JP2010066478A - Noise suppressing device and noise suppressing method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To maintain high sound quality, while effectively suppressing noise. <P>SOLUTION: A noise suppressing device 10 includes: a Fourier transformation section 2 and 6 for performing Fourier transformation on a mixed measurement signal including a voice signal and a noise signal, and a noise estimation signal which is an estimated noise signal; a mask function calculation section 8 for calculating a mask function H(f, t) which is a subtraction coefficient on the basis of the mixed measurement signal and the noise estimation signal, on which Fourier transformation is performed by the Fourier transformation section 2 and 6; and a subtraction processing section 9 for calculating a voice estimation signal which is estimated as the voice signal by subtracting the noise estimation signal from the mixed measurement signal by using the mask function H(f, t) calculated by the mask function calculation section 8. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a noise suppression device and a noise suppression method, and more particularly to a noise suppression device and a noise suppression method that can realize high sound quality while suppressing noise.

2. Description of the Related Art Conventionally, a speech recognition apparatus using a spectrum subtraction method that extracts only a speech signal from a mixed observation signal in which another speech signal, an environmental noise signal, or the like is mixed with a user speech signal that is a target signal is known ( For example, see Patent Document 1). This spectrum subtraction method is a method of estimating a target audio signal by subtracting a power spectrum of a noise signal separately estimated from a power spectrum of a mixed observation signal mixed with noise. When performing this subtraction process, a coefficient called a subtraction coefficient is multiplied by the power spectrum of the noise signal, and the power spectrum is corrected.
JP 2007-248534 A

By the way, when the audio signal s (t), the noise signal n (t), and the mixed observation signal x (t) are used, the following equation (4) is generally established.
x (t) = s (t) + n (t) (4) Formula

At this time, if the noise estimation signal estimated separately is n1 (t), the speech estimation signal obtained by the spectrum subtraction method can be expressed by the following equation (5).

  In the above equation (5), X (f, t) and N1 (f, t) are signals obtained by performing short-time Fourier transform on x (t) and n1 (t), respectively. Β is a subtraction coefficient, and angle (Y) is a function that outputs the phase angle of the complex number Y.

  Therefore, according to the above equation (5), it can be seen that the degree of residual noise in the speech estimation signal s1 (t) that is the output signal depends on the value of the subtraction coefficient β. For example, if the subtraction coefficient β is increased to improve the suppression performance, the sound quality is deteriorated. On the other hand, if the subtraction coefficient β is decreased to suppress the suppression performance, the sound quality is improved. In this way, since there is a tradeoff between suppression performance and sound quality, it is difficult to set an optimal subtraction coefficient β that satisfies both suppression performance and sound quality at the same time.

  The present invention has been made to solve such problems, and mainly provides a noise suppression device and a noise suppression method capable of maintaining high sound quality while effectively suppressing noise. Objective.

  One aspect of the present invention for achieving the above object includes a Fourier transform unit that performs Fourier transform on a mixed observation signal including a speech signal and a noise signal, and a noise estimation signal that is an estimated noise signal, Based on the mixed observation signal Fourier-transformed by the Fourier transform unit and the noise estimation signal, a mask function computing unit that computes a mask function to be a subtraction coefficient, and the mask function computed by the mask function computing unit And a subtraction processing unit that calculates a speech estimation signal that is estimated to be a speech signal by subtracting the noise estimation signal from the mixed observation signal. According to this aspect, it is possible to maintain high sound quality while effectively suppressing noise.

  In this aspect, the mask function calculation unit may calculate the mask function including two certainty factors represented by a ratio between the Fourier-transformed noise estimation signal and the mixed observation signal.

Furthermore, in this one aspect, the mask function calculation unit may calculate the certainty factors P _f (t) and P _b (f) using the equations (1) and (2), respectively.

  In this aspect, the mask function calculation unit may calculate the mask function H (f, t) using equation (3).

  In this aspect, the image processing apparatus may further include a Fourier inverse transform unit that performs an inverse Fourier transform on the speech estimation signal calculated by the subtraction processing unit.

