JP2004266726A - Echo canceller, method for calculating filter coefficient for echo canceling processing, and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an echo canceller capable of more quickening adaptation of a filter coefficient for an initial lock period while suppressing cost increase. <P>SOLUTION: The echo canceller performs an arithmetic operation of the filter coefficient ω(n) of an adaptive filter by selecting a method of increasing an adaptation speed up to the initial lock period and performs the arithmetic operation by selecting a method of improving adaptation accuracy after a lapse of the initial lock period. Concretely, the echo canceller replaces a learning coefficient μ being an arithmetic parameter of the filter coefficient ω(n) with a prescribed value f for the initial lock period (steps S1 to S4) and carries out the arithmetic operation on the basis of Equation (3) after a lapse of the initial lock period (steps S5 to S8). <P>COPYRIGHT: (C)2004,JPO&NCIPI

Description

但し、ｎは時刻，ｘ（ｎ）はスピーカより出力される音声信号（Ａ／Ｄ変換される前のデジタル信号），ｙ（ｎ）はマイクより入力された音声信号（ＩＩＲフィルタ２の出力信号），ｅ（ｎ）はエコー除去信号（適応フィルタ３より出力されるエコー信号成分除去後の信号），αはリーキー積分の忘却係数，βは残留係数，μは学習係数である。そして、エコー除去信号ｅ（ｎ）は（２）式で演算される。
ｅ（ｎ）＝ｙ（ｎ）−ｘ（ｎ）^Ｔ・ｗ（ｎ） …（２）
【０００８】
また、学習係数μは、パワー計算部５の出力結果に基づき、学習係数計算部４において（３）式で演算される。
μ＝Ｍ／（１＋ＥＰＷＲ／（ＸＰＷＲ＋１）） ・・・（３）
ここで、Ｍは係数、ＥＰＷＲはエコー除去信号ｅ（ｎ）の電力値であり、ＸＰＷＲは出力音声信号ｘ（ｎ）の電力値である。尚、斯様な構成の従来技術は、例えば特許文献１に開示されている。
【０００９】
【特許文献１】
特開２００１−１４８６４４
【００１０】
【発明が解決しようとする課題】
ところで、以上の処理については、以下のような問題があった。即ち、ダブルトーク状態が発生すると、基本的に学習は正常に行なわれなくなるため、学習自体を停止させるか、或いは学習係数μを極力小さい値に設定することが好ましい。そして、ダブルトーク状態では（３）式の右辺における電力値ＥＰＷＲが急激に大きくなるので、学習係数μは小さい値になるように演算される。
【００１１】
しかしながら、例えば、装置に対して電源を投入した時のようなエコーキャンセル処理の開始時である所謂初期引込み期間においては、各演算を行なうためのフィルタ係数などのバッファ値は０クリアされた状態にあることから、その値に基づいて得られるエコー除去信号の電力値ＥＰＷＲも大きい値になる。すると、学習係数μは小さい値になることから、図９に示すように、フィルタ係数ωがエコーパスの伝達関数に適応した目標値に到達して収束するまでに極めて長い時間を要してしまう。
