JP2014150367A - Echo suppression gain estimation method, echo cancellation device using the same, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an echo suppression gain estimation method that is robust to a variation of an echo path.SOLUTION: The echo suppression gain estimation method includes an echo power correction value calculation step and an echo suppression gain calculation step. The echo power correction value calculation step inputs a pickup signal spectrum Yand an echo power estimate |D^|, and determines an echo power correction value Φfrom a least square solution that minimizes a difference between a power |Y|of the pickup signal spectrum Yand the product of the echo power estimate |D^|multiplied by the echo power correction value Φ. The echo suppression gain calculation step inputs the pickup signal spectrum Y, the echo power correction value Φand the echo power estimate |D^|, and calculates a gain coefficient for echo suppression that decreases as the echo power correction value Φincreases.

Description

  The present invention relates to an echo canceller used in a communication conference system having an acoustic reproduction system, an echo suppression gain estimation method applied to the echo canceller, and a program therefor.

  Echo suppression processing based on short-time spectral amplitude (STSA) estimation estimates a gain coefficient that suppresses echo assuming no correlation between echo and near-end talker speech, and echoes in the amplitude frequency domain Is described in Non-Patent Document 1, for example. On the other hand, Patent Document 1 discloses a technique for improving the gain coefficient estimation method described in Non-Patent Document 1 (hereinafter referred to as “echo suppression gain estimation method”) in order to improve speech quality after echo suppression. Has been proposed. Here, a functional configuration example of the echo cancellation apparatus 10 using the echo suppression gain estimation method disclosed in Patent Document 1 is shown in FIG. 10 and its operation will be briefly described.

  The echo cancellation apparatus 10 includes a reproduction signal frequency analysis unit 81, a sound pickup signal frequency analysis unit 82, an echo power estimation unit 83, a similarity coefficient calculation unit 12, an echo suppression gain calculation unit 14, a multiplication unit 85, And a frequency synthesizer 86. The reproduction signal x (k) is a signal having a discrete value at a sampling frequency of 16 kHz, for example, and is converted into an acoustic signal by the speaker 1. Note that an AD converter that converts the reproduction signal into discrete values and a DA converter that converts the discrete values into continuous values are omitted. Here, k is a sample point number indicating a discrete time of a predetermined interval, and the reproduction signal x (k) and the sound collection signal y (k) are digital signals.

The reproduction signal frequency analysis unit 81 collects 256 discrete values of the reproduction signal to form one frame, performs frequency analysis by 1/2 overlap addition, and converts the frequency range up to 8 kHz in units of frames into 128 reproduction signal spectra X Convert to _{ω, i} . i is a frame number, and ω is a frequency spectrum number (0 to 127) obtained at intervals of 64 Hz in this example.

The collected sound signal y (k) collected by the microphone 2 is superimposed on the near-end speaker signal s (k) by the echo d (k) generated by reproducing the reproduced signal x (k) from the speaker 1. Signal. The collected sound signal y (k) is converted into a collected sound signal spectrum Y _{ω, i} by the collected sound signal frequency analysis unit 82 in the same manner as the reproduction signal x (k). Y _{ω, i} is Y _{ω, i} = D _{ω, i} + S _{ω, i} , D _{ω, i} is an echo spectrum, and S _{ω, i} is a near-end speaker signal spectrum.

The echo power estimator 83 calculates an echo power estimated value | D _{ω, i} ^ | ² shown in Expression (1) with the reproduction signal spectrum X _{ω, i} and the collected sound signal spectrum Y _{ω, i} as inputs. ^ Represents an estimated value, but the notation is correct as shown in the equations and figures.

Here, H _{ω, i} is the spectrum of the echo path that goes from the speaker 1 to the microphone 2, and | H _{ω, i} ^ | ² is the estimated value of the acoustic coupling amount. min {,} is a function for selecting the minimum value. Therefore, | H _{ω, i} ^ | ² is the estimated value of the smaller acoustic coupling amount between adjacent frames.

The echo suppression gain calculation unit 14 receives the sound pickup signal spectrum Y _{ω, i} , the echo power estimated value | D _{ω, i} ^ | ^2, and the similarity coefficient | r _ω | G _{ω, i} is output. The gain coefficient G _{ω, i} takes a real value between 0 and 1, and takes a small value when there are many echo components in the collected sound signal spectrum Y _{ω, i} , and a large value when there are many components other than the echo components. .

