JP4903842B2 - Adaptive filter and echo canceller having the same - Google Patents

Description

  The present invention relates to an adaptive filter that generates a pseudo echo by calculating a received voice signal with a tap coefficient, and an echo canceller that cancels an echo using the pseudo echo generated by the adaptive filter.

  Along with changes in call formats and hardware, we have a greater chance of using microphones. The microphone acquires not only the necessary sound, but also the sound output from the speaker and the indirect wave (echo) reflected from the wall surface of the room. Sounds late and is very annoying. In hands-free telephones and the like, a specific frequency component in an echo is amplified and howling (oscillation phenomenon) may occur.

  An echo canceller is a system for canceling this echo. The echo canceller updates the adaptive filter tap coefficient to be equal to the transfer function from the speaker to the microphone, and uses the tap coefficient to process the input signal (hereinafter referred to as “pseudo echo”) to the microphone. Subtract from the collected audio signal. As a result, the signal output from the echo canceller (hereinafter referred to as “error signal”) is the one in which the echo component is suppressed.

The adaptive filter generates a pseudo echo by calculating the input speech at the discrete time n with the tap coefficient w _n [i]. The adaptive filter updates the tap coefficient w _n [i] using a predetermined adaptive algorithm. Note that n is a sample time index, and i is an argument (index) representing the tap position of the adaptive filter.

The recurrence formula for updating the tap coefficient w _n [i] is generally expressed by the following formula (1). In Equation (1), w _n [i] is a tap coefficient before adaptation, w _{n + 1} [i] is a tap coefficient after adaptation, μ is a step size for adjusting the convergence speed of the tap coefficient, and Δw _n [i]. Is an update factor. The update coefficient Δw _n [i] varies depending on the type of adaptive algorithm.

  An adaptive filter generally receives an input signal equal to a system input signal input to a system with an unknown transfer function, outputs an output signal by filtering the input signal, and taps in the filter processing The coefficient is defined as a filter that updates the tap coefficient to be equal to the transfer function of the system based on the error signal that is the difference between the system output signal output from the system and the output signal and the input signal. .

  By the way, the adaptive estimation operation of the adaptive filter is normally performed in a single talk state in which only the far-end speaker's voice signal exists.

However, in the case of a double talk state in which the speech signal of the near-end speaker (near-end signal) and the speech signal of the far-end speaker (far-end signal) are superimposed, the adaptive algorithm used so far is Since the influence of the voice (additive noise) of the speaker at the end side is not taken into consideration, if the adaptive filter is updated in this state, the adaptive error of the adaptive filter increases, and the tap coefficient w _n [i] diverges. There is a risk of it.

  In order to solve this problem, the conventional technology is provided with a double-talk detection circuit that detects whether or not a double-talk state is detected. In the case of a double-talk state, the adaptive estimation operation of the adaptive filter is stopped or an adaptive algorithm is used. Measures such as reducing the step size μ are taken (see Patent Document 1, Patent Document 2, and Patent Document 3).

JP 2007-274714 A JP 2008-78973 A JP 2008-98929 A

  However, in the method of providing a double talk detection circuit as in the prior art, it takes time to obtain an optimum threshold value for determining whether or not a double talk state is present. In addition, it is difficult to accurately detect the double talk state, and it is difficult to maintain good convergence characteristics of the adaptive filter depending on the conventional technology.

  The present invention has been made in view of the above points, and provides an adaptive filter and an echo canceller having the adaptive filter capable of preventing an increase in an adaptive error in a double talk state without providing a double talk detection circuit. The purpose is to do.

The adaptive filter of the present invention receives an input signal x [n] equal to a system input signal input to a system having an unknown transfer function, and filters the input signal x [n] to output the output signal y [n]. n ”, an error signal e [n] that is a difference between the system output signal d [n] output from the system and the output signal y [n], and the input signal x [n]. And a tap coefficient setting unit that updates the tap coefficient w _n [i] in the filter processing so as to be equal to the transfer function of the system, where n is a sample time index, i Is an argument representing the tap position of the adaptive filter), and the tap coefficient setting unit adds the product of the step size 2μ and the update coefficient Δw _n [i] to the tap coefficient w _n [i] and updates the tap coefficient. w _{n +} Generates a [i], the updated coefficient [Delta] w n _[i], generated by dividing the numerator term in a normalized denominator, the error signal e of the numerator term, the input signal x [n-i] [N], and the normalized denominator term is a positive constant c that is the mean square root of the product of the square of the input signal x [n] and the square of the error signal e [n]. The configuration is generated by adding .

