JP2000330597A - Noise suppressing device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a robust noise suppressing device which can take out clear speech against variations in noise power and spectrum. SOLUTION: A suppression coefficient a corresponding to two parameters of an S/N and frequency components is stored in a suppression coefficient data table 13. When a speech signal containing noise is inputted it is converted into signal in a frequency domain by a spectrum converting means 11. The S/N presuming means 12 presumes an S/N from the noise spectrum. A suppression amount presuming means 14 takes out a desired suppression coefficient from the suppression coefficient data table 13, and performs presumption of the noise spectrum to be suppressed. A noise suppressing means 15 suppresses the noise components by using the presumed noise spectrum to the output of the spectrum converting means 11.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a noise suppression device capable of clearly extracting only an audio signal from a signal containing stationary noise having a strong whiteness or non-stationary noise whose power frequency characteristic gradually changes. It concerns the device.

[0002]

2. Description of the Related Art Wiener filters and spectral subtraction methods are known as noise suppression methods for suppressing stationary noise. In either suppression method, in order to adjust the amount of noise suppression, it is general to multiply the estimated noise spectrum by a suppression coefficient α which is a scaling variable. The function of the Wiener filter is expressed by equation (1).

(Equation 1) Note that S (ω) represents a voice spectrum, X (ω) represents an input signal spectrum, N (ω) represents a noise spectrum, and ∧ (hat) on the symbol of the spectrum signal represents the meaning of estimation. Also, {} is a symbol indicating the absolute value of the signal in the frequency domain.

[0005] The configuration of a conventional noise suppression device employing such a system will be described with reference to FIGS.
FIG. 15 is a configuration diagram of a conventional noise suppression device using a general Wiener filter. This noise suppression device includes an FFT 21 for immediately analyzing the frequency of an input signal, a power spectrum conversion unit 22 for converting an input signal on a frequency axis into a power spectrum, a suppression amount estimation unit 14 for estimating a noise suppression amount for each spectrum component, A Wiener filter estimating unit 23 for estimating various filter parameters from the noise suppression amount estimated by the amount estimating unit 14, a filter coefficient deriving unit 24 for specifically deriving the filter coefficient, and calculating the input signal using the obtained filter coefficient. By doing so, it is configured to include the filtering operation means 25 for suppressing noise.

As a specific example, in the Wiener filter estimating means 23, the suppression coefficient α is a constant value, and the characteristic H (ω) of the Wiener filter is expressed by equation (2).

(Equation 2)

The filter coefficient deriving means 24 obtains a filter coefficient by, for example, IFFT. The filtering operation unit 25 performs a convolution operation on the input signal X (t) in the time domain using the filter coefficient obtained from the filter coefficient derivation unit 24.

FIG. 16 shows a configuration of FIG. 15 in which the suppression coefficient α = 1
FIG. 9 is a characteristic diagram showing a noise suppression effect when the frequency is fixed to. Specifically, the average voice spectrum for 5 seconds is S (ω), voice and noise are mixed at an S / N ratio of 5 dB, the input signal spectrum input to the device is X1 (ω), and the input signal is FIG. Signal spectrum X after suppression by the noise suppression device of FIG.
2 (ω), the spectral distance X1 before suppression
(Ω) −S (ω) and spectral distance X2 (ω) after suppression
−S (ω) is shown in a graph. As the spectral distance is closer to 0 dB, it means that the noise is suppressed and the voice is faithfully reproduced. From this figure, it can be seen that the conventional method can always obtain a noise suppression effect of about 10 dB regardless of the frequency band. Thus, a certain amount of noise can be suppressed in all frequency bands.

[0007]

However, the noise suppressor having the above-mentioned structure cannot be said to be able to efficiently suppress the noise component when the audio level or noise level to be extracted changes. In FIG. 16, for example, 400
In the frequency band above Hz, the spectral distance after noise suppression is negative. This is because the noise spectrum N (ω) used for the estimation is pulled too much more than necessary, resulting in speech spectrum distortion. In particular, in the speech recognition function, if there is a spectrum distortion of the speech, the recognition rate decreases. To address these issues,
There is a method of reducing the value of the suppression coefficient α, that is, lightly reducing noise to eliminate spectral distortion. However, when the suppression coefficient α is reduced, the sense of noise increases in terms of audibility, and the S / N ratio after suppression also deteriorates. That is, there is a contradiction that the noise suppression effect is reduced as a whole.

[0008] The reason why the suppression cannot be performed efficiently is that the value to be suppressed changes depending on the frequency component, or the valley of the spectrum seen at several frequencies in FIG. For example, suppression works even at a good frequency.

FIG. 17 is a characteristic diagram showing respective spectral distances obtained when noise suppression is performed by a conventional method on three types of data in which speech and noise are mixed at different S / N ratios. Here, the suppression coefficient is α = 0.5. Although the spectral distance after noise suppression is no longer negative, the spectral distance is 0 at any S / N ratio.
It can be seen that noise is not efficiently suppressed depending on the S / N ratio because the noise does not reach around dB. That is, it can be seen from this figure that it is necessary to adjust the noise suppression amount according to the S / N ratio.

SUMMARY OF THE INVENTION The present invention has been made in view of such a conventional problem, and realizes a noise suppression system in which the amount of suppression is optimally changed according to the S / N ratio, and eliminates noise. An object of the present invention is to provide a noise suppression device capable of extracting a desired sound without causing sound deterioration.

[0011]

In order to solve the above-mentioned problems, the invention of claim 1 of the present application provides a spectrum conversion means for converting an input speech signal containing noise from a time domain to a frequency domain, and the spectrum conversion means. S / N estimating means for estimating the S / N ratio of the input audio signal from the converted output of the means;
A desired suppression coefficient is extracted from the suppression coefficient data table based on the S / N ratio estimated by the S / N estimation means, wherein the suppression coefficient data table stores a plurality of suppression coefficients for controlling the amount of noise suppression. A suppression amount estimating unit for estimating a noise spectrum to be suppressed; and a noise suppressing unit for suppressing a noise component included in a conversion output of the spectrum converting unit based on the suppression noise spectrum estimated by the suppression amount estimating unit. And characterized in that:

