JPH1062460A - Signal separator - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a signal separator which can extract an anti-dominant signal using the basic frequency of a dominant signal to be extracted easily. SOLUTION: A signal inputted from an input section 11 is analyzed at a frequency analyzing section 12 and the basic frequency of a dominant signal is extracted at a basic frequency extracting section 13. Convolving operation of the spectrum and the Bartlett window several times as long as the basic frequency is performed at an operating section 14. Spectral envelope is then calculated at an amplitude determining section 15 for the anti-dominant signal of the input signal subjected to frequency analysis. The signal is then synthesized again at a synthesizing section 16 according to a spectral envelope thus determined and converted at the output section 17 into a sound signal before being outputted.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a signal separating apparatus, and more particularly, to a case where a plurality of audio signals such as voices are mixed, a spectrum of the signal is used by using a fundamental frequency of a signal having a fundamental frequency of one of the signals. And a signal separating apparatus that separates a signal by extracting a spectrum of another competing signal by flattening the signal.

[0002]

2. Description of the Related Art When one fundamental frequency is known or can be extracted from a signal in which two signals having different spectral characteristics are mixed, conventionally, as means for separating the two signals, for example, an enhancement of speech is used. by A
daptive Filtering, RH franzier and others Proc. IEEE. In
t. Conf. on Acoust. Speech and signal processing A
Comb filters such as those described in pr. 12-14 pp. 251-253 (1976) have been widely used.

FIG. 13 is a schematic block diagram of a conventional signal separating device using a comb filter. In FIG.
An input signal is input from an input unit 1 and a fundamental frequency extracting unit 2
And the fundamental frequency of one of the signals mixed in the input signal is extracted. The extracted fundamental frequency is supplied to the comb filter design unit 3, and a comb filter is created such that only a frequency band corresponding to an integral multiple of the extracted fundamental frequency becomes a pass band. Then, filtering is performed by the designed filtering unit 4, and output from the output unit 5.

[0004]

However, in an actual sound environment, a voiced sound having a harmonic structure and a large energy often masks a target signal. Therefore, in order to extract a target signal, an anti-dominant signal is used. There are many opportunities to seek. However, in the above-described conventional structure, even in such a case, it is necessary to directly obtain the properties of the anti-dominant signal such as the fundamental frequency, which is actually difficult.

[0005] Therefore, a main object of the present invention is to provide a signal separation device capable of extracting an anti-dominant signal by using a fundamental frequency of a dominant signal which is easy to extract.

[0006]

The invention according to claim 1 is
A signal separation device that separates an input signal, a frequency analysis unit that performs a frequency analysis of the input signal, and a fundamental frequency extraction unit that determines a fundamental frequency of a dominant signal among frequencies included in the frequency-analyzed input signal. A convolution operation means for performing a convolution operation of a Bartlett window and a spectrum having a length several times the extracted fundamental frequency, and an amplitude determination means for calculating a spectrum envelope of an anti-dominant signal of the frequency-analyzed input signal, It is provided with a synthesizing means for re-synthesizing a signal based on the obtained spectrum envelope.

According to a second aspect of the present invention, there is provided a signal separating apparatus for separating respective signals from an input signal including a dominant signal and an anti-dominant signal, wherein the frequency analyzing means performs a frequency analysis of the input signal; A fundamental frequency extracting means for finding a fundamental frequency of a dominant signal among signals included in the analyzed input signal, and a convolution operation for performing a convolution operation of a Bartlett window having a length several times the extracted fundamental frequency and a spectrum. Means, amplitude determining means for determining the envelope of the anti-dominant signal by differentiating the result of the convolution operation several times, and synthesizing means for synthesizing the signal from the determined envelope.

[0008] The invention according to claim 3 further includes output means for converting the recombined signal into an audio signal and outputting the audio signal.

According to a fourth aspect of the present invention, the convolution operation means of the first aspect includes a first convolution operation means for convolving a window function having a length several times the fundamental frequency and a real part of the spectrum, A window function having a length several times the frequency and second convolution operation means for convolving the imaginary part of the spectrum, wherein the amplitude determination means smoothes the peak of the spectrum convolved by the first convolution operation means First amplitude determining means for obtaining, for each frequency, a difference between the envelope connected to the waveform and the envelope obtained by smoothly connecting the spectrum dip, and an envelope and spectrum obtained by smoothly connecting the peaks of the spectrum convolved by the second convolution operation means. And a second amplitude determining means for obtaining, for each frequency, a difference between envelopes obtained by smoothly connecting the dips.

