall time or only one of them is as follows. That is, first, when scanning the entire analysis target signal at once, it is generally considered that focusing on both the rise time and the fall time improves the separation accuracy. However, if the duration of the signal is unknown,
Not necessarily. In this embodiment, the time length to be read for analysis is not particularly defined with respect to the time length of the entire observation target signal. Therefore, when the analysis time length is divided into short lengths and the analysis is repeated continuously in time, there is no guarantee that both the rise and the fall can always be observed within the analysis time. For example, if the analysis time is relatively short with respect to the duration of the signal, the frequency components are temporally separated, and only one of the rise and fall may be observed within the same analysis time.

Second, the apparent rise time or fall time is apparent due to the detection accuracy of the sound pressure level at the time of sampling the signal from the signal source, and the rise and fall characteristics of the sound pressure level of the signal source. May shift. If the temporal rise of the signal magnitude is extremely slow or the detection accuracy is insufficient for the observation signal level, a time later than the true signal rise time is detected as the rise time. If the rise time is detected with all frequency components shifted by the same time width, the effect on the analysis is small, but originally,
Even if the same signal source has a specific rise in apparent rise time, there is a concern that a specific frequency component may be erroneously recognized as a frequency component of a signal from another signal source. Similarly, when analyzing the fall of a signal, it may be erroneously recognized that the fall is faster than the true signal fall time.

When the frequency component of a single signal source is extracted by paying attention only to the frequency component where both the rise time and the fall time coincide, when the first or second situation occurs, excessively There is a danger that components will be shaken off, which may lead to deterioration of the quality of the reproduced signal.

Therefore, in the present embodiment, when either one of the rise time and the fall time coincides, the frequency component of the single signal source is identified. The first reason for shortening the analysis time length in the present embodiment is to reduce the used memory capacity, but the second reason is that the analysis is performed continuously, with a slight time delay from the reading of the observation signal. This is because a signal analysis result can be obtained, and an almost real-time analysis environment can be obtained.

In this embodiment, attention is also paid to Bregman's rule (2), and the frequency components are temporally divided according to the representative magnitude of the signal. An object of the present invention is to be able to separate a signal component of another signal source when the signal component is superimposed on the same frequency band. Applying Bregman's rule (2), it is considered that the sound pressure level basically changes asymptotically within the same signal source. That is, even if the signal magnitude suddenly increases or decreases even within the same frequency band, if a discontinuity is observed in the signal magnitude,
Judgment is made of a signal from another signal source, and the signal is temporally separated and then reconstructed.

Further, the present embodiment can cope with a signal source which shows a sudden rise in sound pressure level like a piano sound.
That is, as described above, if the representative magnitude of the signal is merely averaged, the sudden rising peak is regarded as a discontinuous point, and the peak position is classified as another signal source.
Here, when Bregman's law (4) is applied, a peak point is also observed in other frequency-resolved components, and a plurality of frequency components are grouped by using this as a representative magnitude of a signal to obtain a frequency of the same signal source. Can be extracted as a component. That is, even when components of different signal sources are superimposed on the same frequency band, these components can be separated by comparison with other frequency components.

By providing such means, the frequency component of a single signal source can be identified and reconstructed to separate the signal of the single signal source.

Furthermore, in this embodiment, even if the signal source is once identified and reconstructed and output by the analysis continuation determining unit 2 and the analysis continuation determining unit 3 shown in FIG. In this case, the frequency resolution is further updated to a higher frequency, and the above processing is performed again. Thus, different single sound sources having the fundamental frequency shifted to the higher frequency can be sequentially identified, reconstructed, and output. This iterative operation
Unless a forced analysis stop input is received from the outside, the processing is continued until a predetermined frequency resolution is reached or no signal component is present. Thereby, all the signal sources included in the observation signal can be identified, and the output signal can be extracted.

