JP2014021438A - Noise suppression device and program thereof - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a noise suppression device and program which can obtain a good result without unnatural noise by noise suppression processing.SOLUTION: A noise suppression device comprises: several suppression processing units outputting noise suppression voice data about input voice data by processing with a noise suppression method different in respectively; a weight calculation unit calculating a weighting factor for each noise suppression method on the basis of the noise suppression voice data output by the several suppression processing units; and a voice integration unit mixing the noise suppression voice data by multiplying the weighting factor for each noise suppression method calculated by the weight calculation unit and the noise suppression voice data output by the several suppression processing units together.

Description

  The present invention relates to audio processing. In particular, the present invention relates to a noise suppression apparatus and program capable of suppressing noise mixed in speech.

  Recording of audio for broadcasting such as television and radio is not necessarily performed in an environment suitable for recording of audio material, including the case of live broadcast. In particular, when relaying from the site of an emergency report, etc., it may be necessary to prepare electric power by in-house power generation, and it is inevitable that various noises are mixed when recording audio. In order to obtain clear sound that can withstand broadcasting even in such a situation, a technique for suppressing mixed noise with high quality is required.

  Of the conventional techniques for suppressing the noise component added to the speech, one of the well-known methods is a spectral subtraction technique. Non-Patent Document 1 describes a technique of spectral subtraction. This method is a method for reducing noise by estimating an average value of a spectrum of noise and subtracting the estimated average value from a spectrum of an input signal in which noise is mixed.

  As another technique, there is a Wiener filter method in which a linear filter that minimizes the mean square error between an original audio signal and an estimated audio signal is configured to obtain an original audio signal from an input signal in which noise is mixed. Non-Patent Document 2 describes the Wiener filter method.

  As another technique, while estimating the SN ratio for each frequency from the average value of the amplitude spectrum of the input signal mixed with noise and the noise estimation spectrum, the average of the short-time amplitude spectrum of the estimated speech signal as the original speech signal There is an MMSE-STSA method that restores the short-time amplitude spectrum so as to minimize the square error. Non-Patent Document 3 describes the MMSE-STSA method.

  As another method, there is a signal subspace method in which an original speech component is estimated by separating an input signal in which noise is mixed into a signal space composed of speech and noise and a noise space composed only of noise. Non-Patent Document 4 describes the signal subspace method.

S. F. Boll, “Suppression of acoustic noise in speech using spectral subtraction”, IEEE Transactions on Acoustics, Speech, & Signal Processing, vol.ASSP-27, no.7, pp. 113-120, 1979 J. S. Lim and A. V. Oppenheim, “All-pole modeling of degraded speech”, IEEE Transactions on Acoustics, Speech, & Signal Processing, vol. 26, no. 3, pp. 197-210, 1978 Y. Ephraim and D. Malah, “Speech enhancement using a minimum-mean square error short-time spectral amplitude estimator”, IEEE Transactions on Signal Processing, vol. 32, no. 6, pp. 1109-1121, 1984 Y. Ephraim and H. L. V. Tres, “A Signal subspace approach for speech enhancement”, IEEE Transactions on Speech and Audio Processing, vol. 3, no. 4, pp. 251-266, 1995

In the method of spectral subtraction described in Non-Patent Document 1, an unnatural noise component (musical noise) may be generated in the processed speech. This is because, when the difference between the spectrum of the noise-mixed speech and the spectrum of the noise estimation value is a negative value, the amplitude spectrum of the noise-suppressed speech is obtained by half-wave rectification in which the value is replaced with zero. As a result, an independent small peak is generated at a random frequency position for each processing frame, which is perceived as musical noise, and the quality of noise-suppressed speech deteriorates. In addition, noise estimation includes errors, and the noise is perceived, resulting in sound quality degradation in the noise-suppressed speech.
Similarly, since the methods described in Non-Patent Documents 2 to 4 depend on the noise component estimation accuracy, sound quality deterioration due to an estimation error is inevitable.
Note that the methods described in Non-Patent Documents 3 and 4 are superior to the methods described in Non-Patent Documents 1 and 2 in suppressing the noise part, but the deterioration of the voice part is easily perceived.
As described above, it is impossible to obtain noise-suppressed speech with high quality by any method, and this problem cannot be solved.

  The present invention has been made on the basis of the above problem recognition, and provides a noise suppression device and a program that can obtain better results than conventional noise suppression methods.

[1] In order to solve the above-described problem, a noise suppression device according to an aspect of the present invention is configured to output a plurality of noise-suppressed speech data by performing processing based on different noise suppression methods on input speech data. A suppression processing unit, a weight calculation unit that calculates a weighting factor for each noise suppression method based on the noise-suppressed speech data output from the plurality of suppression processing units, and each calculated by the weight calculation unit A speech integration unit that multiplies each of the noise-suppressed speech data output from the plurality of suppression processing units by a weighting coefficient for the noise suppression method, and mixes the noise-suppressed speech data.
Noise is mixed in the voice input to the noise suppression device. With the above configuration, each suppression processing unit outputs noise-suppressed voice data by a different noise suppression method. Each noise-suppressed voice data includes a voice component (clean speech) and a noise component, and the noise component is suppressed by each noise suppression method. Note that the use of different noise suppression methods also changes the way in which noise components are suppressed. The voice integration unit mixes these noise-suppressed voice data. Such mixing reduces the energy of the noise component. Note that the noise-suppressed voice data mixed by the voice integration unit may be time domain signal data or frequency domain signal data.
Moreover, according to said structure, when a speech integration part mixes noise suppression audio | voice data, it weights according to a noise suppression method. Therefore, the noise-suppressed voice data can be mixed at a better ratio according to the noise suppression method, and the effect of noise suppression is improved.

[2] Further, according to one aspect of the present invention, in the noise suppression device, the weight calculation unit calculates a correlation coefficient between the noise-suppressed speech data output from the plurality of suppression processing units. The noise suppression method having a higher correlation with other noise suppression methods is calculated such that the value of the weighting coefficient is increased.
It can be said that the greater the correlation coefficient between the noise-suppressed speech data, the stronger the speech (clean speech) component in the noise-suppressed speech data than the noise component. Therefore, the noise-suppressed voice data obtained by the noise suppression method in which the voice component is stronger than the noise component can be mixed with a stronger (higher weight). Therefore, the effect of noise suppression is further improved. As an example, the weighting factor for a certain noise suppression method is proportional to the sum of cross-correlation coefficients between that noise suppression method and other noise suppression methods (summation for other noise suppression methods). Is calculated.

