JP2014096027A - Parameter estimation device, echo cancel device, parameter estimation method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a parameter estimation technique capable of integrally updating a nonlinear parameter and a linear parameter although suppressing an operation quantity small when estimating an output signal of a nonlinear system and an output signal of a linear system.SOLUTION: A parameter estimation device includes: a nonlinear parameter application part which finds a nonlinear system output estimated value using a nonlinear parameter; a differential signal generation part which finds a value equivalent to a partial differential value as to the nonlinear parameter of the nonlinear system output estimated value; a frequency region linear parameter multiplication part which finds a transmission system output estimated value by multiplying the nonlinear system output estimated value in a frequency region by the linear parameter; an error calculation part which finds the difference between an output signal of a transmission system and a transmission system output estimated value; and a mixed type parameter update part which updates the nonlinear parameter using the difference of the frequency region, partial differential value, and nonlinear system output estimated value, and updates the linear parameter in the frequency region.

Description

  The present invention relates to a cascade connection between a system in which an electric signal, sound, and vibration are propagated, having a nonlinear characteristic (hereinafter, “nonlinear system”) and a system having a linear characteristic (hereinafter, “linear system”). The present invention relates to a technique for estimating an unknown parameter of the transmission system when it can be considered.

  Estimating the characteristics of an unknown transmission system through which an electric signal, sound, or vibration propagates is useful for predicting the output of the transmission system with respect to an arbitrary input to the transmission system. The problem of estimating a transmission system is often handled as a problem in which a model that can be expressed with a finite parameter is given to the transmission system in advance and the value of the parameter is estimated. The model given in advance is selected based on the physical features (relationship between input and output) of the transmission system to be estimated. For example, when the relationship between input and output can be regarded as linear, a linear FIR (Finite Impulse Response) filter or IIR (Infinite Impulse Response) filter is selected as a model of a transmission system in the discrete time domain.

  Here, there is a problem of estimating the parameters of the transmission system 1 as shown in FIG. The transmission system 1 has a characteristic in which the relationship between input and output is non-linear. Further, in the transmission system 1, the nonlinear system 2 and the linear system 3 are separable and can be provided as a cascade connection model. The characteristic of the nonlinear system 2 can be regarded as a hard clip characteristic that does not depend on past inputs and outputs, and the characteristic of the linear system 3 can be regarded as a finite-length impulse response.

<Model of nonlinear system 2>
When the nonlinear system 2 in FIG. 1 has a hard clip characteristic that does not depend on past inputs and outputs, the relationship between the input signal x (n) and the output signal s (n) uses a threshold value a (> 0). Using the hard clip function f expressed as

And given. Here, n is a discrete time, and the output signal s (n) is related to the input signal x (n) as being dependent only on the value at the same discrete time n. Here, the threshold value a is an unknown parameter to be estimated. Hereinafter, the estimated value of the threshold value a is also referred to as a nonlinear parameter a ^.

<Model of linear system 3>
If the input signal to the linear system 3 in FIG. 1 is the output signal s (n) of the nonlinear system 2, the output signal y (n) of the linear system 3 is

, Where the finite-length impulse responses h ₀ , h ₁ ,..., H _L are unknown parameters to be estimated. Hereinafter, the impulse response h ₀ of finite length, h _1, ..., linear estimates of h _L parameter _{h ^ 0, h ^ 1,} ..., also referred to as h ^ _L. However, the order L is an integer greater than or equal to 0, and the output signal s (n) is an internal signal that cannot actually be observed directly.

From the above, the parameters to be estimated in the transmission system 1 are a in equation (1) and h ₀ , h ₁ ,..., H _L in equation (2).

<Parameter estimation device 6 according to the prior art>
Non-Patent Document 1 discloses a parameter estimation method in such a framework. FIG. 2 is a configuration example for realizing a conventional parameter estimation method based on Non-Patent Document 1. Hereinafter, the parameter estimation device 6 will be described.

  In the drawings used for the following description, constituent parts having the same function and steps for performing the same process are denoted by the same reference numerals, and redundant description is omitted. In the following description, the symbol “^” or the like used in the text should be described immediately above the character just before, but it is described immediately after the character due to restrictions on text notation. In the formula, these symbols are written in their original positions. Further, the processing performed for each element of a vector or matrix is applied to all elements of the vector or matrix unless otherwise specified.

  In the discrete time n, the following processing is performed in each part of the parameter estimation device 6.

  The nonlinear parameter application unit 601 receives the time domain input signal x (n), and uses the input signal x (n) and the nonlinear parameter a ^ (n) to calculate the time domain input signal x (n The nonlinear system output estimated value s ^ (n), which is an estimated value of the output signal of the nonlinear system 2 for), is obtained and output.