[0010]

FIG. 1 is a block diagram showing an embodiment of the present invention. In FIG. 1, an input signal input from an input unit 11 is provided to a frequency analysis unit 12, where a frequency analysis of the input signal is performed. The frequency-analyzed input signal is provided to a fundamental frequency extraction unit 13 and a convolution operation unit 14. The fundamental frequency extracting unit 13 finds a fundamental frequency of a dominant signal among the frequencies included in the frequency-analyzed input signal. The convolution operation unit 14 performs a convolution operation between the Bartlett window having a length several times the extracted fundamental frequency and the spectrum. The amplitude determiner 15 calculates the spectrum envelope of the anti-dominant signal of the frequency-analyzed input signal, and supplies the calculated spectrum envelope to the synthesizer 16. Synthetic unit 1
6 re-synthesizes the signal according to the obtained spectrum envelope, and supplies it to the output unit 17. The output unit 17 converts the recombined signal into an audio signal and outputs the signal.

FIG. 2 shows two signals f1 (t) and f2 (t).
FIG. 3 shows the spectrum of the signal f1 (t), and FIG. 4 shows the spectrum of the signal f2.
5 shows the spectrum of (t), and FIG. 5 shows the spectrum of the input signal of FIG. 2 and the result of convolution of the Bartlett function having a window length twice as long as the fundamental frequency of the dominant signal included in the input signal.

FIG. 6 shows a spectrum obtained by Fourier-transforming other two signals having different fundamental frequencies, FIG. 7 shows a spectrum of one signal constituting FIG. 6, and FIG. 8 shows another spectrum constituting the signal of FIG. 9 shows the spectrum of the input signal shown in FIG. 6 and the result of convolution of the Bartlett function having a window length twice as long as the fundamental frequency of one signal included in the input signal. is there.

Next, the operation of one embodiment of the present invention will be described with reference to FIGS. The input unit 11 is shown in FIG.
An input signal as shown in FIG. This input signal f (t) has two signals f1 (t) and f2
(T) is included and is expressed by the following equation (1).

[0014]

(Equation 1)

The frequency analysis unit 12 performs a frequency analysis of an input signal using Fourier transform. The analysis result is expressed by the following equation (2), and the amplitude value is expressed by the following equation (3). Fourier transform of a signal in which two different signals are mixed results in a spectrum as shown in FIG. 2, and the spectra of the two signals f1 (t) and f2 (t) are, as shown in FIGS. The spectrum runs at integer multiples of the fundamental frequency.

[0016]

(Equation 2)

The fundamental frequency extracting unit 13 extracts the fundamental frequency of one of the input signals, and the convolution operation unit 14 convolves the spectrum with a window function having twice the length of the fundamental frequency. The result is expressed by the following equation (4).

[0018]

(Equation 3)

Now, when a Bartlett window function having a window length twice as long as the spectrum interval of FIG. 3 is convolved with the spectrum of FIG. 2, when the envelope of F1 (ω) is sufficiently smooth, , F1 (ω) are interpolated and only the envelope appears. However, since the relationship between the spectral interval and the window length of the Bartlett window function as the interpolation function is not appropriate for the line spectrum based on F2 (ω), the influence of the line spectrum affects the envelope after interpolation as shown in FIG. Appears. At this time, the interval Δd between the envelope peak and the dip is F1
If the amplitude of the line spectrum of (ω) is constant, it is a constant value.

Similarly, when a Bartlett function having a window length twice as long as the spectrum interval in FIG. 7 is convoluted, a function as shown in FIG. 9 is obtained. From FIG. 9 and equation (4) above, it can be seen that the peak-dip interval Δd seen in FIG. 5 is multiplied by Bn to become B (ω) · Δd. Further, from the expression (4), the influence of the spectrum An of the signal f1 (t) is equally added to the peak and the dip as a bias. As a result, the difference B (ω) · Δd between the peak of the spectral function and the dip after the convolution operation is similar to the spectral envelope of the signal f2 (t), and the position of the peak of the envelope is F2 (ω). It can be seen that this corresponds to the position of the line spectrum of ()).

The amplitude determining section 15 obtains, for each frequency, a difference between envelopes obtained by smoothly connecting dips of an envelope spectrum obtained by smoothly connecting peaks of a spectrum obtained as a result of convolution. Then, the synthesizing unit 16 performs a Fourier inverse transform on the spectrum having the obtained envelope difference as the amplitude envelope to obtain a time signal, and outputs the time signal.
Converts the signal into an acoustic signal and outputs it.