In this manner, the signals of the respective sound sources are reproduced independently. However, by changing the signal as shown in the flowchart of FIG. 2, an arbitrary number of signal sources can be synthesized and output simultaneously. Will be possible. That is, the reconstructed single sound source signals are sequentially stored in the reconstructed signal storage unit, and only the signal of the designated signal source is synthesized and output.

For example, assume that a reconstructed signal output result is obtained from a mixed signal of signals from four musical instruments A, B, C, and D.
When the reconstructed signal output result of each instrument and the storage unit storage number corresponding to each instrument are known, by inputting the storage number of the reconstructed signal storage unit to the reconstructed signal combining unit as a composite signal instruction,
For example, a synthesized signal output obtained by synthesizing only the signals of the musical instrument A and the musical instrument B can be obtained. If the musical instruments A, B and C are designated, a synthesized signal excluding only the signal of the musical instrument D can be obtained. Needless to say, if a single sound source is designated, for example, in the case of a song with the performance of a musical instrument, only the singing voice can be extracted.

FIG. 3 shows a flowchart of the analysis processing, and FIG. 4 shows a flowchart of the time / frequency analysis unit of FIG. For example, at the beginning of the analysis, the flow shown in FIG. 1 is used, and the frequency resolution storage unit shown in FIG. 1 and FIG.
The frequency resolution for decomposing the observation signal into a single signal source is preserved. At the stage when the analysis by the flow of FIG.
The parallel processing of the analysis shown in the flow in FIGS. 3 and 4 is executed. First, the signal distribution unit shown in FIG. 3 distributes a predetermined number of analysis signals to at least two or more time-frequency analysis units (n) shown in FIG. Next, a predetermined frequency resolution is obtained from the frequency resolution unit shown in FIG. 4 to analyze the signal. When the number of sound sources is limited, such as ensemble of musical instruments, or when it is desired to separate only audio signals in a state where the number of speakers is limited, a repetitive loop is not used, so that the analysis load is reduced. Processing can be sped up by parallel processing.

Hereinafter, this embodiment will be described more specifically.
Basically, the analysis according to the present embodiment only needs to obtain three pieces of information: (1) information decomposed into frequency components, (2) information on a time axis, and (3) information on a signal magnitude. The analysis method is not limited to this. As one method, the above-mentioned information (1), (2) and (3) can be obtained by filtering with a filter bank in which the band is finely divided. As another method, for example, when FFT is used, the information of (1) and (3) can be easily obtained. Although the information of (2) cannot be obtained, it is not a simple FFT, but a short-time FFT is used continuously in time, and the results for each time section are sequentially compared,
It is possible to determine when a certain frequency component rises and when it falls based on the presence or absence of the spectral component.

An analysis method using the wavelet transform will be described in detail. By using the wavelet transform,
The following effects are obtained. That is, first, since the frequency resolution moves in a binary manner, the conversion result is equivalent to that obtained by filtering with a filter bank in which the center frequency is set every two times.
Since the wavelet transform performs not only a frequency analysis but also a time / frequency analysis like a Fourier transform, time information is stored as a result of frequency decomposition. Therefore, if frequency components appear simultaneously on adjacent levels, the harmonic structure can be easily searched. Second, by utilizing the fact that the basis functions for frequency decomposition have orthogonality, the decomposition / reconstruction algorithm is determined, so that the signal source can be accurately reproduced from the decomposed components.

The wavelet transform is generally defined by the following equation (1). That is,

[0033]

(Equation 1)

[0034]

(Equation 2) Here, a is a scale parameter, and b is a translate. Analyzing wavelet which is a signal function f (x) to be analyzed and a basis function for time / frequency analysis.
(X) is convolved and integrated. The feature of the wavelet transform is that, when the analyzing wavelet satisfies the condition of Equation 2 and scans on the time axis (when the value of translate b is increased), the scale parameter a indicating the frequency resolution Automatically changes and is continuously developed into frequency components. The function of automatically changing the scale parameter is called a zoom-in / zoom-out function for a signal to be analyzed. Thus, even if the frequency component to be analyzed is unknown, the minimum resolution suitable for the frequency component is automatically obtained, and the step of setting the initial frequency resolution in the flow of FIGS. 1 and 2 can be omitted. This is particularly effective when the frequency component of the signal to be analyzed is unknown.