  The echo canceller of the present invention subtracts the output signal y [n] output from the adaptive filter from the adaptive filter and the system output signal d [n] to generate the error signal e [n]. And a container.

  According to the present invention, by controlling the convergence speed based on the power of the far-end signal and the power of the near-end signal, the optimum convergence speed can be automatically maintained, and a good and stable convergence characteristic can be obtained. be able to. And according to this invention, in the case of a double talk state, a convergence speed can be reduced and it can prevent that an adaptation error increases.

The block diagram which shows the structure of the digital communication apparatus containing the echo canceller which concerns on one embodiment of this invention The block diagram which shows the internal structure of the adaptive filter which concerns on one embodiment of this invention The figure which shows the structure of the arithmetic circuit in the case of Formula (2) of the tap coefficient setting part of the adaptive filter which concerns on one embodiment of this invention The figure which shows the structure of the arithmetic circuit in the case of Formula (3) of the tap coefficient setting part of the adaptive filter which concerns on one embodiment of this invention The figure which shows the structure of the arithmetic circuit in the case of Formula (4) of the tap coefficient setting part of the adaptive filter which concerns on one embodiment of this invention The figure which shows the simulation result of the adaptive filter which concerns on one embodiment of this invention The figure which shows the structure of the system at the time of the simulation of the adaptive filter which concerns on one embodiment of this invention The figure which shows the structure of the arithmetic circuit in the case of Formula (6) of the tap coefficient setting part of the adaptive filter which concerns on one embodiment of this invention The figure which shows the structure of the arithmetic circuit in the case of Formula (7) of the tap coefficient setting part of the adaptive filter which concerns on one embodiment of this invention The figure which shows the result of the computer simulation performed in order to demonstrate the effectiveness of the adaptive filter which concerns on one embodiment of this invention

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a digital communication apparatus including an echo canceller according to an embodiment of the present invention.

  As shown in FIG. 1, the digital communication apparatus includes a digital / analog (D / A) converter 101, a power amplifier 102, a speaker 103, a microphone 104, a microphone amplifier 105, and an analog / digital (A / D). ) It is mainly composed of a converter 106 and an echo canceller 107. The echo canceller 107 includes an adaptive filter 108 and a subtractor 109.

  The signal transmitted from the communication partner device is subjected to processing such as demodulation and decoding in a receiving circuit (not shown) to become a digital received voice signal x [n]. The digital reception voice signal x [n] is an input signal of the adaptive filter 108.

  The digital / analog converter 101 converts the digital received audio signal x [n] into an analog audio signal. The analog received audio signal is amplified by the power amplifier 102 and output from the speaker 103 as audio.

  The analog transmission audio signal input to the microphone 104 is amplified by the microphone amplifier 105 and input to the analog / digital converter 106. Note that the input sound s [n] of the near-end speaker and the reproduced sound g [n] from the speaker 103 are input to the microphone 104 as an echo. The input speech of the near-end speaker corresponds to additive noise that disturbs the convergence of the adaptive filter 108.

  The analog / digital converter 106 converts the analog transmission voice signal into a digital transmission voice signal d [n].

The echo canceller 107 updates the tap coefficient w _n [i] of the adaptive filter 108 to be equal to the impulse response of the transfer function from the digital / analog converter 101 to the analog / digital converter 106. Since the power amplifier 102 and the microphone amplifier 105 normally have flat frequency characteristics, w _n [i] is equal to the impulse response of the transfer function from the speaker 103 to the microphone 104. Then, the echo canceller 107 subtracts the pseudo echo y [n] obtained by processing the digital reception voice signal x [n] using the tap coefficient w _n [i] from the digital transmission voice signal d [n]. The suppressed error signal e [n] is output.

