JPH08160994A - Noise suppression device - Google Patents

Abstract

PURPOSE: To provide a noise suppression device which can eliminate noise from sound superimposed with noise by less storage capacity and operation quantity. CONSTITUTION: When sound is inputted sometimes in environment in which an environment noise exists, a noise spectrum dating section 13 obtains a present noise spectrum based on the output of a Fourier analysis section 12 and a previous noise spectrum in a noise spectrum storage section 14. The noise spectrum updating section 13 switches a previous noise spectrum in the noise spectrum storage section 14 and obtained present noise spectrum based on discrimination of a noise signal discriminating section 15, outputs it as an estimated noise spectrum, and a previous noise spectrum in the noise spectrum storage section 14 is updated by the present noise spectrum.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise suppressing device for suppressing a noise signal superimposed on a voice signal.

[0002]

2. Description of the Related Art In recent years, since a device using voice as an interface has been used in various environments, it has become important to be able to use voice as an interface even in a noisy environment. In addition, as such a device, it has become possible to manufacture an inexpensive device by improving the hardware capability of a DSP (digital signal processor) or the like, and voice has come to be used as a man-machine interface for various devices. It was for that reason,
A method for reducing environmental noise superimposed on a voice signal by digital processing has been attracting attention.

A conventional noise suppressor will be described below with reference to the drawings. Regarding the conventional noise suppressor, its method is described in IEEE Tra published in April 1979.
nsaction on Acoustics, speech, and signal processing
vol.ASSP-27 No.2pp.113-120 `` Supression of Acoust
ic Noise Speech Using Spectral Subtraction ”is disclosed as a spectral subtraction method.

This spectrum subtraction method will be briefly described with reference to the conventional noise suppression device shown in FIG. FIG. 5 is a block diagram showing the configuration of a conventional noise suppression device. In FIG. 5, 51 is a cutout unit, 52 is a Fourier analysis unit, 53 is a noise spectrum average calculation unit, 5
4 is a noise spectrum memory, 55 is a noise signal determination unit, 5
6 is an amplitude subtraction unit, 57 is a half-wave rectification unit, 58 is an inverse Fourier transform unit, and 59 is a waveform reproduction unit.

The operation of the conventional noise suppressor having the above components will be described below with reference to the flowchart shown in FIG. In step # 61, the input speech is subjected to a window in the clipping unit 51, which can obtain frequency resolution accuracy required for Fourier analysis later, and a fixed length of a section of about 20 ms with a period of about 10 ms. Cut out the waveform in the frame. In this way, to cut out the waveform of a certain section from the continuous waveform,
As a window condition, the waveforms do not become discontinuous when they are combined and connected in the later combining process, and the accuracy of Fourier analysis can be secured, and the original waveforms are overlapped by 50%. Selected analysis window. The following Hanning window is used as this window function. Here, when the number of data of Fourier analysis is N, the weighting W (n) for the n-th data after windowing is expressed as shown in Expression 1.

[0006]

[Equation 1]

At step # 62, the Fourier analysis unit 5
In 2, the voice waveform of the frame to be analyzed is subjected to frequency analysis (FFT analysis) by Fourier analysis. Since the analysis result is composed of a real number part and an imaginary number part, these are decomposed into an amplitude spectrum and a phase spectrum. The amplitude spectrum is output to the noise spectrum memory 54, the noise signal determination unit 55, and the amplitude subtraction unit 56. On the other hand, since the phase spectrum is re-synthesized with the voice waveform, the inverse Fourier transform unit 58
Is output to

In step # 63, the noise signal determining section 5
At 5, it is determined whether the input signal is a signal in which a noise signal is superimposed on a voice signal or a signal containing only a noise signal. This determination is performed based on the power information of the input signal. A voice section having a voice is set at a high power, and a noise section having no voice is set at a low power. The information "present" and "absent" of this voice is output to the noise spectrum memory 54.

In step # 64, the noise spectrum memory 54 stores the noise spectrum of the past several 10 ms. In the noise section, the amplitude spectrum from the Fourier analysis unit 52 is stored. In the voice section, the noise spectrum memory 54 does not store the amplitude spectrum from the Fourier analysis unit 52. Therefore, the latest noise spectrum is always stored in the noise spectrum memory 54.