The multiplication unit 85 multiplies the collected sound signal spectrum Y _{ω, i} by a gain coefficient G _{ω, i} . Since the gain coefficient G _{ω, i} when there are many echo components becomes a small value, the output signal of the multiplication unit 85 becomes the near-end speaker signal spectrum estimated value S _{ω, i} ^ with the echo component suppressed. The near-end speaker signal spectrum estimated value S _ω ^ corresponding to each frequency component ω is re-synthesized by the frequency synthesizer 86 into a time-domain output signal s ^ (k).

The similarity coefficient calculation unit 12 receives the reproduction signal spectrum X _{ω, i} and the collected sound signal spectrum Y _{ω, i} as input, and reproduces the inner product <X _{ω, i} , Y _{ω, i} > by, for example, Expression (3) norm value ‖X signal spectrum _{omega, i} ‖ a norm value of the collected sound signal spectrum ‖Y _{ω, i} ‖, for example to calculate each by the formula (4) and (5).

Here, * represents a complex conjugate. ε is a forgetting factor satisfying 0 <ε ≦ 1, and determines an exponential decay time constant. For example, ε = 0.016. As ε approaches 1, the value depends on (weights) the current reproduction signal spectrum X _{ω, i} and the collected sound signal spectrum Y _{ω, i} . In addition, you may use Formula (6)-(8) which paid its attention also to the adjacent frequency.

Here, M _{1 to} M ₂ represent a predetermined frequency range. Expressions (6) to (8) are forms in which the correlation in the time axis direction is obtained after obtaining the correlation from the adjacent frequency. The order of correlation may be reversed. Further, instead of Expression (3), Expression (9) obtained from an absolute value that does not consider the phase component may be used.

The similarity coefficient calculation unit 12 calculates the similarity coefficient | r _ω | according to Expression (10) after calculating the respective values according to Expressions (3) to (5).

Here, the basis for deriving the equation (10) for calculating the similarity coefficient | r _ω | will be described.

For example, in the Wiener filter method (hereinafter referred to as “WF method”), the echo coefficient is suppressed by estimating the gain coefficient G _{ω, i} that minimizes the evaluation amount e _ε of Equation (11).

When equation (11) is differentiated by _{Gω, i} , it can be expressed by the following equation.

When the gain coefficient G _{ω, i} is obtained from an equation in which equation (12) is set to 0, equation (13) is obtained.

When the equation (13) is modified, the gain coefficient G _{ω, i} can be expressed by the equation (14).

Here, r _ω indicates the complex coherence of the echo spectrum D _{ω, i} and the near-end speaker signal spectrum S _{ω, i} . If the unknown vector D _{ω, i} is eliminated from r _ω using D _{ω, i} = H _{ω, i} ^* X _{ω, i} , it can be expressed by equation (15).

Here, * represents a complex conjugate. As can be seen from equation (15), even if D _{ω, i} is eliminated, it is difficult to obtain r _ω because there is an unknown variable H _{ω, i} . However, it is noted that H _{ω, i} can be erased as shown in the equation (16) if the absolute value.

The gain coefficient is given as shown in the following equation.

As is clear from the equation (17), the gain coefficient G _{ω, i} becomes smaller as the similarity coefficient | r _ω | approaches 1 and the echo is suppressed. At the same time, the gain coefficient obtained in this way operates to reduce the loss of the near-end speaker signal s (i). That is, the echo is suppressed using the short-time spectrum amplitude without requiring a long time to satisfy <D _{ω, i} , S _{ω, i} > = 0, so that the inner product value is not 0 is an error. Generation of musical noise that occurs can be suppressed.

Japanese Patent No. 4787851

Sakauchi, S., Haneda, Y., Kataoka, A., “Echo suppression processing based on STSA estimation, gain emphasis method” (1), 1vol.J88-A, no.6, Jun.2005, p695-703

  The performance of the echo canceller is related to the estimation accuracy of the echo power. The accuracy of the estimated echo power value depends on the estimated speed of the acoustic coupling amount (power spectrum of the impulse response of the echo path). If the estimated speed is slow, the echo path fluctuates, and the echo cancellation performance drops significantly for a while from the time the acoustic coupling changes, causing problems such as echo remaining unerased or musical noise. . However, in general, in order to avoid the influence of disturbance such as a near-end speaker signal, the amount of acoustic coupling is estimated based on analysis of long-time data of the collected sound signal, so it is difficult to increase the estimated speed. .