【００１２】
このように、フィルタ係数ωの収束が遅くなるということは、エコーパスの伝達関数に適応するのが遅くなることであるから、初期引込み期間においてはエコー成分が除去しきれず、音声認識処理や通話等が困難となる不具合がある。
【００１３】
斯様な問題を解決するためには、例えば、フィルタ係数をバックアップしておくことも考えられるが、そのためのメモリが必要となるため、システムコストが上昇することになる。
【００１４】
本発明は上記事情に鑑みてなされたものであり、その目的は、コストアップを抑えて、初期引込み期間におけるフィルタ係数の適応をより速くすることができるエコーキャンセラ装置、及びエコーキャンセル処理におけるフィルタ係数計算方法、並びに、エコーキャンセル処理機能を実現するためのプロセッサによって実行されるコンピュータプログラムを提供することにある。
【００１５】
【課題を解決するための手段】
請求項１記載のエコーキャンセラ装置によれば、適応フィルタのフィルタ係数の演算を、初期引込み期間までは適応速度を向上させる方式で実行し、初期引込み期間の経過後は、適応精度を向上させる方式で実行するように切替える。即ち、各処理期間における演算方式を適切に切替えるだけであるから、フィルタ係数を記憶させるメモリを用意する必要がない。そして、初期引込み期間ではフィルタ係数を高速に適応させてエコーキャンセル処理を実効あるものとすることができる。
【００１６】
請求項２記載のエコーキャンセラ装置によれば、フィルタ係数の演算パラメータである学習係数を、エコー除去信号の電力値が大きな値となる初期引込み期間においては後述の演算結果よりも大となる一定値に置き換え、初期引込み期間の経過後は、エコー除去信号と送信信号との電力比の逆数に基づいて演算する。斯様に構成すれば、初期引込み期間における学習係数は、前記演算の結果よりも大きな値となるので、フィルタ係数を、エコーパスの伝達関数に対して高速に適応させることができる。
【００１７】
請求項３記載のエコーキャンセラ装置によれば、初期引込み期間においては学習係数を前記演算の式に前記電力比の逆数を乗じて求め、初期引込み期間の経過後は、前記演算により求める。斯様に構成すれば、初期引込み期間におけるフィルタ係数の演算に電力比が及ぼす影響（即ち、フィルタ係数の適応速度の低下）を打ち消すことができ、フィルタ係数を、エコーパスの伝達関数に対して高速に適応させることができる。
【００１８】
請求項４記載のエコーキャンセラ装置によれば、初期引込み期間においてダブルトーク状態が発生している場合は、フィルタ係数の演算を、適応精度を向上させる方式で実行する。即ち、初期引込み期間であっても、同時にダブルトーク状態が発生している場合は、フィルタ係数の収束速度を低下させて適応精度を優先しなければ適切な学習を行なうことはできない。従って、請求項４のように構成すれば適切な学習が可能となる。
【００１９】
【発明の実施の形態】
（第１実施例）
以下、本発明の第１実施例について図１乃至図４を参照して説明する。尚、図８と同一部分には同一符号を付して説明を省略し、以下異なる部分についてのみ説明する。図３は図８相当図であるが、学習係数計算部４が学習係数計算部１１に置き換わっているだけであり、その他の構成は図７と同様である。
【００２０】
図４は、エコーキャンセラ装置１２の全体構成図である。エコーキャンセラ装置１２は、マイク１３より入力される音声信号をＡ／Ｄ変換するＡ／Ｄ変換部１４，上述した適応フィルタ３，学習係数計算部１１，パワー計算部５に対応する処理をプログラム（コンピュータプログラム）１５に基づいて実行するＤＳＰ（Ｄｉｇｉｔａｌ Ｓｉｇｎａｌ Ｐｒｏｃｅｓｓｏｒ，プロセッサ）１６，スピーカ１７によって外部に出力される音声信号をＤ／Ａ変換するためのＤ／Ａ変換部１８等を備えている。
【００２１】
そして、エコーキャンセラ装置１２は、エコーキャンセル処理した音声信号を例えばカーナビゲーション装置や携帯電話機等に転送し、また、それらによって与えられる音声信号を受けてスピーカ１７により外部に出力するようになっている。
【００２２】