In step # 65, the noise spectrum average calculation unit 53 calculates the average spectrum of the amplitude spectra of the past several ten frames stored in the noise spectrum memory 54 as the amplitude information of the noise estimation amount, and utters the voice. Is the amplitude spectrum of the noise immediately before, that is, the noise spectrum. Thus, the amount of memory used to calculate the noise spectrum is the number of window lengths analyzed by the Fourier analysis × the number of storage frames.

In step # 66, the amplitude subtraction unit 56 calculates the noise amplitude for each frequency of the noise spectrum average calculation unit 53 from the amplitude corresponding to each frequency of the speech signal on which noise is superimposed from the Fourier analysis unit 52. , The voice-only amplitude is calculated by subtracting for each frequency.

In step # 67, the half-wave rectification unit 57 sets the amplitude to 0 for the frequency at which the amplitude is negative as a result of the subtraction by the amplitude subtraction unit 56. Step # 68
As a result, the inverse Fourier transform unit 58 synthesizes a speech waveform from which noise is removed from the amplitude corresponding to each frequency from the half-wave rectification unit 57 and the phase corresponding to each signal from the Fourier analysis unit 52 (IFFT). By synthesizing), a noise-free speech waveform for one frame is generated.

In step # 69, the waveform reproducing section 59 performs the reverse processing of the clipping processing in the clipping section 51 on the noise-free speech waveform for one frame from the inverse Fourier transforming section 58, and steps it by 50%. The speech waveforms of one frame are overlapped and added to output the original continuous waveform.

[0014]

However, in the conventional noise suppressing apparatus as described above, since the noise spectrum average calculating unit 53 calculates the estimated amount of noise, it is necessary to calculate the noise spectrum as the amplitude information of the noise. However, in order to calculate this, the amplitude spectra for the past several tens of frames are required, and therefore, it is necessary to store ten or more frequency feature amounts for the past several tens of frames per frame. However, there is a problem that a large amount of memory is required.

Further, there has been a problem that an average value for the past several tens of frames has to be obtained, and a very large amount of calculation is required. An object of the present invention is to solve the above problems, and an object thereof is to provide a noise suppression device capable of removing noise from a voice on which noise is superimposed with a small storage capacity and a small amount of calculation.

[0016]

In order to achieve the above object, a noise suppressing device according to claim 1 performs windowing on an input signal at a predetermined time interval, and at the time interval from the input signal. A cutout unit that cuts out a waveform in a frame having a length corresponding to the windowing, a Fourier analysis of the waveform in the frame from the cutout unit, and a Fourier analysis unit that decomposes into an amplitude spectrum and a phase spectrum, and the Fourier Based on the amplitude spectrum from the analysis unit, a noise signal determination unit that determines whether the input signal is only noise or voice with superimposed noise, and according to the determination of the noise signal determination unit, from the Fourier analysis unit A noise spectrum updating unit that outputs a noise spectrum corresponding to the noise obtained based on the amplitude spectrum of A noise spectrum storage unit that stores a noise spectrum, an amplitude subtraction unit that subtracts the noise spectrum stored in the noise spectrum storage unit from the amplitude spectrum decomposed by the Fourier analysis unit, and the amplitude subtraction unit A half-wave rectification unit that rectifies the result of the subtraction by half-wave, inverse Fourier transform is performed based on the amplitude spectrum after the half-wave rectification by the half-wave rectification unit, and the phase spectrum from the Fourier analysis unit, and a voice waveform is reproduced. An inverse Fourier transform section for generating a continuous speech waveform by synthesizing the speech waveform of each frame reproduced by the inverse Fourier transformation section, and the noise spectrum updating section for the Fourier analysis section. Based on the amplitude spectrum decomposed from the waveform in the frame and the noise spectrum in the noise spectrum storage unit. Then, the noise spectrum of the waveform in the frame is calculated, and when the determination that the input signal of the noise signal determination unit is noise is detected, the noise spectrum obtained by the calculation is output, and When it is detected that the input signal of the noise signal determination unit is voice, the noise spectrum in the noise spectrum storage unit is output.

[0017]

According to the structure of claim 1, when the voice is occasionally inputted in the environment where the environmental noise exists, the noise spectrum updating section is arranged to output the Fourier analysis section and the previous section in the noise spectrum storage section. Based on the noise spectrum and
Find the current noise spectrum.