The signal storage unit 602 receives the nonlinear system output estimated value s ^ (n) and a partial differential value s ^ '(n) described later, and goes back to L samples in the past, and L + 1 nonlinear system output estimated values s. ^ (n), s ^ (n-1), ..., s ^ (nL) and L + 1 partial differential values s ^ '(n), s ^' (n-1), ..., s ^ Accumulate '(nL). S ^ (n) = [s ^ (n), s ^ (n-1),…, s ^ (nL)] ^T , S ^ '(n) = [s ^' (n), s ^ '(n-1), ... , s ^' (nL)] and ^T. Here, ^T represents transposition.

The linear parameter multiplication unit 603 extracts L + 1 non-linear system output estimation values s ^ (n), s ^ (n-1),..., S ^ (nL) from the signal storage unit 602, and , L + 1 nonlinear system output estimates s ^ (n), s ^ (n-1), ..., s ^ (nL) and L + 1 linear parameters h ^ ₀ (n), h ^ ₁ (n), ..., h ^ _L (n) are all multiplied to obtain the transmission system output estimate y ^ (n), which is the estimated value of the output signal from the estimation target transmission system. Output.
y ^ (n) = h ^ ₀ (n) s ^ (n) + h ^ ₁ (n) s ^ (n-1) +… + h ^ _L (n) s ^ (nL) (4)
Is output. In addition, H ^ (n) = If the title and _{[h ^ 0 (n),} h ^ 1 (n), ..., h ^ L (n)] T, equation (4),
y ^ (n) = H ^ (n) ^T S ^ (n) (5)
This corresponds to a vector inner product operation.

The error calculation unit 604 receives the output signal y (n) from the transmission system to be estimated and the transmission system output estimated value y ^ (n), obtains the difference as in the following equation, and calculates the error signal e (n ) Is output.
e (n) = y (n) -y ^ (n) (6)

  The differential signal generation unit 605 obtains and outputs a partial differential value s ^ '(n) of the nonlinear system output estimated value s ^ (n) with respect to the nonlinear parameter a ^ (n) as in the following equation. In the following equation, a nonlinear system output estimated value s ^ (n) may be used instead of the input signal x (n).

  The parameter update unit 606 receives the error signal e (n), and extracts the partial differential value S ^ ′ (n) and the nonlinear system output estimated value S ^ (n) from the signal storage unit 602.

  Here, for the error signal e (n),

And the gradient for the nonlinear parameter a ^ (n) and the linear parameter H ^ (n)

And obtained.
Therefore, the parameter update unit 606 updates the nonlinear parameter a ^ (n) and the linear parameter H ^ (n) according to the following expression, and outputs a new update value at time n.

The step sizes μ _a and μ _H are selected within the ranges of 0 ≦ μ _a ≦ 1 and 0 ≦ μ _H ≦ 1, respectively.

A. Stenger, W. Kellermann, “Nonlinear acoustic echo cancellation with fast converging memoryless preprocessor”, IEEE International Conference on Acoustics, Speech, and Signal Processing, 2000. ICASSP '00, 2000, Vol. 2, II805-II808.

  A transmission system 1 in which a nonlinear system 2 and a linear system 3 are cascade-connected as shown in FIG. 1 is modeled by equations (1) and (2), and their nonlinear parameters a ^ (n) and linear parameters H ^ In the conventional technique in which (n) is sequentially updated according to Equation (11) and Equation (12), respectively, the equation (4) using the linear parameter H ^ (n) is calculated in the calculation of the transmission system output estimate y ^ (n). ) Or a convolution operation in the time domain corresponding to Equation (5) requires a large amount of calculation. It is known that a convolution operation in a time domain with a large amount of calculation can be efficiently executed in a frequency domain. On the other hand, the calculation of the nonlinear system output estimated value s ^ (n) corresponding to the equation (3) using the nonlinear parameter a ^ (n) is more efficient when calculated in the time domain. Therefore, after calculating the nonlinear system output estimated value s ^ (n) in the time domain, the transmission system output estimated value y ^ (n) may be calculated in the frequency domain. However, in this case, in the sequential update of the nonlinear parameter a ^ (n) and the linear parameter H ^ (n), formulas (11) and (12) that are assumed to be executed in the time domain should be applied as they are. I can't.

  The present invention is a parameter that can update the nonlinear parameter and the linear parameter uniformly while keeping the amount of calculation when estimating the output signal s (n) of the nonlinear system 2 and the output signal y (n) of the linear system 3 small. The purpose is to provide estimation technology.

  In order to solve the above problems, according to the first aspect of the present invention, the parameter estimation device includes a nonlinear system that can be regarded as a hard clip characteristic independent of past inputs and outputs, and a linear system that can be regarded as a finite-length impulse response. The threshold of the hard clip characteristic and the impulse response of a finite length of the transmission system that can be regarded as a cascade connection are estimated. The parameter estimation device includes a nonlinear parameter application unit that obtains a nonlinear system output estimated value that is an estimated value of a nonlinear system output signal with respect to a time domain transmission system input signal using a nonlinear parameter that is a threshold estimated value, A differential signal generation unit for obtaining a value corresponding to a partial differential value for a nonlinear parameter of a system output estimated value, and a time domain nonlinear system output estimated value and a value corresponding to a partial differential value are each converted to a frequency domain, Multiplying the frequency domain transform unit to obtain the value corresponding to the nonlinear system output estimate and partial differential value, the linear parameter that is an estimate of the impulse response of finite length in the frequency domain, and the nonlinear system output estimate in the frequency domain , A frequency domain linear parameter multiplier for obtaining a transmission system output estimation value, which is an estimation value of the frequency domain of the transmission system output signal, an output signal of the transmission system, Update the nonlinear parameter using the error calculation unit that calculates the difference from the transmission system estimated value, the frequency domain difference, the value corresponding to the partial differential value, and the nonlinear system output estimated value, and the linear parameter in the frequency domain. A mixed parameter updating unit for updating.