  The present invention has been made in view of such problems, and is an echo suppression gain estimation method that is resistant to echo path fluctuations and can suppress the occurrence of musical noise, and an echo cancellation apparatus and program using the same. The purpose is to provide.

The echo suppression gain estimation method of the present invention includes a reproduction signal frequency analysis stage, a sound pickup signal frequency analysis stage, an echo power estimation stage, an echo power correction value calculation stage, and an echo suppression gain calculation stage. The reproduction signal frequency analysis stage converts the reproduction signal into a reproduction signal spectrum _{Xω, i} in the frequency domain. In the collected signal frequency analysis step, the collected signal is converted into a collected signal spectrum Y _{ω, i} in the frequency domain. In the echo power estimation step, the echo power estimated value | D _{ω, i} ^ | ² in the entire frequency region is calculated by using the reproduction signal spectrum X _{ω, i} and the collected sound signal spectrum Y _{ω, i} as inputs. In the echo power correction value calculation stage, the collected sound signal spectrum Y _{ω, i} and the echo power estimated value | D _{ω, i} ^ | ² are input, and the power of the collected sound signal spectrum Y _{ω, i} | Y _{ω, i} | ^2, the echo power estimate | D _{ω, i} ^ | ² to obtain the echo-power correction value [Phi _i from the least-squares solution that minimizes the difference between the value obtained by multiplying the echo-power correction value [Phi _i. Echo suppression gain calculating step, the sound collection signal spectrum Y _{omega, i} and the echo-power correction value [Phi _i and echo power estimate | D _{ω, i} ^ | as inputs ² and, the echo-power the gain coefficient for suppressing an echo Calculation is performed so that the correction value Φ _i becomes smaller as the value increases.

According to the echo suppression gain estimation method of the present invention, the echo power estimated value | D _{ω, i} ^ | ² is dynamically adjusted by the echo power correction value Φ _i . The echo power correction value Φ _i takes a value in the range of 0 to infinity, and the echo power estimated value | D _{ω, i} ^ | ² calculated from the entire frequency region is actually calculated from the entire frequency region. If it is smaller than the total energy of the echo power, it operates to take a large value. As a result, by multiplying the echo power correction value Φ _i by the echo power estimated value | D _{ω, i} ^ | ² , the gain coefficient becomes small and the echo is suppressed.

When the total energy of the echo power estimated value | D _{ω, i} ^ | ² calculated from the entire frequency region is larger than the total energy of the actual echo power calculated from the entire frequency region, the echo power correction value Φ _i Works to take a small value. As a result, by multiplying the echo power correction value Φ _i by the echo power estimated value | D _{ω, i} ^ | ² , the gain coefficient is increased, and the operation is performed to reduce the loss of the near-end speaker signal.

  Therefore, compared with the conventional method, it is possible to perform echo suppression that is robust against fluctuations in the echo path and has little deterioration in sound quality.

The figure which shows the function structural example of the echo cancellation apparatus 100 using the echo suppression gain estimation method of this invention. The figure which shows the operation | movement flow of the echo suppression gain estimation method of this invention. The figure which shows the function structural example of the echo power correction value calculation part 110. The figure which shows the function structural example of the echo suppression gain calculation part 120. FIG. The figure which shows the operation | movement flow of the echo suppression gain estimation method including the operation | movement flow of the echo suppression gain calculation part 120. The figure which shows the function structural example of the echo power correction value calculation part 210. FIG. It is a figure which shows the signal used for evaluation experiment, (a) is a reproduction signal x (i), (b) is a figure which shows an example of a near-end speaker signal s (i). It is a figure which shows the signal used for evaluation experiment, (a) is the sound collection signal y (i), (b) and (c) are the output signals e (i) by a conventional method, (d) is the output by this invention. The figure which shows an example of signal e (i). The figure which shows the example of the evaluation result of the subjective quality evaluation which compared the sound quality of the output signal e (i) with the conventional method. The figure which shows the function structure of the conventional echo cancellation apparatus 10. FIG.

  Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

[Basic idea of the invention]
Prior to the description of the embodiment, the basic idea of the dynamic adjustment method of the echo power estimated value | D _{ω, i} ^ | ² of the present invention will be described. The present invention assumes that the echo power spectrum | D _{ω, i} ^ | ² and the power spectrum of the near-end speaker signal | S _{ω, i} ^ | ² are statistically uncorrelated, and the power spectrum of the collected sound signal | Y _{omega, i} | ^2, the echo power estimate | D _{ω, i} ^ | ² to determine the echo-power correction value [Phi _i from the least-squares solution that minimizes the difference between the value obtained by multiplying the echo-power correction value [Phi _i The echo power correction value Φ _i thus obtained is multiplied by the estimated echo power value | D _{ω, i} ^ | ² to achieve echo suppression that is robust against echo path fluctuations.

In contrast to the conventional gain coefficient calculation formula (formula (17)), the present invention obtains the gain coefficient by the following formula. Note that ^p in G _{ω, i} ^p is a symbol for distinguishing from the above-described prior art gain coefficient G _{ω, i} .

Here, ω represents a frequency number, i represents a frame number, and Φ _i represents an echo power correction value. The echo power correction value Φ _i is obtained by minimizing the evaluation amount e _ε expressed by the following equation.

Here, the subscript F means each value (norm, inner product) in the frequency axis direction. When equation (19) is differentiated by Φ _i , it can be expressed by equation (20).

  Equation (21) with Equation (20) set to 0

When the echo power correction value Φ _i is obtained from the above, equation (22) is obtained. Here, the frequency axis direction of the subscript F means each value (norm, inner product) of the sample sequence in the ω direction (ω = 0 to L−1).

  Practically, the inner product and norm may be calculated in the range over the past N frames in addition to the entire frequency region as shown in equations (23) and (24). N is a number of about 1 to 200, for example.

Here, L represents the total number of frequency numbers. When the frequency analysis length (FFT points) is, for example, 256 points, L is 128, which is a half analysis point that does not exceed the Nyquist frequency. From the equations (23) and (24), the echo power correction value Φ _i can be expressed by the equation (25).

Hereinafter, the operation of Expression (22) will be considered. Assuming that the power spectrum | S _{ω, i} | ^{2 of the} near-end speaker signal and the echo power estimated value | D _{ω, i} ^ | ² are uncorrelated, Equation (22) can be approximated as shown in the following equation.

Furthermore, assuming that the spectral structure of the echo power spectrum | D _{ω, i} | ² and the echo power estimated value | D _{ω, i} ^ | ² is the same as the equation (26), as shown in the equation (27): Can be approximated.

Thus the echo-power correction value [Phi _i is the echo power spectrum | D _{omega, i} | ² and the echo power estimate | D _{ω, i} ^ | expressed by energy ratio between ^2. Equation (27) is actually calculated from the entire frequency region rather than the energy of the echo power estimated value | D _{ω, i} ^ | ² used to obtain the echo power correction value Φ _i, with reference to Equation (24). This indicates that the larger the energy of the echo power spectrum, the larger the echo power correction value.

From the equation (27), when the energy of the echo power spectrum | D _{ω, i} | ^{2 and} the energy of the echo power estimated value | D _{ω, i} ^ | ² are equal, the echo power correction value Φ _i is 1. Further, when the estimated echo power value | D _{ω, i} ^ | ² is smaller in energy than the echo power spectrum | D _{ω, i} | ² , the echo power correction value Φ _i is a value larger than 1. Further, when the echo power estimated value | D _{ω, i} ^ | ² has larger energy than the echo power spectrum | D _{ω, i} | ² , the echo power correction value Φ _i operates so as to be a value smaller than 1. .

As described above, since the echo power correction value Φ _i is updated in a short time of the past N frames, the echo power estimated value | D _{ω, i} ^ | ² can be adjusted at high speed. As a result, the echo suppression gain estimation method of the present invention can realize echo suppression that is robust against fluctuations in the echo path.