The noise spectrum updating unit switches the previous noise spectrum in the noise spectrum storage unit and the obtained current noise spectrum based on the judgment of the noise signal judging unit, and outputs it as an estimated noise spectrum. Update the previous noise spectrum to the current noise spectrum.

In this updating, the noise spectrum updating unit always automatically follows the fluctuation of the input environmental noise based on the judgment of the noise signal judging unit, and is not affected by sudden noise. Perform noise estimation.

[0020]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A noise suppressing device according to an embodiment of the present invention will be described below with reference to the drawings. The noise suppressing device according to the first embodiment of the present invention will be described below.

FIG. 1 is a block diagram showing the configuration of the noise suppressing apparatus of the first embodiment. In FIG. 1, 11 is a cutout unit, 12 is a Fourier analysis unit, 13 is a noise spectrum update unit, 14 is a noise spectrum storage unit, 15 is a noise signal determination unit, 16 is an amplitude subtraction unit, 17 is a half-wave rectification unit, Reference numeral 18 is an inverse Fourier transform unit, and 19 is a waveform reproducing unit.

The operation of the noise suppressor having the above components will be described below with reference to the flowchart of FIG. In step # 21, the input speech is subjected to a window in the clipping unit 11 that can obtain frequency resolution accuracy necessary for Fourier analysis later, and a period of about 10 ms and a fixed length of a section of about 20 ms are set. Cut out the waveform in the frame. Since the waveform of a certain section is cut out from the continuous waveform in this way, the condition of the window does not become a discontinuous waveform when combined and connected in a later combining process, and the accuracy of Fourier analysis can be secured.
Select analysis windows with 50% overlap so that the original waveform is obtained. As the window function, the Hanning window represented by the weighting W (n) in Expression 1 is used as in the conventional case.

In step # 22, the Fourier analysis unit 1
In 2, the voice waveform of the frame to be analyzed is subjected to frequency analysis (FFT analysis) by Fourier analysis. Since the analysis result is composed of a real number part and an imaginary number part, these are decomposed into an amplitude spectrum and a phase spectrum. The amplitude spectrum is output to the noise spectrum updating unit 13, the noise signal determining unit 15, and the amplitude subtracting unit 16. On the other hand, since the phase spectrum is resynthesized with the speech waveform, the inverse Fourier transform unit 18
Is output to

In step # 23, the noise signal determination unit 1
In step 5, in the input signal, only the voice section is discriminated from the voice section having the voice signal on which the noise signal is superimposed and the noise section having only the noise signal having no voice. This determination is performed based on the power information of the input signal. A voice section having a voice is set at a high power, and a noise section having no voice is set at a low power. As shown in Expression 2, when the ratio of the power of the current analysis frame and the power of the previous analysis frame is larger than the threshold value THp of the power ratio, it is determined that the voice has started.

[0025]

[Equation 2]

Further, as shown in Expression 3, a state smaller than the threshold value THp is represented by Expression 4 as a threshold value TH of duration.
When it lasts longer than d, it is determined that the voice ends.

[0027]

(Equation 3)

[0028]

[Equation 4]

The information "present" or "absent" of the voice is output to the noise spectrum updating section 13. In step # 24, the noise spectrum updating unit 13 estimates the environmental noise from the signal such that the voice is occasionally input in the environment where the stationary environmental noise exists. The noise spectrum updating unit 13 updates the noise signal N (i) stored in the noise spectrum storage unit 14 by the noise spectrum updating unit 13 when a signal without voice is received from the noise signal determining unit 15. The noise signal N (i) is calculated using the amplitude information Xd (i) corresponding to each frequency of the output of the Fourier analysis unit 12 and E (i) described later, as in Expression 5. In this way, the amount of memory used to update the noise spectrum corresponding to the noise signal N (i) is the number of window lengths analyzed by Fourier analysis.

[0030]

(Equation 5)

The constant μ used in the noise spectrum updating unit 13 changes according to the expression 6 according to the result of the noise signal determining unit 15. As in Equation 5, the noise signal N (i)
Changes while adapting to the input noise while noise is being input, and holds the adaptive noise signal up to immediately before, while speech with noise superimposed is being input. To realize this, the constant μ is set to a non-zero constant when the noise signal determination unit 15 determines that the input signal is noise, and is set to zero when voice is input.