次に、本実施例の作用について図１及び図２をも参照して説明する。図１は、ＤＳＰ１５によって実行されるエコーキャンセル処理のうち、主に学習係数計算部１１と、適応フィルタ３に関する処理内容を示すフローチャートである。例えば、エコーキャンセラ装置１２に電源が投入されて起動すると（スタート）、学習係数計算部１１は、学習係数μの初期値ｆを設定する（ステップＳ１）。
【００２３】
ここでの初期値ｆは、エコーキャンセル処理の演算に使用される全てのパラメータがゼロクリアされている状態で得られるエコー除去信号ｅ（ｎ）の電力値ＥＰＷＲと、出力音声信号ｘ（ｎ）の電力値ＸＰＷＲとに基づいて、（３）式で演算される学習係数μよりも十分大きな値を選択する。
【００２４】
それから、適応フィルタ３は、フィルタ係数ω（ｎ）を演算すると（ステップＳ２）、その演算結果ω（ｎ）と前回の演算結果ω（ｎ−１）との差の絶対値を求め、その差の絶対値が所定値ｄ未満か否かを判断する（ステップＳ３）。所定値ｄは、フィルタ係数ω（ｎ）が略適応状態となることで、演算結果の差が十分小さくなったものと判断するのに適切な値とする。
【００２５】
ステップＳ３で「ＮＯ」と判断すると、適応フィルタ３は、フィルタ係数ω（ｎ）をω（ｎ−１）に代入してから（ステップＳ４）ステップＳ２に戻る。従って、ステップＳ３で「ＹＥＳ」と判断するまでは、ステップＳ２〜Ｓ４のループを回り続ける。
【００２６】
以上のループを回っている期間（この期間を、初期引込み期間とする）では、学習係数μが比較的大なる初期値ｆに設定されていることにより、フィルタ係数ω（ｎ）の値は図２に示すように急激に上昇する。また、フィルタ係数ω（ｎ）がエコーパスの伝達関数に適応する目標係数に近付くと、その目標係数を中心として比較的大きな振れ幅で振動する。
【００２７】
そして、以上のループを回っている間にフィルタ係数ω（ｎ）が目標係数付近に収束することで、ステップＳ３において「ＹＥＳ」と判断すると、適応フィルタ３は、ステップＳ４と同様にフィルタ係数ω（ｎ）をω（ｎ−１）に代入する（ステップＳ５）。それから、学習係数計算部１１は、学習係数μを（３）式によって従来と同様に演算する（ステップＳ６）。
【００２８】
また、適応フィルタ３は、ステップＳ６によって得られた学習係数μに基づいてフィルタ係数ω（ｎ）を演算する（ステップＳ７）。それから、エコーキャンセル処理が終了か否かを判断し（ステップＳ８）、終了しなければ（「ＮＯ」）ステップＳ５に戻って処理を継続する。
【００２９】
以上のように本実施例によれば、エコーキャンセラ装置１２は、適応フィルタ３のフィルタ係数ω（ｎ）の演算を、初期引込み期間までは適応速度を向上させる方式で実行し、初期引込み期間の経過後は、適応精度を向上させる方式で実行するように切替える。具体的には、フィルタ係数ω（ｎ）の演算パラメータである学習係数μを、初期引込み期間においては一定値ｆに置き換え、初期引込み期間の経過後は、（３）式に基づいて演算するようにした。
【００３０】
従って、各処理期間における演算方式を適切に切替えるだけであるから、フィルタ係数ω（ｎ）を記憶させるメモリを用意する必要がない。そして、初期引込み期間ではフィルタ係数ω（ｎ）を高速に適応させてエコーキャンセル処理を実効あるものとすることができ、その後は、従来と同様の適応精度を確保することが可能となる。
【００３１】
尚、本実施例の手法によれば、初期引き込み期間におけるフィルタ係数ω（ｎ）の振動振幅は比較的大きくなるが、目標係数により速く収束させることで、当該期間におけるエコーキャンセル機能は従来に比較してより有効に作用することになる。
【００３２】
（第２実施例）
図５は本発明の第２実施例を示すものであり、第１実施例と同一部分には同一符号を付して説明を省略し、以下異なる部分についてのみ説明する。第２実施例では、図３に示す構成の学習係数計算部１１による処理内容が第１実施例とは若干異なるだけである。
【００３３】
図５に示すフローチャートでは、ステップＳ１に代わってステップＳ９が配置されている。そのステップＳ９においては、学習係数μを（４）式によって演算する。
【００３４】

即ち、（４）式は、（３）式に電力比ＥＰＷＲ／ＸＰＷＲを乗じたものである。そして、ステップＳ４を実行するとステップＳ９に戻るようになっている。その他の処理は第１実施例と同様である。
【００３５】