In the case where the wavelet transform is used, the frequency component decomposition unit in FIG. 1 corresponds to a portion for executing the wavelet transform according to Equation 1. Normally, in the wavelet transform, a minimum value of a scale parameter is automatically determined when an analysis target signal is scanned on a time axis by an automatic zoom-in / zoom-out function. FIG. 5 shows a schematic diagram of time / frequency analysis by wavelet transform. Level-1 indicates the frequency component closest to the high frequency. Less than,
Each time the level decreases, the center frequency of the frequency band within the level moves in a binary manner as described above.

If the center frequency of each level in the wavelet transform matches the fundamental frequency and the overtone frequency of the speech or acoustic signal to be analyzed, the level-2 and -2 in FIG.
As shown at 3 or levels -4 and -5, a result is obtained in which the harmonic components are developed between the levels. However, when the center frequency of each level and the frequency of each harmonic component of the signal to be analyzed do not match, the frequency components are not concentrated on a specific level and are scattered uncorrelated between the levels. If you select the component, you cannot get a clue as to whether to reconstruct the original sound. Even if the levels are reconfigured on an ad hoc basis, there is a high possibility that components of other sound sources are included, and there is no guarantee that the original signal source before being expanded to each level can be accurately reconfigured.

Therefore, in the present embodiment, the following method is adopted instead of automatically determining the value of the scale parameter of the analyzing wavelet from Equation 1 as in the ordinary wavelet transform. That is, as a result of frequency analysis, frequency components appear between adjacent levels, and it is determined by the harmonic structure determination unit (FIG. 1) whether or not harmonic components are distributed to each level. Here, since the wavelet transform is used, it is possible to easily determine whether or not components appear concentrated at adjacent levels as shown in FIG.

When the components are scattered uncorrelatedly, the frequency resolution changing unit (FIG. 1) is used. The minimum value of the scale parameter obtained as a result of automatically zooming in in the first signal scan is stored in the frequency resolution storage unit, and this minimum value is used for the second and subsequent wavelet transforms. The minimum value of the scale parameter that is permitted to be used in an appropriate step width without permission is gradually increased. This allows
The frequency resolution changing operation can be easily realized.

If this operation is perceived by a filter bank having a configuration in which the center frequency is doubled, the filter bank initially configured has the center frequency of each level shifted to a higher frequency side than the frequency component of the signal source to be separated. However, it is shown that the band is sequentially shifted by the repetitive operation by the harmonic structure determination unit, and the optimum filter bank for the signal to be separated is automatically generated.

In the conventional separation method using the comb filter or the template matching, a comb filter or a template having innumerable characteristics must be prepared. On the other hand, in this embodiment, basically only one comb filter or a template is required. Only the analyzing wavelet, the processing only needs to change the parameter value. As a result, in this embodiment, the amount of information prepared by the system is significantly reduced, and a highly flexible system capable of blind separation of mixed signals can be constructed.

When the wavelet transform is used, the level is reconstructed by the inverse wavelet transform formula shown in Expression 3, and
The original single sound source signal can be reconstructed from the mixed sound.

[0042]

(Equation 3)

[0043]

(Equation 4) Equation 4 is an admissible condition for defining the right side of Equation 3.

FIG. 5A shows the mixed observation signal obtained during the observation time t _n , which is scanned by the wavelet transform during the time t _n , and in this embodiment, the waves of levels -1 to -6 are obtained. Decomposed into let components. In wavelet transform,
Since the frequency expansion from a certain level to a lower level moves in a binary manner, the frequency components of level-2 and level-3 have a harmonic structure relationship, and similarly, level-4 and level-
The fifth frequency component also has a harmonic structure relationship.