According to a second aspect of the present invention, there is provided a frame processing means for dividing an input audio signal containing noise into a unit of a fixed number of samples, performing frame conversion and outputting audio data, and converting the audio data output from the frame processing means into audio data. Spectrum conversion means for obtaining a frequency spectrum for the input signal type determination means for generating an n-level evaluation index indicating the degree of noise likeness or sound likeness on a frame basis from audio data output from the frame processing means, Power spectrum conversion means for converting the output of the spectrum conversion means into a power spectrum, and a noise spectrum learning means for learning a noise spectrum based on an output from the power spectrum conversion means and an output from the input signal type determination means. , The output from the power spectrum conversion means and the noise spectrum Based on the output from the learning means, a spectral component by S / N estimation means for estimating the S / N ratio for each frequency component of the input audio signal, the input audio signal S /
A suppression coefficient data table in which a plurality of suppression coefficients for controlling the amount of noise suppression are stored by using two of the N ratio and the frequency component as variables, and each frequency component output from the S / N estimating means for each spectrum component. A suppression coefficient determining unit that retrieves a desired suppression coefficient from the suppression coefficient data table based on the S / N ratio of each noise, and a noise spectrum that is output from the noise spectrum learning unit, and is determined by the suppression coefficient determination unit. A suppression amount estimating unit for estimating a noise spectrum to be suppressed using the suppressed coefficient, a suppression noise spectrum estimated by the suppression amount estimating unit, and a power spectrum output from the power spectrum converting unit. Noise suppression means for suppressing a noise component included in the input audio signal.

According to a third aspect of the present invention, there is provided a frame processing means for dividing an input audio signal including noise into a unit of a fixed number of samples, performing frame conversion and outputting audio data, and converting the audio data output from the frame processing means into audio data. Spectrum conversion means for obtaining a frequency spectrum for the input signal type determination means for generating an n-level evaluation index indicating the degree of noise likeness or sound likeness on a frame basis from audio data output from the frame processing means, Power spectrum conversion means for converting the output of the spectrum conversion means into a power spectrum, and a noise spectrum learning means for learning a noise spectrum based on an output from the power spectrum conversion means and an output from the input signal type determination means. Using the evaluation index as a variable, the suppression coefficient for controlling the amount of noise suppression is duplicated. The number of stored suppression coefficient data tables and the suppression coefficient determination means for extracting a desired suppression coefficient from the suppression coefficient data table based on the evaluation index output from the input signal type determination means, and the noise spectrum learning means A suppression amount estimating unit for estimating a noise spectrum to be suppressed by referring to the noise spectrum to be output and using the suppression coefficient determined by the suppression coefficient determining unit; and a suppression noise estimated by the suppression amount estimating unit. A noise suppression unit that suppresses a noise component included in the input voice signal based on a spectrum and a power spectrum output from the power spectrum conversion unit.

According to a fourth aspect of the present invention, there is provided a frame processing means for dividing an input audio signal containing noise into a predetermined number of samples, converting the data into frames, and outputting audio data; Spectrum conversion means for obtaining a frequency spectrum for the input signal type determination means for generating an n-level evaluation index indicating the degree of noise likeness or sound likeness on a frame basis from audio data output from the frame processing means, Power spectrum conversion means for converting the output of the spectrum conversion means into a power spectrum, and a noise spectrum learning means for learning a noise spectrum based on an output from the power spectrum conversion means and an output from the input signal type determination means. , The output from the power spectrum conversion means and the noise spectrum S / N estimating means for each spectral component for estimating the S / N ratio for each frequency component of the input audio signal based on the output from the learning means, the n-stage evaluation index, and the S / N of the input audio signal A suppression coefficient data table in which a plurality of suppression coefficients for controlling the amount of noise suppression are stored with three ratios and frequency components as variables; and a frequency component for each frequency component output from the S / N estimating means for each spectrum component. S
/ N ratio and a suppression coefficient determination unit for extracting a desired suppression coefficient from the suppression coefficient data table based on the evaluation index output from the input signal type determination unit, and a noise spectrum output from the noise spectrum learning unit. And a suppression amount estimating unit for estimating a noise spectrum to be suppressed using the suppression coefficient determined by the suppression coefficient determining unit; a suppression noise spectrum estimated by the suppression amount estimating unit; And a noise suppression unit for suppressing a noise component included in the input audio signal based on a power spectrum output from the spectrum conversion unit.

According to a fifth aspect of the present invention, there is provided a frame processing means for dividing an input audio signal containing noise into a unit of a fixed number of samples, performing frame conversion and outputting audio data, and converting the audio data output from the frame processing means into audio data. A spectrum converting means for obtaining a frequency spectrum, a power spectrum converting means for converting an output of the spectrum converting means into a power spectrum, and noise likeness or voice in frame units from data of the power spectrum output from the power spectrum converting means. An input signal type determining unit for each spectrum component that generates an n-level evaluation index indicating the degree of likelihood for each frequency component, an output from the power spectrum converting unit, and an output from the input signal type determining unit for each spectral component. Noise spectroscopy to learn the noise spectrum based on Means, a suppression coefficient data table storing a plurality of suppression coefficients for controlling the amount of noise suppression using two of the evaluation index and the frequency component as variables, and output from the spectrum component input signal type determination means. On the basis of the evaluation index for each frequency component, a suppression coefficient determination unit that extracts a desired suppression coefficient from the suppression coefficient data table, and a noise spectrum output from the noise spectrum learning unit, and the suppression coefficient determination unit A suppression amount estimating unit for estimating a noise spectrum to be suppressed using the determined suppression coefficient; a suppression noise spectrum estimated by the suppression amount estimating unit; and a power spectrum output from the power spectrum converting unit. Noise suppression means for suppressing a noise component included in the input audio signal. It is an butterfly.

According to a sixth aspect of the present invention, there is provided a frame processing means for dividing an input audio signal containing noise into a unit of a fixed number of samples, performing frame conversion and outputting audio data, and converting the audio data output from the frame processing means into audio data. A spectrum converting means for obtaining a frequency spectrum, a power spectrum converting means for converting an output of the spectrum converting means into a power spectrum, and noise likeness or voice in frame units from data of the power spectrum output from the power spectrum converting means. An input signal type determining unit for each spectrum component that generates an n-level evaluation index indicating the degree of likelihood for each frequency component, an output from the power spectrum converting unit, and an output from the input signal type determining unit for each spectral component. Noise spectroscopy to learn the noise spectrum based on Means, and a power spectrum of an input signal in a frame output from the power spectrum conversion means and a power spectrum of an estimated noise in the frame output from the noise spectrum learning means, and an S / N ratio in the frame is calculated. S / N by representative value or spectral component
S / N estimating means for each spectral component for estimating a ratio, and a suppression coefficient data table storing a suppression coefficient for controlling the amount of noise suppression using the S / N ratio, the evaluation index, and the frequency component as variables. From the suppression coefficient data table based on the S / N ratio output from the S / N estimating means for each spectral component and the evaluation index for each frequency component output from the input signal determining means for each spectral component. With reference to a suppression coefficient determination unit that extracts a desired suppression coefficient and a noise spectrum output from the noise spectrum learning unit, estimation of a noise spectrum to be suppressed is performed using the suppression coefficient determined by the suppression coefficient determination unit. Suppression amount estimating means to be performed, a suppression noise spectrum estimated by the suppression amount estimating means, and output from the power spectrum converting means. Based on the word spectrum and is characterized by comprising a noise suppression means for suppressing a noise component contained in the input speech signal.