  In order to solve the above problems, according to the second aspect of the present invention, the parameter estimation method includes a nonlinear system that can be regarded as a hard clip characteristic independent of past inputs and outputs, and a linear system that can be regarded as a finite-length impulse response. The threshold of the hard clip characteristic and the impulse response of a finite length of the transmission system that can be regarded as a cascade connection are estimated. The parameter estimation method includes a nonlinear parameter applying step for obtaining a nonlinear system output estimated value that is an estimated value of a nonlinear system output signal with respect to a time domain transmission system input signal using a nonlinear parameter that is a threshold estimated value, A differential signal generation step for obtaining a value corresponding to the partial differential value for the nonlinear parameter of the system output estimated value, and a time domain nonlinear system output estimated value and a value corresponding to the partial differential value are each converted to the frequency domain, and the frequency domain Multiplying the frequency domain transformation step to find the value corresponding to the nonlinear system output estimate and partial differential value of, the linear parameter that is an estimate of the impulse response of finite length in the frequency domain, and the nonlinear system output estimate in the frequency domain , Frequency domain linear parameter multiplication step to determine the transmission system output estimate, which is the frequency domain estimate of the output signal of the transmission system The nonlinear parameter is updated using the error calculation step for obtaining the difference between the transmission system output signal and the transmission system output estimate, the frequency domain difference, the value corresponding to the partial differential value, and the nonlinear system output estimate. And a mixed parameter update step of updating the linear parameter in the frequency domain.

  According to the present invention, the output signal s (n) of the nonlinear system 2 is estimated in the time domain, the output signal y (n) of the linear system 3 is estimated in the frequency domain, and the calculation efficiency is improved. In addition, the linear parameters can be updated uniformly.

The figure for demonstrating a transmission system. The functional block diagram of the parameter estimation apparatus using a prior art. The functional block diagram of the parameter estimation apparatus which concerns on 1st embodiment. The figure which shows the processing flow of the parameter estimation apparatus which concerns on 1st embodiment. The functional block diagram of the parameter estimation apparatus which concerns on the modification of 1st embodiment. The figure which shows the processing flow of the parameter estimation apparatus which concerns on the modification of 1st embodiment. The functional block diagram at the time of diverting the parameter estimation apparatus of 1st embodiment to an echo cancellation apparatus. The functional block diagram at the time of diverting the parameter estimation apparatus of the modification of 1st embodiment to an echo cancellation apparatus.

  Hereinafter, embodiments of the present invention will be described.

<Points of first embodiment>
In this embodiment, in order to efficiently fuse the processing in the nonlinear system with high calculation efficiency in the time domain and the processing in the linear system with high calculation efficiency in the frequency domain, the estimation parameters of the nonlinear system are in the form of the time domain. While maintaining the estimation parameters of the linear system as a form in the frequency domain, we derived an update formula that can update both in a unified manner. The method will be described.

  Formulas (11) and (12), which are estimation parameter update formulas in the time domain, are generalized as the following formulas.

Here, update of nonlinear parameters and linear parameters suitable for performing estimation calculation of output signal s (n) of nonlinear system 2 in the time domain and estimation calculation of output signal y (n) of linear system 3 in the frequency domain. Consider the specific form of the expression. First, with respect to Equation (13) and Equation (14), the step size μ _a = μ _H = 1 is satisfied, and the updated values Δa ^ (n) and ΔH ^ (n) satisfy the input / output relationship for the corresponding time n. If we do this, the following relationship is obtained.

On the other hand, an error signal vector consisting of N error signals e (n)
E (n) = [e (n), e (n-1), ..., e (n-N + 1)] (16)
Is converted to the frequency domain using discrete Fourier transform, discrete cosine transform, filter bank, etc., the relationship corresponding to equation (15) is as follows (at least approximately: Can be considered to hold).