The adaptive filter 108 operates the digital received speech signal x [n] at the discrete time n with the tap coefficient w _n [i] to generate a pseudo echo y [n]. Further, the adaptive filter 108 updates the tap coefficient w _n [i] by a predetermined adaptive algorithm based on the error signal e [n] and the digital received speech signal x [n] output from the subtractor 109. Note that i is an argument (index) representing the tap position of the adaptive filter.

  The subtractor 109 subtracts the pseudo echo y [n] from the digital transmission audio signal d [n] output from the analog / digital converter 106 to obtain an error signal e [n]. The error signal e [n] is subjected to processing such as encoding and modulation in a transmission circuit (not shown), and is transmitted to a communication partner apparatus.

  Next, the internal configuration of the adaptive filter 108 will be described with reference to the drawings. In the following description, the digital reception voice signal x [n] is referred to as an input signal, and the pseudo echo y [n] is referred to as an output signal.

  FIG. 2 is a block diagram showing the internal configuration of the adaptive filter 108. As shown in FIG. 2, the adaptive filter 108 includes a filter unit 201 and a tap coefficient setting unit 202.

The filter unit 201 is an FIR filter or the like, and generates an output signal y [n] by calculating (filtering) the input signal x [n] with the tap coefficient w _n [i].

The tap coefficient setting unit 202 sets the tap coefficient w _n [i] by a predetermined adaptive algorithm using the input signal x [n] and the error signal e [n], and filters the tap coefficient w _n [i]. Output to the unit 201. The present invention is characterized by an arithmetic circuit of an adaptive algorithm in the tap coefficient setting unit 202.

  Hereinafter, an adaptive filter according to the present invention when an NLMS (Normalized Least Mean Square) algorithm is used as the adaptive algorithm will be described. The NLMS algorithm is an adaptive algorithm that performs processing in the time domain, performs sequential calculation each time a new signal sample value is input, and gradually converges the tap coefficient to an optimum value.

When the NLMS algorithm is used, the update recurrence formula of the tap coefficient w _n [i] of the present invention is expressed by any of the following formulas (2) to (4). In Equations (2) to (4), w _n [i] is the tap coefficient before adaptation (time n), w _{n + 1} [i] is the tap coefficient after adaptation (time n + 1), and 2μ is the tap coefficient. A step size for adjusting the convergence speed, x [n] is an input signal, e [n] is an error signal, and c is a small positive constant that prevents the fractional denominator value from becoming zero. The symbol 表 す represents the time average of the signal.

Note that the time average is obtained by a moving average using a finite number of samples in the past or an integration process using a forgetting factor. The moving average can be calculated with an N-tap FIR filter in which all tap coefficients are 1 / N, where N is the average data length. The integration process using the forgetting factor α is defined by the following equation (5). Expression (5) shows a case where the time average is obtained for the input signal f [n].

  In any of the equations (2) to (4), unlike the conventional NLMS algorithm, normalization processing is performed using not only the power of the input signal x [n] but also the power of the error signal e [n]. .

  Expression (2) performs normalization processing using the average value of the sum of the power of the input signal x [n] and the power of the error signal e [n]. Expression (3) performs normalization processing by the square root of the average value of the products of the power of the input signal x [n] and the power of the error signal e [n]. The processing of the formula (2) and the processing of the formula (3) can obtain almost the same effect.

  In Expression (4), in order to omit the calculation of the square root of Expression (3), normalization processing is performed using an average value of absolute values of products of the input signal x [n] and the error signal e [n]. Since the product of the input signal x [n] and the error signal e [n] is a power dimension, the square root operation is unnecessary, and the processing of the equation (4) is easier than the equation (3).

FIG. 3 is a diagram illustrating a configuration of an arithmetic circuit in the case of the expression (2) of the tap coefficient setting unit 202. FIG. 4 is a diagram illustrating a configuration of an arithmetic circuit in the case of the equation (3) of the tap coefficient setting unit 202. FIG. 5 is a diagram illustrating a configuration of an arithmetic circuit in the case of the equation (4) of the tap coefficient setting unit 202. 3, 4, and 5, (+) represents an adder, and (×) represents a multiplier. Further, (·) ⁻¹ represents an inverse arithmetic circuit for obtaining the inverse of the input signal. Further, √ (•) represents a square root arithmetic circuit for obtaining the square root of the input signal. Further, | · | represents an absolute value calculation circuit for obtaining the absolute value of the input signal. Z- ⁱ represents a delay circuit that performs i-sample delay processing.