According to a seventh aspect of the present invention, there is provided a pseudo audio data recording means for storing pseudo audio data having a fixed frame length in advance as audio data in a time domain, and the pseudo audio data recording means is provided only when an input audio signal is determined to be noise. Audio mixing means for mixing pseudo audio data output from audio data recording means and the input audio signal at a plurality of different S / N ratios and storing the mixed data; and a plurality of S / N ratios output from the audio mixing means. First noise suppressing means for performing noise suppression on the voice data mixed in the first and second, and spectrum comparing means for comparing the spectrum of the voice data output from the first noise suppressing means with the spectrum of the pseudo voice data. A suppression coefficient that controls the amount of noise suppression based on the comparison result output from the spectrum comparison means, and stores the updated suppression coefficient. A table, the removed suppression coefficient from the suppression coefficient data table, and the second noise suppression unit for performing noise suppression of the input audio signal, is characterized in that it comprises a.

According to the invention of claim 8 of the present application, the first and second noise suppression means of the noise suppression device according to claim 7 are replaced with a suppression coefficient from the noise suppression device according to any one of claims 1 to 6. The configuration is such that the data table is removed.

In particular, according to the configuration of the first aspect, the data table of the suppression coefficient is provided, and the S table for each frequency component of the input signal is provided.
By performing the processing of estimating the / N ratio and determining the suppression coefficient, it is possible to perform noise suppression corresponding to the sound quality of the input signal level and the noise level and the level change.

In particular, according to the configuration of the second aspect, noise and speech are distinguished for each frame, the noise spectrum is learned only when the noise is determined, and the current signal for each frequency component is obtained from the input signal power spectrum of the current frame. S / of frame
By performing the process of estimating the N ratio and determining the suppression coefficient, the suppression amount can be adjusted, and the accuracy of suppression is improved.

In particular, according to the configuration of the third aspect, by performing processing of discriminating noise and speech for each frame and determining the suppression coefficient based on the discrimination, the noise is unconditionally increased when the frame is determined to be noise. When a frame is determined to be speech and is determined to be speech, suppression can be performed so that only speech is extracted, and the efficiency of suppression is improved.

In particular, according to the structure of the fourth aspect, the effects of the second and third aspects can be simultaneously obtained.

In particular, according to the configuration of claim 5, noise and speech are distinguished for each frequency component in the frame, and a process of determining a suppression coefficient based on the distinction is performed. It is possible to greatly suppress the condition and to suppress only the voice when the frequency component is a voice, so that the suppression can be performed more accurately.

In particular, according to the structure of claim 6, the effects of claims 2 and 5 can be simultaneously obtained.

In particular, according to the configuration of the seventh and eighth aspects, the data table of the suppression coefficient is constantly updated by learning,
Noise suppression with better adaptability can be performed.

[0026]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A noise suppression apparatus according to the present invention eliminates a sense of noise by adaptively changing a noise suppression component to be estimated according to an S / N ratio and a frequency characteristic of an input signal. This is to extract clear voice while preventing deterioration of the sound. As a result, even when the S / N ratio is low, the noise suppression capability is enhanced, and only the speech spectrum components are extracted without leaving unpleasant sounds. Furthermore, by providing the noise suppression device of the present invention for a speech recognition device whose recognition rate is greatly reduced due to deterioration of the speech spectrum, the speech recognition rate can be improved. Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(Embodiment 1) A noise suppression apparatus according to Embodiment 1 of the present invention will be described with reference to FIGS. FIG. 1 is a block diagram showing a basic configuration of the noise suppression device according to the first embodiment. This noise suppression device includes a spectrum conversion unit 11, an S / N estimation unit 12, a suppression coefficient data table 13, a suppression amount estimation unit 14, and a noise suppression unit 15.

It is assumed that the input signal contains speech and stationary noise, or non-stationary noise that gradually changes from speech. For example, it is assumed that a driver utters a voice to a voice response car navigation device which is a voice response device mounted on a vehicle. At this time, since the road noise is mixed into the voice of the driver, the S / N ratio of the voice signal is deteriorated. In the following embodiment, it is assumed that such an input signal is input to the noise suppression device.

The spectrum converting means 11 is a converting means for converting an input signal including voice and noise (non-voice) from a time domain to a frequency domain. The S / N estimating means 12 discriminates speech and non-speech by analyzing the spectrum distribution of the input signal in the frequency domain, the time change of the amplitude of the input signal, or the time change of the low amplitude component and the high amplitude component. , S / N ratio. It is technically difficult to distinguish between speech and noise, and it is possible to identify the pronunciation of a particular speaker among many speakers, or to recognize the speaker's speech when sudden non-stationary noise occurs during speech. It is especially difficult. Here, since the above-described environment is targeted, the S / N ratio can be estimated. Assuming that the estimated noise spectrum is N (ω) and the spectrum of the input signal is X (ω), the S / N ratio is estimated using, for example, the following equation (3).

(Equation 3) However, the present invention is not limited to the equation (3), and may be an S / N ratio calculation method based on the following equation (4).

(Equation 4)

The suppression coefficient data table 13 has different suppression coefficients α (S / N,
ω) is a table for storing values. Note that the suppression coefficient α is a magnification variable that is multiplied by the estimated noise spectrum. The suppression amount estimating means 14 calculates the suppression coefficient data table 1
3 is an estimating means for selecting an optimal suppression coefficient α from 3 and estimating a noise spectrum to be suppressed. Noise suppression means 15
Is a suppression amount estimating means 1 for an input signal in the frequency domain.
4 is a suppression unit that actually performs noise suppression using the estimated noise spectrum output by the output unit 4.

FIG. 2 is a block diagram of a noise suppression device when a Wiener filter is used for the noise suppression means 15 in FIG. This noise suppression device is the S / N of FIG.
In addition to the estimation unit 12, the suppression coefficient data table 13, and the suppression amount estimation unit 14, an FFT 21 for immediately analyzing the frequency of the input signal, a power spectrum conversion unit 22 for converting an input signal in the frequency domain into a power spectrum, a suppression amount estimation A Wiener filter estimating means 23 for estimating various filter parameters from the noise suppression amount estimated by the means 14, a filter coefficient deriving means 24 including an IFFT and specifically deriving filter coefficients, and an input using the obtained filter coefficients. It is configured to include a filtering calculation unit 25 that suppresses noise by calculating with a signal.