In the above embodiment, the frequency analysis unit 12 uses the Fourier transform. However, the present invention is not limited to this. For example, the frequency analysis unit 12 may perform frequency analysis using an LPC, a cepstrum, a narrow band filter, or the like. Is also good. The detected spectrum may be a power spectrum instead of an amplitude spectrum. Further, if the obtained dominant signal has a harmonic structure, the same processing may be performed by obtaining its fundamental frequency to extract an anti-dominant signal.

FIG. 10 is a block diagram showing another embodiment of the present invention. This embodiment separates a complex signal by separately processing a real part and an imaginary part of a frequency characteristic. To this end, the frequency analysis unit 12 performs a frequency analysis of the input signal using the Fourier transform,
Output real and imaginary parts. The real part is the first convolution operation unit 1
41 and the imaginary part is provided to a second convolution operation unit 142. First and second convolution operation units 141,
142 convolves the imaginary part and the real part of the spectrum with the window function having a length twice as long as the fundamental frequency, and gives the result to the first amplitude determining unit 151 and the second amplitude determining unit 152. The first and second amplitude determination units 151 and 152 determine, for each frequency, the difference between the envelope that smoothly connects the peaks of the convolved spectrum and the envelope that smoothly connects the dip of the spectrum. The spectrum having the real part and the imaginary part obtained by the first and second amplitude determination units 151 and 151 is provided to the synthesis unit 16 and the inverse Fourier transform is performed. The other parts of the input unit 11, the fundamental frequency extracting unit 13, and the output unit 17 are configured in the same manner as in FIG.

Next, the operation will be described. Input unit 11
, The input signal expressed by the expression (1) described above is input, the frequency analysis unit 12 analyzes the frequency of the input signal using Fourier transform, and the result is expressed by the expression (2) described above. . In the equation (2), A = ar + iai,
Assuming that b = br + ibi, the real part and the imaginary part are expressed by equations (5) and (6), respectively.

[0025]

(Equation 4)

The first convolution operation section 141 convolves the real part of the spectrum with the window function having twice the length of the fundamental frequency according to the following equation (7).

[0027]

(Equation 5)

On the other hand, the second convolution operation unit 152 performs convolution of the window function having a length twice as long as the fundamental frequency and the imaginary part of the spectrum according to the following equation (8).

[0029]

(Equation 6)

The first amplitude determining unit 151 obtains, for each frequency, a difference between an envelope that smoothly connects the peaks of the spectrum convolved by the first convolution operation unit 141 and an envelope that smoothly connects the dips of the spectrum. The second amplitude determination unit 152 obtains, for each frequency, a difference between an envelope that smoothly connects the peaks of the spectrum obtained as a result of the convolution in the second convolution operation unit 142 and an envelope that smoothly connects the dips of the spectrum. Then, the synthesis unit 16 obtains a time signal by performing an inverse Fourier transform of the spectrum having the real part and the imaginary part obtained by the first and second amplitude determination units 151 and 152,
The output unit 17 converts an output signal of the configuration unit 16 into an audio signal and outputs the audio signal.

Next, another example of the embodiment shown in FIG. 1 will be described. The spectral envelope of the dominant signal f1 (t) is F1 (ω), the order of the highest harmonic is N1, and the spectral envelope of the antidominant signal f2 (t) is F2.
(Ω), the order of the highest harmonic is N2, and the respective fundamental frequencies are ω ₁ and ω ₂ . Each spectrum has a harmonic structure and can be described by a line spectrum. At this time, the spectrum F (ω) of the signal obtained by combining the signals f1 (t) and f2 (t) is expressed by the following equation (9).

[0032]

(Equation 7)

Here, if f1 (t) and f2 (t) do not have harmonic components of a common frequency, they are expressed by the following equation (10).

[0034]

(Equation 8)

On the other hand, when the signals f1 (t) and f2 (t) have harmonic components of a common frequency, the amplitude of the spectrum of the harmonic is interfered by the phase relationship. For now, it is assumed that there is no common harmonic component.