[3] Further, according to an aspect of the present invention, in the noise suppression device, the weight calculation unit is configured to perform the noise suppression method based on the noise suppression speech data output from the plurality of suppression processing units. For noise-suppressed speech data, an adaptive filter coefficient that uses the noise-suppressed speech data as a desired signal by another noise suppression method is calculated. The larger the calculated adaptive filter coefficient value, the larger the weighting factor value. It is characterized by calculating as follows.
It can be said that the larger the calculated adaptive filter coefficient is, the stronger the speech (clean speech) component is compared to the noise component in the noise-suppressed speech data. Therefore, the noise-suppressed voice data obtained by the noise suppression method in which the voice component is stronger than the noise component can be mixed with a stronger (higher weight). Therefore, the effect of noise suppression is further improved. As an example, the weighting factor for a noise suppression method is such that the noise suppression method is proportional to the sum of the adaptive filter coefficients for which the other noise suppression method is the desired signal (summation for other noise suppression methods). Is calculated.

[4] Further, according to one aspect of the present invention, in the above-described noise suppression device, a frequency characteristic calculation unit that calculates frequency characteristic data based on the noise-suppressed speech data output from each of the plurality of suppression processing units; An amplitude characteristic calculation unit that calculates amplitude characteristic data based on the frequency characteristic data, and the weight calculation unit calculates a weighting coefficient for each noise suppression method based on the amplitude characteristic data. The voice integration unit mixes the noise-suppressed voice data by multiplying the amplitude characteristic data by the weighting coefficient and mixing them.
With this configuration, the time domain signal data output from the suppression processing unit can be converted into frequency domain signal data, and the audio integration unit can mix the audio signals in the frequency domain. The mixed frequency domain audio signal may be converted back to the time domain audio signal as appropriate. The frequency characteristic calculation unit calculates frequency characteristic data from the noise-suppressed voice data by performing Fourier transform. Note that both the frequency characteristic data and the amplitude characteristic data described above are noise-suppressed speech data processed by the noise suppression method by the suppression processing unit.

  [5] Further, according to one aspect of the present invention, a plurality of suppression processing steps for outputting noise-suppressed speech data by performing processing based on different noise suppression methods on waveform data of speech input to a computer, A weight calculation process for calculating a weighting coefficient for each noise suppression method based on the noise-suppressed speech data output in the suppression processing process, and a weight for each noise suppression method calculated by the weight calculation process A program for executing a process of a voice integration process of multiplying each noise-suppressed voice data output from the plurality of suppression processing processes by a coefficient and mixing the noise-suppressed voice data.

  According to the present invention, a better noise suppression result can be obtained than when any one of the conventional noise suppression methods is used alone.

It is a block diagram which shows the function structure of the noise suppression apparatus by the 1st Embodiment of this invention. It is a block diagram which shows the detailed functional structure of the weight calculation part by the embodiment. It is a graph (voice waveform) which shows the example of the result of the noise suppression by the noise suppression apparatus of the embodiment. It is a graph (voice spectrum) which shows the example of the result of the noise suppression by the noise suppression apparatus of the embodiment. It is a graph which shows the time change of the weighting coefficient value which the weight calculation part calculated at the time of the noise suppression process by the noise suppression apparatus of the embodiment. It is a graph which shows the time change of the cross correlation coefficient which the cross correlation coefficient calculation part calculated at the time of the noise suppression process by the noise suppression apparatus of the embodiment. It is a block diagram which shows the function structure of the noise suppression apparatus by 2nd Embodiment. It is a graph (voice waveform) which shows the example of the result of the noise suppression by the noise suppression apparatus of the embodiment. It is a graph (voice spectrum) which shows the example of the result of the noise suppression by the noise suppression apparatus of the embodiment. It is a block diagram which shows the function structure of the weight calculation part by 3rd Embodiment.

Next, embodiments of the present invention will be described with reference to the drawings.
In addition, in the following description, when a mathematical expression or an expression in a mathematical expression is referred to, when a hat “^” is added on a certain expression (characters such as variables), the expression and “^” Is enclosed in square brackets. For example, [x ^] indicates that a hat is added on the letter x. When a tilde is added on a certain expression (character), the expression and “˜” are enclosed in square brackets. For example, [x˜] indicates that a tilde is added on the letter x. When an absolute value symbol is attached to a certain expression, the expression is enclosed by a vertical bar “|”. For example, | x | represents that an absolute value symbol is attached to the character x.

[First Embodiment]
FIG. 1 is a block diagram illustrating a functional configuration of the noise suppression device according to the first embodiment. As shown in the figure, the noise suppression apparatus 1 includes a voice input unit 11, a waveform cutout unit 12, I (I is an integer of 2 or more) suppression processing units 14-1 to 14-I, and noise-suppressed speech. The matrix storage unit 15, the weight calculation unit 16, the voice integration unit 17, the waveform superposition unit 18, and the voice output unit 19 are configured.

The voice input unit 11 acquires voice signal data from the outside.
The waveform cutout unit 12 cuts out the voice acquired by the voice input unit 11 into an appropriate analysis frame.
Each of the suppression processing units 14-1 to 14-I applies a noise suppression method to the input speech and outputs noise-suppressed speech data. The suppression processing units 14-1 to 14-I apply the noise suppression method for each extracted analysis frame. Each of the suppression processing units 14-1 to 14-I uses I types (I ≧ 2) of noise suppression methods having different properties. Here, each of the suppression processing units 14-1 to 14-I performs noise suppression processing using a plurality of noise suppression methods having different properties from each other, thereby obtaining the effects of the present embodiment. A better effect is obtained when I ≧ 3. Examples of the noise suppression method to be used include the aforementioned spectral subtraction, Wiener filter method, MMSE-STSA method, and signal subspace method. Other noise suppression methods may be used.

The noise-suppressed speech matrix storage unit 15 stores data of noise-suppressed speech that is a result of processing by the suppression processing units 14-1 to 14-I. Specifically, the noise suppression speech matrix storage unit 15 stores data in the form of a noise suppression speech matrix configured by arranging the noise suppression speech vectors generated by the suppression processing units 14-1 to 14-I. . The noise suppression speech vector is data for each analysis frame.
The weight calculation unit 16 reads the noise-suppressed speech data output from the suppression processing units 14-1 to 14-I and temporarily stored in the noise-suppressed speech matrix storage unit 15, and based on this data, each noise suppression Calculate the weighting factor for the modulo. The weight calculation unit 16 calculates a weight coefficient so that the noise suppression result after mixing is optimized. The weight calculation unit 16 calculates the above weight coefficient for each analysis frame. Details of the weighting factor calculation method will be described later.

The voice integration unit 17 mixes the noise suppression voice data output from the suppression processing units 14-1 to 14-I. Further, when mixing the noise-suppressed voice data, the voice integration unit 17 weights the mixing ratio for each noise suppression method using a weighting coefficient corresponding to each noise suppression method. The voice integration unit 17 mixes the noise-suppressed voice data for each analysis frame. The weight coefficient is calculated by the weight calculation unit 16.
Based on the voice waveform data for each analysis frame mixed by the voice integration unit 17, the waveform superposition unit 18 generates voice waveform data that is shifted by the shift width of the analysis frame and superimposed. Needless to say, the speech waveform data generated by the waveform superimposing unit 18 is a speech waveform that has been subjected to noise suppression processing by a plurality of noise suppression methods and mixed based on weighting factors.
The audio output unit 19 outputs the audio generated by the waveform superimposing unit 18 to the outside.