  FIG. 1 shows a functional configuration example of an echo canceling apparatus 100 according to the present invention. The operation flow is shown in FIG. The echo cancellation apparatus 100 includes a reproduction signal frequency analysis unit 81, a sound pickup signal frequency analysis unit 82, an echo power estimation unit 83, an echo power correction value calculation unit 110, an echo suppression gain calculation unit 120, and a multiplication unit 85. And a frequency synthesizer 86. The echo canceling apparatus 100 is realized by reading a predetermined program into a computer composed of, for example, a ROM, a RAM, a CPU, and the like, and executing the program by the CPU.

The reproduction signal frequency analysis unit 81 converts the reproduction signal x (i) into a reproduction signal spectrum _{Xω, i} in the frequency domain (step S81). The collected sound signal frequency analysis unit 82 converts the collected sound signal y (i) into a collected sound signal spectrum Y _{ω, i} in the frequency domain (step S82).

The echo power estimator 83 calculates the echo power estimated value | D _{ω, i} ^ | ² using the reproduction signal spectrum X _{ω, i} and the collected sound signal spectrum Y _{ω, i} as inputs (step S83). The echo power correction value calculator 110 receives the collected sound signal spectrum Y _{ω, i} and the echo power estimated value | D _{ω, i} ^ | ² as input, and the power of the collected sound signal spectrum Y _{ω, i} | Y _{ω, i} The echo power correction value Φ _i is obtained from the least squares solution that minimizes the difference between | ² and the echo power estimated value | D _{ω, i} ^ | ² multiplied by the echo power correction value Φ _i (step S110). .

The echo suppression gain calculation unit 120 receives the collected sound signal spectrum Y _{ω, i} , the echo power correction value Φ _i and the echo power estimated value | D _ω ^ | ² as input, and a gain coefficient G _{ω, i} for suppressing the echo. ^p is calculated so that the smaller the echo power correction value Φ _i is, the smaller the value is (step S120). Note that the processing steps of the multiplication unit 85 and the frequency synthesis unit 86 are omitted.

  The echo cancellation apparatus 100 is similar to the conventional echo cancellation apparatus 10 (FIG. 10) except that the similarity coefficient calculation unit 12 is replaced with an echo power correction value calculation unit 110 and the echo suppression gain calculation unit 14 is replaced with an echo suppression gain calculation unit 120. Is the same configuration as the prior art. The echo power correction value calculator 110 and the echo suppression gain calculator 120 of this new configuration will be described in more detail.

[Echo power correction value calculation section]
FIG. 3 shows a more specific functional configuration example of the echo power correction value calculation unit 110. The echo power correction value calculation unit 110 includes an inner product calculation unit 110a, a norm calculation unit 110b, a division unit 110c, a register 110d, and a register 110e.

The inner product calculation means 110a receives the echo power estimated value | D _{ω, i} ^ | ² output from the echo power estimation unit 83 and the sound collection signal spectrum Y _{ω, i} output from the sound collection signal frequency analysis unit 82, For example, the inner product is calculated by Expression (28).

The norm calculation unit 110b receives the echo power estimated value | _{Dω, i} ^ | ² output from the echo power estimation unit 83 as an input, and calculates the norm by, for example, Expression (29).

  Here, α is an enhancement coefficient, and the value is stored in the register 110e. The enhancement coefficient α is a positive number, for example, 2. The inner product may be calculated by equation (30), and the norm may be calculated by equation (31).

  Here, ε is a forgetting factor that satisfies 0 <ε ≦ 1, and determines an exponential decay time constant. For example, ε = 0.16. As ε approaches 1, it becomes a (weighted) value that depends on the calculated value of the current frame (the first term on the right side of Equation (30) and Equation (31)).

The dividing unit 110c calculates the echo power correction value Φ _i using the equation (32) with the inner product calculated by the inner product calculating unit 110a and the norm calculated by the norm calculating unit 110b as inputs.