以上のように第２実施例によれば、学習係数計算部１１は、初期引込み期間においては学習係数μを（３）式に電力比ＥＰＷＲ／ＸＰＷＲを乗じて求め、初期引込み期間の経過後は、学習係数μを（３）式によって求めるので、初期引込み期間におけるフィルタ係数ω（ｎ）の演算に電力比ＥＰＷＲ／ＸＰＷＲが及ぼす影響（即ち、適応速度の低下）を打ち消すことができ、フィルタ係数ω（ｎ）を、エコーパスの伝達関数に対して高速に適応させることができる。
【００３６】
（第３実施例）
図６及び図７は本発明の第３実施例を示すものであり、第１実施例と同一部分には同一符号を付して説明を省略し、以下異なる部分についてのみ説明する。第２実施例では、図３に示す構成の学習係数計算部１１を学習係数計算部１９に置き換え、ダブルトークディテクタ（ダブルトーク検出手段）１９を加えた構成である。
【００３７】
ダブルトークディテクタ２０は、パワー計算部５より得られるエコー除去信号と送信信号との電力比ＥＰＷＲ／ＸＰＷＲが所定値を超えた場合にダブルトーク状態の発生を検出し、検出信号ＤＴを学習係数計算部１９に出力するように構成されている。
【００３８】
次に、第３実施例の作用について図７をも参照して説明する。図７に示すフローチャートにおいて、学習係数計算部１９はステップＳ４の実行後にステップＳ１０に移行して、ダブルトークディテクタ２０が検出信号ＤＴを出力しているか否かを判断する。そして、検出信号ＤＴを出力していなければ（「ＮＯ」）ステップＳ２に戻り、ダブルトーク状態が検出され、検出信号ＤＴが出力されていれば（「ＹＥＳ」）ステップＳ５に移行する。
【００３９】
即ち、ステップＳ２〜Ｓ４のループを回っている初期引込み期間にある場合でも、ダブルトーク状態が検出された時はステップＳ５に移行して、学習係数μを（３）式によって演算する。ダブルトーク状態が発生した場合は、フィルタ係数ω（ｎ）の収束速度を低下させて適応精度を優先しなければ適切な学習を行なうことができないからである。
【００４０】
以上のように第３実施例によれば、学習係数計算部１９は、初期引込み期間においてダブルトーク状態が発生している場合は、フィルタ係数ω（ｎ）の演算を、（３）式に基づき適応精度を向上させる方式で実行するので、状況に応じて適切な学習を行なうことができる。
【００４１】
本発明は上記し且つ図面に記載した実施例にのみ限定されるものではなく、以下のような変形または拡張が可能である。
第２実施例において、ダブルトークディテクタ２０の機能を、学習係数計算部１９の内部に持たせても良い。
【図面の簡単な説明】
【図１】本発明の第１実施例であり、エコーキャンセル処理のうち、主に学習係数計算部と、適応フィルタに関する処理内容を示すフローチャート
【図２】エコーキャンセル処理の開始からの時間経過に伴うフィルタ係数ωの変化を示す図
【図３】エコーキャンセラの概略的な構成を示す機能ブロック図
【図４】エコーキャンセラ装置の全体構成図
【図５】本発明の第２実施例を示す図１相当図
【図６】本発明の第３実施例を示す図３相当図
【図７】図１相当図
【図８】従来技術を示す図３相当図
【図９】図２相当図
【符号の説明】
３は適応フィルタ、１１は学習係数計算部、１２はエコーキャンセラ装置、１５はプログラム（コンピュータプログラム）、１６はＤＳＰ、１９は学習係数計算部、２０はダブルトークディテクタ（ダブルトーク検出手段）を示す。[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention realizes an echo canceller device that performs echo cancellation processing by an adaptive filter using an NLMS algorithm based on a digitized output acoustic signal, a filter coefficient calculation method in the echo cancellation processing, and an echo cancellation processing function. And a computer program executed by a processor for executing the program.