  The voice input unit 11 acquires a voice signal from the outside. The voice input unit 11 performs AD (analog-to-digital) conversion when voice is acquired as an analog signal. Then, the voice input unit 11 supplies the digitized voice data to the waveform cutout unit 12. This audio data is, for example, data having a sampling frequency of 16 kHz and a quantization bit number of 16 bits (bits). Note that noise is mixed in the voice acquired by the voice input unit 11.

  Data that the voice input unit 11 supplies to the waveform cutout unit 12 is represented as noise-mixed voice y (n). Here, n is a time-series sample number, and y (n) is the sample value. The noise-mixed speech y (n) is composed of an additive noise model of the following equation (1) by speech (no noise) speech x (n) and noise d (n).

  The waveform cutout unit 12 cuts out the noise-containing voice y (n) acquired from the voice input unit 11 for each appropriate analysis frame. For example, the analysis window width N is set to 256 samples (about 16 milliseconds (= 256/16 kHz)), and the shift width is set to 128 samples (about 8 milliseconds), which is half the window width (N / 2). Note that the analysis window width may be set to different values as appropriate. The data of the nth sample in the cut out mth frame is represented as y (m, n). For simplicity, the noise-mixed speech vector cut out in the mth frame is represented by y as in the following equation (2).

  In addition, "T" attached | subjected to the right shoulder of the right side of Formula (2) represents transposition. Also, white bold “R” represents a set of real numbers. That is, the vector y is an N-dimensional column vector whose elements are real numbers. Then, the waveform cutout unit 12 supplies the data of the vector y to each of the suppression processing units 14-1 to 14-I.

  Each of the suppression processing units 14-1 to 14-I processes a given noise-mixed speech vector y by a unique noise suppression method. The i-th (i = 1, 2,..., I) suppression processing unit 14-i performs processing for obtaining noise-suppressed speech by using its own noise suppression method Fi. At this time, it is desirable that the suppression processing units 14-1 to 14-I use I noise suppression methods having different properties. When the noise suppression method Fi is regarded as a certain function, the processing by the suppression processing units 14-1 to 14-I is expressed by the following expression (3).

  In Expression (3), [x ^] i is a noise-suppressed speech vector calculated by the suppression processing unit 14-i using the noise-mixed speech vector y as an input. [X ^] is a noise-suppressed speech matrix (N rows and I columns) having row vectors from the noise-suppressed speech vectors [x ^] 1 to [x ^] I. Note that it is assumed that each noise-suppressed speech vector [x ^] i is available. I is a set of noise suppression method numbers. That is, I = {1, 2,..., I}.

  Each of the suppression processing units 14-1 to 14 -I writes the data of the noise suppression speech vector [x ^] i calculated by itself into the noise suppression speech matrix storage unit 15.

  Next, the weight calculation unit 16 obtains a weight coefficient for speech integration based on the obtained noise-suppressed speech matrix. A weight vector corresponding to each noise suppression method Fi is defined as wi, and a column vector w having these weight coefficients as elements is defined as in the following Expression (4).

A method for calculating the vector w having the weight coefficient wi as an element will be described in detail later.
Next, the speech integration unit 17 multiplies the noise-suppressed speech matrix [X ^] read from the noise-suppressed speech matrix storage unit 15 by the vector w of the weight coefficient wi corresponding to each noise suppression method, thereby integrating noise suppression. A speech vector [x˜] is calculated. That is, the integrated noise-suppressed speech vector [x˜] is expressed by the following equation (5). That is, the vector [x˜] corresponds to a product obtained by multiplying each noise-suppressed speech vector [x ^] i by the weighting factor wi and mixing (a product-sum type using the weighting factor).

The integrated noise-suppressed speech vector [x˜] calculated by Expression (5) is an N-dimensional column vector.
Equation (6) below represents the relationship between the speech vector x to be obtained and the error vector ei between each noise-suppressed speech vector [x ^] i. E is a matrix (N rows and I columns) having error vectors e1, e2,.

  From the equation (6), the following equation (7) can be obtained. That is, each noise-suppressed speech vector [x ^] i can be expressed by Expression (7).

  Substituting Equation (7) into Equation (3), the integrated noise-suppressed speech vector [x˜] represented by Equation (5) is a matrix of error and the speech vector x to be obtained as shown in Equation (8) below. It can be expressed as E.

  If the weighting coefficient vector w that minimizes the second term (product of the error matrix and the weighting coefficient vector) on the right side in Equation (8) is obtained, the integrated noise-suppressed speech vector [x˜] becomes the speech vector x to be obtained. Will approach. However, in reality, the speech vector x and the error matrix E to be obtained are unknown. Therefore, in order to obtain the optimum weight coefficient vector wopt, an optimization method related to the weight coefficient vector w is used. The optimization method used in this embodiment uses a correlation coefficient.

FIG. 2 is a block diagram illustrating a detailed functional configuration of the weight calculation unit. As illustrated, the weight calculation unit 16 includes a cross-correlation coefficient calculation unit 201, a cross-correlation coefficient addition unit 202, and a weight coefficient normalization unit 203.
The cross-correlation coefficient calculation unit 201 calculates a cross-correlation coefficient between these vectors based on noise-suppressed speech vectors obtained by the respective noise suppression methods. That is, the cross-correlation coefficient calculation unit 201 calculates a cross-correlation coefficient between noise suppression methods for each analysis frame.
The cross-correlation coefficient adding unit 202 adds a cross-correlation coefficient between the noise suppression method and another noise suppression method for a certain noise suppression method for all the other noise suppression methods (takes a summation). ). This value is the basis of the weight coefficient value for the noise suppression method.
The weight coefficient normalization unit 203 normalizes the weight coefficient value for each noise suppression method calculated by the cross correlation coefficient addition unit 202. Specifically, the weight coefficient normalization unit 203 adjusts so that the sum of the weight coefficients for all noise suppression methods becomes 1, for example.
Details of the processing of each unit will be described below.

Hereinafter, the procedure for calculating the weighting factor will be described with reference to the block diagram of FIG.
First, the cross-correlation coefficient calculation unit 201 obtains a cross-correlation coefficient between each noise-suppressed speech vector [x ^] i by each noise suppression method. The cross-correlation coefficient xcori, j between the noise-suppressed speech vectors [x ^] i and [x ^] j is calculated by the following equation (9).