The echo power correction value Φ _i is a value obtained by multiplying the square root magnitude | D _{ω, i} ^ | of the echo power estimated value to the α power and a value obtained by multiplying the magnitude of the collected sound signal spectrum | Y _{ω, i} | by the α power. This is a value obtained by dividing the inner product value by the value of the square root of the estimated echo power value | D _{ω, i} ^ |

Further, from the relationship between the above formulas (28) and (29), the echo power correction value Φ _i is the α power of the square root magnitude | D _{ω, i} ^ | The value obtained by multiplying the value obtained by multiplying the magnitude | Y _{ω, i} | by the α-th power in a predetermined range over the entire frequency region and the past several frames (formula (28)) is the echo power estimated value. The square root | D _{ω, i} ^ | is the value obtained by dividing the value of 2α to the power of 2α by a value obtained by adding all values calculated in a predetermined range over the entire frequency region and the past several frames (formula (29)). I can say that.

[Echo suppression gain calculator]
FIG. 4 shows a more specific functional configuration example of the echo suppression gain calculation unit 120. The echo suppression gain calculation unit 120 includes an echo power adjusting unit 120a, a subtracting unit 120b, a dividing unit 120c, a register 120d that records a constant C that eliminates an echo cancellation residue, and a register 120e that records an enhancement coefficient β. Prepare.

The echo power adjusting means 120a receives the echo power estimated value | D _{ω, i} ^ | ² and the echo power correction value Φ _i as input, and squares the echo power estimated value | D _{ω, i} ^ | The echo power adjustment value is calculated by multiplying each other and multiplying that value by a constant C (step S120a in FIG. 5). By including this echo power adjustment process, the echo component can be sufficiently suppressed.

The subtracting unit 120b receives the echo power adjustment value and the collected sound signal spectrum Y _{ω, i} as input, and obtains the echo power adjustment value from a value obtained by raising the magnitude | Y _{ω, i} | Subtract (Step S120b).

The dividing unit 120c divides the output signal of the subtracting unit 120b by the value obtained by raising the magnitude | Y _{ω, i} | of the collected sound signal spectrum by the power of the enhancement coefficient β, and outputs a gain coefficient G _{ω, i} ^p (step). S120c). That is, the echo suppression gain calculator 120, the gain coefficient G _omega by calculating the equation _(33), and outputs the _i ^p.

Here, the enhancement coefficient β is assumed to be a positive number, for example, 2 which is the same value as α. The constant C is assumed to be a positive number, for example, 1.0. And emphasis coefficient beta, the constant C is to adjust the gain factor G _omega, the _i ^p to a suitable value. When the enhancement coefficient β = 2, the expression is close to a Wiener filter.

Gain factor G _omega obtained as described _{above, i} ^p operates the echo canceller 100 to reduce the defects of the near-end talker signal s (i). Therefore, it is possible to perform echo suppression with less musical noise generation, which is more robust to echo path fluctuations than the conventional method.

FIG. 6 shows a functional configuration example of the echo power correction value calculation unit 210 in which an upper limit and a lower limit are provided for the magnitude of the echo power correction value Φ _i . The echo power correction value calculation unit 210 includes an inner product calculation unit 110a, a norm calculation unit 110b, a division unit 110c, a register 110d, a register 110e, and an upper / lower limit adjustment unit 210a. The echo power correction value calculation unit 210 is different from the echo power correction value calculation unit 110 in that it includes upper and lower limit adjustment means 210a.

The upper / lower limit adjustment unit 210a outputs an echo power correction value Φ _i ′ in which an upper limit R _max and a lower limit R _min are set as shown in the following equation with respect to the echo power correction value Φ _i calculated by the division unit 110c.

  The echo power correction value calculation unit 110 described in the first embodiment is based on the premise that the spectral envelope of the echo power is maintained when the echo path changes or changes. In other words, it is assumed that the spectral structure of the echo and its estimated value match.

When this assumption is broken, the echo power correction value cannot be obtained correctly. Therefore, the echo power correction value calculation unit 210 prevents the echo power correction value from becoming too small or too large. For example, R _max = 100 dB and R _min = 0 dB.

  The echo canceling apparatus 100 ′ of the present invention including the echo power correction value calculation unit 210 has an effect of stabilizing the echo suppression performance in an actual scene where the vectors of the echo spectrum and the near-end speaker signal spectrum are not completely orthogonal. Play.