  On the other hand, one aspect of the present invention for achieving the above object is a Fourier transform step of performing a Fourier transform on a mixed observation signal including a speech signal and a noise signal and a noise estimation signal estimated to be a noise signal. And a mask function calculation step for calculating a mask function to be a subtraction coefficient based on the mixed observation signal and the noise estimation signal Fourier-transformed in the Fourier transform step, and a mask function calculation step calculated by the mask function calculation step. A subtraction processing step of calculating a speech estimation signal estimated as a speech signal by subtracting the noise estimation signal from the mixed observation signal using the mask function. It may be.

  According to the present invention, it is possible to maintain high sound quality while effectively suppressing noise.

  The best mode for carrying out the present invention will be described below with reference to the accompanying drawings. FIG. 1 is a block diagram showing a system configuration of a noise suppression apparatus according to an embodiment of the present invention.

  The noise suppression apparatus 10 according to the present embodiment includes a mixed observation signal input unit 1, a first discrete Fourier transform unit 2, a phase calculation unit 3, a first spectrum calculation unit 4, a noise estimation signal input unit 5, A second discrete Fourier transform unit 6, a second spectrum computation unit 7, a mask function computation unit 8, a subtraction processing unit 9, a discrete Fourier inverse transform unit 11, and a speech estimation signal output unit 12 are provided. .

  The noise suppression apparatus 10 includes, as main hardware configurations, a CPU (Central Processing Unit) that performs control processing, arithmetic processing, and the like, and a ROM (Read Only) that stores control programs, arithmetic programs, and the like executed by the CPU. Memory) and RAM (Random Access Memory) that temporarily stores processing data and the like. Also, the first discrete Fourier transform unit 2, the phase calculation unit 3, the first spectrum calculation unit 4, the second discrete Fourier transform unit 6, the second spectrum calculation unit 7, the mask function calculation unit 8, the subtraction processing unit 9, and the discrete For example, the inverse Fourier transform unit 11 may be realized by a program stored in the ROM and executed by the CPU.

  A mixed observation signal x (t) including an audio signal and a noise signal is input to the input terminal of the mixed observation signal input unit 1. Here, the voice signal is a target signal such as a user voice, and the noise signal is so-called noise such as ambient voice and environmental noise. Further, the mixed observation signal x (t) may be subjected to linear filtering processing to enhance the voice. The mixed observation signal input unit 1 outputs the input mixed observation signal x (t) to the first discrete Fourier transform unit 2.

  The first discrete Fourier transform unit 2 performs a well-known Fourier transform on the input mixed observation signal x (t) and calculates a Fourier-transformed mixed observation signal X (f, t). The first discrete Fourier transform unit 2 then outputs the Fourier-transformed mixed observation signal X (f, t) to the phase computation unit 3, the first spectrum computation unit 4, and the mask function computation unit 8.

The phase calculation unit 3 uses the following equation (6) based on the Fourier-transformed mixed observation signal X (f, t) from the first discrete Fourier transform unit 2 to calculate the phase angle (X (f, t) ) Is calculated.
angle (X (f, t)) = arctan (A / B) (6)

  Here, X (f, t) = A + Bi (i is a complex number, and A and B are arbitrary real numbers). The phase calculation unit 3 outputs the calculated phase angle (X (f, t)) to the subtraction processing unit 9.

The first spectrum calculation unit 4 calculates a power spectrum | X (f, t) | ² of the mixed observation signal X (f, t) subjected to Fourier transform from the first discrete Fourier transform unit 2, and a mask function calculation unit 8 is output.

  The noise estimation signal n1 (t), which is an estimated noise signal, is input to the input terminal of the noise estimation signal input unit 5. A known noise estimation algorithm can be used for estimating the noise signal. The noise estimation signal input unit 5 outputs the input noise estimation signal n1 (t) to the second discrete Fourier transform unit 6.

  The second discrete Fourier transform unit 6 performs a well-known Fourier transform on the input noise estimation signal n1 (t), and calculates a Fourier estimated noise estimation signal N1 (f, t). Then, the second discrete Fourier transform unit 6 outputs the noise estimation signal N1 (f, t) obtained by Fourier transform to the mask function computation unit 8 and the second spectrum computation unit 7.

The second spectrum calculation unit 7 calculates a power spectrum | N1 (f, t) | ² of the noise estimation signal N1 (f, t) subjected to Fourier transform from the second discrete Fourier transform unit 6, and a mask function calculation unit 8 is output.