The Wiener filter estimating means 23 calculates the estimated noise spectrum N estimated by the suppression amount estimating means 14.
The characteristic H (ω) of the Wiener filter is obtained from (ω) and the spectrum X (ω) of the input signal obtained from the power spectrum converting means 22. This characteristic H (ω) is described by the following equation (5).

(Equation 5)

The filtering operation means 25 operates on the input signal with the filter coefficients obtained from the filter coefficient deriving means 24 to suppress noise contained in the input signal. An example of the implementation of the filter coefficient is a multiplication coefficient in each multiplier of the FIR filter. An example of the configuration of the filtering operation unit 25 is a convolution operation circuit including a plurality of delay units, multipliers, and adders.

FIG. 3 is a block diagram of the noise suppression device when spectrum subtraction is used for the noise suppression means 15 in FIG. This noise suppression device inputs a speech signal in frame units. In addition to the S / N estimation means 12, the suppression coefficient data table 13, the suppression amount estimation means 14, the FFT 21, the power spectrum conversion means 22, the phase conversion means 31, a difference circuit 32, a spectrum restoration means 33, an IFFT 34, and a waveform connection means 35.

The phase converter 31 is a converter for calculating the phase information of the input signal. The difference circuit 32 calculates the output of the suppression amount estimating unit 14, that is, the product of the estimated noise spectrum and the suppression coefficient α (S / N, ω), and the power spectrum of the input signal obtained from the power spectrum converting unit 22. This is a circuit for calculating the difference. The spectrum restoring unit 33 is a restoring unit that restores a suppressed spectrum from the phase information obtained from the phase conversion unit 31 and the suppressed power spectrum information obtained from the difference circuit 32. IFFT34
Is means for converting the restored spectral signal after suppression into a signal in the time domain. The waveform connection unit 35 is a connection unit that connects the audio signal restored on a frame basis to the frame.

In the first embodiment, instead of the S / N estimating means 12, an input signal type judging means for judging the noise likeness or the sound likeness of the input signal and outputting the judgment result to the suppression amount estimating means 14 is provided. You may. The S / N estimating means 12 and the input signal type determining means described above may be used in combination. The suppression coefficient data table 13 has three variables of frequency and S / N ratio, two variables of frequency and input signal type, and three of frequency, S / N ratio and input signal type. Variables are conceivable, and any of them may be a data table using a set of variables.

FIG. 4 is a characteristic diagram showing spectrum distances before and after suppression when noise suppression is performed by the noise suppression device of the first embodiment. In FIG. 4, compared to FIG. 16, it can be seen that excessive noise suppression, particularly in the high frequency range, that is, voice distortion is reduced.

FIG. 5 is a characteristic diagram showing a spectral distance when noise suppression is performed by the noise suppression device of the first embodiment using noise and speech mixed at different S / N ratios. Unlike the characteristics shown in FIG. 17, in FIG. 5, the spectral distance is near 0 dB at any S / N ratio. As described above, in the present embodiment, it can be seen that noise can be efficiently suppressed even when the S / N ratio is different.

As described above, by having the data table of the suppression coefficient and performing the processing of estimating the S / N ratio for each frequency component of the input signal and determining the suppression coefficient, the sound quality of the input signal and the noise can be improved. It is possible to perform robust (high adaptive ability to change) noise suppression corresponding to the level change. The noise suppression device according to the present embodiment capable of performing high-precision noise suppression in this way is also effective as an input preprocessing unit for a speech recognition device.

(Embodiment 2) Next, a noise suppressing apparatus according to Embodiment 2 of the present invention will be described with reference to FIGS. The same parts as those in the conventional example and the first embodiment are denoted by the same reference numerals, and detailed description is omitted. FIG. 6 is a configuration diagram of the noise suppression device of the present embodiment. This noise suppression device includes an FFT 21, a power spectrum conversion unit 22,
In addition to the suppression coefficient data table 13, a frame processing unit 61, an input signal type determination unit 62, a noise spectrum learning unit 63, an S / N estimation unit for each spectral component 64,
It is configured to include a suppression coefficient determination unit 65, a suppression amount estimation unit 66, and a noise suppression unit 67.

The frame processing means 61 is a processing means for cutting out an input audio signal as a frame of a fixed length. The input signal type determination means 62
This is a determination unit that inputs one frame signal, distinguishes noise-likeness and voice-likeness of audio signals included in each frame, and outputs the determination result. Here, the determination may be performed by dividing the likelihood of noise and the likelihood of voice into n stages. The judgment method is to take the input signal in frame units,
By analyzing the spectrum distribution of the input signal in the frequency domain, the time change of the amplitude of the input signal, or the time change of the low amplitude component and the high amplitude component, it is determined whether the input signal is voice or non-voice.
When the input signal is such that a voice uttered by a driver in the vehicle is superimposed on road noise or engine noise, the above determination is technically easy.

The noise spectrum learning means 63 is a learning means for learning a power spectrum of a frame determined to be noise by the input signal type determining means 62 and temporarily storing a learning result. S / N estimating means 64 for each spectral component
S / N estimating means for estimating the S / N ratio for each frequency component in the frame from the power spectrum of the frame obtained by the power spectrum converting means 22 and the noise spectrum learned and accumulated by the noise spectrum learning means 63 It is.

The suppression coefficient determining means 65 selects an optimum suppression coefficient from the suppression coefficient data table 13 based on the spectral component S / N output from the spectral component S / N estimating means 64 and estimates the amount of suppression. The determination means is provided to the means 66. The suppression amount estimation unit 66 is an estimation unit that estimates a noise spectrum to be suppressed from the noise spectrum obtained from the noise spectrum learning unit 63 and the suppression coefficient determined by the suppression coefficient determination unit 65. The noise suppression unit 67 performs suppression for actually performing noise suppression on the input signal based on the power spectrum of the input signal obtained from the power spectrum conversion unit 22 and the noise suppression amount estimated by the suppression amount estimation unit 66. Means.

FIG. 7 is a block diagram of a noise suppression device when a Wiener filter is used for the noise suppression means 67 in FIG. This noise suppression device comprises a noise suppression unit 6
As an internal configuration of 7, a Wiener filter estimating unit 23, an IFFT 34, an impulse response control unit 77, and a convolution operation unit 78 are provided as indicated by dotted lines.