[0002]
[Prior art]
For example, many car navigation devices (car navigation systems) have a function of outputting the contents of navigation instructions by voice. Some car navigation systems have a function of performing destination recognition, calling a map screen, and the like by performing voice recognition processing on a voice generated by a user. Further, in order to perform such processing, a configuration in which a car navigation system is combined with a hands-free type portable telephone device and the above-described voice processing is performed together with the telephone function may be adopted.
[0003]
In such a configuration, since the speaker as the transmitting unit and the microphone as the receiving unit of the mobile phone device are located close to each other in an open state, the audio signal output from the speaker goes around the microphone (echo path). It is assumed that When an audio signal wraps around in this way, the signal is positively amplified and feedback occurs, or the received voice is transmitted again to the other party to generate an echo, so that the echo signal component can be removed by the echo canceller function. Is being done.
[0004]
The echo canceller estimates the acoustic characteristics of the echo path and generates a pseudo echo signal based on the audio signal to be output from the speaker. Then, the echo signal component is removed by subtracting the generated pseudo echo signal from the audio signal input to the microphone.
[0005]
As one of the methods for estimating the acoustic characteristics of the echo path, there is a method using an adaptive filter based on an NLMS (Normalized Least Mean Square) algorithm. FIG. 8 is a functional block diagram showing a schematic configuration of an echo canceller using this adaptive filter. The echo canceller 1 includes an IIR (Infinite Impulse Response) filter 2, an adaptive filter 3, a learning coefficient calculator 4, a power calculator 5, a high-pass filter (HPF) 6, and the like.
[0006]
The IIR filter 2 arranged on the input side limits the band of the input signal and removes the offset component. The high-pass filter 6 disposed on the output side removes an offset component included in the output signal.
[0007]
The filter coefficient ω for each tap in the adaptive filter 3 is calculated by equation (1).

Here, n is time, x (n) is an audio signal output from a speaker (digital signal before A / D conversion), and y (n) is an audio signal input from a microphone (output signal of IIR filter 2). ), E (n) are echo removal signals (signals after removal of echo signal components output from the adaptive filter 3), α is a forgetting coefficient of leaky integration, β is a residual coefficient, and μ is a learning coefficient. Then, the echo removal signal e (n) is calculated by equation (2).
e (n) = y (n) −x (n) ^T · w (n) (2)
[0008]
Further, the learning coefficient μ is calculated by the learning coefficient calculating section 4 according to the equation (3) based on the output result of the power calculating section 5.
μ = M / (1 + EPWR / (XPWR + 1)) (3)
Here, M is a coefficient, EPWR is a power value of the echo removal signal e (n), and XPWR is a power value of the output audio signal x (n). Note that a conventional technique having such a configuration is disclosed in, for example, Patent Document 1.
[0009]
[Patent Document 1]
JP-A-2001-148644
[0010]
[Problems to be solved by the invention]
By the way, the above processing has the following problems. That is, when the double talk state occurs, learning is basically not normally performed. Therefore, it is preferable to stop the learning itself or set the learning coefficient μ to a value as small as possible. Then, in the double talk state, the power value EPWR on the right side of the equation (3) sharply increases, so that the learning coefficient μ is calculated to be a small value.
[0011]
However, for example, in a so-called initial pull-in period at the start of the echo canceling process such as when the power is turned on to the apparatus, buffer values such as filter coefficients for performing each operation are cleared to 0. Therefore, the power value EPWR of the echo removal signal obtained based on the value also becomes a large value. Then, since the learning coefficient μ becomes a small value, as shown in FIG. 9, it takes an extremely long time for the filter coefficient ω to reach and converge to a target value adapted to the transfer function of the echo path.