At the time of level-2 and level-3, the time t _na
To t _nb, there is a common frequency component indicating the rise and fall times, and the level -4 and the level -5 have the time t _nc to
is t _{n d} the frequency component representing a common rise and fall times between. Therefore, t _na to t of level-2 and level-3
Frequency components between _nb can be regarded as the same signal source is considered a component, components same level -4 and level -5 t _nc ~t any single signal source frequency components followed by between _nd.

The effect of using not only the frequency analysis but also the time analysis by the rise / fall time determination unit is as follows. That is, the times t _{na to} t _nb are changed from the times t _{nc to} t
_nd , and that the _fall time t _nb and the rise time t _nc almost coincide with each other can be easily analyzed.

When reconstructing the frequency components, it is necessary to make a subtle difference in the reconstruction method depending on the characteristics of the signal to be analyzed and the analysis result desired to be obtained. In the case of the present embodiment, the rise and fall times are different, but since the components from level-2 to level-5 are all in a harmonic structure, there is a possibility that the same sound source emits sounds of different pitches. .
If the analyst knows that this signal is the sound of a musical instrument, there are two methods of reconstruction, depending on the purpose of the analysis.

If the user wants to hear the reproduced signal as the sound of a musical instrument, the following is performed. That is, when the levels -2 to -5 are collectively reconstructed from the time t _na to t _nd , the effect of cutting off the weak frequency components when extracting the rise time and the fall time is small, and generally, the quality is high. We can expect reproduction sound. In FIG. 5, the frequency components at levels -2 and -3 are
Although shown in time t _na ~t _nb, not necessarily can deny that weak frequency component is continuing the time t _nb later. In general, the presence or absence of a high frequency component is often perceived by a human as a difference in tone, so it can be considered that the quality of reproduced sound is improved even with a weak frequency component. The low frequency side level -4, in -5, is regarded that there is no significant frequency components in the time t _na ~t _nb (for example, in the case of piano, generally indicates a rise of rapid sound pressure level.)
It is determined that the quality of the reproduced sound has little effect whether or not it is used for the reconstruction.

If pitch information is desired, the following is performed. That is, when signals are continuously output from the same musical instrument, it is considered that a melody is being played. When obtaining pitch information and automatically transcribing, level-2, -3 and level-4,-
By playing back the five segments separately, each is reconstructed as an independent pitch. This indicates that the function of the pre-processing of the process of converting the acoustic information into a code can be realized.

In the example of FIG. 5, the preceding levels -2, -3
Has a lower signal level and the subsequent levels -4 and -5 have a higher signal level. Therefore, if this is a voice, the sounds at levels -2 and -3 are consonants, and the levels -4 and -5 are consonants.
-5 can be determined to be a vowel. Also in this case, if priority is given to the quality of the reproduced sound, as in the first case, level-2
To -5 are collectively reconstructed from time t _na to t _nd . Because a consonant and a vowel usually have a transitional part called a crossover, and the vowel waveform is different between a case where the vowel is pronounced alone and a case where the consonant is pronounced continuously. The reason for this is that subtle changes occur, so that continuous continuous reproduction is perceived as a more natural sound.

In the case where the purpose is to perform text coding by a voice input system or the like, the load on the voice recognition unit is reduced by reproducing consonants and vowels separately.

To summarize the above, the high-frequency component of level -1 which is uncorrelated with time is regarded as, for example, a noise component, and the low-frequency component of level -6 is regarded as, for example, some kind of signal distortion. , Level-2 to -5, it is possible to separate and extract a single signal source from the noise environment.