In the equation (9), E [] is an expected value. That is, the cross-correlation coefficient calculated by Equation (9) is obtained by dividing the covariance of the noise-suppressed speech vectors [x ^] i and [x ^] j by their standard deviations. The noise-suppressed speech vectors [x ^] i and [x ^] j are obtained by using noise suppression methods having different properties. Therefore, the cross-correlation coefficient xcori, j between the noise-suppressed voices is expected to be high in the speech section, and the cross-correlation coefficient xcori, j between the noise-suppressed voices is expected to be low in the non-speech period (noise section). The
The cross-correlation coefficient calculation unit 201 calculates cross-correlation coefficients for all i and j combinations where i, jεI, and i ≠ j. Then, the cross-correlation coefficient calculating unit 201 passes the obtained cross-correlation coefficient to the cross-correlation coefficient adding unit 202.

  Next, the cross-correlation coefficient adding unit 202 uses the cross-correlation coefficient xcori, j calculated by the cross-correlation coefficient calculating unit 201 to use the weight coefficient [w ^] for each noise-suppressed speech vector [x ^] i. i is calculated. The weighting coefficient [w ^] i is calculated by the following equation (10).

  Here, n is an index for setting the degree of the weight coefficient, and for example, n = 2. Note that the value of n may be appropriately changed depending on the setting. As shown in Expression (10), the weighting coefficient [w ^] i is based on the sum of the correlation coefficients related to the noise-suppressed speech vector [x ^] i. In other words, the weighting coefficient [w ^] i is based on the sum of the correlation coefficients between the noise-suppressed speech vectors [x ^] i and [x ^] j with respect to j (where i ≠ j). To the nth power.

  Next, the weight coefficient normalization unit 203 normalizes the weight coefficient obtained by Expression (10) so that the weight coefficient vector [w ^] satisfies Expression (4). The normalized weight coefficient vector is expressed by Equation (11).

  Then, the weight coefficient vector [w ^] obtained in this way is set as the optimum weight coefficient vector wopt. That is, wopt = [w ^]. The weight calculation unit 16 outputs the weight coefficient vector wopt obtained in this way.

  Returning to the block diagram of FIG. 1, the speech integration unit 17 applies the optimum weight coefficient vector wopt supplied from the weight calculation unit 16 to the equation (5), that is, substitutes wopt for w in the equation (5). Thus, an optimal integrated noise suppression speech vector is calculated. The speech integration unit 17 calculates an optimal integrated noise suppression speech vector [x˜] opt by the following equation (12).

  In equation (12), m represents the index of the frame, and n represents the index of the sample in the frame. Also, as shown in the following equation (13), c is a constant obtained by averaging the cross-correlation coefficients of equation (9), and is used to set the degree of suppression of non-speech intervals (noise intervals). Use.

  K in the equation (13) is an index for setting the degree of constant, and for example, k = 2. Note that the value of k may be appropriately changed depending on the setting.

  The waveform superimposing unit 18 shifts the time waveform [x˜] opt (m, n) calculated by the equation (12) by the shift width for each frame and superimposes the noise suppressed speech [x˜] ( n).

  3 and 4 are graphs showing examples of results of noise suppression by the noise suppression device of the present embodiment. (A) to (g) in FIG. 3 each show a speech waveform, where the horizontal axis represents time and the vertical axis represents amplitude. Moreover, (a) to (g) in FIG. 4 each show a speech spectrum, the horizontal axis is time, and the vertical axis is frequency. The unit of the horizontal axis in FIGS. 3 and 4 is second. The unit of the vertical axis in FIG. 4 is hertz. FIG. 4 represents the time transition of the strength of the component for each frequency as the gray scale density, and the darker the color (that is, the closer to black), the stronger the component. In this example, each of the suppression processing units 14-1, 14-2, and 14-3 performs the suppression process using three types of noise suppression methods. 3 and 4, (a) is clean speech, (b) is additional noise, (c) is a noise-mixed speech that is input to the noise suppression device, and (d) is the noise suppression device of this embodiment. (E) is a noise-suppressed speech by the noise suppression method 1 (output from the suppression processing unit 14-1), and (f) is a noise-suppressed speech by the noise suppression method 2 (suppression processing unit 14-2). And (g) show examples of noise-suppressed speech (output from the suppression processing unit 14-3) by the noise suppression method 3.

Compared with the noise suppression methods 1 to 3, it can be seen that the present method suppresses the degradation of the speech section and effectively suppresses the noise in the non-speech section (noise section). For example, in FIGS. 3 and 4, (d) the noise part in the noise suppression result according to the present embodiment is as small as (g) noise suppression method 3, and (d) the voice part of the present embodiment is (e). The noise suppression method 1 is clear and there is little information loss and distortion. In FIG. 4, (g) noise suppression method 3 has lost high-frequency information (upper graph) in the speech portion, but (d) in this embodiment, (e) noise suppression method 1 is used. Information remains. Although such a difference can be confirmed from the graph, the evaluation using the objective evaluation value will be described later.

  FIG. 5 is a graph showing the time change of the weighting coefficient value calculated by the weight calculation unit 16 during the noise suppression processing by the noise suppression device of the present embodiment. In each of FIGS. 4A to 4C, the horizontal axis represents time, and the vertical axis represents the value of the weighting factor. The unit of the horizontal axis is second. The values of the weighting factors shown in FIG. 5 correspond to the noise suppression processing execution results shown in FIGS. In this graph, the weighting factor values corresponding to the suppression processing units 14-1, 14-2, and 14-3 are w1 in (a), w2 in (b), and w3 in (c), respectively. It shows that the larger the value of the weight coefficient, the greater the contribution to the integrated noise-suppressed speech. As shown in the graph, the contribution of the suppression processor 14-2 (noise suppression method 2) is large in the speech section, and the contribution of the suppression processor 14-3 (noise suppression method 3) is large in the non-speech section (noise section). . The increase in the contribution of the different noise suppression methods depending on whether the speech section is a non-speech section is a result that matches the speech spectrum shown in FIG.

  FIG. 6 is a graph showing the time change of the cross-correlation coefficient calculated by the cross-correlation coefficient calculation unit 201 during the noise suppression processing by the noise suppression apparatus of the present embodiment. In each of the drawings (a) to (d), the horizontal axis represents time, the vertical axis of (a) to (c) is the cross-correlation value, and the vertical axis of (d) is the cross-correlation value according to equation (13). Represents the value of a constant obtained by averaging. The unit of the horizontal axis is second. The graph described as “M1: M2” in (a) represents the time change of the cross-correlation coefficients xcor1, 2 between the noise suppression methods 1 and 2. The graph described as “M2: M3” in (b) represents the time change of the cross-correlation coefficient xcor2, 3 between the noise suppression methods 2 and 3. The graph described as “M3: M1” in (c) represents the time change of the cross-correlation coefficient xcor3,1 between the noise suppression methods 3 and 1. In addition, the graph described as “c” in (d) shows a temporal change in the degree of suppressing the non-voice section (noise section). All the cross-correlation coefficients show a high value in the speech section and a low value in the non-speech section (noise section). Then, the effect of noise suppression can be enhanced by the degree c of suppressing the non-voice section (noise section).