[0032]

(Equation 6)

As step # 25, the amplitude subtraction unit 16 calculates the noise spectrum storage unit from the amplitude Xd (i) corresponding to each frequency of the voice signal on which the noise is superimposed from the Fourier analysis unit 12 as shown in Expression 7. The amplitude N (i) of the noise for each frequency stored in 14 is subtracted for each frequency to calculate the amplitude E (i) of only the voice.

[0034]

(Equation 7)

In step # 26, the half-wave rectification unit 17 sets the amplitude to 0 for the frequency at which the amplitude subtraction unit 16 subtracts the amplitude. Step # 27
As described above, the inverse Fourier transform unit 18 synthesizes a noise-free speech waveform from the amplitude spectrum corresponding to each frequency from the half-wave rectification unit 17 and the phase spectrum corresponding to each signal from the Fourier analysis unit 12. By performing (IFFT synthesis), a noise-free speech waveform for one frame is generated.

In step # 28, the waveform reproducing section 19 performs the reverse processing of the clipping processing in the clipping section 11 on the noise-free speech waveform of the frame of the inverse Fourier transform section 18 to obtain 50% by 1 frame. The audio waveforms of are overlapped and added to output the original continuous waveform.

By the above operation, the noise can be removed from the voice on which the noise is superimposed with a small storage capacity and a small amount of calculation. A noise suppressing device according to the second embodiment of the present invention will be described below.

FIG. 3 is a block diagram showing the configuration of the noise suppressing apparatus of the second embodiment. In FIG. 3, the basic part is the same as that of the first embodiment, but as the initial value when the noise suppression device is operated, the first spectrum of the Fourier analysis part 32 at the start of operation is set to the noise spectrum storage part 34. It is configured to set the output as the initial value of the noise spectrum.

With this configuration, a value close to the input environmental noise is stored in the noise spectrum storage unit 34 from the start of the operation. By the above operation, the time when the noise spectrum of the noise spectrum storage unit 14 becomes the input environmental noise can be shortened, and the noise removing operation can be performed immediately after the operation starts.

A noise suppressing apparatus according to the third embodiment of the present invention will be described below. FIG. 4 is a block diagram showing the configuration of the noise suppression apparatus of the third exemplary embodiment. 4, the basic part is the same as that of the first embodiment, but the configuration of the noise spectrum updating unit 43 is different. That is, in the third embodiment, the noise signal determination unit 15 of the first embodiment is deleted.

This structure can be implemented for the following reasons. In the first embodiment, μ in Expression 5 is a constant that determines the tracking speed for updating the noise spectrum, and this constant μ is set based on the judgment of the input voice of the noise signal judgment unit 15 as shown in Expression 6. , The noise spectrum is constantly updated according to Expression 5 by following all the input voices, including the voice section in which noise is superimposed.

Here, when the voice section is long, the constant μ is set so that the noise spectrum update follows the input voice based on the determination of the noise signal determination unit 15, and the noise spectrum is updated according to the equation (5). It

On the other hand, in the case of an input voice having a length of about a word (about 0.5 seconds), based on the determination of the noise signal determination unit 15, a constant μ that prevents the noise spectrum update from following the input voice. Is set and the noise spectrum is not updated because it does not follow Equation 5.

Therefore, when only a voice having a length of about a word (about 0.5 seconds) is always input, the noise spectrum update should not always follow the input voice.
Since the constant μ that does not update the noise spectrum is set based on the determination by the noise signal determination unit 15, the noise signal determination unit 15 may be omitted from the beginning.

With this configuration, the noise spectrum updating unit 43 causes the noise spectrum updating operation to follow only the voice for a short time, which has little influence on the updating of the noise spectrum.

By the above operation, when the input signal is a voice signal for a short time such as a word length,
Even if the noise / voice determination is not performed, the noise removal operation can be performed as in the first embodiment, and even if the noise signal determination unit 15 is deleted, the noise removal operation is reliably performed. be able to.

[0047]

As described above, according to the present invention, when voice is occasionally input in an environment in which environmental noise exists, the noise spectrum updating unit outputs the output of the Fourier analysis unit and the noise spectrum storage unit. Based on the previous noise spectrum in the section, the current noise spectrum is obtained, the previous noise spectrum in the noise spectrum storage section and the obtained current noise spectrum are switched based on the determination of the noise signal determination section, and the estimated noise It is possible to output as a spectrum and update the previous noise spectrum in the noise spectrum storage unit to the current noise spectrum.

