JPS6252492B2 - - Google Patents

Description

[Detailed description of the invention]

[Technical Field of the Invention] The present invention provides an echo canceller that uses a training signal to identify echo characteristics caused by mismatching of a balanced circuit associated with 2-wire to 4-wire conversion in a preset manner, and cancels the echo on the receiving side. Regarding. [Technical background of the invention and its problems] Figure 1 shows a general configuration example of an echo canceller using a transversal filter with a variable tap coefficient. 1 is a memory that stores the tap coefficient to be applied to the transversal filter 1, 3 is a circuit that determines this tap coefficient, 4 is a hybrid transformer, and 5 is an A/R that samples the echo path input signal and converts it into a digital value. D converter, 6 is an A/D converter that samples the echo signal from the echo path and converts it into a digital value, 7 is the pseudo echo signal of the echo path output from the transversal filter 1 and the A/D converter. A subtracter 8 subtracts the echo signal output from the subtracter 7, and a D/A converter 8 converts the output of the subtracter 7, that is, the residual signal after canceling the echo signal, into an analog value and outputs the analog value. The operation of this echo canceller will be briefly explained. The sample value of the echo path input signal x _(t) at time kT is x _k , and the echo signal e _(t) from the echo path is
Let e _k be the sample value at time kT. Similarly, here all signals sampled every T time are considered. Now, if the impulse response of the echo path is h _i , the echo signal e _k is It can be expressed as On the other hand, N tap transversal
The pseudo echo signal e _k ' which is the output of filter 1 is expressed as follows, assuming that the tap coefficients are C ₀ , C ₁ , ..., C _N-1. becomes. Therefore, if the tap coefficients are C ₀ = h ₀ , C ₁
= h ₁ ,..., C _N-1 = h _N-1 , the residual signal r _k of the output of the subtractor 7 is It can be seen that the echo signal is canceled.
However, it is assumed that N is selected to be sufficiently large so that h _i ≈0 (iN). From the above, it can be seen that in the echo canceller, the tap coefficient of the transversal filter 1 should be made equal to the impulse response of the echo path. The simplest way to find the impulse response of a reverberant path is to input an impulse into the reverberant path and observe the response from the reverberant path.
However, since echo signals usually include line noise, one-time observation is insufficient. In order to accurately determine the impulse response, it is necessary to repeatedly input impulses into the echo path and average the observed results, but this poses the problem of increasing the time required to determine the tap coefficients. Another method for determining tap coefficients is the steepest descent method. This uses the echo path input signal x _k and the residual signal r _k after canceling the echo signal with an echo canceller, and calculates the tap coefficient as C ^k _i =C ^k-1 _i +αr _k x _ki …(4 ). however,
C ^k _i is the tap coefficient C _i after the kth correction is completed.
shows. Although this method does not require a special training signal sequence and therefore follows adaptively even when the echo path characteristics change, it takes a long time for the initial convergence of the tap coefficients when the line is connected. It has also been proposed to use a Kalman filter to speed up the convergence speed of tap coefficients (Itakura, Nishikawa, “Echo canceller algorithm using Kalman filter and its simplification”, Journal of the Institute of Electronics and Communication Engineers, Vel J62). -A No.1 P50). However, this method is not practical because the Kalman filter algorithm requires a large amount of calculation. [Object of the Invention] An object of the present invention is to provide an echo canceller that can perform initial training of tap coefficients in a short time without performing complicated arithmetic operations. [Summary of the invention] The outline of this invention will be explained using FIG. 2. In Figure 2, h is N-1 from the 0 point of the echo path.
This is a vector representation of the impulse response up to the point in time, and is h=(h ₀ , h ₁ , h ₂ , . . . h _N-1 ) ^t (t represents transposition). Further, C is a vector representation of the tap coefficient of the transpersal filter in the echo canceller, and is C=(C ₀ , C ₁ , C ₂ , . . . C _N-1 ) ^t . n _k is a signal component that cannot be removed, and