The mask function calculation unit 8 includes the power spectrum | X (f, t) | ² of the mixed observation signal X (f, t) from the first spectrum calculation unit 4 and the noise estimation signal N1 from the second spectrum calculation unit 7. Based on the power spectrum | N1 (f, t) | ^{2 of} (f, t), a soft mask function (mask function) H (f, t) corresponding to the subtraction coefficient β is calculated.

  Here, the subtraction coefficient β is a coefficient that determines the suppression performance when suppressing noise. For example, when the subtraction coefficient β is increased, as shown in Equation (7) of the subtraction processing unit 9 described later, the suppression performance is increased. On the other hand, if the subtraction coefficient β is lowered, the suppression performance is lowered.

First, the mask function calculation unit 8 first calculates the power spectrum | N1 (f, t) | ^{2 of} the noise estimation signal N1 (f, t) and the power spectrum | X (f, t) of the mixed observation signal X (f, t). ) | The ^two certainty factors P _f (t) and P _b (f) represented by the ratio to 2 are calculated using the following equations (1) and (2), respectively.

Next, the mask function calculation unit 8 calculates the soft mask function H (f, t) using the following equation (3) based on the calculated certainty factors P _f (t) and P _b (f). calculate.

In the above formula (3), for example, 1 is set as I. For the minimum subtraction coefficient δ _o and the maximum subtraction coefficient δ _m , optimum values that give the best suppression performance and sound quality described later are experimentally obtained and set.

Thus, the power spectrum of the noise estimation signal N1 (f, t) | N1 (f, t) | ^2, the power spectrum of the mixed observation signals X (f, t) | X (f, t) | 2 and the By using the two certainty factors P _f (t) and P _b (f) expressed by the ratio, the soft mask function H (f, t) that simultaneously satisfies the suppression performance and the sound quality is set optimally and automatically. be able to. The mask function calculation unit 8 outputs the calculated soft mask function H (f, t) to the subtraction processing unit 9.

  The subtraction processing unit 9 subtracts the noise estimation signal N1 (f, t) from the mixed observation signal X (f, t) using the soft mask function H (f, t) calculated by the mask function calculation unit 8. Thus, the estimated speech signal s1 (f, t) estimated as the speech signal is calculated.

More specifically, the subtraction processing unit 9 performs the phase angle (X (f, t)) from the phase calculation unit 3, the mixed observation signal X (f, t) and the noise estimation signal N1 (f , T) and the soft mask function H (f, t) calculated by the mask function calculation unit 8, the speech estimation signal s 1 (f, t) is calculated using the following equation (7). .

  In the above equation (7), γ is set to an arbitrary optimum value that optimizes the suppression performance and sound quality, as will be described later.

  The subtraction processing unit 9 outputs the calculated speech estimation signal s 1 (f, t) to the discrete Fourier inverse transform unit 11.

  The discrete Fourier inverse transform unit 11 performs inverse Fourier transform on the input speech estimation signal s1 (f, t), and calculates a speech estimation signal s1 (t) obtained by inverse Fourier transform. Then, the discrete Fourier inverse transform unit 11 outputs the speech estimated signal s1 (t) subjected to the Fourier inverse transform to the speech estimated signal output unit 12.

  The speech estimation signal output unit 12 outputs the speech estimation signal s1 (t), which is the final output signal output from the discrete Fourier inverse transform unit 11, from the output terminal.

By the way, in the conventional noise suppression apparatus, when there is no correlation between the noise signal and the voice signal and it can be estimated that the noise estimation signal completely matches the noise signal, for example, 1 is set to the subtraction coefficient β. Can be set. In this case, the speech estimation signal s1 (f, t) can be expressed by the following equation (8).

However, it is very difficult to accurately estimate the noise signal. In practice, the noise signal is mixed into the speech estimation signal (output signal) due to the error, or the speech estimation signal is deleted. . Here, it is assumed that the noise estimation signal is expressed by the following equation (9).
N1 (f, t) = 0.7 × N1 (f, t) (9) Formula

In this case, the speech estimation signal s1 (f, t) can be expressed by the following equation (10). Note that the speech estimation signal s1 (f, t) actually contains errors due to estimation processing in addition to amplitude modulation.