FIG. 6 shows an example of a waveform when a mixed signal of two sounds having different fundamental frequencies is subjected to wavelet transform.
FIG. 6A shows an observation signal to be observed, and FIGS.
(5) shows a development result by the analyzing wavelet ψ _n (x), and (6) to (11) show a development result by the analyzing wavelet _{ｍ m} (x). The problem here is that using an analyzing wavelet with the same minimum frequency resolution, unless the frequency components of each other have an overtone relationship, the two tones will work well with the frequency components forming the harmonic structure in one operation. It cannot be expanded.

In order to identify the harmonic structure of each sound source individually, the frequency resolution changing unit and the harmonic structure judging unit shown in FIG. 1 sequentially change the minimum value of the scale parameter of the analyzing wavelet and optimize it. Then, the same rise and fall time frequency component extraction unit extracts a frequency component. An extraction result reconstruction unit reconstructs the extraction result and outputs a reproduced signal.

When the minimum scale parameter is a ⁻¹ = n, levels −2n and −3n are obtained by the expansion using the analyzing wavelet _{ｎ n} (x).

Next, by executing the iterative loop again, when the minimum scale parameter a ⁻¹ is a ⁻¹ = m: (n <m), the analog of the minimum frequency resolution different from _{ｎ n} (x) is obtained. Rising Wavelet _{ｍ m} (x)
Levels -3m to -5m were obtained.

In FIG. 6, in this embodiment, a plurality of signal sources having different fundamental frequencies could be separated by using the analyzing wavelets having different minimum frequency resolutions.

In FIG. 6, for example, levels -3m to -5
Even if the signal source is distributed over a wide frequency band like m
This shows that it can be extracted as the frequency component of the same signal source.
This indicates that even if the sound source is a piano, for example, and a chord is played and the band is wide, it can be separated from other signal sources. Since many of the conventional separation methods correspond only to single melody separation, it can be seen that this embodiment is an extremely effective method.

FIG. 7 shows an example in which a mixed signal of signals from signal sources having similar harmonic structures but different fundamental frequencies is developed. Different frequency components of signal sources appear at levels -2, -3, -4 and levels -4, -5, respectively.

Attention is paid to level-4 in which the frequency components of the high-frequency signal and the low-frequency signal appear to overlap. When the Bregmn rule (1) is applied to the level-4 frequency component, the rise / fall time determination unit calculates
It can be determined that the component constituting the signal source on the high frequency side in the level-4 is between the rising and falling times t _{na to} t _nc which are almost the same as the components on the high frequency side, levels −2 and −3. Similarly, level components constituting the source of the low frequency side in -4 can be determined to be in between another lower frequency rise and fall times of the component levels -5 t _nb ~t _nd. According to this determination result, at level-4, the time t _nb -t
_It can be seen that the high frequency side component and the low frequency side component overlap between _nc, and it becomes a problem whether to include this component on the high frequency side or on the low frequency side.

Here, rule (4) is applied in addition to Bregman's rule (1). The signal magnitude comparison / determination unit and the signal magnitude grouping unit shown in FIG.
regman's rule (1) and rule (4) apply.
7, the frequency components are observed between the level -4 time t _na ~t _nd, a representative magnitude of the signal before the time point t _nb, e.g. amplitude peak - is represented click width h _4,
Representative magnitude of the signal after time t _nb amplitude peak - when a representative to click width l _4, when the size comparison section of the signal to determine a significant difference in representative magnitude of the two signals , Level-4 can be discriminated as a frequency component of another signal source before and after the time t _nb on the time axis. Let it be expressed as the following equation. That is,

[0062]

｛H ₄ ,｝ <｛l ₄ ,｝ Next, the grouping unit according to the signal magnitude is applied to the level-
The representative magnitudes of the signals at each of the four high-frequency and low-frequency levels are compared. When the representative magnitude of the signal of each level is smaller on the high frequency side, it is expressed as the following equation. That is,

[0063]

{H ₂ , h ₃ , h ₄ , {<{l ₄ , l ₅ }} Thus, the same frequency band is first time-divided according to the signal magnitude, and then shown in equations 5 and 6. If such a condition can be found, the frequency components in the section where the components overlap, such as between the time t _{nb and} t _nc of the level-4, are shifted to the frequency component group having the larger signal representative magnitude. Considered included.