[0012]
Thus, the slower convergence of the filter coefficient ω means that the adaptation to the transfer function of the echo path becomes slower, so that the echo component cannot be completely removed during the initial pull-in period, and the speech recognition processing, speech communication, etc. There is a defect that makes it difficult.
[0013]
In order to solve such a problem, for example, it is conceivable to back up the filter coefficients. However, since a memory for the filter coefficients is required, the system cost increases.
[0014]
The present invention has been made in view of the above circumstances, and an object of the present invention is to suppress an increase in cost and to quickly adapt a filter coefficient in an initial drop-in period. It is an object of the present invention to provide a calculation method and a computer program executed by a processor for realizing an echo cancellation processing function.
[0015]
[Means for Solving the Problems]
According to the echo canceller device of the first aspect, the calculation of the filter coefficient of the adaptive filter is executed by a method for improving the adaptation speed until the initial pull-in period, and after the initial pull-in period elapses, the adaptive accuracy is improved. Switch to execute. That is, it is only necessary to appropriately switch the operation method in each processing period, and it is not necessary to prepare a memory for storing the filter coefficients. Then, in the initial pull-in period, the filter coefficients can be adapted at high speed to make the echo cancellation processing effective.
[0016]
According to the echo canceller device of the second aspect, the learning coefficient, which is a calculation parameter of the filter coefficient, is set to a constant value that is larger than a calculation result described later in the initial pull-in period in which the power value of the echo removal signal is large. After the elapse of the initial pull-in period, the calculation is performed based on the reciprocal of the power ratio between the echo removal signal and the transmission signal. With such a configuration, the learning coefficient during the initial pull-in period has a value larger than the result of the calculation, so that the filter coefficient can be quickly adapted to the transfer function of the echo path.
[0017]
According to the echo canceller device of the third aspect, during the initial pull-in period, the learning coefficient is obtained by multiplying the equation of the calculation by the reciprocal of the power ratio, and after the initial pull-in period has elapsed, the learning coefficient is obtained by the calculation. With such a configuration, the influence of the power ratio on the calculation of the filter coefficient during the initial pull-in period (that is, a decrease in the adaptation speed of the filter coefficient) can be canceled, and the filter coefficient can be made faster with respect to the transfer function of the echo path. Can be adapted to
[0018]
According to the echo canceller of the fourth aspect, when the double talk state occurs during the initial pull-in period, the calculation of the filter coefficient is executed by a method for improving the adaptation accuracy. In other words, even during the initial pull-in period, if a double talk state is occurring at the same time, proper learning cannot be performed unless the convergence speed of the filter coefficients is reduced to give priority to the adaptation accuracy. Therefore, if configured as in claim 4, appropriate learning can be performed.
[0019]
BEST MODE FOR CARRYING OUT THE INVENTION
(First embodiment)
Hereinafter, a first embodiment of the present invention will be described with reference to FIGS. Note that the same parts as those in FIG. 8 are denoted by the same reference numerals, and the description thereof will be omitted. FIG. 3 is a diagram corresponding to FIG. 8, but the learning coefficient calculation unit 4 is simply replaced by the learning coefficient calculation unit 11, and the other configuration is the same as that of FIG.
[0020]
FIG. 4 is an overall configuration diagram of the echo canceller device 12. The echo canceller device 12 performs processing corresponding to the A / D converter 14 for A / D-converting the audio signal input from the microphone 13, the adaptive filter 3 described above, the learning coefficient calculator 11, and the power calculator 5 as a program ( The system includes a DSP (Digital Signal Processor) 16 which is executed based on a computer program 15, a D / A converter 18 for D / A converting an audio signal output to the outside by a speaker 17, and the like.