  Next, the objective evaluation value of the processing result by the noise suppression apparatus of this embodiment will be described. There are various methods for objectively evaluating the noise suppression method, but a method with little deviation from the subjective evaluation result is preferable. Here, a frequency-weighted segmental SNR (hereinafter referred to as “fwSNRseg”) is used as an objective evaluation value. fwSNRseg can be calculated by the following equation (14).

In Expression (14), Bj is a weight for the jth (j = 1, 2,..., K) frequency band. K is the number of frequency bands, for example, K = 25. M is the total number of frames of the signal. | X (m, j) | is the amplitude of the filter bank of the jth frequency band of the mth frame of the clean speech. | [X ^] (m, j) | is the amplitude of the filter bank in the jth frequency band of the mth frame of the noise-suppressed signal. In this fwSNRseg, since the segmental SNR is weighted for each auditory frequency band, the correlation with the result of the subjective auditory test is high. The higher the evaluation value of fwSNRseg, the higher the evaluation. The objective evaluation value fwSNRseg is also described in the following reference.
References: Tribolet, J., Noll, P., McDermott, B., and Crochiere, RE “A study of complexity and quality of speech waveform coders.” Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing, 586-590, 1978.

  The evaluation results of the present embodiment using the above fwSNRseg are as shown in Table 1 below.

  Table 1 shows objective evaluation values (fwSNRseg) between clean speech, each noise suppression method, and the proposed method in the present embodiment. Also from the result of this objective evaluation value, it can be seen that the noise suppression result according to the present embodiment is of higher quality than the case where each of the noise suppression methods 1 to 3 is used alone.

  As described above, when mixing noise-suppressed speech obtained by multiple noise suppression methods with different properties in the time domain, the noise from each noise suppression method is calculated using the weighting coefficient calculated using the correlation coefficient. By weighting the suppressed speech, the effects of energy reduction of the noise component and energy amplification of the speech portion can be accurately obtained, and high-quality noise-suppressed speech can be easily obtained.

[Second Embodiment]
Next, a second embodiment will be described. In addition, description is abbreviate | omitted about the matter similar to the above-mentioned embodiment, and it demonstrates centering on the matter peculiar to this embodiment.
FIG. 7 is a block diagram illustrating a functional configuration of the noise suppression device according to the second embodiment. Note that functional blocks that perform the same processing as in the previous embodiment are assigned the same reference numerals as those in the previous embodiment. As shown in the figure, the noise suppression apparatus 2 includes a voice input unit 11, a waveform cutout unit 12, a frequency characteristic calculation unit 22, a phase characteristic calculation unit 24, and I (I is an integer of 2 or more) suppression. Processing units 14-1 to 14-I, I frequency characteristic calculation units 25-1 to 25-I, I amplitude characteristic calculation units 26-1 to 26-I, and a noise suppression amplitude characteristic matrix storage unit 35, a weight calculation unit 36, a voice integration unit 37, a frequency characteristic calculation unit 38, a voice waveform calculation unit 39, a waveform superposition unit 18, and a voice output unit 19.

The voice input unit 11 and the waveform cutout unit 12 have the same functions as those in the first embodiment.
The frequency characteristic calculation unit 22 calculates the frequency characteristic data by Fourier transform based on the voice data (noise-mixed voice) cut out by the waveform cutout unit 12.
The phase characteristic calculation unit 24 calculates phase characteristic data based on the frequency characteristic data obtained by the frequency characteristic calculation unit 22.

The suppression processing units 14-1 to 14-I have the same functions as the suppression processing unit in the first embodiment. The suppression processing units 14-1 to 14-I use noise suppression methods having different properties.
The frequency characteristic calculation units 25-1 to 25-I calculate the frequency characteristic data by Fourier transform based on the noise-suppressed speech data calculated by the suppression processing units 14-1 to 14-I, respectively.
The amplitude characteristic calculation units 26-1 to 26-I calculate amplitude characteristic data based on the frequency characteristic data obtained by the frequency characteristic calculation units 25-1 to 25-I, respectively.
The noise suppression amplitude characteristic matrix storage unit 35 stores the amplitude characteristic data of the noise suppression speech obtained by the amplitude characteristic calculation units 26-1 to 26-I. Specifically, the noise suppression amplitude characteristic matrix storage unit 35 stores data in the form of a noise suppression amplitude characteristic matrix configured by arranging the amplitude characteristic vectors respectively generated by the amplitude characteristic calculation units 26-1 to 26-I. To do.

The weight calculation unit 36 reads the amplitude characteristic data of the noise-suppressed speech stored in the noise suppression amplitude characteristic matrix storage unit 35, and calculates a weight coefficient for each noise suppression method based on this data. The weight calculation unit 16 calculates a weight coefficient so that the noise suppression result after mixing is optimized.
The voice integration unit 37 mixes the amplitude characteristic data output from the amplitude characteristic calculation units 26-1 to 26-I. At this time, the voice integration unit 37 performs weighting of the mixing ratio for each noise suppression method using a weighting coefficient corresponding to each noise suppression method. The voice integration unit 37 mixes the noise-suppressed voice data for each analysis frame. The weight coefficient is calculated by the weight calculation unit 16. In the present embodiment, the data mixed by the sound integration unit 37 is data of a frequency domain sound signal.

The frequency characteristic calculation unit 38 is based on the amplitude characteristic data after mixing (data mixed with optimum weighting) output from the voice integration unit 37 and the phase characteristic data of the input voice calculated by the phase characteristic calculation unit 24. The frequency characteristic data of the sound after mixing is calculated.
The voice waveform calculation unit 39 obtains time waveform data of noise-suppressed voice by inverse Fourier transform based on the frequency characteristic data obtained by the frequency characteristic calculation unit. This time waveform data is data for each analysis frame.
The waveform superimposing unit 18 and the audio output unit 19 have the same functions as those in the first embodiment.

Details of the processing procedure performed by the noise suppression device 2 will be described below. Also in this embodiment, the sampling frequency is 16 kHz and the number of quantization bits is 16 bits. The length N of the noise-containing speech vector is 256 (about 16 milliseconds).
The voice input unit 11 acquires voice from outside. The waveform cutout unit 12 cuts out a speech waveform for each analysis frame.

  The suppression processing units 14-1 to 14 -I perform processing to suppress noise using the noise suppression methods Fi having different properties with respect to the noise-mixed speech vector y extracted by the waveform cutting unit 12 for each frame. By this processing, the suppression processing units 14-1 to 14-I output noise-suppressed speech vectors [x ^] i (i = 1, 2,..., I). The length of each noise-suppressed speech vector is also N (256).