In FIG. 3, the tap coefficient setting unit 202 calculates the numerator term of the update coefficient Δw _n [i] by the delay circuit 301 and the multiplier 302 using the input signal x [n] and the error signal e [n]. , Multiplier 303, multiplier 304, adder 305, averaging circuit 306, and adder 307 calculate the normalized denominator term of the update coefficient Δw _n [i]. Then, the tap coefficient setting unit 202 calculates the update coefficient Δw _n [i] by dividing the numerator term by the normalized denominator term by the reciprocal arithmetic circuit 308 and the multiplier 309. Then, the tap coefficient setting unit 202 uses the multiplier 310, the adder 311, and the delay circuit 312 to update the tap coefficient w _n [i] by multiplying the update coefficient Δw _n [i] by the step size 2 μ. The subsequent tap coefficient w _{n + 1} [i] is generated.

In FIG. 4, the tap coefficient setting unit 202 calculates the numerator term of the update coefficient Δw _n [i] using the delay circuit 401 and the multiplier 402 using the input signal x [n] and the error signal e [n]. , Multiplier 403, multiplier 404, multiplier 405, averaging circuit 406, square root calculation circuit 407, and adder 408 calculate the normalized denominator term of the update coefficient Δw _n [i]. Then, the tap coefficient setting unit 202 calculates an update coefficient Δw _n [i] by dividing the numerator term by the normalized denominator term by the reciprocal arithmetic circuit 409 and the multiplier 410. Then, the tap coefficient setting unit 202 uses the multiplier 411, the adder 412, and the delay circuit 413 to update the tap coefficient w _n [i] by multiplying the update coefficient Δw _n [i] by the step size 2μ. The subsequent tap coefficient w _{n + 1} [i] is generated.

In FIG. 5, the tap coefficient setting unit 202 calculates the numerator of the update coefficient Δw _n [i] using the delay circuit 501 and the multiplier 502 using the input signal x [n] and the error signal e [n]. The multiplier 503, the absolute value calculation circuit 504, the averaging circuit 505, and the adder 506 calculate the normalized denominator term of the update coefficient Δw _n [i]. Then, the tap coefficient setting unit 202 calculates the update coefficient Δw _n [i] by dividing the numerator term by the normalized denominator term by the reciprocal arithmetic circuit 507 and the multiplier 508. Then, the tap coefficient setting unit 202 uses the multiplier 509, the adder 510, and the delay circuit 511 to update the tap coefficient w _n [i] by multiplying the update coefficient Δw _n [i] by the step size 2μ. The subsequent tap coefficient w _{n + 1} [i] is generated.

  Next, the adaptive filter of the present invention when the FLMS (Fast Least Mean Square or Frequency domain Least Mean Square) algorithm is used as the adaptive algorithm will be described. The FLMS algorithm is an adaptive algorithm in which block processing is performed in the frequency domain using discrete Fourier transform periodically every several samples, and tap coefficients gradually converge to an optimum value. The FLMS algorithm may be referred to as an FBLMS (Fast Block Least Mean Square or Frequency domain Block Least Mean Square) algorithm.

When the FLMS algorithm is used, the update recurrence formula of the discrete Fourier transform W _n [i] of the tap coefficient of the present invention is expressed by the following formula (6) or formula (7). In Equations (6) and (7), W _n [k] is the discrete Fourier transform of the tap coefficient w _n [i], and W _{n + L} [k] is the discrete Fourier transform of the tap coefficient w _{n + L} [i]. , Μ is a step size for adjusting the convergence speed of the tap coefficient, X [k] is a discrete Fourier transform of the input signal x [n], E [k] is a discrete Fourier transform of the error signal e [n], and c is A small positive constant that prevents the value of the denominator of the fraction from becoming zero, and p is a small positive constant that prevents the fractional numerator from becoming an extremely small value. K represents a frequency. L represents a cycle for performing the coefficient update process. Each time a new signal sample value is input with L = 1, the calculation may be performed sequentially. The symbol 表 し represents the time average of the signal, and the symbol ^* represents the complex conjugate.