Further, depending on the subtraction coefficient β set during system operation, the suppression performance (noise suppression performance) and sound quality for the speech estimation signal s1 (f, t) may vary, for example, as shown in Table 1 below. I understand.

  As shown in Table 1, it can be seen that the optimum value of the subtraction coefficient β depends on the accuracy of the noise estimation signal. Also, since the audio signal is a non-stationary signal and the accuracy of the noise estimation signal changes from moment to moment, the optimum value of the subtraction coefficient β also changes accordingly.

  Therefore, the noise suppression apparatus 10 according to the present embodiment uses the soft mask function H (f, t) to improve the suppression performance and maintain a high sound quality as will be described later. β can be set optimally and automatically.

  FIG. 2 is a flowchart illustrating an example of a processing flow of a noise suppression method performed by the noise suppression apparatus according to the present embodiment.

  As shown in FIG. 2, the first discrete Fourier transform unit 2 performs a Fourier transform on the input mixed observation signal x (t), and calculates a Fourier-transformed mixed observation signal X (f, t). (Fourier transform process) (step S100).

  The second discrete Fourier transform unit 6 performs a well-known Fourier transform on the input noise estimation signal n1 (t), and calculates a Fourier-transformed noise estimation signal N1 (f, t) (Fourier transform). Process) (step S101).

  Next, the phase calculation unit 3 calculates the phase angle (X (f, t)) based on the mixed observation signal X (f, t) subjected to the Fourier transform from the first discrete Fourier transform unit 2 (step). S102).

The first spectrum calculation unit 4 calculates the power spectrum | X (f, t) | ² of the mixed observation signal X (f, t) subjected to the Fourier transform from the first discrete Fourier transform unit 2 (step S103). ).

Further, the second spectrum calculation unit 7 calculates the power spectrum | N1 (f, t) | ² of the noise estimation signal N1 (f, t) subjected to the Fourier transform from the second discrete Fourier transform unit 6 (step S104). ).

Thereafter, the mask function calculation unit 8 calculates the certainty factors P _f (t) and P _b (f) (step S105), and based on the calculated certainty factors P _f (t) and P _b (f). Then, the soft mask function H (f, t) is calculated (mask function calculation step) (step S106).

  The subtraction processing unit 9 calculates the speech estimation signal s1 (f, t) using the soft mask function H (f, t) calculated by the mask function calculation unit 8 (subtraction processing step) (step S107). Then, the discrete Fourier inverse transform unit 11 performs Fourier inverse transform on the speech estimation signal s1 (f, t) (step S108), and the speech estimation signal output unit 12 performs speech inversely transformed speech estimation signal s1 ( t) is output from the output terminal (step S109).

  3A and 3B show an example of a comparison test result between the noise suppression device 10 according to the present embodiment and a conventional noise suppression device. FIG. 3A shows the suppression performance of each noise suppression device, and shows the relationship between the phase angle (horizontal axis) and the SNR (S / N ratio) value (vertical axis). FIG. 3B shows the sound quality of each noise suppression device, and shows the relationship between the phase angle (horizontal axis) and the CD value (vertical axis).

  In FIGS. 3A and 3B, a solid line (1) is actually observed voice data that is not subjected to noise suppression. A solid line (2) is voice data when noise suppression is performed by a conventional noise suppression apparatus using the well-known Beamformer method. A solid line (3) is voice data when noise suppression is performed by a conventional noise suppression device set to a subtraction coefficient β = 2. A solid line (4) is voice data when noise suppression is performed by a conventional noise suppression device set to a subtraction coefficient β = 5. A solid line (5) is voice data when noise suppression is performed by the noise suppression device 10 according to the present embodiment.

  As shown in FIG. 3A, data (solid line (5)) when noise suppression is performed by the noise suppression apparatus 10 according to this embodiment is data (solid line (2) to ( Compared with 4)), the SNR value is high over the entire phase region, and noise suppression is satisfactorily performed. Further, as shown in FIG. 3B, the data (solid line (5)) when noise suppression is performed by the noise suppression apparatus 10 according to the present embodiment has a relatively low CD value over the entire phase region. Is maintained at a high level, and the sound quality is maintained. That is, the noise suppression device 10 according to the present embodiment can maintain high sound quality while exhibiting high suppression performance as compared with the conventional noise suppression device.