However,

ここで、E(ω_k,i)、H^(ω_k,i)、S^(ω_k,i),S^'(ω_k,i)はk番目の離散周波数ω_kに対応するE(n),H^(n),S^(n),S^’(n)の周波数成分を表し（ただし、k=0,1,…,N-1）、iはブロック番号、式(16)のように、信号をN個の要素によるブロック単位としてまとめる時間間隔（例えば、シフト幅であり、1以上N以下の離散時間に対応する時間間隔）毎に割り当てられる番号であり、Δa^(i)はブロックiにおける非線形パラメータa^(i)の時間領域の更新量に相当し、ΔH^(ω_k,i)はブロックiにおける線形パラメータH^(ω_k,i)の周波数領域における更新量に相当する。

Where E (ω _k , i), H ^ (ω _k , i), S ^ (ω _k , i), S ^ '(ω _k , i) are E corresponding to the _kth discrete frequency ω _k (n), H ^ (n), S ^ (n), S ^ '(n) represents the frequency component (where k = 0,1, ..., N-1), i is the block number, As in (16), it is a number assigned to each time interval (for example, a time interval corresponding to a discrete time of 1 to N) that summarizes the signal as a block unit of N elements, and Δa ^ (i) corresponds to the amount of time domain update of nonlinear parameter a ^ (i) in block i, and ΔH ^ (ω _k , i) is in the frequency domain of linear parameter H ^ (ω _k , i) in block i. It corresponds to the update amount.

  Now, equation (17) is an indefinite equation with a smaller number of equations than unknowns, and there is a degree of freedom in how to give the solution. In this embodiment, the weighted minimum norm solution of Expression (17) is adopted as the update amount of each parameter. For this purpose, a weight matrix W is introduced into Equation (17),

And here

It is. For equation (18), a weighted minimum norm solution is obtained by the following equation. However, ^* represents conjugate transposition.

  However, for equation (19)

If the inverse matrix calculation is performed as it is, a large amount of calculation is required. Therefore, in this embodiment,
Inverse matrix theorem
A = B ^-1 + CD ^-1 C ^*
A ^-1 = B-BC (D + C ^* BC) ^-1 C ^* B (20)
To make the above inverse matrix operation more efficient.
First,

And in equation (20),

ここで、S^S^^*は対角行列であるため、対角要素の逆数を取ることで、簡単に逆行列演算が可能である。式(21)を式(19)に代入すると、次式のように、更新量が得られる。

Here, since S ^ S ^ ^* is a diagonal matrix, the inverse matrix operation can be easily performed by taking the reciprocal of the diagonal element. When equation (21) is substituted into equation (19), the update amount is obtained as in the following equation.

Here, α is a scalar quantity. From this, the update equation of the nonlinear parameter is

It becomes. Note that real (A) is a function that returns only the real part of the complex number A. Since the nonlinear parameter a ^ (i) is a real value, only the real part of the complex number α is used for updating. Also, the linear parameter update formula is independent for each discrete frequency ω _k ,

It becomes. Here, the step size μ _h is 0 ≦ μ _h ≦ 1, and a different value may be given for each discrete frequency ω _k or may be given so as to change with time. Note that the linear parameter may be updated by the following equation instead of the equation (24).

Note that δh is a positive constant for preventing division by zero. Hereinafter, a configuration example of the parameter estimation apparatus will be shown.

<Parameter estimation apparatus according to the first embodiment>
FIG. 3 is a functional block diagram of the parameter estimation apparatus 10 according to the first embodiment, and FIG. 4 is a diagram showing a processing flow thereof.

  The parameter estimation apparatus 10 includes a nonlinear parameter application unit 110, a differential signal generation unit 120, a signal accumulation unit 130, a frequency domain conversion unit 140, a linear parameter multiplication unit 150, a frequency domain conversion unit 160, an error calculation unit 170, and a mixed parameter update. Part 180.

The parameter estimation device 10 receives the input signal x (n) and the output signal y (n) of the transmission system 1, receives the hard clip characteristic threshold a of the nonlinear system 2, and the finite-length impulse response h ₀ of the linear system 3. , h ₁ , ..., h _L , the nonlinear parameter a ^ (i), which is an estimate of the threshold a, and the frequency domain estimate of the finite-length impulse response h ₀ , h ₁ , ..., h _L Output a linear parameter H ^ (ω _k , i). Hereinafter, the processing content of each part is demonstrated.

  In the discrete time n, the nonlinear parameter application unit 110, the differential signal generation unit 120, and the signal storage unit 130 of the parameter estimation device 10 perform the following processing.

<Nonlinear Parameter Application Unit 110>
The nonlinear parameter application unit 110 receives the nonlinear parameter a ^ (i) and the time domain input signal x (n), and uses these values to calculate the time domain by the following equation or an equivalent equation or an approximate equation. A nonlinear system output estimated value s ^ (n), which is an estimated value of the output signal of the nonlinear system 2 with respect to the input signal x (n), is obtained (s1) and output to the differential signal generation unit 120 and the signal storage unit 130.

  Note that a ^ (i) is a value received from the mixed parameter update unit 180 for each block i (for example, for each shift width).

<Differential signal generation unit 120>
The differential signal generation unit 120 uses a value s ^ ′ corresponding to the partial differential value of the nonlinear parameter a ^ (i) of the nonlinear system output estimated value s ^ (n) by, for example, the following expression or an equivalent expression or an approximate expression: (n) is obtained (s3) and output to the signal storage unit 130. As shown in FIG. 3, the nonlinear system output estimated value s ^ (n) is received, and in the following equation, the nonlinear system output estimated value s ^ (n) may be used instead of the input signal x (n). Then, the nonlinear output estimation value s ^ (n) can be partially differentiated with respect to the nonlinear parameter a ^ (i) to obtain the value s ^ '(n).