In this updating, the noise spectrum updating section always automatically follows the fluctuation of the input environmental noise based on the judgment of the noise signal judging section, and is not affected by the sudden noise. Noise estimation can be performed.

Therefore, the noise can be removed from the voice on which the noise is superimposed with a small storage capacity and a small amount of calculation. Further, from the start of the operation, the noise spectrum storage unit can store a value close to the input environmental noise.

Therefore, the time during which the noise spectrum of the noise spectrum storage unit becomes the input environmental noise can be shortened, and the noise removing operation can be performed immediately after the start of the operation.

Furthermore, the noise spectrum updating unit can follow the noise spectrum updating operation for a short time speech which has little influence on the updating of the noise spectrum.

Therefore, when the input signal is a voice signal for a short time, the noise removal operation can be performed even if the determination as to whether it is noise or voice is not performed, and the noise signal determination unit is deleted. Even then, the noise removal operation can be reliably performed.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a noise suppressing device according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing the operation of the noise suppression apparatus of the same embodiment.

FIG. 3 is a configuration diagram of a noise suppressing device according to a second embodiment of the present invention.

FIG. 4 is a configuration diagram of a noise suppressing device according to a third embodiment of the present invention.

FIG. 5 is a configuration diagram of a conventional noise suppression device.

FIG. 6 is a flowchart showing the operation of the conventional noise suppression device.

[Explanation of symbols]

 11, 31, 41 Cutout unit 12, 32, 42 Fourier analysis unit 15, 35 Noise signal determination unit 13, 33, 43 Noise spectrum update unit 14, 34, 44 Noise spectrum storage unit 16, 36, 46 Amplitude subtraction unit 17 , 37, 47 Half-wave rectifying unit 18, 38, 48 Inverse Fourier transform unit 19, 39, 49 Waveform reproducing unit

Claims (3)

[Claims] 1. A cutting-out unit that performs windowing on an input signal at a predetermined time interval and cuts out a waveform in a frame having a length corresponding to the windowing at the time interval from the input signal, and from the cutting-out unit. Fourier analysis of the waveform in the frame of, the Fourier analysis unit that decomposes into an amplitude spectrum and a phase spectrum, based on the amplitude spectrum from the Fourier analysis unit, the input signal is only noise And a noise spectrum updating unit that outputs a noise spectrum corresponding to the noise obtained based on the amplitude spectrum from the Fourier analyzing unit, according to the determination of the noise signal determining unit, A noise spectrum storage unit that stores the noise spectrum output from the noise spectrum updating unit, and a Fourier analysis unit. An amplitude subtraction unit that subtracts the noise spectrum stored in the noise spectrum storage unit from the decomposed amplitude spectrum, a half-wave rectification unit that half-wave rectifies the result of the subtraction by the amplitude subtraction unit, and Inverse Fourier transform based on the amplitude spectrum after the half-wave rectification by the half-wave rectification unit and the phase spectrum from the Fourier analysis unit, an inverse Fourier transform unit for reproducing a voice waveform,
A waveform reproducing unit for generating a continuous voice waveform by synthesizing the voice waveforms of the respective frames reproduced by the inverse Fourier transform unit, and the noise spectrum updating unit from the waveform in the frame by the Fourier analyzing unit. The noise spectrum of the waveform in the frame is calculated based on the decomposed amplitude spectrum and the noise spectrum in the noise spectrum storage unit, and the determination that the input signal of the noise signal determination unit is noise is detected. In this case, the noise spectrum obtained by the calculation is output, and when the determination that the input signal of the noise signal determination unit is a voice is detected, the noise spectrum in the noise spectrum storage unit is output. Noise suppressor configured as described above. 2. The noise suppression apparatus according to claim 1, wherein the noise spectrum storage unit is configured to store the amplitude spectrum decomposed by the Fourier analysis unit as an initial value at the start of operation. 3. The noise spectrum updating unit sets a constant for updating processing to a value that does not update the noise spectrum in the noise spectrum storage unit for a voice input having a length of about a word, and the noise signal determining unit. 2. The noise spectrum calculated according to the amplitude spectrum decomposed from the waveform in the frame by the Fourier analysis unit and the noise spectrum in the noise spectrum storage unit is output regardless of the determination of 1. Noise suppressor.