In FIG. 7, the average peak value of the signal is used as the representative magnitude of the signal. For example, in the case where the signal source on the low frequency side is a signal that shows a sharp rise in sound pressure level like a piano. By applying Bregman's law (4), two signal sources are superimposed on the same band as in level-4 by extracting and grouping frequency components showing sharp rising at each level at the same time. Even discrimination is possible.

In the case of FIG. 7, when reconstructing the signal source, the reconstruction on the high frequency side requires time t with respect to levels -2 and -3.
The components between _{na and} t _nc are used, and the components between times t _{na and} t _nb are used for level-4. In the reconstruction on the low frequency side, components between times t _{nb and} t _nd are used for levels −4 and −5.

Although the high-frequency component is included in the overlapped period between the times t _{nb and} t _nc , reconstructing the high-frequency component as the low-frequency component is apparently strict in terms of signal processing. It looks as if it lacks. However, assuming that a person listens to the reconstructed signal as a sound, the sound with a low sound pressure level on the high frequency side is masked by the sound with a high sound pressure level on the low frequency side and is hardly perceived aurally due to the human hearing characteristics. Therefore, quality degradation of the reconstructed signal can be avoided.

[0067]

As can be easily understood from the above description, according to the present invention, the signals of the original signal sources can be individually or arbitrarily selected from the mixed signals of the signals from the plurality of signal sources. Can be combined to obtain a continuous separated signal.

[Brief description of the drawings]

FIG. 1 is a diagram showing a schematic configuration and a flowchart of an embodiment of the present invention.

FIG. 2 is a diagram showing a schematic configuration and a flowchart for synthesizing and outputting signals from an arbitrary number of signal sources.

FIG. 3 is a diagram showing a schematic configuration and a flowchart of signal analysis.

FIG. 4 is a diagram showing a schematic configuration and a flowchart of a time / frequency analysis unit.

FIG. 5 is a schematic diagram of a processing result of a signal in which noise / distortion is superimposed on a signal source according to the present embodiment.

FIG. 6 is a schematic diagram of a signal processing result by analyzing wavelets having different frequency resolutions according to the present embodiment.

FIG. 7 is a schematic diagram of a processing result of a mixed signal partially superimposed temporally according to the present embodiment.

Claims (20)