<Signal Storage Unit 130>
The signal accumulating unit 130 receives the nonlinear system output estimated value s ^ (n) and the value s ^ '(n), and goes back to the past of L samples, and L + 1 nonlinear system output estimated values s ^ (n) , s ^ (n-1), ..., s ^ (nL) and L + 1 values s ^ '(n), s ^' (n-1), ..., s ^ '(nL) Accumulate (s5).

  For each block i (for example, for each shift width), the frequency domain conversion units 140 and 160, the linear parameter multiplication unit 150, the error calculation unit 170, and the mixed parameter update unit 180 of the parameter estimation device 10 perform the following processing.

<Frequency domain converters 140 and 160>
The frequency domain transform unit 140 obtains a nonlinear system output estimated value S ^ (n) = [s ^ (n), s ^ (n-1),... S ^ (nL)] from the signal storage unit 130 and a value S. ^ '(n) = [s ^' (n), s ^ '(n-1),…, s ^' (nL)] are extracted and converted to the frequency domain, respectively, to estimate the nonlinear system output in the frequency domain The value S ^ (ω _k , i) and the value S ^ '(ω _k , i) are obtained (s7), and the nonlinear system output estimated value S ^ (ω _k , i) and the value S ^' (ω _k , i i) is output to the mixed parameter updating unit 180, and the nonlinear system output estimated value S ^ (ω _k , i) is output to the linear parameter multiplying unit 150. Note that as a method for converting to the frequency domain, discrete Fourier transform, discrete cosine transform, filter bank, and the like are conceivable.

Similarly, the frequency domain converter 160 receives the output signal y (n) of the transmission system 1 and outputs L + 1 output signals y (n), y (n-1) of the L + 1 transmission systems 1 to a signal storage unit (not shown). ,..., Y (nL) are accumulated, and for each block i, the output signals y (n), y (n-1),. To obtain the output signal Y (ω _k , i) of the transmission system 1 in the frequency domain (s11) and output it to the error calculator 170.

<Linear parameter multiplication unit 150>
The linear parameter multiplier 150 receives the linear parameter H ^ (ω _k , i) and the frequency domain nonlinear system output estimate S ^ (ω _k , i), and for each discrete frequency ω _k as shown in the following equation. By multiplying these values, a transmission system output estimated value Y ^ (ω _k , i) that is an estimated value of the output signal Y (ω _k , i) of the transmission system 1 in the frequency domain is obtained (s9), and the error calculation unit 170 is calculated. Output to.
Y ^ (ω _k , i) = H ^ (ω _k , i) S ^ (ω _k , i)
Since the calculation amount of this equation is smaller than the calculation amounts of Equation (4) and Equation (5), the parameter estimation device 10 can reduce the calculation amount as compared with the conventional technique. With such a configuration, the output signal s (n) of the nonlinear system 2 can be estimated in the time domain, and the output signal y (n) of the linear system 3 can be estimated in the frequency domain.

<Error calculation unit 170>
The error calculation unit 170 receives the output signal Y (ω _k , i) of the transmission system 1 in the frequency domain and the transmission system output estimated value Y ^ (ω _k , i), and outputs the discrete frequency ω _k as shown in the following equation. The difference is obtained as an error signal E (ω _k , i) every time (s 13) and output to the mixed parameter updating unit 180.
E (ω _k , i) = Y (ω _k , i) -Y ^ (ω _k , i)

<Mixed parameter update unit 180>
The mixed parameter updating unit 180 receives the error signal E (ω _k , i) in the frequency domain, the value S ^ ′ (ω _k , i), and the nonlinear system output estimated value S ^ (ω _k , i). Is used to update the nonlinear parameter a ^ (i), the linear parameter H ^ (ω _k , i) is updated in the frequency domain (s15), and the updated nonlinear parameter a ^ (i + 1) is updated. The updated linear parameter H ^ (ω _k , i + 1) is output to the nonlinear parameter application unit 110 to the linear parameter multiplication unit 150.

  Specifically, it is updated according to equations (23) and (24).

と零除算防止のための正の定数δ（大きくてもH^S^'^*[S^S^^*]^-1H^S^'の平均値を超えない程度）を分母に与えてもよい。このような更新式を用いることで、非線形パラメータを時間領域での形式、線形系の推定パラメータを周波数領域での形式として保持しておきながら、両者を統一的に更新できる。また、重み係数wは正数であり、非線形パラメータの推定に重点を置く場合は、１より小さい値を与え、線形パラメータの推定に重点を置く場合は、１より大きい値を与える。利用者は、非線形系及び線形系の推定パラメータの更新重みを自由に制御することができる。

And a positive constant δ to prevent division by zero (even if it is large, it does not exceed the average value of H ^ S ^ ' ^* [S ^ S ^ ^* ] ^-1 H ^ S ^') may be given to the denominator . By using such an update formula, both can be updated uniformly while maintaining the nonlinear parameters in the time domain format and the linear system estimation parameters in the frequency domain format. Further, the weighting factor w is a positive number, and a value smaller than 1 is given when emphasizing the estimation of the nonlinear parameter, and a value larger than 1 is given when emphasizing the estimation of the linear parameter. The user can freely control the update weights of the estimation parameters of the nonlinear system and the linear system.