  Both the equations (6) and (7) are different from the conventional FLMS algorithm, not only the discrete Fourier transform X [k] of the input signal x [n] but also the discrete Fourier transform of the error signal e [n]. Normalization processing is also performed using E [k].

In equations (6) and (7), X [k] is obtained by the following equation (8), and E [k] is obtained by the following equation (9). In Expressions (8) and (9), N is the tap length of the adaptive filter, and DFT represents a discrete Fourier transform.

The discrete Fourier transform W _{n + L} [k] of the tap coefficient in the frequency domain updated by the equations (6) and (7) is subjected to a discrete inverse Fourier transform as shown in the following equation (10), and its real part To obtain the time domain tap coefficient w _{n + L} [i]. In Equation (10), IDFT represents a discrete inverse Fourier transform. REAL represents the process of extracting the real part of the argument. Usually, a fast Fourier transform algorithm is used for DFT / IDFT computation.

The tap coefficient length obtained by Expression (10) is 2N from wn _{+ L} [0] to wn _{+ L} [2N-1], but wn _{+ L} [N] to wn _{+ L} [2N-1] are discarded, w _{n + L} [0] to w _{n + L} [N−1] are used as tap coefficients.

  Expression (6) is obtained by applying the normalization process based on the absolute value of the product of the input signal x [n] and the error signal e [n] in Expression (4) to the FLMS algorithm that performs block processing in the frequency domain. Then, normalization processing is performed in the frequency domain using the absolute value of the product of the discrete Fourier transform X [k] of the input signal x [n] and the discrete Fourier transform E [k] of the error signal e [n].

  Equation (7) is obtained by adding processing for improving the convergence speed from the initial state at the time of system startup to Equation (6), and is the product of the output signal y [n] of the adaptive filter and the error signal e [n]. Is obtained by dividing the absolute value of the average value of the signals by the absolute value of the product of the discrete Fourier transform X [k] of the input signal x [n] and the discrete Fourier transform E [k] of the error signal e [n]. Process. However, the average value of the product of y [n] and e [n] is calculated by integration processing using the forgetting coefficient of the above equation (5), and the initial value must be a value other than zero. The initial value may be about the same as the average power of the input signal x [n].

Here, the product of y [n] and e [n] can be expressed as in the following formula (11). Note that * in the expression (11) is an operator representing convolution.

In Equation (11), s [n] and x [n] are signals from different signal sources, and both are uncorrelated, so if the average interval is sufficiently long, s [n] and x [n ] * W _n [i], that is, the correlation between s [n] and x [n] * w _n [i] is almost zero. Therefore, the absolute value of the average value of the products of y [n] and e [n] is expressed as the following Expression (12).

  Unless the adaptive filter converges completely and g [n] = y [n], the estimation error (g [n] −y [n]) and y [n] have a correlation with g [n]. And the absolute value of the average of the products of y [n] and e [n] shown in Expression (12) is a positive value.

  Since the estimation error is large immediately after the start of the operation of the adaptive filter, the value of Expression (12) has a large value immediately after the start of the operation of the adaptive filter, but approaches 0 as the adaptive operation progresses and converges. Therefore, in the adaptive filter coefficient updating process based on the equation (7), immediately after the start when the estimation error becomes a large value, it is equivalent to the value of the step size μ being increased, and the convergence speed is accelerated. . This process not only improves the convergence speed immediately after the start of the operation of the adaptive filter, but also when the impulse response of the identification target system (acoustic system) fluctuates rapidly, and the error signal e [n] increases rapidly. Has the effect of speeding up the convergence of the adaptive filter.

  Actually, a speech signal is used as the input signal x [n], a simulation is performed in the adaptive filter, and the absolute value of the average value of the products of y [n] and e [n] shown in Expression (12) is obtained. As shown in FIG. In FIG. 6, the horizontal axis represents time (unit: sample), and the vertical axis represents the absolute value of the average value of the products of y [n] and e [n]. Also, in FIG. 6, Equation (6) is used for the coefficient update algorithm of the adaptive filter. The system configuration at the time of simulation is shown in FIG. 7, and different human voice data recorded at a sampling frequency of 8 kHz was used for the input signals x [n] and s [n]. Further, the averaging process of the equation (12) is performed by the integral calculation defined by the equation (5) using the forgetting factor, and the value of the forgetting factor α is set to 0.02. In addition, the initial value when performing the calculation of Expression (5) is the power value of the first 300 samples of the input signal x [n]. In FIG. 7, h [i] is the tap coefficient of the filter to be identified that represents the transfer function of the acoustic system.