[Claims] 1. A signal separation method in which an observation signal separates a mixed signal of signals from a plurality of signal sources, wherein the mixed signal is analyzed continuously in time and separated for each single signal source. And a reconstruction step for reconstructing a signal of a desired signal source from a separation result of the separation step. 2. The signal separation method according to claim 1, wherein the analysis separation step includes a storage step of temporarily storing a signal reconstructed as a signal of a single signal source. 3. The analysis and separation step includes a distribution step of distributing the mixed signal to an arbitrary number.
3. The signal separation method according to 1. 4. The analyzing and separating step includes: a frequency decomposition step of decomposing the mixed signal into frequency components with an arbitrary frequency resolution; a harmonic structure determining step of determining whether a harmonic structure exists between the frequency components; When it is determined by the harmonic structure determination step that the frequency component is not in the harmonic structure, an analysis condition changing step of sequentially changing analysis conditions such as frequency resolution until a harmonic structure is obtained, and the frequency decomposition step The signal separation method according to claim 1, further comprising: a frequency component extracting step of extracting the decomposed frequency components; and a storing step of storing the frequency resolution. 5. The rising / falling time determining step in which the frequency separation is performed on the time axis to determine the rising and falling times of the frequency component from the magnitude of the signal. 5. The method according to claim 4, further comprising: a comparing step of comparing a rise time and a fall time between each other, and an extraction step of extracting a frequency component in a time segment in which at least one of the rise time and the fall time matches both. The signal separation method as described. 6. The analysis / separation step scans frequency components decomposed into the same frequency band on a time axis, and when a discontinuity is observed in a representative magnitude of the signal, the representative magnitude of the signal is determined. An analysis target frequency component extracting step of classifying and extracting an analysis target frequency component in a continuous time, and a signal source for comparing the divided frequency component with a frequency component of an adjacent band and determining the frequency component of a different signal source The signal separation method according to claim 5, further comprising a determining step. 7. The reconstructing step reconstructs a frequency component in which a harmonic structure between frequency components is determined, and the frequency component is determined to have at least one of a rise time and a fall time equal to each other. The signal separation method according to claim 5, wherein 8. The method according to claim 6, wherein in the reconstructing step, a harmonic structure between the frequency components is determined, and the frequency components extracted based on the representative magnitude of the frequency components are reconstructed for each signal source. The signal separation method as described. 9. The signal separation method according to claim 7, wherein the analysis separation step includes a step of continuing / stopping the analysis by referring to at least one of a predetermined analysis condition and an analysis result. 10. The signal separation method according to claim 7, wherein a wavelet transform is used for frequency decomposition and frequency component reconstruction in an analysis process. 11. A signal separation apparatus for separating a mixed signal of observation signals from a plurality of signal sources, wherein the separated signal is analyzed continuously and temporally and separated for each single signal source. And a reconstructing means for reconstructing a signal of a desired signal source from the separation result of the separating means. 12. The signal separating apparatus according to claim 11, wherein said analyzing and separating means includes a storage means for temporarily storing a signal reconstructed as a signal of a single signal source. 13. The signal separation device according to claim 11, wherein the analysis separation means includes a distribution means for distributing the mixed signal to an arbitrary number. 14. A frequency decomposition means for decomposing the mixed signal into frequency components with an arbitrary frequency resolution, a harmonic structure determination means for determining the presence or absence of a harmonic structure between the frequency components, When a frequency component is determined not to be in the harmonic structure by the harmonic structure determining means, an analysis condition changing means for sequentially changing analysis conditions such as frequency resolution until a harmonic structure is obtained, and the frequency decomposing means. Frequency component extracting means for extracting the decomposed frequency component,
The signal separation device according to claim 11, further comprising storage means for storing the frequency resolution. 15. Rising and falling time determining means for scanning the frequency components subjected to frequency decomposition on the time axis to determine the rise and fall times of the frequency components from the magnitude of the signal, A comparison means for comparing a rise time and a fall time between each other, and an extraction means for extracting a frequency component in a time section in which at least one of the rise time and the fall time coincides with each other. A signal separation device as described in the above. 16. The analysis / separation means scans the frequency components decomposed into the same frequency band on the time axis, and when a discontinuity is observed in the representative magnitude of the signal, the representative magnitude of the signal is determined. Analysis target frequency component extracting means for classifying and extracting the analysis target frequency component in a continuous time, and a signal source for comparing the divided frequency component with a frequency component of an adjacent band and determining the frequency component of a different signal source The signal separation device according to claim 15, further comprising a determination unit. 17. The reconstructing means reconstructs a frequency component for which a harmonic structure between frequency components is determined and for which the frequency component is determined to have at least one of a rise time and a fall time, and outputs The signal separation device according to claim 15, wherein: 18. The method according to claim 16, wherein said reconstructing means determines a harmonic structure between frequency components, and reconstructs a frequency component extracted based on a representative magnitude of the frequency component for each signal source. A signal separation device as described in the above. 19. The signal separation device according to claim 17, wherein the analysis separation means includes means for continuing or stopping the analysis by referring to at least one of a predetermined analysis condition and an analysis result. 20. The wavelet transform according to claim 17, wherein a wavelet transform is used for frequency decomposition and frequency component reconstruction in an analysis process.
9. The signal separation device according to 8.