  Here, for the weight coefficient w,

And may be given to change over time. Here, N is the frequency division number.

Further, each discrete frequency ω _k may be configured by a (small) FIR filter having a length M (having M coefficients) instead of the scalar value of the linear parameter H ^ (ω _k , i). At this time, the weight coefficient w is

Can also be given. While the norm of the linear parameter H ^ (ω _k , i) is a value around 1, the magnitude of the nonlinear parameter a ^ (i) depends on the amplitude, and the possible range is the linear parameter H ^ (ω _{It is} very large compared to _{k 1} , i). By giving the weighting coefficient w to change with time according to the value of the nonlinear parameter a ^ (i), it is possible to more appropriately set the parameter to be preferentially estimated. In other words, if the value of the nonlinear parameter a ^ (i) is large, the value of the weighting factor w is reduced to focus on the estimation of the nonlinear parameter, and if it is small, the value of the weighting factor w is increased to focus on the estimation of the linear parameter. Put. Such a configuration enables computationally efficient and highly accurate estimation processing.

The parameter estimation device 10 may output the nonlinear parameter a ^ (i + 1) and the linear parameter H ^ (ω _k , i + 1) as estimated values for each block i, or the nonlinear parameter a ^ After the values of (i + 1) and linear parameters H ^ (ω _k , i + 1) have converged, the nonlinear parameters a ^ (i + 1) and linear parameters H ^ (ω _k , i + 1) may be output as an estimated value. For example, it is determined that convergence has occurred in the following case.

(1) When the difference before and after each update of the non-linear parameter a ^ (i) and the linear parameter H ^ (ω _k , i) is less than a certain value (the two parameters until the two differences are less than a certain value) May be updated, or when the difference of either one becomes less than a certain value, only the relevant parameter may be updated)
(2) When the above processes s1 to s15 are repeated a certain number of times (for example, 100 times) (3) When either (1) or (2) is satisfied

Further, the parameter estimation device 10 may convert the linear parameter H ^ (ω _k , i + 1) into the time domain and output it.

<Effect>
According to the parameter estimation device according to the present embodiment, the calculation efficiency can be improved by calculating the nonlinear system output estimation value in the time domain and calculating the linear system output estimation value in the frequency domain. Further, the nonlinear parameters and the linear parameters can be updated uniformly by the equations (22) to (24). In addition, by uniformly handling the error signal in the frequency domain, the nonlinear system output estimated value, and the value corresponding to the partial differential value, an update equation that can freely control the update weights of the estimation parameters of the nonlinear system and the linear system is provided. By deriving, it is possible to perform estimation processing that is computationally efficient and highly accurate.

<Modification>
FIG. 5 is a functional block diagram of a modification of the first embodiment, and FIG. 6 is a diagram showing a processing flow thereof. Only parts different from the first embodiment will be described.

  The parameter estimation device 20 includes a nonlinear parameter application unit 110, a differential signal generation unit 120, a signal accumulation unit 130, a frequency domain conversion unit 140, a linear parameter multiplication unit 150, an error calculation unit 270, a time domain conversion unit 191, and a frequency domain conversion unit. 193 and the mixed parameter update unit 180. That is, the point including the time domain conversion unit 191 and the frequency domain conversion unit 193, the point not including the frequency domain conversion unit 160, and the input / output of the error calculation unit 270 are different from the first embodiment. For each block i, the error calculator 270, time domain converter 191 and frequency domain converter 193 perform the following processing.

<Time domain conversion unit 191>
The time domain conversion unit 191 receives the transmission system output estimated value Y ^ (ω _k , i) from the linear parameter multiplication unit 150, converts it into the time domain, and transmits the time domain transmission system output estimation value Y ^ (n) = [ y ^ (n), y ^ (n-1),..., y ^ (nL)] are obtained (s21) and output to the error calculation unit 270. In addition, as a conversion method to the time domain, a method corresponding to the conversion method of the frequency domain conversion unit 140 may be used.