  The frequency characteristic calculators 25-1 to 25-I obtain frequency characteristic vectors [X ^] i for the noise-suppressed speech vectors [x ^] i supplied from the suppression processors 14-1 to 14-I, respectively. . Note that the frequency characteristic calculators 25-1 to 25-I generate an appropriate window function (for example, a Hamming window whham (n) = 0.54−0.46 cos (2πn / N) based on each noise-suppressed speech vector. A frequency characteristic vector [X ^] i is calculated by performing discrete Fourier transform (FFT) on the signal cut out by multiplying (n = 1,..., N)). The number of FFT points is N.

  The amplitude characteristic calculators 26-1 to 26 -I take the absolute values of the frequency characteristic vectors [X ^] i supplied from the frequency characteristic calculators 25-1 to 25 -I, respectively. An amplitude characteristic vector | [X ^] i | The amplitude characteristic calculation units 26-1 to 26-I write the calculated amplitude characteristic vectors | [X ^] i | into the noise suppression amplitude characteristic matrix storage unit 35, respectively.

  The noise suppression amplitude characteristic matrix storage unit 35 stores a noise suppression amplitude characteristic matrix [X ^] having an amplitude characteristic vector | [X ^] i | of noise-suppressed speech by each noise suppression method as a row vector. That is, a series of processing steps for calculating the noise suppression amplitude characteristic matrix [X ^] is expressed by the following equation (15).

In Expression (15), white bold “C” represents a set of complex numbers. Re () represents taking the real part of the complex number, and Im () represents taking the imaginary part of the complex number. Also, whmm is the window function described above.
The weight calculation unit 36 calculates a weight coefficient described below.
The voice integration unit 37 obtains an integrated noise suppression amplitude characteristic vector | [X˜] | using the weight calculated by the weight calculation unit 36. Specifically, it is as follows. That is, a diagonal component is obtained by multiplying the noise suppression amplitude characteristic matrix [X ^] by the amplitude characteristic vector of each noise-suppressed speech by a matrix W having the weight coefficient vector wi corresponding to each noise suppression method Fi as a column vector. As a result, the integrated noise suppression amplitude characteristic vector | [X˜] | is obtained as in the following Expression (16).

  The obtained integrated noise suppression amplitude characteristic vector | [X˜] | is mixed by multiplying each bin | [X ^] i [l] | by the weight coefficient wi [l] of the amplitude characteristic vector of each noise-suppressed speech. It corresponds to a thing. In equation (16), diag [] represents taking a column vector obtained by extracting the diagonal components of the matrix. The weight coefficient vector wi is an N-dimensional row vector. The matrix W has I rows and N columns.

As in the case of the first embodiment, the weight calculation unit 36 obtains an optimum weight coefficient matrix Wopt by an optimization method such as using a cross-correlation coefficient. The procedure for calculating the weight will be described below.
First, the weight calculation unit 36 obtains the cross-correlation coefficient between the amplitude characteristic vectors | [X ^] i | of the noise-suppressed speech by the respective noise suppression methods as in the following equation (17). That is, the cross-correlation coefficient xcori, j is obtained by dividing the covariance of the amplitude characteristic vectors | [X ^] i | and | [X ^] j |

  Next, the weight calculation unit 36 uses the obtained cross-correlation coefficient xcori, j, and the weight coefficient vector for the amplitude characteristic vector | [X ^] j | [W ^] i is calculated. Here, n is an index for setting the degree of the weight coefficient, and for example, n = 2.

  The calculation shown in Expression (18) is to add the correlation coefficient related to the amplitude characteristic vector | [X ^] j | of each noise-suppressed speech, and is common to each bin. Then, the weight calculation unit 36 obtains a weighting coefficient matrix [W ^] from the weighting coefficient vector [w ^] j.

  The weight calculation unit 36 performs normalization according to the following equation (20) so that the weight coefficient matrix [W ^] satisfies the equation (16). The right side of equation (20) is

  The weighting coefficient matrix [W ^] obtained in this way is set as an optimum weighting coefficient matrix Wopt. That is, Wopt = [W ^].

  The speech integration unit 37 applies the optimum weighting coefficient matrix Wopt to the equation (16), that is, W = Wopt, as shown in the following equation (21), the optimum integrated noise suppression amplitude characteristic vector | [ X˜] opt |.

On the other hand, the frequency characteristic calculation unit 22 obtains the frequency characteristic vector Y from the noise-containing speech vector y cut out for each frame.
Then, the phase characteristic calculation unit 24 calculates the phase characteristic vector ∠Y based on the frequency characteristic vector Y.
The calculation of the phase characteristic vector ∠Y is performed by the following equation (22).

  Then, the frequency characteristic calculation unit 38 uses the optimum integrated noise suppression amplitude characteristic vector | [X˜] opt | and the phase characteristic vector ∠Y of the noise-mixed speech to express the frequency characteristic as shown in the following equation (23). The vector [X˜] opt is calculated.

  Then, the speech waveform calculation unit 39 takes the inverse Fourier transform (IFFT) of the frequency characteristic [X˜] opt obtained above, and the speech time waveform [x˜] opt (m, n) (n = 1,..., N) is obtained. Here, m represents the frame, and n represents the index of the sample in the frame.

Then, the waveform superimposing unit 18 divides the time waveform [x˜] opt (m, n) by the hamming window whmm (n) and divides by the hamming window whmm (n) to obtain an appropriate window function (for example, the Hanning window whann (n ) = 0.5−0.5 cos (2πn / T), and the data multiplied by the window function is shifted by the soft width for each frame and superimposed, thereby noise-reduced speech [x˜] (n) Get.
The voice output unit 19 outputs the noise-suppressed voice calculated by the waveform superposition unit 18 to the outside.

8 and 9 are graphs showing examples of results of noise suppression by the noise suppression device of the present embodiment. (A) to (g) of FIG. 8 each show a speech waveform, with the horizontal axis representing time and the vertical axis representing amplitude. Moreover, (a)-(g) of FIG. 9 shows an audio | voice spectrum, respectively, a horizontal axis is time and a vertical axis | shaft is a frequency. The unit of the horizontal axis in FIGS. 8 and 9 is second. The unit of the vertical axis in FIG. 9 is hertz. FIG. 9 shows the time transition of the strength of the component for each frequency as the gray scale, as in FIG. 8 and FIG. 9 are similar to FIGS. 3 and 4, respectively, (a) clean speech, (b) additional noise, (c) noise-mixed speech to be input to the noise suppression device, (d) An example of speech with noise suppressed according to the embodiment, (e) noise-suppressed speech by noise suppression method 1 (f) is noise-suppressed speech by noise suppression method 2, and (g) noise-suppressed speech by noise suppression method 3.
Compared with the noise suppression methods 1 to 3, it can be seen that the method of the present embodiment also suppresses the degradation of the speech segment and effectively suppresses the noise in the non-speech segment (noise segment).

  Next, the objective evaluation value of the processing result by the noise suppression apparatus of this embodiment will be described. As the objective evaluation value, fwSNRseg is used as in the first embodiment. The evaluation results of this embodiment using fwSNRseg are as shown in Table 2 below.