  As is apparent from FIG. 6, the absolute value of the average value of the products of y [n] and e [n] has a large value until 10000 samples after the start of operation, but is almost 0 after 10000 samples. The value is within 3 or less. In this case, it is appropriate that the value of the constant p in the equation (7) is from 0.3 to a fraction of that.

FIG. 8 is a diagram illustrating a configuration of an arithmetic circuit in the case of the equation (6) of the tap coefficient setting unit 202. FIG. 9 is a diagram illustrating a configuration of an arithmetic circuit in the case of the equation (7) of the tap coefficient setting unit 202. 8 and 9, (+) represents an adder, and (x) represents a multiplier. Further, (·) ⁻¹ represents an inverse arithmetic circuit for obtaining the inverse of the input signal. Further, | · | represents an absolute value calculation circuit for obtaining the absolute value of the input signal. Z- ⁱ represents a delay circuit that performs i-sample delay processing. DFT represents a Fourier transform circuit that performs discrete Fourier transform. IDFT represents an inverse Fourier transform circuit that performs discrete inverse Fourier transform. CONJ represents a complex conjugate arithmetic circuit for obtaining the complex conjugate of the input signal. ADD ZERO represents an insertion circuit that performs processing for adding N zeros in the above equation (9) to the input number sequence. REAL represents an extraction circuit that outputs only the real part of the input complex signal. In FIG. 8 and FIG. 9, the configuration for discarding the extra tap coefficients w _{n + 1} [N] to w _{n + 1} [2N−1] on the right side of the equation (10) is omitted.

In FIG. 8, the tap coefficient setting unit 202 obtains an input signal vector X [k] by performing a discrete Fourier transform on an input signal x [n] by a Fourier transform circuit 801, and an error is obtained by an insertion circuit 802 and a Fourier transform circuit 803. The signal e [n] is subjected to discrete Fourier transform to obtain an error signal vector E [k]. Then, the tap coefficient setting unit 202 uses the complex conjugate arithmetic circuit 804 and the multiplier 805 to calculate the numerator of the update vector ΔW _n [k] using X [k] and the error signal vector E [k]. A multiplier 806, an absolute value calculation circuit 807, and an adder 808 calculate a normalized denominator term of the update vector ΔW _n [k]. Then, the tap coefficient setting unit 202 calculates an update vector ΔW _n [k] by dividing the numerator term by the normalized denominator term by the reciprocal arithmetic circuit 809 and the multiplier 810. Then, the tap coefficient setting unit 202 adds the product obtained by multiplying the update vector ΔW _n [k] by the step size μ to the tap coefficient vector W _n [k] by the multiplier 811, the adder 812, and the delay circuit 813. An updated tap coefficient vector W _{n + L} [k] is generated. Then, the tap coefficient setting unit 202 performs discrete inverse Fourier transform on W _{n + L} [k] by the inverse Fourier transform circuit 814 to obtain w _{n + L} [i].

In FIG. 9, the tap coefficient setting unit 202 obtains an input signal vector X [k] by performing a discrete Fourier transform on an input signal x [n] by a Fourier transform circuit 901, and an error is obtained by an insertion circuit 902 and a Fourier transform circuit 903. The signal e [n] is subjected to discrete Fourier transform to obtain an error signal vector E [k]. Then, the tap coefficient setting unit 202 uses the output signal y [n], the error signal e [n], the input signal vector X [k], and the error signal vector E [k] to generate a complex conjugate arithmetic circuit 904, a multiplier. 905, a multiplier 906, an averaging circuit 907, an absolute value calculation circuit 908, an adder 909 and a multiplier 910 calculate the numerator term of the update vector ΔW _n [k], and a multiplier 911, an absolute value calculation circuit 912, and an addition A normalized denominator term of the update vector ΔW _n [k] is calculated by the calculator 913. Then, the tap coefficient setting unit 202 calculates an update vector ΔW _n [k] by dividing the numerator term by the normalized denominator term by the reciprocal arithmetic circuit 914 and the multiplier 915. Then, the tap coefficient setting unit 202 adds the product obtained by multiplying the update vector ΔW _n [k] by the step size μ to the tap coefficient vector W _n [k] by the multiplier 916, the adder 917, and the delay circuit 918. An updated tap coefficient vector W _{n + L} [k] is generated. Then, the tap coefficient setting unit 202 performs discrete inverse Fourier transform on W _{n + L} [k] by the inverse Fourier transform circuit 919 to obtain w _{n + L} [i].