[0021]
The echo canceller device 12 transfers the sound signal subjected to the echo cancellation processing to, for example, a car navigation device, a mobile phone, or the like, receives the sound signal given by the signal, and outputs the received sound signal to the outside through the speaker 17. .
[0022]
Next, the operation of the present embodiment will be described with reference to FIGS. FIG. 1 is a flowchart mainly showing the learning coefficient calculation unit 11 and the adaptive filter 3 in the echo cancellation processing executed by the DSP 15. For example, when the echo canceller device 12 is powered on and started (start), the learning coefficient calculation unit 11 sets an initial value f of the learning coefficient μ (step S1).
[0023]
The initial value f here is the power value EPWR of the echo removal signal e (n) obtained in a state where all the parameters used for the operation of the echo cancellation processing are cleared to zero, and the output audio signal x (n). Based on the power value XPWR, a value sufficiently larger than the learning coefficient μ calculated by the equation (3) is selected.
[0024]
Then, when calculating the filter coefficient ω (n) (step S2), the adaptive filter 3 calculates the absolute value of the difference between the calculation result ω (n) and the previous calculation result ω (n−1), and calculates the difference. It is determined whether or not the absolute value of is smaller than a predetermined value d (step S3). The predetermined value d is an appropriate value for judging that the difference between the calculation results has become sufficiently small when the filter coefficient ω (n) is substantially in the adaptive state.
[0025]
If “NO” is determined in step S3, the adaptive filter 3 substitutes the filter coefficient ω (n) for ω (n−1) (step S4) and returns to step S2. Therefore, the loop of steps S2 to S4 is continued until it is determined “YES” in step S3.
[0026]
In the period in which the above loop is being performed (this period is referred to as an initial pull-in period), the value of the filter coefficient ω (n) is set as shown in FIG. As shown in FIG. When the filter coefficient ω (n) approaches a target coefficient adapted to the transfer function of the echo path, the filter oscillates with a relatively large amplitude around the target coefficient.
[0027]
When the filter coefficient ω (n) converges to the vicinity of the target coefficient during the loop described above, if “YES” is determined in step S3, the adaptive filter 3 determines the filter coefficient ω similarly to step S4. (N) is substituted for ω (n-1) (step S5). Then, the learning coefficient calculation unit 11 calculates the learning coefficient μ according to equation (3) in the same manner as in the related art (step S6).
[0028]
The adaptive filter 3 calculates a filter coefficient ω (n) based on the learning coefficient μ obtained in step S6 (step S7). Then, it is determined whether or not the echo canceling process has been completed (step S8), and if not ("NO"), the process returns to step S5 to continue the process.
[0029]
As described above, according to the present embodiment, the echo canceller device 12 executes the calculation of the filter coefficient ω (n) of the adaptive filter 3 in a manner that improves the adaptive speed until the initial pull-in period, and performs the calculation in the initial pull-in period. After the lapse of time, switching is performed so as to execute the adaptive accuracy improvement method. Specifically, the learning coefficient μ, which is a calculation parameter of the filter coefficient ω (n), is replaced with a constant value f during the initial pull-in period, and after the initial pull-in period has elapsed, the calculation is performed based on equation (3). I made it.
[0030]
Accordingly, since only the operation method in each processing period is appropriately switched, there is no need to prepare a memory for storing the filter coefficient ω (n). Then, in the initial pull-in period, the filter coefficient ω (n) can be adapted at a high speed to make the echo cancellation process effective, and thereafter, the same adaptation accuracy as in the related art can be secured.
[0031]
According to the method of this embodiment, the oscillation amplitude of the filter coefficient ω (n) during the initial pull-in period is relatively large, but by converging more quickly to the target coefficient, the echo cancellation function during the period can be compared with the conventional one. Will work more effectively.