[Results of evaluation experiment]
The performance of the echo suppression gain estimation method of the present invention was compared with that of the conventional method by applying it to the short-time spectral amplitude echo suppression processing. The conventional method is an echo suppression method with FDCC (Frequency-Domain Cross Correlation) (reference: C. Faller and C. Tournery, “Robust acoustic echo control using a simple echo path model,” Process. ICASSP2006, vol. .5, pp.281-284, May 2006.). The performance was compared by the simulation test and the subjective quality evaluation method MUSHRA method. The test conditions were a sampling frequency of 16 kHz, a frequency analysis length of 256 points, frequency analysis synthesis by 1/2 overlap addition, and a room with an echo path reverberation time of 300 ms.

  FIG. 7 shows the reproduced signal x (k) (a) and the near-end speaker signal s (k) (b) used in the evaluation experiment. The horizontal axis represents time “second”, and the vertical axis represents amplitude. In order to simulate a sudden increase in echo, the echo d (k) (see FIG. 1) was rapidly increased by about 20 dB at 12 seconds. The A section of 0 to 4 seconds and the B section of 4 to 8 seconds are single talk states of only the reproduction signal x (k) and only the near-end speaker signal s (k). The C section of 8 to 12 seconds and the D section of 12 to 16 seconds are in a double talk state.

  FIG. 8 shows an output signal s ^ (k) (see FIG. 1) obtained by the conventional method and the present invention. The horizontal axis is the time “second”, the vertical axis is the amplitude, and the time axis is the same time as in FIG. (A) is a sound pickup signal y (k) picked up by a microphone, and (b) is an output signal s ^ (k when the number of frames for estimating the acoustic coupling amount by the conventional method is 10 (F = 10). ), (C) are output signals s (k) when the same number of frames is 200 (F = 200) in the conventional method, and (d) is the same number of frames (F = 200) in the present invention. Is an output signal s ^ (k).

  An echo can be seen in section A of the collected sound signal y (k), and it can be seen that the echo is suppressed by any of the methods (b), (c), and (d). What should be noted is the waveform obtained by each method immediately after 12 seconds when the echo is suddenly increased. Comparing these waveforms, it can be seen that the waveform of the output signal s (k) of the present invention (d) is closest to the waveform of the near-end speaker signal s (k). This result indicates that the echo suppression gain estimation method of the present invention can suppress the echo most even immediately after the echo path is changed.

  FIG. 9 shows the result of evaluating the sound quality of the output signal s ^ (k) of each method by the subjective quality evaluation method. (A) shows the evaluation result of C section, (b) shows the evaluation result of D section. The horizontal direction is each method, and the vertical direction is the subjective value.

  The subjective value in the C section before changing the echo path has the highest score of the conventional method F = 200, and the present invention was 65.58 which is almost equal to the score. On the other hand, in the D section immediately after changing the echo path, the present invention showed the highest score of 22.75. In the section D, the score of the conventional method F = 10 with a small number of frames is better than that of the conventional method F = 200 with a large number of frames.

  Further, since the score of the present invention is better than that of the conventional method F = 10, the echo suppression gain estimation method of the present invention can cope with echo path fluctuations quickly and accurately and is robust to echo path fluctuations. I understand that.

  As described above, by using the echo suppression gain estimation method of the present invention, it is robust against echo path fluctuations, and the generation of musical noise can be reduced while achieving high echo suppression performance.

  In addition, the method and apparatus of this invention are not limited to the above-mentioned embodiment, In the range which does not deviate from the meaning of this invention, it can change suitably. Further, the processes described in the above method and apparatus are not only executed in time series according to the order of description, but also may be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Good.

  Further, when the processing means in the above apparatus is realized by a computer, the processing contents of functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-
R (Recordable) / RW (ReWritable) or the like can be used as a magneto-optical recording medium, MO (Magneto Optical disc) or the like as a semiconductor memory, EEP-ROM (Electronically Erasable and Programmable Read Only Memory) or the like as a semiconductor memory.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (8)