The results of computer simulation performed to verify the effectiveness of the adaptive filter of the present invention described above are shown below. The system configuration at the time of simulation is the same as that of FIG. 7, and different speakers read out the same sentence to the input signal x [n] and s [n] corresponding to additive noise, and the sampling frequency is 8 kHz. The data which was recorded in the above and normalized so that the absolute value of the amplitude is 1 or less was used. The identification target FIR filter and adaptive filter are both 300 taps. The tap coefficient h [i] of the identification target filter was set as in the following formula (13). The range of the argument i is 0 ≦ i <300.

FIG. 10 shows a change in the estimation error amount D [n] of the tap coefficient h [i] of the filter to be identified represented by Expression (14). In FIG. 10, the horizontal axis represents time, and the vertical axis represents the error amount D [n].

  10A shows the case where the prior art NLMS algorithm is used, FIG. 10B shows the case where equation (2) is used, FIG. 10C shows the case where equation (3) is used, and FIG. ) Shows the case where equation (4) is used, FIG. 10 (e) shows the case where equation (6) is used, and FIG. 10 (f) shows the case where equation (7) is used. 10 (e) and 10 (f) use a block processing algorithm in the frequency domain using DFT / IDFT, the tap coefficient is updated every 300 samples, which is the same as the tap length of the adaptive filter. ing. In any of the processing examples, all the average calculations use the integral calculation using the forgetting coefficient of the above formula (5), and the value of the forgetting coefficient is 0.005 in FIGS. 10 (a) to 10 (d). In (f), it is set to 0.05. In addition, the value of the constant c for preventing the fractional denominator from being 0 is set to 0.0001.

  As shown in FIG. 10A, in the NLMS algorithm of the prior art, if the value of the step size μ is extremely small as 0.00001, the convergence is hardly affected by the double talk, but the convergence speed is very high. Become slow. On the other hand, in the conventional NLMS algorithm, convergence becomes faster when μ = 0.00002 or more, but a stable convergence state cannot be maintained due to the influence of double talk.

  On the other hand, as is clear from the comparison between FIG. 10 (a) and FIGS. 10 (b) to 10 (d), the adaptive filter coefficient update algorithm in the time domain according to the present invention is used, both of which are conventional NLMS. Convergence characteristics better than the algorithm can be obtained.

  Further, as shown in FIG. 10E, the adaptive filter coefficient update algorithm of Equation (6) in the frequency domain according to the present invention can obtain better convergence characteristics than the time domain processing.

  As shown in FIG. 10 (f), the adaptive filter coefficient update algorithm of the equation (7) in the frequency domain according to the present invention can further improve the convergence speed immediately after the start-up than FIG. 10 (e). . Based on the simulation result of FIG. 6, the value of the constant p in the equation (7) was set to 0.2.

  Thus, in the present invention, the conventional NLMS algorithm performs normalization processing using only the power of the input signal, whereas the adaptive filter NLMS algorithm of the present invention uses the power of the input signal and the error signal. Normalization processing is performed using both of the powers.

  As the power of the additive noise added to the desired signal increases, the power of the error signal also increases. By normalizing using the error signal power as in the present invention, the convergence speed of the adaptive algorithm is automatically lowered when the power of additive noise is large, thereby preventing deterioration of convergence characteristics due to the influence of noise. Can do.

  In the echo canceller, when the double talk state is entered, the adaptive error of the adaptive filter increases and the power of the error signal also increases. Therefore, as in the present invention, by performing the normalization process using the power of the error signal, the convergence speed of the adaptive filter is reduced in the double talk state, and the adaptation error increases rapidly due to the double talk. Can be prevented.