<Error calculation unit 270>
The error calculation unit 270 outputs the output signal y (n) of the transmission system 1 and the transmission system output estimated value Y ^ (n) = [y ^ (n), y ^ (n-1), ..., y ^ (nL). The difference is obtained as an error signal e (n) at each discrete time n as shown in the following equation (s23), and is output to the frequency domain transform unit 193.
e (n) = y (n) -y ^ (n)
For example, the error calculation unit 270 stores output signals y (n), y (n−1),..., Y (nL) of L + 1 transmission systems 1 in a signal storage unit (not shown) and transmits the signal. Every time system output estimate Y ^ (n) is received, an error signal vector E (n) consisting of L + 1 error signals e (n), e (n-1), ..., e (nL) is obtained. . Also, the output signal y (n) of the transmission system 1 is accumulated for the shift width, and every time the transmission system output estimated value Y ^ (n) is received, the error signal e (n) is obtained for the shift width, An error signal vector E (n) composed of error signals e (n), e (n−1),..., E (nL) for one block together with the error signal obtained before that is sent to the frequency domain transform unit 193. It is good also as a structure to output.

<Frequency domain conversion unit 193>
The frequency domain transform unit 193 receives the error signal vector E (n), transforms it into the frequency domain, obtains the frequency domain error signal E (ω _k , i) (s25), and outputs it to the mixed parameter update unit 180. .

  With such a configuration, the same effect as that of the first embodiment can be obtained.

<Second embodiment>
The parameter estimation device according to the first embodiment and its modification can be diverted to, for example, an echo canceller. 7 and 8 are functional block diagrams when the parameter estimation device of the first embodiment and its modification is diverted to an echo canceller, respectively.

  The echo cancellers 30 and 40 subtract the pseudo echo signal obtained by estimating the echo that circulates into the microphone 302 in the same room and the acoustic signal reproduced from the speaker 202 from the sound pickup signal of the microphone. The signal from which the echo is canceled is output.

  As shown in FIGS. 7 and 8, when a signal is reproduced by the amplifier 201 or the speaker 202 when the reproduction signal x (n) to be reproduced by the speaker 202 is gained by the amplifier 201 and reproduced from the speaker 202. There is. That is, the transfer characteristic until the reproduced signal x (n) to be reproduced by the speaker 202 passes through the amplifier 201 and the speaker 202 that generate distortion and a signal emitted from the speaker 202 is obtained is regarded as nonlinear. It can be assumed that the nonlinear system 2 is configured. In other words, the nonlinear system 2 includes the speaker 202 and the amplifier 201. Further, the acoustic signal emitted from the speaker 202 passes through the indoor reverberation path 301 to the microphone 302, and the transfer characteristic until the sound pickup signal y (n) is obtained is regarded as linear. It can be assumed that it is composed. In other words, the linear system 3 includes an echo path 301 and a microphone 302.

In this case, the parameter estimation device 10 or 20 shown in FIG. 3 or 5 can operate as the echo cancellation device 30 or 40 shown in FIG. 7 or FIG. 8, respectively, and the input signal x (n ) Is read (replaced) as a reproduction signal x (n) for reproduction by the speaker 202, and an output signal y (n) as a sound collection signal y (n) collected by the microphone 302 in the same space as the speaker 202. Therefore, the duplicated explanation is omitted. The echo canceling device 30 or 40 does not need to output the nonlinear parameter a ^ (i + 1) or the linear parameter H ^ (ω _k , i) to the outside, but sends the error signal e (n) to the transmission signal. Output as. For this reason, in the configuration of the echo canceller 30 of FIG. 7, a time domain conversion unit 310 that converts E (ω _k , i) into e (n) is newly provided.

The collected sound signal y (n) includes the utterance voice of the near-end speaker in addition to the echo originating from the transmission system. When the near-end speaker's voice and echo are mixed, the error signal e (n) obtained by subtracting the transmission system output estimate y ^ (n), which is a pseudo echo signal, from y (n) is The voice of the near-end speaker is mainly used as a transmission signal to the communication partner. At the same time, however, the near-end speaker's voice present in the error signal e (n) becomes a factor that disturbs the estimation results of the nonlinear parameters and linear parameters. Therefore, the following measures are necessary separately. If the value of the step size mu _a or mu _h a fixed value gives sufficiently small value. If the presence of the near-end speaker's voice can be detected, control is performed so that the step size μ _a or μ _h is reduced only when the near-end speaker's voice is detected (for example, μ _a = μ _h = 0).

<Other variations>
The present invention is not limited to the above-described embodiments and modifications. For example, the various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. In addition, it can change suitably in the range which does not deviate from the meaning of this invention.

<Program and recording medium>
The parameter estimation device and echo canceling device described above can also be operated by a computer. In this case, each process of a program for causing a computer to function as a target device (a device having the functional configuration shown in the drawings in various embodiments) or a process procedure (shown in each embodiment) is processed by the computer. A program to be executed by the computer may be downloaded from a recording medium such as a CD-ROM, a magnetic disk, or a semiconductor storage device or via a communication line into the computer, and the program may be executed.