As shown in the table, the method according to the second embodiment gives better results than the original noise suppression methods 1 to 3. Further, the method according to the second embodiment has obtained better results than the method according to the first embodiment.
In the second embodiment, the time domain audio signals output from the suppression processing units 14-1 to 14-I are converted into frequency domain signals, and weights are obtained by obtaining cross-correlation values between the frequency domain signals. Coefficients were calculated and frequency domain signals were mixed based on the weighting coefficients. As described above, when mixing noise-suppressed speech obtained by multiple noise suppression methods with different properties in the frequency domain, the noise from each noise suppression method is calculated using the weighting coefficient calculated using the optimization method. By weighting the suppressed speech, the effects of energy reduction of the noise component and energy amplification of the speech portion can be accurately obtained, and high-quality noise-suppressed speech can be precisely obtained.

[Third Embodiment]
Next, a third embodiment will be described. In addition, description is abbreviate | omitted about the matter similar to the above-mentioned embodiment, and it demonstrates centering on the matter peculiar to this embodiment. The noise suppression device according to the present embodiment has a configuration similar to that of the noise suppression device according to the first embodiment, and is different in the calculation method of the weight coefficient. That is, the noise suppression apparatus according to the present embodiment has a configuration in which the weight calculation unit 16 in the functional block diagram of FIG. 1 is replaced with a weight calculation unit 56 described below.

  Prior to the calculation of the weight coefficient by the weight calculation unit 56, the waveform cutout unit 12 cuts out the input noise-containing speech for each appropriate analysis frame, as in the processing in the first embodiment. In addition, for the data cut out for each frame, each of the suppression processing units 14-1 to 14-I obtains each noise-suppressed speech by I noise suppression methods having different properties.

  FIG. 10 is a block diagram illustrating a functional configuration of the weight calculation unit according to the present embodiment. As illustrated, the weight calculation unit 56 includes an adaptive filter coefficient calculation unit 221, an adaptive filter coefficient addition unit 222, and a weight coefficient normalization unit 223. In order to obtain the optimum weight coefficient vector wopt corresponding to each noise suppression method, the weight calculation unit 56 uses an adaptive filter. There are various adaptive filter methods. In this embodiment, a normalized LMS algorithm is used as an example. The normalized LMS algorithm normalizes the coefficient correction term of the LMS algorithm with the filter state vector norm.

The adaptive filter coefficient calculation unit 221 uses, as a desired signal, an adaptive filter coefficient that uses a noise-suppressed speech vector obtained by another noise suppression method as a desired signal with respect to a noise-suppressed speech vector obtained by a certain noise suppression method based on the noise-suppressed speech vector obtained by each noise suppression method Ask for.
The adaptive filter coefficient adding unit 222 adds, for all noise suppression methods, an adaptive filter coefficient that uses another noise suppression method as a desired signal with respect to the noise suppression method for all the other noise suppression methods. . This value is the basis of the weight coefficient value for the noise suppression method.
The weighting factor normalization unit 223 normalizes the weighting factor value for each noise suppression method calculated by the adaptive filter factor addition unit 222. Specifically, the weight coefficient normalization unit 223 performs adjustment so that the sum of the weight coefficients for all noise suppression methods is, for example, 1.

Hereinafter, a procedure for calculating a weighting coefficient using an adaptive filter will be described.
First, the adaptive filter coefficient calculation unit 221 first obtains the noise-suppressed speech vector [x ^] obtained from each noise suppression method by the suppression processing units 14-1 to 14-I from the noise-suppressed speech matrix storage unit 15 (FIG. 1). The data of i (i = 1,..., I) is read and the adaptive filter coefficient is obtained. Specifically, the adaptive filter coefficient calculation unit 221 uses the noise suppression speech vector [x ^] j (i ≠ j) by another noise suppression method as the desired signal for the noise suppression speech vector [x ^] i. The coefficient hk + 1 (i, j) is obtained. The adaptive filter coefficient hk + 1 (i, j) is calculated by the following equation (24).

  The adaptive filter coefficient calculation unit 221 obtains the converged adaptive filter coefficient by the recurrence formula of Expression (24). Here, α is a step size parameter that determines the degree of convergence of the adaptive filter, and β is a stabilization parameter that prevents division by zero. The adaptive filter coefficient is updated for each sample. As a premise, noise suppression methods having different properties are used in the suppression processing units 14-i and 14-j. The adaptive filter coefficient hk + 1 (i, j) corresponds to the degree of cross-correlation between the noise-suppressed speech. Therefore, it is expected that the adaptive filter coefficient hk + 1 (i, j) is high in the voice section and the adaptive filter coefficient hk + 1 (i, j) is low in the non-voice section (noise section).

  Next, the adaptive filter coefficient adding unit 222 uses each of the adaptive filter coefficients hk + 1 (i, j) obtained in Expression (24) and performs each noise-suppressed speech vector [x ^] i according to Expression (25) below. The weight coefficient [w ^] i for is calculated.

  As shown in Expression (25), the obtained weighting coefficient [w ^] i is obtained by adding the adaptive filter coefficients related to each noise-suppressed speech vector [x ^] i for all j (where i ≠ j). It is.

  Then, the weight coefficient normalization unit 223 performs normalization according to Expression (26) so that the weight coefficient vector [w ^] satisfies Expression (4).

The weighting coefficient vector [w ^] obtained in this way is set as the optimum weighting coefficient vector wopt. That is, wopt = [w ^].
The processing after the optimum weight coefficient vector wopt is calculated by the weight calculation unit 56 is the same as that in the first embodiment. That is, the speech integration unit 17 (FIG. 1) applies the optimum weighting coefficient vector wopt to the equation (5) (w = wopt), and according to the following equation (27), the optimum integrated noise suppression speech vector [x ~] Get opt.

In equation (27), m represents the index of the frame, and n represents the index of the sample in the frame.
Then, the waveform superimposing unit 18 superimposes the time waveform [x˜] opt (m, n) by shifting the shift width for each frame. Thereby, noise-suppressed speech [x˜] (n) is obtained.

  Next, the objective evaluation value of the processing result by the noise suppression apparatus of this embodiment will be described. As the objective evaluation value, fwSNRseg is used as in the first and second embodiments. The evaluation results of this embodiment using fwSNRseg are as shown in Table 3 below.

As shown in the table, the method according to the third embodiment gives better results than the original noise suppression methods 1 to 3. In addition, the method according to the third embodiment gives better results than the methods according to the first and second embodiments (results shown in Tables 1 and 2 respectively).
As described above, when mixing noise-suppressed speech obtained by multiple noise suppression methods with different properties in the time domain, the noise suppression from each noise suppression method is performed using the weighting coefficient calculated using an adaptive filter. By weighting the voice, the effects of energy reduction of the noise component and the energy amplification of the voice part can be accurately obtained, and high-quality noise-suppressed voice can be easily obtained.