  Further, if the double talk state is eliminated and the power of the error signal e [n] is reduced, the convergence speed is automatically increased.

  In addition, since the present invention is a processing method that does not assume the assumption of stationarity of the input signal, the assumption of whiteness / normal distribution characteristics, the assumption of correlation characteristics, etc., actual audio signal input in which these assumptions are difficult to hold. Good convergence characteristics can be obtained. The precondition for the processing of the present invention is that the ratio of the power of the input signal x [n] in FIG. 1, that is, the voice of the speaker of the far-end terminal and the power of the input voice s [n] varies slowly. The only assumption is that It is self-evident that the voice power of different speakers fluctuates uncorrelated, and this assumption holds without problems.

  Note that only the processing of the above equation (7) and FIG. 9 performs the speeding up of the convergence speed based on the assumption regarding the correlation characteristic of the input signal, but it is shown that the assumption is sufficiently established in actual speech. This is demonstrated by 6 simulation results.

  As described above, according to the present invention, the optimum convergence speed is automatically maintained according to the power of the far-end signal and the power of the near-end signal, so that a good and stable convergence characteristic can be obtained.

  When the present invention is compared with the conventional method of controlling the adaptive speed using the double talk detection circuit, the present invention does not require a threshold setting process for determining whether or not it is in the double talk state. Even if the environment changes, there is an advantage that it is not necessary to obtain and reset an optimum threshold value each time.

  In the above-described embodiment, the echo canceller having the adaptive filter has been described. The adaptive filter of the present invention is not limited to the echo canceller, and can be used for other devices.

  The present invention is suitable for use in an echo canceller, a howling canceller, and the like of a bidirectional communication system (wireless telephone, wired telephone, interphone, video conference system, etc.).

DESCRIPTION OF SYMBOLS 101 Digital / analog converter 102 Power amplifier 103 Speaker 104 Microphone 105 Microphone amplifier 106 Analog / digital converter 107 Echo canceller 108 Adaptive filter 109 Subtractor 201 Filter part 202 Tap coefficient setting part 301,312,401,413,501,511 , 813, 918 Delay circuit 302, 303, 304, 309, 310, 402, 403, 404, 405, 410, 411, 502, 503, 508, 509, 805, 806, 810, 811, 905, 906, 910, 911, 915, 916 Multiplier 305, 307, 311, 408, 412, 506, 510, 808, 812, 909, 913, 917 Adder 306, 406, 505, 907 Average circuit 308, 40 , 507, 809, 914 Reciprocal arithmetic circuit 407 Square root arithmetic circuit 504, 807, 908, 912 Absolute value arithmetic circuit 801, 803, 901, 903 Fourier transform circuit 802, 902 Insertion circuit 804, 904 Complex conjugate arithmetic circuit 814, 919 Inverse Fourier transform circuit

Claims (2)

A filter unit that inputs an input signal x [n] equal to a system input signal input to a system with an unknown transfer function, and outputs an output signal y [n] by filtering the input signal x [n]. When,
Based on the error signal e [n], which is the difference between the system output signal d [n] output from the system and the output signal y [n], and the input signal x [n], the transfer function of the system A tap coefficient setting unit for updating the tap coefficient w _n [i] in the filtering process so as to be equal to
(N is a sample time index, i is an argument indicating the tap position of the adaptive filter),
The tap coefficient setting unit
Tap coefficient updated by adding the step size 2μ and update the coefficient [Delta] w _n tap coefficient multiplied by the _{[i] w n [i]} w n + 1 to generate a [i],
Generating the update factor Δw _n [i] by dividing the numerator term by the normalized denominator term;
Generating the numerator term by multiplying the input signal x [n−i] by the error signal e [n];
The normalized denominator term is generated by adding a positive constant c to the average square root of the product of the square of the input signal x [n] and the square of the error signal e [n] .
Adaptive filter. The adaptive filter of claim 1 ;
A subtracter that subtracts the output signal y [n] output from the adaptive filter from the system output signal d [n] to generate the error signal e [n];
Echo canceller with

Images