It consists of an echo signal generated by the impulse response after time N on the echo path and line noise. v ₀ , v ₁ , v ₂ , . . . v _k-1 are training signal sequences of sequence length K. In order to simplify the subsequent mathematical expression, a vector V _k =(v _k , v _k-1 , v _k-2 , . . . v _k-N+1 ) ^t in which the training signal sequence is written and arranged is defined. The echo canceller according to the present invention is Least
It is based on the Squares method. The Least Squares method is the sum of the residual signal power. C is determined so as to minimize . Formula (5)
If we partially differentiate with respect to C and set it as 0, we get Therefore, the tap coefficient that minimizes P is however, It can be found by To find tap coefficient C using equation (7)
We must calculate an N×N-order inverse matrix called V ^-1 . However, since the amount of calculation required for this increases in proportion to N ² , the achievable value of N is extremely limited. Therefore, the usual Least Squares method does not calculate the above equation directly, but is modified to calculate it sequentially. However, when training tap coefficients using a constant training signal sequence of v ₀ , v ₁ , ..., v _K-1, V ^-1 becomes a constant matrix, so this V ^-1 can be calculated in advance and used as, for example, If you store it in memory, you can simply refer to the contents of this memory during training. This makes it possible to significantly reduce the amount of calculation. The present invention is based on such a principle. In order to show that tap coefficients can be found correctly using this training method, let us first consider ignoring the noise n _k in the echo path. Assuming that the impulse response h of the echo path does not change during training, then from e _k = h ^t V _k That is, if the training signal sequences v ₀ , v ₁ , ..., v _k-1 are selected so that V ^-1 exists (for this, they must be at least K2N-1), then by the above training method, It can be seen that the tap coefficient can be calculated correctly. Next, after the training is completed using this training method, the tap coefficient of the echo canceller is fixed, and the residual signal power when the echo canceler is used to cancel the echo by the audio signal sequence {x _k } is calculated. think of. The expected value of the residual signal power in this case is ε=σ ² {1+1/K-N+1 Trace(RV ^-1 )}...(10
) becomes. However, it is assumed that the voice signal sequence {x _k } and the line noise sequence {n _k } are uncorrelated. The audio signal sequence {x _k } is a stationary sequence, and R is its covariance matrix Further, {n _k } is a white noise sequence, and σ ² is its power. Here, assuming that Trace (RV ^-1 ) takes a certain value (≫1) that is independent of the training signal sequence length K, then from equation (10) when K is relatively short, the residual signal power ε is K It can be seen that it decreases in inverse proportion to -N+1. Also, if K→∞, C→h, ε→σ ²
Therefore, it can be seen that the optimal tap coefficient can be obtained by the above training method. by the way Training signal sequences v ₀ , v ₁ ,...v _k-
If ₁ is selected, equation (10) becomes as follows. ² is the average power of the audio signal sequence {x _k }. {x _k }
Since it is assumed that is stationary, the suffix k has been omitted. If the average power ^δ2 of the training signal sequence is equal to the average power ² of the audio signal sequence, ε=σ ² (1+N/K-N+1)...(13) According to this, the residual signal power ε can be reduced to 2σ ^2. The training signal sequence length required to reduce or 2N−
1, which is extremely short. In fact, equation (13) is exactly the same as the relationship between the number of tap corrections and the residual signal power when a Kalman filter is used, and the training time required by the above training method is extremely short. Furthermore, it can be seen from equation (13) that the residual signal power ε does not depend on the covariance matrix R of the audio signal sequence.
This is a very convenient property when determining the length K of the training signal sequence. That is, by specifying the ratio λ=ε/σ ² of ε and σ ² , K can be easily determined using equation (13). Furthermore, if V is made a diagonal matrix as shown in equation (11), the number of multiplications can be further reduced, and the elements of V ^-1 can also be stored in one word of memory, or as described later. It is only necessary to hold it as part of the coefficients of the multiplier. Therefore, it is desirable that the matrix V be a diagonal matrix with equal diagonal elements as shown in equation (11). Next, the training signal sequences v ₀ , v ₁ ,...v _k-1
Let me explain how this occurs. For example, as a training signal sequence,