The Wiener filter estimating unit 23 calculates the estimated noise spectrum N estimated by the suppression amount estimating unit 66.
The characteristic H (ω) of the Wiener filter is estimated from (ω) and the spectrum X (ω) of the input signal obtained from the power spectrum converting means 22. IFFT34
Is a means for converting the filter characteristic output from the Wiener filter estimating means 23 into a control signal in the time domain (control signal of an impulse response). The impulse response control unit 77 is a control unit for controlling the audio signal to change smoothly between frames using a control signal of an impulse response obtained for each frame. Specifically, there is a method in which when changing from the impulse response sequence in the immediately preceding frame to the impulse response sequence in a new frame and changing the sequence, the ratio is changed by a certain ratio without changing 100%. As an example. This is effective for connecting data between a plurality of frames without feeling uncomfortable.

The convolution operation means 78 is an operation means for performing a convolution operation on an input signal using a filter coefficient appropriately output from the impulse response control means 77. The suppression coefficient data table 13 includes two values of the frequency and the S / N ratio.
What is necessary is just to prepare the data table which uses one as a variable. Note that, instead of the spectral component-specific S / N estimating means 64,
S / N estimating means for estimating a representative value of the S / N ratio for each frame may be used. In this case, a data table having only one S / N ratio as a variable may be prepared.

FIG. 8 is an explanatory diagram showing a speech recognition rate when noise is suppressed by the noise suppression device of the second embodiment. Five speakers each uttered 100 place names as 100 words, and S / N ratio -2d
Noise was added so as to be B to 10 dB, and seven types of speech signals were generated for each word, and the speech signals were evaluated. Then, the recognition rate of the place name for each audio signal was examined, and the average value was plotted. The recognition rate before noise suppression is indicated by R3, the recognition rate after noise suppression by the conventional method is indicated by R2, and the recognition rate after noise suppression by the method in the present embodiment is indicated by R1. According to this, in a region having a particularly low S / N ratio, for example, in a speech signal having an S / N ratio of 0 dB or less, the speech recognition rate is significantly improved when noise suppression is performed as compared with the case where noise suppression is not performed. To improve. In addition to this effect, it can be seen that the speech recognition rate by the noise suppression method of the present embodiment is further improved as compared with the value by the conventional noise suppression method.

As described above, in the second embodiment, noise likeness and speech likeness are distinguished for each frame, and a noise spectrum is learned only when it is determined to be noise. By performing the processing of estimating the S / N ratio of the current frame and determining the suppression coefficient, it is possible to adjust the amount of suppression robustly and improve the accuracy of suppression.

(Embodiment 3) Next, a noise suppressing apparatus according to Embodiment 3 of the present invention will be described with reference to FIG. FIG. 9 is a configuration diagram of the noise suppression device according to the present embodiment, and all the components are in accordance with the noise suppression device of the second embodiment shown in FIG. Here, the output of the input signal type determination means 62 is sent to the suppression coefficient determination means 65 so that the suppression coefficient α is determined by the input signal type determination means 62 instead of the spectral component-specific S / N estimation means 64 of FIG. It is configured to be input. Further, although the case where a Wiener filter is used as the noise suppression means 67 indicated by the dotted line is shown, other noise suppression means may be used. In addition, the input signal type determination means 62 may be a two-step determination such as noise or voice, or the degree may be n.
The determination may be made in stages. Here, as the suppression coefficient data table 13, a table having two variables of the frequency and the input signal type may be prepared.

As described above, by performing processing of discriminating the likelihood of noise and speech for each frame and determining the suppression coefficient based on the discrimination, the input signal is greatly suppressed unconditionally when the frame is likely to be noise. When the frame is likely, it is possible to suppress the input signal so that only the voice is extracted, and the efficiency of the suppression is improved.

When the variables are set by the S / N ratio as in the second embodiment, the size of the data table may be increased due to the size of the variables, which may be disadvantageous in terms of hardware implementation. In the case of the present embodiment, n in the n-stage determination of noise or voice may be an arbitrary number, and even if it is set to the minimum of 2, there is no problem, and this is advantageous in terms of hardware implementation. Such a noise suppression device is particularly effective as an input preprocessing unit for a voice recognition function mounted on a small device such as a mobile phone or an IC card.

(Embodiment 4) Next, a noise suppressing apparatus according to Embodiment 4 of the present invention will be described with reference to FIG. FIG. 10 is a configuration diagram of a noise suppression device according to the present embodiment, and all components are the same as those in FIG. 7 of the second embodiment. However, the present embodiment uses two pieces of information from the input signal type determination means 62 and the spectral component-specific S / N estimation means 64 as inputs to the suppression coefficient determination means 65 for determining the suppression coefficient α. It is a feature. FIG. 10 shows a configuration in which a Wiener filter is employed as the noise suppression means 67, but other noise suppression means may be used. As the suppression coefficient data table 13, a data sheet having three variables of frequency, input signal type, and S / N ratio may be prepared.

The S / N estimating means for each spectral component 6
Instead of 4, an S / N estimating means for newly estimating the S / N ratio for each frame may be used. In this case, as the suppression coefficient data table 13, n data sheets using only the S / N ratio as variables may be prepared for each input signal type.

As described above, in this embodiment, the likelihood of noise and the like of speech are determined for each frame in n stages.
Only when it is determined to be noise, the noise spectrum is learned. Also, from the power spectrum in the frame,
The S / N ratio for each frequency component is always estimated, and a data sheet is selected in light of the n-stage determination of the audio type of the frame. That is, by performing the process of determining the suppression coefficient by using two variables, the S / N ratio and the frequency, the suppression amount can be adjusted more robustly than in the second and third embodiments, and the suppression can be performed with high accuracy.

(Embodiment 5) Next, a noise suppressing apparatus according to Embodiment 5 of the present invention will be described with reference to FIG. FIG. 11 is a configuration diagram of the noise suppression device according to the present embodiment, and all components are the same as those in FIG. 7 of the second embodiment. However, in order to determine the frame of the input signal in n stages of noise likeness and voice likeness for each spectral component, an input signal type determining unit 111 for each spectral component is provided. As input to the suppression coefficient determining means 65 for determining the suppression coefficient α, information from the input signal type determining means 111 for each spectrum component is used.

The difference from the third embodiment is that the input signal type judging means 62 shown in FIG. 9 judges the type of noise, voice, etc. in n stages on a frame basis. The feature is that type determination such as noise or voice of n stages is made for each frequency component in the frame. FIG. 11 shows a configuration in which a Wiener filter is employed as the noise suppression means 67, but other noise suppression means may be used. As the suppression coefficient data table 13, a table having two variables of the frequency and the input signal type may be prepared.