[Fourth Embodiment]
Next, a fourth embodiment will be described. In addition, description is abbreviate | omitted about the matter similar to the above-mentioned embodiment, and it demonstrates centering on the matter peculiar to this embodiment.

In the first embodiment described above, the time-domain audio signals output from the suppression processing units 14-1 to 14-I are mixed based on the weighting factor. At this time, the cross-correlation value of the data (noise-suppressed speech vector) output from the suppression processing units 14-1 to 14-I was taken, and the weighting coefficient was obtained based on the cross-correlation value.
In the second embodiment described above, the time-domain audio signals output from the suppression processing units 14-1 to 14-I are converted into frequency-domain signals, and cross-correlation values between the frequency-domain signals are obtained. The weighting factor was calculated by the above, and signals in the frequency domain were mixed based on this weighting factor.
In the above-described third embodiment, the time-domain audio signal is mixed based on the weighting factor. However, in the embodiment, the weighting coefficient is calculated by taking an adaptive filter value between the data (noise-suppressed speech vectors) output from the suppression processing units 14-1 to 14-I.
The fourth embodiment has a configuration having the characteristics of the second embodiment and the third embodiment. That is, an adaptive filter value is calculated between signals in the frequency domain, and a weighting coefficient is calculated based on the adaptive filter value. Then, the frequency domain signals are mixed based on the calculated weighting factor.

  That is, the noise suppression device according to the present embodiment has a configuration similar to that of the functional block diagram shown in FIG. 7, and only the weight coefficient calculation method by the weight calculation unit is different. The weight calculation unit according to the present embodiment uses another amplitude characteristic vector as a desired signal for each amplitude characteristic vector based on the amplitude characteristic vector of the noise-suppressed speech output from the amplitude characteristic calculation units 26-1 to 26-I. Calculate the adaptive filter value. Then, the weight calculation unit calculates the sum of adaptive filter values using other amplitude characteristic vectors (other noise suppression methods) as desired signals for each amplitude characteristic vector, and further makes the total sum of the weighting coefficients be 1. Normalize. The voice integration unit mixes the amplitude characteristic vectors while performing weighting using the obtained weighting coefficient. Then, the mixed frequency-domain noise-suppressed speech signal is converted back to a time-domain signal, the waveforms are superimposed for each time window, and the obtained noise-suppressed speech is output.

  Note that the function of the noise suppression device in each of the above-described embodiments may be realized by a computer. In that case, a program for realizing this function may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be read into a computer system and executed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer-readable recording medium” refers to a storage device such as a flexible disk, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, a “computer-readable recording medium” dynamically holds a program for a short time, like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory inside a computer system serving as a server or a client in that case may be included, and a program that holds a program for a certain period of time. The program may be a program for realizing a part of the above-described functions, or may be a program that can realize the above-described functions in combination with a program already recorded in a computer system.

[Modification]
You may make it implement this invention with the following modifications.
In the first to fourth embodiments, a weighting coefficient is set by calculating a cross-correlation value or an adaptive filter coefficient value and extracting a waveform having a high correlation between signals obtained by two different noise suppression methods. Instead, in the modification, an evaluation function using a weighting factor as a parameter is appropriately set, and an evaluation function value is calculated for a noise suppression result (a result of mixing signals from a plurality of different noise suppression methods). Then, a parameter for which the evaluation function value is optimal is obtained. The parameter to be obtained is a multidimensional vector (I × 1 dimension or I × N dimension), and the parameter is optimized using, for example, the steepest descent method.

  As described above, a plurality of embodiments of the present invention have been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes a design and the like within a scope not departing from the gist of the present invention. .

  The present invention can be used for audio processing in general. As an example, the present invention can be used for devices that record audio for broadcast programs and the like.

1, 2 Noise suppressor
DESCRIPTION OF SYMBOLS 11 Voice input part 12 Waveform cut-out part 14-1-14-I Suppression process part 15 Noise suppression voice matrix memory | storage part 16,36,56 Weight calculation part 17,37 Voice integration part 18 Waveform superposition part 19 Voice output part 22 Frequency Characteristic calculation unit 24 Phase characteristic calculation unit 25-1 to 25-I Frequency characteristic calculation unit 26-1 to 26-I Amplitude characteristic calculation unit 35 Noise suppression amplitude characteristic matrix storage unit 38 Frequency characteristic calculation unit 39 Speech waveform calculation unit 201 Correlation coefficient calculation unit 202 Cross correlation coefficient addition unit 203 Weight coefficient normalization unit 221 Adaptive filter coefficient calculation unit 222 Adaptive filter coefficient addition unit 223 Weight coefficient normalization unit

Claims (5)

A plurality of suppression processing units that output noise-suppressed voice data by performing processing based on different noise suppression methods for input voice data, and
A weight calculating unit that calculates a weighting coefficient for each noise suppression method based on the noise-suppressed speech data output from the plurality of suppression processing units;
Voice integration for mixing the noise-suppressed voice data by multiplying each noise-suppressed voice data output from the plurality of suppression processing sections by a weighting coefficient for each noise suppression method calculated by the weight calculation section And
A noise suppression device comprising: The weight calculation unit calculates a correlation coefficient between the noise-suppressed speech data output from the plurality of suppression processing units, and the noise suppression method having a higher correlation with other noise suppression methods, Calculating so that the value of the weighting factor is large,
The noise suppression apparatus according to claim 1. The weight calculation unit is configured to request the noise-suppressed speech data by another noise suppression method for the noise-suppressed speech data by each noise suppression method based on the noise-suppressed speech data output from the plurality of suppression processing units. Calculating an adaptive filter coefficient as a signal, and calculating the value of the weighting coefficient to be larger as the calculated value of the adaptive filter coefficient is larger;
The noise suppression apparatus according to claim 1. A frequency characteristic calculation unit that calculates frequency characteristic data based on the noise-suppressed speech data output from each of the plurality of suppression processing units;
An amplitude characteristic calculator that calculates amplitude characteristic data based on the frequency characteristic data;
Further comprising
The weight calculation unit calculates a weight coefficient for each noise suppression method based on the amplitude characteristic data,
The speech integration unit mixes the noise-suppressed speech data by multiplying the amplitude characteristic data by the weighting factor and mixing.
The noise suppression device according to any one of claims 1 to 3, wherein On the computer,
A plurality of suppression processing steps for outputting noise-suppressed speech data by performing processing by different noise suppression methods on the waveform data of the input speech,
A weight calculation process for calculating a weighting coefficient for each noise suppression method based on the noise-suppressed voice data output in the plurality of suppression processes;
Speech integration for mixing the noise-suppressed speech data by multiplying each noise-suppressed speech data output from the plurality of suppression processing steps by a weighting coefficient for each noise suppression method calculated by the weight calculating step process,
Program to execute the process.