〔Effect of the invention〕

According to the present invention, a digital echo canceler using a tap coefficient transversal filter does not require complicated arithmetic operations.
The Least Squares method allows initial training of tap coefficients to be performed within a short time. Furthermore, after initial training, it is easy to adaptively modify the tap coefficients using the steepest descent method. [Embodiment of the Invention] FIG. 5 shows the configuration of an echo canceller according to an embodiment of the invention. In the figure,
The switch 9 is for switching the echo path input signal to either the audio signal or the training signal.
Normally, this switch 9 is set to the audio signal side, and is turned to the training signal side only when training the tap coefficients. The training signal generator 10 includes a memory 11 for storing a training sequence and a D/A converter 12 for converting it into an analog signal. Transversal filter 1 is shift register 1
4, a multiplier 15, and a total adder circuit 16, and the shift register 14 receives the echo path input signal x.
(t) is stored in the form of a digital value via the A/D converter 5. The capacity of this shift register 14 is assumed to be N. The switch 13 is operated so that the latest N samples are always stored in the shift register 14. The shift register 17 is a tap coefficient memory and stores tap coefficients denoted C ₀ , C ₁ , . . . , C _N-1 . During each sample period T, the shift register 14 and the shift register 17 make one round, and during that time, the multiplier 15 and the total adder circuit 16 generate a pseudo echo signal. is required. Echo signal e(t) from the echo path
is converted into a digital value e _k by the A/D converter 6, and subtracted from the pseudo echo signal e _k ' by the subtracter 7. The residual signal r _k output from the subtracter 7 is converted into an analog signal by the D/A converter 8 and output. The tap coefficient determining means includes a switch 18, first and second multipliers 19 and 20, an adder 21, and a memory 22.
and a matrix calculation circuit 23. That is, during training, the switch 18 is set to e.
It is turned to _{the k} side and e _k is applied to the first multiplier 19 . Multiplier 19 multiplies e _k and N samples v _k , which are in shift register 14 during sample period T.
V _k-1 , ..., v _k-N+1 are sequentially multiplied. This result is multiplied by the coefficient 1/K-N+1 by the second multiplier 20, and further multiplied by the shift register 1 by the adder 21.
7 and stored in the shift register 17 again. With this operation, the shift register 17 will be However, each element of the vector V _k =(v _k , v _k-1 , . . . , v _k-N+1 ) ^t is stored. memory 22
teeth, Each element of the inverse matrix V ^-1 is memorized. At the end of the training, in the matrix calculation circuit 23, C from the shift register 17 is multiplied by V ^-1 from the memory 22, and the multiplication result V ^-1 ·C is stored in the shift register 17 again and becomes the final tap coefficient. is found. After training, the switch 18 is turned off or connected to the r _k side. If the switch 18 is turned off, no further modification of the tap coefficient will be performed. Furthermore, if the switch 18 is connected to the r _k side, the tap coefficient will be adaptively corrected by the steepest descent method. FIG. 6 shows another embodiment of the present invention, in which the matrix V is a diagonal matrix with equal diagonal elements. In this embodiment, most of the configurations are the same as in FIG. 5. Therefore, in FIG. 6, the same parts as those in FIG. 5 are designated by the same reference numerals, and detailed explanation thereof will be omitted. 6 differs from FIG. 5 in that the memory 22 and matrix calculation circuit 23 are omitted. Since the matrix V is a diagonal matrix with equal diagonal elements, these circuits are unnecessary. In this case, matrix V
The information of 1/δ ² , which is an element of the inverse matrix V ^-1 of
is held as part of the coefficient 1/(K-N+1) ^δ2 of the multiplier 20. Further, as the training signal sequence in this case, for example, the chirp sequences u ₀ , u ₁ , . . . u _k-1 described above may be used.