As described above, by performing processing of discriminating the likelihood of noise and the likelihood of speech for each frequency component in the frame and determining the suppression coefficient based on the discrimination, when the frequency component is likely to be noise, the input signal is unconditionally greatly increased. Can be suppressed, and the input signal can be suppressed so as to extract only the voice when the frequency component is likely to be a voice, and the suppression can be performed more accurately than in the third embodiment.

(Embodiment 6) Next, a noise suppressing apparatus according to Embodiment 6 of the present invention will be described with reference to FIG. FIG. 12 is a configuration diagram of a noise suppression device according to the present embodiment, in which all components are the same as in FIG. 7 of the second embodiment. However, as inputs to the suppression coefficient determining means 65 for determining the suppression coefficient α, the input signal type determining means 111 for each spectral component and the S / N estimating means 6 for each spectral component are used.
The fourth embodiment is similar to the fourth embodiment in that information from the fourth embodiment is used. In the fourth embodiment, the input signal type determination unit 62 for each spectrum component performs the n-stage type determination of noise, voice, and the like in frame units. In the means 111,
The feature is that type determination such as noise-likeness and voice-likeness of n stages is made for each frequency component in the frame. FIG. 12 shows a configuration in which a Wiener filter is employed as the noise suppressing means 67, but other noise suppressing means may be used.

As the suppression coefficient data table 13, a table having three variables of frequency, input signal type, and S / N ratio may be prepared. In addition, S / N for each spectral component
Instead of the estimating means 64, the S / N is newly added in frame units.
S / N estimating means for estimating the representative value of the ratio may be used.

As described above, the processing of discriminating noise likeness and speech likeness for each frequency component in the frame, calculating the S / N ratio for each spectrum component, and determining the suppression coefficient based on these two data is performed. Thus, when the frequency component is likely to be noise, it is possible to unconditionally suppress the noise component, and when the frequency component is likely to be speech, it is possible to suppress such that only the voice is extracted, and the suppression can be performed more accurately than in the fourth embodiment. For example, if such a noise suppression device is provided before the speech recognition device, the above-described word recognition rate becomes 7%.
It improved from 5% to 82%.

(Embodiment 7) Next, a noise suppressing apparatus according to Embodiment 7 of the present invention will be described with reference to FIGS. FIG. 13 shows Embodiment 7 of the present invention.
FIG. 2 is a block diagram showing a basic configuration of a noise suppression device in FIG. This noise suppression device is provided with a pseudo voice data recording unit 13.
1, voice mixing means 132, first noise suppression means 133,
Spectrum comparing means 134, second noise suppressing means 13
5. It includes the suppression coefficient data table 13.

The pseudo sound data recording means 131 is a data recording means for storing pseudo sound data in the time domain.
The sound mixing means 132 is a mixing means for judging the type of the input signal and, when judging it as noise, taking out pseudo sound data from the pseudo sound data recording means 131 and mixing sounds at several S / N ratios. Note that it is not always necessary to determine whether or not noise is present, and the pseudo signal may always be mixed with the input signal. Further, a result obtained by learning and averaging noise on a spectrum may be returned to a time waveform and mixed.

The noise suppressing means 133 is the first noise suppressing means for suppressing the noise mixed with the respective S / N ratios with a constant suppression coefficient α, and the noise suppressing means according to any one of the first to sixth embodiments. The configuration is the same as that of the device except for the suppression coefficient data table. The spectrum comparing unit 134 compares the spectrum of the pseudo audio data with the spectrum of the audio data for each S / N ratio, which is noise-suppressed by the first noise suppressing unit 133, and evaluates the evaluation amount as the average spectral distance in the frame. Is a comparison means for evaluating the amount of suppression by using. Note that the comparison and evaluation method need not be the above-described spectral distance, and may be any method that can evaluate the state of suppression of each spectral component.

Noise suppression means 1 as second noise suppression means
35 is the same as that of the noise suppression apparatus of any one of the first to sixth embodiments except that the suppression coefficient data table is removed, and based on the comparison result of the spectrum comparison means 134, the optimum suppression coefficient Select α,
This is a means for suppressing the noise of the input signal. Also in this case, more robust noise suppression can be performed.

FIG. 14 is a configuration diagram of a noise suppression device showing a more specific configuration example of the seventh embodiment. This noise suppression device includes a noise suppression unit 133 and a noise suppression unit 1 shown in FIG.
13 are combined into one, and the noise suppressing means 136 indicated by a dotted line in FIG.
33 and the noise suppression means 135. Noise suppression means 1
At 36, the FFT 141 is the first FFT in the first noise suppression means 133 in FIG. FFT1
Reference numeral 42 denotes a second FFT that divides the noise-suppressed voice data into frames again and obtains a power spectrum of each frame.
It is. The data table determining unit 143 is a determining unit that determines the value of the suppression coefficient α for each S / N ratio from the spectral distance. For the components other than the above in FIG. 14,
This is a combination of the components of the second embodiment shown in FIG. 7 with the pseudo audio data recording unit 131 and the audio mixing unit 132 shown in FIG.

Such a configuration is advantageous in terms of hardware mounting. However, it is not necessary to combine the noise suppressing means 133 and the noise suppressing means 135 into one.
Two noise suppression means may be employed. By changing the suppression coefficient for each frame as in this example, more robust noise suppression can be realized.

As described above, by constantly calculating and using the data table of the suppression coefficient adaptively, it is possible to perform noise suppression corresponding to a change in the noise spectrum shape and level.

[0068]

As described above, according to the noise suppressing apparatus of the first, second, fourth and sixth aspects, the S / N ratio for each frequency component of the input signal is estimated by having the data table of the suppression coefficient. By performing the process of determining the suppression coefficient, it is possible to perform robust noise suppression corresponding to a change in the sound quality or level of an input signal or noise. These noise suppression devices capable of performing high-precision noise suppression in this way are also effective as input preprocessing means for a speech recognition function.

According to the noise suppressing apparatus of the third aspect, by performing processing of distinguishing noise likeness and speech likeness for each frame and determining the suppression coefficient by the distinction, unconditional processing is performed for a frame like noise. In the case of a frame that seems to be a voice, it is possible to suppress the voice so that only the voice is extracted, and the efficiency of the suppression is improved. When compared with the noise suppression device according to the second aspect, the size of the data table is increased due to the size of the variable area called the S / N ratio, whereas n in the n-stage determination of noise or voice may be small. This is advantageous in terms of hardware mounting.