As described above, in the noise suppression apparatus 10 according to the present embodiment, the mask function calculation unit 8 includes the power spectrum | N1 (f, t) | ² of the noise estimation signal N1 (f, t) and the mixed observation signal X (f, The certainty factors P _f (t) and P _b (f) expressed by the ratio of t) to the power spectrum | X (f, t) | ² are calculated. Then, the mask function calculation unit 8 calculates the soft mask function H (f, t) based on the calculated certainty factors P _f (t) and P _b (f). Further, the subtraction processing unit 9 calculates the speech estimation signal s1 (f, t) using the calculated soft mask function H (f, t).

  As a result, the soft mask function H (f, t), which is the subtraction coefficient β, can be optimally and automatically set so as to maintain high sound quality while exhibiting high suppression performance. That is, it is possible to maintain high sound quality while effectively suppressing noise.

  Note that the noise suppression device 10 according to the present embodiment may be applied to, for example, a speech recognition system that recognizes speech, and is applied to any system that removes a noise signal from a mixed observation signal including a speech signal and a noise signal. Is possible.

  Moreover, although the best mode for carrying out the present invention has been described using one embodiment, the present invention is not limited to such one embodiment, and within the scope not departing from the gist of the present invention, Various modifications and substitutions can be added to the above-described embodiments.

It is a block diagram which shows the system configuration | structure of the noise suppression apparatus which concerns on one Embodiment of this invention. It is a flowchart which shows an example of the processing flow of the noise suppression method by the noise suppression apparatus which concerns on one Embodiment of this invention. (A) It is a figure which shows the suppression performance by each noise suppression apparatus, and is a figure which shows the relationship between a phase and a SNR value. (B) It is a figure which shows the sound quality by each noise suppression apparatus, and is a figure which shows the relationship between phase angle and CD value.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Mixed observation signal input part 2 1st discrete Fourier transform part 3 Phase calculation part 4 1st spectrum calculation part 5 Noise estimation signal input part 6 2nd discrete Fourier transform part 7 2nd spectrum calculation part 8 Mask function calculation part 9 Subtraction process 10 Noise suppression device 11 Discrete Fourier inverse transform unit 12 Speech estimation signal output unit

Claims (6)

A Fourier transform unit that performs Fourier transform on a mixed observation signal including a speech signal and a noise signal, and a noise estimation signal that is an estimated noise signal;
A mask function calculation unit that calculates a mask function to be a subtraction coefficient based on the mixed observation signal and the noise estimation signal Fourier-transformed by the Fourier transform unit;
A subtraction processing unit that calculates a speech estimation signal to be estimated as a speech signal by subtracting the noise estimation signal from the mixed observation signal using the mask function calculated by the mask function computation unit. A noise suppression device characterized by that. The noise suppression device according to claim 1,
The said mask function calculating part calculates the said mask function containing two reliability represented by ratio of the said noise estimated signal and the mixed observation signal which were Fourier-transformed, The noise suppression apparatus characterized by the above-mentioned. The noise suppression device according to claim 2,
The said mask function calculating part calculates the said reliability _Pf (t) and _Pb (f), respectively using the following (1) Formula and (2) Formula, The noise suppression apparatus characterized by the above-mentioned.

N1 (f, t): Fourier-transformed noise estimation signal X (f, t): Fourier-transformed mixed observation signal The noise suppression device according to claim 3,
The mask function computing unit calculates the mask function H (f, t) by using the following equation (3).

δ _o : Minimum subtraction coefficient δ _m : Maximum subtraction coefficient The noise suppression device according to claim 2,
A noise suppression apparatus, further comprising: a Fourier inverse transform unit that performs an inverse Fourier transform on the speech estimation signal calculated by the subtraction processing unit. A Fourier transform step of performing a Fourier transform on the mixed observation signal including the audio signal and the noise signal, and the noise estimation signal estimated to be a noise signal;
A mask function calculation step of calculating a mask function to be a subtraction coefficient based on the mixed observation signal and the noise estimation signal that are Fourier-transformed in the Fourier transformation step;
A subtraction processing step of calculating a speech estimation signal to be estimated as a speech signal by subtracting the noise estimation signal from the mixed observation signal using the mask function calculated in the mask function computation step. The noise suppression method characterized by the above-mentioned.

Images