A Bartlett window function having a window length twice as long as the fundamental frequency ω ₁ of the signal f1 (t) is convolved in the frequency domain in the above equation (10). When the spectrum of f1 (t) is sufficiently smooth,

[0037]

(Equation 9)

Where:

[0039]

(Equation 10)

Therefore, the first term is approximately | S
₁ (ω) |. Next, the second term

[0041]

[Equation 11]

With regard to the above, its properties will be observed in detail. Term of the equation (13) of the middle are arranged a line spectrum omega ₂ intervals. When ω ₁ <ω ₂ , the following equation (14) is represented by FIG.
1, and when ω ₁ > ω ₂ , equation (14) becomes a spectrum as shown in FIG.

[0043]

(Equation 12)

FIG. 12A shows a spectrum of a harmonic complex tone having a fundamental frequency of 153 Hz, and FIG. 12B shows a spectrum obtained by smoothing the spectrum with a Bartlett window having a window length of 200 Hz. FIG. 12A shows a fundamental frequency of 100 Hz.
FIG. 12B shows a spectrum obtained by smoothing the spectrum with a Bartlett window having a window length of 256 Hz.

It can be seen that both of these are functions having local maxima at intervals of ω ₂ . Furthermore, if f2 (t) is a periodic signal having a harmonic structure, the slope of the line connecting the minimum value and the maximum value is constant regardless of the frequency. In particular, ω ₁ >
When the ω _2, the minimum point is present at the midpoint between the local maximum points. Assuming that the maximum value is M _x , the minimum value is M _n , and the fundamental frequency is ω ₁ , when ω ₁ > ω ₂ , the slope of the line connecting the minimum value and the maximum value in Expression (14) is It is given by equation (15).

2 (M _x −M _n ) / ω ₂ (15) Assuming that | F 2 (ω) | = C ₁ ω + C ₂ (16), the (13) th minimum value and the maximum value are connected. The gradient of the line segment is expressed by the following equation (17). The gradient of the line segment connecting the minimum value and the maximum value is accompanied by a constant bias C _{1 Mn} , but | F2 (ω)
It can be seen that the amplitude of | is reflected in the slope.

[0047]

(Equation 13)

When ω ₁ <ω ₂ , the local minimum points are distributed in a certain range centered at the midpoint of the local maximum point, so the gradient of the line connecting the local minimum value and the local maximum value is ω ₁ > ω ₂ However, since the slope is constant irrespective of the frequency, it can be seen that the amplitude of | F2 (ω) | is reflected in the slope as in the case of Expression (17).

The above operation is performed by the convolution operation unit 14 shown in FIG. Next, the operation of the amplitude determining unit 15 will be described. Equation (11) is differentiated twice with respect to ω. here,

[0050]

[Equation 14]

In other words,

[0052]

(Equation 15)

Is as follows. Where | F1 (ω) |, | F2
When (ω) | changes sufficiently smoothly with respect to G (ω), the expression (21) is approximately obtained.

[0054]

(Equation 16)

Here, G (ω) is a function having a maximum value at the frequency where the line spectrum of F2 (ω) exists.
At that frequency, the first derivative of G (ω), dG (ω) / dω =
The second derivative of 0, G (ω) is d ² G (ω) / dω ² = −∞.

In the actual calculation, if the difference is used instead of the derivative, Δ ² (G (ω)) becomes a finite value,
The magnitude reflects the slope of the line connecting the maximum value and the minimum value. Eventually, the spectrum of f2 (t) to be obtained is expressed by the following equation (22).

[0057]

[Equation 17]

By the way, from FIG. 12, when ω ₁ > ω ₂ , since the interval of the spectrum of the non-convolution function is smaller than the window length of the convolution function, the peaks of the minimum frequency and the maximum frequency are not clear. I understand. As described above, the spectral component adjacent to the band where the amplitude value of the spectrum is 0 cannot be extracted under the condition of Expression (22). In particular, a lack of a component corresponding to the fundamental frequency may cause a large change in timbre. Therefore, in order not to lose the fundamental frequency, a DC component equal to the low-frequency spectrum amplitude was added, and then the spectrum was smoothed.

[0059]

As described above, according to the present invention, the frequency analysis of the input signal is performed, the fundamental frequency of the dominant signal among the frequencies included in the input signal is determined, and several times the determined basic frequency is obtained. The convolution operation of the Bartlett window with the length and the spectrum is performed, the spectrum envelope of the anti-dominant signal of the input signal is calculated, and the obtained spectrum envelope signal is re-synthesized, so that the dominant signal of the mixed signal is obtained. Is flattened to extract the characteristics of the spectrum of the anti-dominant signal, and an excellent signal separation device capable of separating two signals can be realized.