According to the noise suppression apparatus of the fifth aspect, noise / voice is distinguished for each frequency component in the frame, and the processing of determining the suppression coefficient based on the distinction is performed, thereby reducing the hardware scale. It is possible to perform the suppression more accurately than the noise suppression device according to the third aspect while keeping the size small.

According to the noise suppression apparatus of the seventh and eighth aspects, the data table of the suppression coefficient is always adaptively updated and used, so that the noise suppression apparatus of the first, second, fourth and sixth aspects is provided. The apparatus can perform noise suppression corresponding to a change in sound quality or level of an input signal or noise.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a basic configuration of a noise suppression device according to Embodiment 1 of the present invention.

FIG. 2 is a block diagram showing a configuration of a noise suppression device using a Wiener filter as a specific example of Embodiment 1 of the present invention.

FIG. 3 is a block diagram illustrating a configuration of a noise suppression device when spectrum subtraction is used as a specific example of Embodiment 1 of the present invention.

FIG. 4 is a characteristic diagram showing a spectral distance for each frequency component before and after noise suppression in the noise suppression device of the first embodiment.

FIG. 5 is a block diagram showing the configuration of each S / S
FIG. 4 is a characteristic diagram showing a spectral distance for each N and each frequency component.

FIG. 6 is a block diagram showing a basic configuration of a noise suppression device according to Embodiment 2 of the present invention.

FIG. 7 is a block diagram showing a configuration of a noise suppression device using a Wiener filter as a specific example of Embodiment 2 of the present invention.

FIG. 8 is an explanatory diagram showing a speech recognition rate when the noise suppression device according to Embodiments 1 and 2 of the present invention is used as preprocessing of the speech recognition device.

FIG. 9 is a block diagram illustrating a configuration of a noise suppression device according to Embodiment 3 of the present invention.

FIG. 10 is a block diagram showing a configuration of a noise suppression device according to Embodiment 4 of the present invention.

FIG. 11 is a block diagram showing a configuration of a noise suppression device according to Embodiment 5 of the present invention.

FIG. 12 is a block diagram illustrating a configuration of a noise suppression device according to Embodiment 6 of the present invention.

FIG. 13 is a block diagram illustrating a configuration of a noise suppression device according to Embodiment 7 of the present invention.

FIG. 14 is a block diagram showing a configuration of a noise suppression device using a Wiener filter as a specific example of the seventh embodiment of the present invention.

FIG. 15 is a block diagram illustrating a configuration example of a conventional typical noise suppression device.

FIG. 16 is a characteristic diagram showing a spectrum distance for each frequency component in the noise suppressor of the conventional example.

FIG. 17 is a characteristic diagram showing a spectrum distance for each S / N and each frequency component in the conventional noise suppression device.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 Spectrum conversion means 12 S / N estimation means 13 Suppression coefficient data table 14 Suppression amount estimation means 15,136 Noise suppression means 21 FFT 22 Power spectrum conversion means 23 Wiener filter estimation means 24 Filter coefficient derivation means 25 Filtering calculation means 31 Phase conversion Means 32 Difference circuit 33 Spectrum restoration means 34 IFFT 35 Waveform connection means 61 Frame processing means 62 Input signal type determination means 63 Noise spectrum learning means 64 S / N estimation means for each spectrum component 65 Suppression coefficient determination means 66 Suppression amount estimation means 67 Noise Suppression means 77 impulse response control means 78 convolution operation means 111 input signal type determination means for each spectrum component 131 pseudo audio data recording means 132 audio mixing means 133 first noise suppression means 134 Vector comparison means 135 second noise suppression unit 141 first FFT 142 second FFT 143 data table determining means

 ────────────────────────────────────────────────── ─── Continuing on the front page (72) Inventor Satoru Ibaraki 1006 Kazuma Kadoma, Kadoma City, Osaka Prefecture Matsushita Electric Industrial Co., Ltd. F-term (reference) 9A001 BB02 BB03 GG01 GG03 HH01 HH16 HH17 HH18 JJ05 KK31

Claims (8)