Claims (8)

The threshold value of the hard clip characteristic and the impulse response of the finite length of a transmission system that can be regarded as a cascade connection of a nonlinear system that can be regarded as a hard clip characteristic independent of past inputs and outputs and a linear system that can be regarded as a finite impulse response. A parameter estimation device for estimating,
Using a nonlinear parameter that is the estimated value of the threshold, a nonlinear parameter application unit that obtains a nonlinear system output estimated value that is an estimated value of the output signal of the nonlinear system with respect to the input signal of the transmission system in the time domain;
A differential signal generation unit for obtaining a value corresponding to a partial differential value for the nonlinear parameter of the nonlinear system output estimated value;
A frequency domain conversion unit that converts the values corresponding to the nonlinear system output estimated value and the partial differential value in the time domain to the frequency domain, respectively, and calculates values corresponding to the nonlinear system output estimated value and the partial differential value in the frequency domain;
Multiplying the linear parameter that is the estimated value of the finite-length impulse response in the frequency domain and the nonlinear system output estimated value in the frequency domain, and the transmission system output estimated value that is the estimated value in the frequency domain of the output signal of the transmission system A frequency domain linear parameter multiplier for obtaining
An error calculation unit for obtaining a difference between the output signal of the transmission system and the estimated value of the transmission system output;
Using the difference in the frequency domain, a value corresponding to the partial differential value, and the nonlinear system output estimation value, updating the nonlinear parameter, and a mixed parameter updating unit that updates the linear parameter in the frequency domain, Including,
Parameter estimation device. The parameter estimation device according to claim 1,
i is a number assigned for each time interval in which signals in the time domain are grouped as a block unit of N elements, and μ _a and μ _h are in a range of 0 ≦ μ _a ≦ 1 and 0 ≦ μ _h ≦ 1, respectively. The selected step size, real (A) is a function that returns only the real part of the complex number A, k = 0,1, ..., N-1, ω _k is the kth discrete frequency, and a ^ (i ) And H ^ (ω _k , i) are the nonlinear and linear parameters, respectively, and S ^ (ω _k , i), S ^ '(ω _k , i) and E (ω _k , i) are in the frequency domain, respectively. In the non-linear system output estimated value, the value corresponding to the partial differential value and the difference, ^* is a conjugate transpose, the weighting factor w is a positive number,

However,

To update the nonlinear parameter a ^ and the linear parameter H ^,
Parameter estimation device. The parameter estimation device according to claim 2, wherein
When the linear parameter H ^ (ω _k , i) is a scalar value, the weight coefficient w is

And when the linear parameter H ^ (ω _k , i) is composed of an FIR filter of length M, the weight coefficient w is

And
Parameter estimation device. The parameter estimation device according to any one of claims 1 to 3,
A second frequency domain transform unit for converting the output signal of the transmission system in the time domain into the frequency domain, and obtaining an output signal of the transmission system in the frequency domain;
In the error calculation unit, a difference between the output signal of the transmission system in the frequency domain and the estimated value of the transmission system output is obtained.
Parameter estimation device. The parameter estimation device according to any one of claims 1 to 3,
Converting the transmission system output estimation value into the time domain, and obtaining a time domain transmission system output estimation value; and
A third frequency domain converter that converts the time domain error signal to the frequency domain and obtains the frequency domain error signal; and
In the error calculation unit, a difference between the output signal of the transmission system in the time domain and the transmission system output estimated value in the time domain is obtained.
Parameter estimation device. An echo canceller including the parameter estimation device according to any one of claims 1 to 5,
The nonlinear system includes a speaker and an amplifier, the linear system includes an echo path and a microphone, the input signal of the transmission system is a reproduction signal of the speaker, and the output signal of the transmission system is a sound pickup of the microphone The echo canceller outputs a time domain error signal e as a transmission signal,
Echo canceler. The threshold value of the hard clip characteristic and the impulse response of the finite length of a transmission system that can be regarded as a cascade connection of a nonlinear system that can be regarded as a hard clip characteristic independent of past inputs and outputs and a linear system that can be regarded as a finite impulse response. A parameter estimation method for estimating,
Applying a nonlinear parameter to obtain a nonlinear system output estimated value that is an estimated value of the output signal of the nonlinear system with respect to an input signal of the transmission system in the time domain, using a nonlinear parameter that is an estimated value of the threshold;
A differential signal generation step for obtaining a value corresponding to a partial differential value for the nonlinear parameter of the nonlinear system output estimated value;
A frequency domain conversion step of converting the values corresponding to the nonlinear system output estimated value and the partial differential value in the time domain to the frequency domain, respectively, and obtaining values corresponding to the nonlinear system output estimated value and the partial differential value in the frequency domain;
Multiplying the linear parameter that is the estimated value of the finite-length impulse response in the frequency domain and the nonlinear system output estimated value in the frequency domain, and the transmission system output estimated value that is the estimated value in the frequency domain of the output signal of the transmission system A frequency domain linear parameter multiplication step to determine
An error calculating step for obtaining a difference between the output signal of the transmission system and the estimated value of the transmission system output;
Using the difference in the frequency domain, the value corresponding to the partial differential value, and the nonlinear system output estimation value, the nonlinear parameter is updated, and the mixed parameter update step of updating the linear parameter in the frequency domain, and Including,
Parameter estimation method.   The program for functioning a computer as each part of the parameter estimation apparatus in any one of Claims 1-5, or each part of the echo cancellation apparatus of Claim 6.

Images