[Brief description of the drawings]

FIG. 1 is a block diagram of one embodiment of the present invention.

FIG. 2 is a diagram illustrating a spectrum of an input signal including two signals f1 (t) and f2 (t).

FIG. 3 is a diagram showing a spectrum of a signal f1 (t).

FIG. 4 is a diagram illustrating a spectrum of a signal f2 (t).

5 is a diagram illustrating a spectrum of the input signal of FIG. 2 and a result of convolution of a Bartlett function having a window length twice as long as a fundamental frequency of one signal included in the input signal.

FIG. 6 is a diagram illustrating a spectrum obtained by Fourier-transforming another two signals having different fundamental frequencies.

FIG. 7 is a diagram showing a spectrum of one signal constituting the signal of FIG. 6;

FIG. 8 is a diagram showing a spectrum of another signal constituting the signal of FIG. 6;

9 is a diagram illustrating a spectrum of the input signal illustrated in FIG. 6 and a result of convolution of a Bartlett function having a window length twice as long as a fundamental frequency of one signal included in the input signal.

FIG. 10 is a block diagram showing another embodiment of the present invention.

FIG. 11 is a diagram showing a spectrum of a harmonic composite sound having a fundamental frequency of 153 Hz and a spectrum smoothed by a Bartlett window.

FIG. 12 is a diagram showing a spectrum of a harmonic complex tone having a fundamental frequency of 100 Hz and a spectrum smoothed by a Bartlett window.

FIG. 13 is a block diagram showing a configuration of a conventional signal separation device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 Input part 12 Frequency analysis part 13 Fundamental frequency extraction part 14 Convolution calculation part 15 Amplitude determination part 16 Synthesis part 17 Output part 141 1st convolution calculation part 142 2nd convolution calculation part 151 1st amplitude determination part 152 2nd The amplitude determination unit

 ──────────────────────────────────────────────────の Continued on the front page (72) Inventor Hani Yahiya 5th Sanraya, Inaya, Seika-cho, Sagara-gun, Kyoto Pref. ATIR Human Information and Communication Research Laboratories Co., Ltd. (72) Inventor Hideki Kawahara Kyoto Soraku No. 5 Sanpani, Seiya-cho, Gunma-cho, Gunma ATI Human Information and Communication Laboratories, Inc.

Claims (4)

[Claims] 1. A signal separation device for separating an input signal, comprising: frequency analysis means for performing frequency analysis of the input signal; and a dominant signal among frequencies included in the input signal analyzed by the frequency analysis means. Fundamental frequency extracting means for obtaining a fundamental frequency of the convolution operation means for performing convolution operation of a Bartlett window and a spectrum having a length several times the fundamental frequency extracted by the fundamental frequency extracting means, and of the frequency-analyzed input signal. Amplitude determining means for calculating the spectrum envelope of the anti-dominant signal,
And a synthesizing means for re-synthesizing the signal based on the determined spectrum envelope. 2. A signal separating apparatus for separating an input signal, comprising: frequency analysis means for performing frequency analysis of the input signal; and a dominant signal among signals included in the input signal frequency-analyzed by the frequency analysis means. Fundamental frequency extracting means for obtaining the fundamental frequency of the convolution operation means, convolution operation means for performing convolution operation of a Bartlett window having a length several times as long as the fundamental frequency extracted by the basic frequency extraction means and spectrum, convolution operation performed by the convolution operation means A signal separating device comprising: an amplitude determining unit that determines the envelope of the anti-dominant signal by differentiating the result several times; and a combining unit that combines the signal from the envelope determined by the amplitude determining unit. 3. The signal separation device according to claim 1, further comprising an output unit that converts the signal re-synthesized by the synthesis unit into an audio signal and outputs the audio signal. 4. The convolution operation means comprises: a first convolution operation means for convolving a window function having a length several times the fundamental frequency and a real part of a spectrum; and a length several times the fundamental frequency. And a second convolution operation means for convolving the imaginary part of the spectrum, the amplitude determination means comprising: an envelope which smoothly connects peaks of the spectrum convolved by the first convolution operation means; First amplitude determining means for obtaining, based on each frequency, a difference between envelopes obtained by smoothly connecting spectral dips; and an envelope and spectrum dip obtained by smoothly connecting peaks of the spectrum convolved by the second convolution operation means. And a second amplitude determining means for obtaining, for each frequency, an envelope difference obtained by smoothly connecting
The signal separation device according to claim 1.