[Claims] 1. A spectrum conversion means for converting an input audio signal containing noise from a time domain to a frequency domain, and an S / N estimation means for estimating an S / N ratio of the input audio signal from a converted output of the spectrum conversion means. A suppression coefficient data table in which a plurality of suppression coefficients for controlling the amount of noise suppression are stored; and a desired suppression coefficient from the suppression coefficient data table based on the S / N ratio estimated by the S / N estimation means. And a noise suppression unit that estimates a noise spectrum to be suppressed, and a noise that suppresses a noise component included in a conversion output of the spectrum conversion unit based on the suppression noise spectrum estimated by the suppression amount estimation unit. And a suppression unit. 2. A frame processing means for dividing an input audio signal containing noise into units of a fixed number of samples, converting the data into frames, and outputting audio data, and a spectrum for obtaining a frequency spectrum for the audio data output from the frame processing means. Conversion means, and n indicating the degree of noise-likeness or sound-likeness on a frame-by-frame basis from audio data output by the frame processing means.
Input signal type determining means for generating a step evaluation index; power spectrum converting means for converting the output of the spectrum converting means into a power spectrum; output from the power spectrum converting means and output from the input signal type determining means Noise spectrum learning means for learning a noise spectrum based on the following: S / N for each frequency component of the input speech signal based on an output from the power spectrum conversion means and an output from the noise spectrum learning means S / N estimating means for each spectrum component for estimating a ratio, and a suppression coefficient storing a plurality of suppression coefficients for controlling a noise suppression amount using two as an S / N ratio and a frequency component of an input speech signal. Based on a data table and an S / N ratio for each frequency component output from the spectrum component-specific S / N estimation means, A suppression coefficient determining unit that extracts a desired suppression coefficient from the compression coefficient data table; and a noise spectrum that is output from the noise spectrum learning unit, and the suppression is performed using the suppression coefficient determined by the suppression coefficient determination unit. A suppression amount estimating unit for estimating a noise spectrum; a noise component included in the input voice signal based on the suppression noise spectrum estimated by the suppression amount estimating unit and a power spectrum output from the power spectrum converting unit. And a noise suppressing unit for suppressing noise. 3. A frame processing means for dividing an input audio signal containing noise into a predetermined number of samples, converting the data into frames, and outputting audio data, and a spectrum for obtaining a frequency spectrum for the audio data output from the frame processing means. Conversion means, and n indicating the degree of noise-likeness or sound-likeness on a frame-by-frame basis from audio data output by the frame processing means.
Input signal type determining means for generating a step evaluation index; power spectrum converting means for converting the output of the spectrum converting means into a power spectrum; output from the power spectrum converting means and output from the input signal type determining means A noise spectrum learning unit that learns a noise spectrum based on the above, a suppression coefficient data table storing a plurality of suppression coefficients that control the amount of noise suppression using the evaluation index as a variable, and the input signal type determination unit. A suppression coefficient determining unit that extracts a desired suppression coefficient from the suppression coefficient data table based on the evaluation index output from the control unit, and a noise spectrum that is output from the noise spectrum learning unit, and is determined by the suppression coefficient determination unit. Suppression amount estimation for estimating the noise spectrum to be suppressed using the reduced suppression coefficient Means, and a noise suppression means for suppressing a noise component included in the input audio signal based on the suppression noise spectrum estimated by the suppression amount estimation means and the power spectrum output from the power spectrum conversion means. A noise suppression device, comprising: 4. A frame processing means for dividing an input audio signal including noise into units of a fixed number of samples, converting the data into frames, and outputting audio data, and a spectrum for obtaining a frequency spectrum for the audio data output from the frame processing means. Conversion means, and n indicating the degree of noise-likeness or sound-likeness on a frame-by-frame basis from audio data output by the frame processing means.
Input signal type determining means for generating a step evaluation index; power spectrum converting means for converting the output of the spectrum converting means into a power spectrum; output from the power spectrum converting means and output from the input signal type determining means Noise spectrum learning means for learning a noise spectrum based on the following: S / N for each frequency component of the input speech signal based on an output from the power spectrum conversion means and an output from the noise spectrum learning means S / N estimating means for each spectral component for estimating a ratio; and a plurality of suppression coefficients for controlling the amount of noise suppression using the n-stage evaluation index, the S / N ratio of the input speech signal, and the frequency component as variables. And the S / N for each frequency component output from the spectral component-specific S / N estimating means. Reference is made to a suppression coefficient determining means for extracting a desired suppression coefficient from the suppression coefficient data table based on the ratio and the evaluation index output from the input signal type determination means, and a noise spectrum output from the noise spectrum learning means. A suppression amount estimating unit that estimates a noise spectrum to be suppressed using the suppression coefficient determined by the suppression coefficient determining unit; a suppression noise spectrum estimated by the suppression amount estimation unit; and the power spectrum conversion. A noise suppression unit configured to suppress a noise component included in the input voice signal based on a power spectrum output from the unit. 5. A frame processing means for dividing an input audio signal containing noise into units of a fixed number of samples, converting the data into frames, and outputting audio data, and a spectrum for obtaining a frequency spectrum for the audio data output from the frame processing means. Conversion means, power spectrum conversion means for converting the output of the spectrum conversion means into a power spectrum, and power spectrum data output from the power spectrum conversion means, from the data of the power spectrum, n stages indicating the degree of noise likeness or sound likeness in frame units A spectrum component input signal type determining means for generating an evaluation index of each frequency component, and a noise spectrum based on an output from the power spectrum converting means and an output from the spectrum component input signal type determining means. Noise spectrum learning means for learning, the evaluation finger Coefficient data table in which a plurality of suppression coefficients for controlling the amount of noise suppression are stored with two variables, namely, a frequency component and a frequency component, and an evaluation for each frequency component output from the spectrum component input signal type determination means. Based on the exponent, a suppression coefficient determining unit that extracts a desired suppression coefficient from the suppression coefficient data table, and a noise spectrum output from the noise spectrum learning unit, and the suppression coefficient determined by the suppression coefficient determination unit is referred to. A suppression amount estimating means for estimating a noise spectrum to be suppressed by using the input sound signal based on a suppression noise spectrum estimated by the suppression amount estimating means and a power spectrum output from the power spectrum converting means. Noise suppression means for suppressing a noise component included in the image signal. Location. 6. A frame processing means for dividing an input audio signal including noise into a unit of a fixed number of samples, performing frame conversion and outputting audio data, and a spectrum for obtaining a frequency spectrum for the audio data output from the frame processing means. Conversion means, power spectrum conversion means for converting the output of the spectrum conversion means into a power spectrum, and power spectrum data output from the power spectrum conversion means, from the data of the power spectrum, n stages indicating the degree of noise likeness or sound likeness in frame units A spectrum component input signal type determining means for generating an evaluation index of each frequency component, and a noise spectrum based on an output from the power spectrum converting means and an output from the spectrum component input signal type determining means. Noise spectrum learning means for learning, and the power Using the power spectrum of the input signal in the frame output from the vector conversion means and the power spectrum of the estimated noise in the frame output from the noise spectrum learning means, a representative value or spectral component of the S / N ratio in the frame S / N estimating means for each spectral component for estimating another S / N ratio, and a suppression coefficient for controlling a noise suppression amount using three of the S / N ratio, the evaluation index, and the frequency component as variables. And S output from the S / N estimating means for each spectral component.
/ N ratio, and an evaluation index for each frequency component output from the spectrum component-based input signal determination means,
A suppression coefficient determining unit that extracts a desired suppression coefficient from the suppression coefficient data table; and a noise spectrum output from the noise spectrum learning unit, and the suppression is performed using the suppression coefficient determined by the suppression coefficient determination unit. A suppression amount estimating unit for estimating a power noise spectrum; a noise included in the input voice signal based on the suppression noise spectrum estimated by the suppression amount estimating unit and a power spectrum output from the power spectrum converting unit. And a noise suppressing unit for suppressing a component. 7. A pseudo-sound data recording means for storing pseudo-speech data having a predetermined frame length as time-domain sound data in advance, and a pseudo-speech data recording means which outputs the pseudo-speech data recording means only when an input sound signal is determined to be noise. Sound mixing means for mixing and storing pseudo sound data and the input sound signal at a plurality of different S / N ratios, and sound data output from the sound mixing means and mixed at a plurality of S / N ratios. A first noise suppression unit that performs noise suppression; a spectrum comparison unit that compares a spectrum of audio data output from the first noise suppression unit with a spectrum of the pseudo audio data; an output of the spectrum comparison unit Based on the comparison result
Update the suppression coefficient to control the amount of noise suppression, a suppression coefficient data table that stores the updated suppression coefficient, and extract the suppression coefficient from the suppression coefficient data table,
Second noise suppression means for suppressing noise of an input voice signal;
A learning-type noise suppression device comprising: 8. The noise suppression device according to claim 1, wherein the first and second noise suppression units are obtained by removing the suppression coefficient data table from the noise suppression device according to claim 1. 7. The noise suppression device according to 7.