JPH0865397A - Echo canceler - Google Patents

Abstract

PURPOSE: To simplify a sum of squares calculation, to reduce the calculation amount and to enable processing in real time when a reception signal is normalized, in an echo canceler updating a filter factor by using a learning identifying method. CONSTITUTION: A reception signal is outputted from a speaker 10, and part enters a microphone 11 via an echo path and part is added to a transmission input signal. A filter 12 estimates the impulse response of the echo path, generates a pseudo echo from the reception signal, adds this echo to the transmission input signal in the opposite phase and cancels the echo. A simplified square sum circuit 15 determines Xn =mXn-1 +X<2> n , where square sum Xn-1 one sample before multiplied by is a constant m (0<m<1) is added to the square X<2> n of the present reception signal Xn in this square sum calculation, performs a reciprocal calculation 14 for this expression and normalizes the reception signal Xn .

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an echo canceller, and more particularly, to an acoustic echo canceller for a device such as a video conference terminal that performs a two-way simultaneous call using a speaker and a microphone.

[0002]

2. Description of the Related Art FIG. 2 is a diagram showing a circuit configuration of a general echo canceller. As is well known, a reception signal is output from a speaker 10, but a part of the reception signal passes through an echo path and a microphone 11 is provided. And add to the transmitted input signal.
This is an echo, and what attenuates this echo is an echo canceller. The echo canceller estimates the impulse response of the echo path, adds the pseudo echo generated by convolving the impulse response to the received signal and adds it to the transmission input signal in antiphase to attenuate the echo. The impulse response of the echo path is represented by the transversal filter 12, and the update of the filter coefficient is usually performed by the learning identification method. The coefficient update in this learning identification method follows the following equation.

[0003]

[Equation 1]

However, h _{i (n)} represents the filter coefficient h _i at time n. α is a control parameter called a step gain, x _n is a received signal at time n, and N is the number of filter taps.

A filter for generating a pseudo echo from a received signal is an adaptive filter that simulates an echo path. In a normal echo canceller, this filter coefficient is updated by the learning identification method. Figure 2 shows this situation. The difference between the learning identification method and ordinary LMS is that
The sum of squares calculation circuit 13 is used, and the received signal is normalized by the sum of squares and used. And for this reason,
The sum of products calculation for the sum of squares calculation is required, but the amount of calculation is about the same as the convolution calculation in the filter, which is the main calculation of the echo canceller, and the multiplication of the received signal and the residual error for coefficient updating. Is.

Further, for the normalization, the reciprocal calculation 14 for converting the sum of squares obtained as described above into the reciprocal is necessary. In general, the echo canceller is often realized by a DSP (digital signal processor). DSP
Since a multiplier is provided as hardware, multiplication can be processed at high speed, but since a divider is not provided, it is necessary to perform reciprocal calculation by iterative calculation. As described above, when the DSP is used, since the reciprocal calculation is an iterative operation, this processing amount also becomes an obstacle for the real-time operation.

[0007]

In the echo canceller using the conventional learning identification method, it is necessary to perform product-sum operation and division (reciprocal calculation) for the number of taps to normalize the received signal. If it is executed as it is using, etc., it will require an extremely large amount of calculation, which will be an obstacle to real-time processing.

The present invention has been made in view of the actual situation as described above, and has been made for the purpose of significantly reducing the processing amount as described above and enabling real-time operation with simpler hardware. is there.

[0009]

In order to solve the above problems, the present invention provides (1) an echo canceller for updating a filter coefficient using a learning identification method, wherein X _n = mX
X _n obtained by _n-1 + x ² _n (where 0 <m <1) is regarded as an approximate value of the sum of squares of the received signal sequence at time n,
The received signal x _n at time n is normalized by multiplying it by the reciprocal of X _n , or (2) in the echo canceller that updates the filter coefficient using the learning identification method, an index is used when the received signal is normalized. It is characterized in that the reciprocal calculation is approximately performed by inverting the sign of the part.

[0010]

In the echo canceller which updates the filter coefficient by using the learning identification method, the sum of squares of one sample before, which is multiplied by a constant, is added to the square of the present received signal to obtain an approximate value of the present sum of squares. In the past, the product-sum operation, which has the number of taps of the filter in the past, is made to be once, and real-time processing is enabled. Also, by reversing the sign of the exponent part during normalization of the received signal, the inverse calculation is performed approximately, and in the process of obtaining the inverse for normalization, iterative calculation is not necessary and common Floating point DS
Make it easy to implement with P.

[0011]

FIG. 1 is a diagram showing a circuit configuration for a simple calculation of the sum of squares according to the present invention (claim 1). In FIG. 1, parts similar to those in FIG. The same reference numerals are attached. Therefore, in the present invention, in the calculation of the sum of squares of the approximate sum of squares calculation circuit 15, the sum of squares before one sample (X _n-1 ) multiplied by a constant (m times, where 0 <m <1) is calculated.
Is added to the square (x ² _n ) of the current received signal (x _n ) to obtain an approximate value X _n of the current sum of squares. The sum operation is done only once.

Further, for the normalization, the reciprocal calculator 14 for reciprocal the sum of squares obtained as described above is required. In general, an echo canceller is often realized by a DSP, but since the DSP has a multiplier as hardware, the multiplication can be processed at high speed, but since no divider is provided, the reciprocal calculation is repeated. Need to be done by. For this reason, the reciprocal calculation that makes the sum of squares the reciprocal number is an iterative operation, and this processing amount also becomes an obstacle for real-time operation. According to the present invention (claim 2), a sign obtained by inverting the sign of the exponent part when X is expressed by a power of 2 is used as an approximate value of the reciprocal of X, whereby repetitive calculation is unnecessary. Can be In a normal floating point DSP, the exponent part and the borrow number part are stored and processed independently, so this method can be realized very easily.

[0013]

As is apparent from the above description, the present invention has the following effects. [Effect corresponding to claim 1] In the process of obtaining the sum of squares for normalization, the product-sum operation, which was conventionally required several times for taps, can be performed once, so that the processing amount can be reduced. Real-time operation becomes possible with a simpler configuration. [Effect corresponding to claim 2] In the process of obtaining the reciprocal for normalization, it is possible to reduce the amount of processing because it is not necessary to perform repetitive operations and can be easily realized by a general floating point DSP. Real-time operation is possible with this configuration.

[Brief description of drawings]

FIG. 1 is a diagram for explaining an embodiment of an echo canceller circuit according to the present invention.

FIG. 2 is a diagram for explaining an example of a conventional echo canceller circuit.

[Explanation of symbols]

10 ... Speaker, 11 ... Microphone, 12 ... Filter, 13 ... Square sum calculation circuit, 14 ... Reciprocal calculation circuit, 15
… Simple sum of squares calculation circuit.

Claims (2)

[Claims] 1. An echo canceller that updates a filter coefficient by using a learning identification method regards X _n obtained by the following equation as an approximate value of a sum of squares of a received signal sequence at time n, and multiplies the reciprocal of this X _n. An echo canceller characterized by normalizing the received signal _xn at time n. X _n = mX _n-1 + x ² _n , where 0 <m <1 2. An echo canceller for updating a filter coefficient using a learning identification method, wherein an inverse number calculation is approximately performed by inverting a sign of an exponent part when normalizing a received signal. Canceller.