[Brief explanation of the drawing]

Figure 1 shows a transversal system with variable tap coefficients.
A diagram showing a general configuration example of a digital echo canceller using a filter, Figure 2 is a mathematical model diagram to explain the function of the echo canceller, and Figures 3a and b are amplitude characteristics of a chirp signal sequence. 4a and 4b are schematic diagrams showing chirp signal sequence patterns, FIG. 5 is a configuration diagram of a digital echo canceller according to an embodiment of the present invention, and FIG. 6 is a diagram showing the chirp signal sequence pattern. FIG. 3 is a configuration diagram of a digital echo canceller according to another embodiment of the present invention. 1... Transversal filter, 2... Tap coefficient memory, 3... Tap coefficient determination circuit, 4
...Hybrid transformer, 5...A/D converter, 6...A/D converter, 7...Subtractor, 8...
D/A converter, 9... switch, 10... training sequence generator, 11... memory, 12...
D/A converter, 13...switch, 14...shift register, 15...multiplier, 16...total addition circuit, 17...shift register (tap coefficient memory), 18...switch, 19...th 1 multiplier, 20... second multiplier, 21... adder, 2
2...Memory, 23...Matrix calculation circuit.

Claims (1)

[Scope of Claims] 1. A transversal filter with variable tap coefficients that creates a pseudo echo signal from an echo path input signal, a tap coefficient memory that stores values of tap coefficients to be given to the transversal filter, and an echo path. a subtracter for subtracting the echo signal from the transversal filter and the pseudo echo signal output from the transversal filter; and a tap coefficient for reducing the echo signal component included in the residual signal output from the subtracter. tap coefficient determining means for determining the value of the tap coefficient in memory; means for generating a training signal sequence having substantially non-zero signal values throughout the training interval; In the echo canceller, the tap coefficient determining means converts sample values of the training signal sequence into v ₀ , v ₁ , v ₂ ,
…When v _k-1 , the matrix (However, v _k = (v _k , v _k-1 , v _k-2 , ... v _k-N+
1 ) ^t , K is the length of the training signal sequence, N is the tap length of the transversal filter, and t
is the transpose) ⁾ , and when e ₀ , e ₁ , e ₂ , ..., e _K are the sample values of the echo signal during training, (However, C=(C ₀ , C ₁ , C ₂ , . . . , C _N-1 ) ^t ). 2. The tap coefficient determining means uses e during training.
A first multiplier that sequentially multiplies k by _{v k ,} v _k-1 , v _k-2 , ..., v _k-N+1 , and a coefficient 1/K- for the output of this first multiplier. a second multiplier for multiplying by N+1; an adder for adding the output of the second multiplier and the output of the tap coefficient memory; and storing the addition result in the tap coefficient memory; and each element of V ^-1 . A memory that stores and holds information in advance, and a matrix calculation circuit that multiplies the output of this memory by the output of the tap coefficient memory at the end of training, and stores the multiplication result in the tap coefficient memory. An echo canceller according to claim 1. 3. The echo canceller according to claim 1, wherein the training signal sequence is selected such that the matrix V is a diagonal matrix with equal diagonal elements. 4. The training signal sequence is selected such that the matrix V is a diagonal matrix with equal diagonal elements, and the tap coefficient determining means determines e _k and v _k , v _k- during training.
₁ , v _k-2 , ..., v _k-N+1
When the elements of the inverse matrix V ^-1 are 1/δ ² , the output of the first multiplier is 1/(K-N+1)δ ²
and an adder that adds the output of the second multiplier and the output of the tap coefficient memory and stores the addition result in the tap coefficient memory. An echo canceller according to claim 1.