A reproduction signal frequency analysis stage for converting the reproduction signal into a reproduction signal spectrum X _{ω, i} in the frequency domain;
A sound collecting signal frequency analysis stage for converting the sound collecting signal into a sound collecting signal spectrum Y _{ω, i} in the frequency domain;
An echo power estimation step of calculating an echo power estimate value | D _{ω, i} ^ | ² in the entire frequency region by using the reproduction signal spectrum X _{ω, i} and the sound pickup signal spectrum Y _{ω, i} as inputs;
Using the collected sound signal spectrum Y _{ω, i} and the estimated echo power value | D _{ω, i} ^ | ² as input _, the power | Y _{ω, i} | ^{2 of the} collected sound signal spectrum Y _{ω, i} and the echo power estimate | D _{ω, i} ^ | and echo power correction value calculation step of determining an echo power correction value [Phi _i ² to the difference between the value obtained by multiplying the echo-power correction value [Phi _i from the least-squares solution that minimizes,
The sound collection signal spectrum Y _{ω, i} , the echo power correction value Φ _i and the echo power estimation value | D _{ω, i} ^ | ² are input, and the gain coefficient for suppressing the echo is the echo power correction value Φ. an echo suppression gain calculation stage for calculating so that _i becomes larger as _i takes a larger value;
An echo suppression gain estimation method comprising: The echo suppression gain estimation method according to claim 1,
The echo power correction value calculation stage is as follows:
The larger the energy of the actual echo power spectrum calculated from the entire frequency region than the energy of the echo power estimated value | D _{ω, i} ^ | ² used for obtaining the echo power correction value Φ _i , the larger the echo power Obtaining a correction value;
An echo suppression gain estimation method characterized by: The echo suppression gain estimation method according to claim 1 or 2,
The echo power correction value calculation stage is as follows:
The magnitude of the square root of the estimated echo power value | D _{ω, i} ^ | raised to the α power of the collected sound signal spectrum | Y _{ω, i} | multiplied by the α power of the entire frequency range and the past N frames A value obtained by adding all the values calculated in a predetermined range was calculated in the predetermined range over the entire frequency region and the past N frames as the value of the square root of the echo power estimated value | D _{ω, i} ^ | Calculating a value obtained by dividing all values by an added value as an echo power correction value Φ _i ;
An echo suppression gain estimation method characterized by: The echo suppression gain estimation method according to any one of claims 1 to 3,
The echo suppression gain calculation step generates an echo power adjustment value obtained by multiplying the value of the square root | D _{ω, i} ^ | of the echo power estimated value to the power of β and the echo power correction value Φ _i. An echo suppression gain estimation method comprising an echo power adjustment process. The echo suppression gain estimation method according to claim 4,
The echo suppression gain calculation stage is
The above echo power adjustment process,
A subtraction process for subtracting the echo power adjustment value from a value obtained by raising the magnitude | Y _{ω, i} |
A division process of dividing the output signal of the subtraction process by a value obtained by raising the magnitude | Y _{ω, i} | of the collected sound signal spectrum to the power of β;
An echo suppression gain estimation method comprising: The echo suppression gain estimation method according to any one of claims 1 to 5,
The echo power correction value calculation stage is as follows:
An echo suppression gain estimation method comprising an upper and lower limit adjustment process for setting an upper limit and a lower limit in the magnitude of the echo power correction value Φ _i obtained by calculation. A reproduction signal frequency analyzer for converting the reproduction signal into a reproduction signal spectrum _{Xω, i} in the frequency domain;
A sound collection signal frequency analysis unit for converting the sound collection signal into a frequency domain sound collection signal spectrum Y _{ω, i} ;
An echo power estimator for calculating an echo power estimate value | D _{ω, i} ^ | ² in the entire frequency region by using the reproduced signal spectrum X _{ω, i} and the collected sound signal spectrum Y _{ω, i} as inputs;
Using the collected sound signal spectrum Y _{ω, i} and the estimated echo power value | D _{ω, i} ^ | ² as input _, the power | Y _{ω, i} | ^{2 of the} collected sound signal spectrum Y _{ω, i} and the echo An echo power correction value calculation unit for obtaining an echo power correction value from a least squares solution that minimizes a difference from a value obtained by multiplying the estimated power value | D _{ω, i} ^ | ² by an echo power correction value Φ _i ;
The sound collection signal spectrum _Yω , the echo power correction value Φ _i, and the echo power estimation value | D _{ω, i} ² are input, and the gain coefficient for suppressing the echo is set to a large value for the echo power correction value Φ _i. An echo suppression gain calculator that calculates the value to be as small as possible,
An echo suppression gain estimation apparatus comprising:   A program for processing the echo suppression gain estimation method according to any one of claims 1 to 6 by a computer.