[0032]
(Second embodiment)
FIG. 5 shows a second embodiment of the present invention. The same parts as those in the first embodiment are denoted by the same reference numerals, and the description thereof will be omitted. Only different parts will be described below. In the second embodiment, the processing contents of the learning coefficient calculation unit 11 having the configuration shown in FIG. 3 are only slightly different from those of the first embodiment.
[0033]
In the flowchart shown in FIG. 5, step S9 is arranged instead of step S1. In step S9, the learning coefficient μ is calculated by the equation (4).
[0034]

That is, equation (4) is obtained by multiplying equation (3) by the power ratio EPWR / XPWR. When step S4 is executed, the process returns to step S9. Other processes are the same as in the first embodiment.
[0035]
As described above, according to the second embodiment, the learning coefficient calculation unit 11 determines the learning coefficient μ by multiplying the equation (3) by the power ratio EPWR / XPWR during the initial pull-in period. , The learning coefficient μ is obtained by the equation (3), so that the effect of the power ratio EPWR / XPWR on the calculation of the filter coefficient ω (n) during the initial pull-in period (that is, the reduction of the adaptation speed) can be canceled. ω (n) can be quickly adapted to the transfer function of the echo path.
[0036]
(Third embodiment)
FIGS. 6 and 7 show a third embodiment of the present invention. The same parts as those in the first embodiment are denoted by the same reference numerals, and description thereof will be omitted. Hereinafter, only different parts will be described. In the second embodiment, the learning coefficient calculator 11 having the configuration shown in FIG. 3 is replaced with a learning coefficient calculator 19, and a double talk detector (double talk detecting means) 19 is added.
[0037]
The double talk detector 20 detects the occurrence of a double talk state when the power ratio EPWR / XPWR of the echo removal signal and the transmission signal obtained from the power calculation unit 5 exceeds a predetermined value, and calculates the detection signal DT as a learning coefficient. It is configured to output to the unit 19.
[0038]
Next, the operation of the third embodiment will be described with reference to FIG. In the flowchart shown in FIG. 7, the learning coefficient calculation unit 19 proceeds to step S10 after executing step S4, and determines whether or not the double talk detector 20 outputs the detection signal DT. If the detection signal DT has not been output (“NO”), the process returns to step S2, and if the double talk state has been detected, and if the detection signal DT has been output (“YES”), the process proceeds to step S5.
[0039]
That is, even in the initial pull-in period during the loop of steps S2 to S4, when the double talk state is detected, the process proceeds to step S5, and the learning coefficient μ is calculated by the equation (3). This is because when the double talk state occurs, appropriate learning cannot be performed unless the convergence speed of the filter coefficient ω (n) is reduced to give priority to the adaptation accuracy.
[0040]
As described above, according to the third embodiment, when the double talk state occurs during the initial pull-in period, the learning coefficient calculation unit 19 calculates the filter coefficient ω (n) based on the equation (3). Since the adjustment is performed in a manner that improves the adaptation accuracy, appropriate learning can be performed according to the situation.
[0041]
The present invention is not limited to the embodiment described above and shown in the drawings, and the following modifications or extensions are possible.
In the second embodiment, the function of the double talk detector 20 may be provided inside the learning coefficient calculator 19.
[Brief description of the drawings]
FIG. 1 is a first embodiment of the present invention, which is a flowchart mainly showing a learning coefficient calculator and processing contents related to an adaptive filter in an echo canceling process. FIG. 3 is a functional block diagram showing a schematic configuration of an echo canceller. FIG. 4 is an overall configuration diagram of an echo canceller device. FIG. 5 is a diagram showing a second embodiment of the present invention. FIG. 6 is a diagram corresponding to FIG. 3 showing a third embodiment of the present invention. FIG. 7 is a diagram corresponding to FIG. 1. FIG. 8 is a diagram corresponding to FIG. Description]
3 is an adaptive filter, 11 is a learning coefficient calculator, 12 is an echo canceller device, 15 is a program (computer program), 16 is a DSP, 19 is a learning coefficient calculator, and 20 is a double talk detector (double talk detecting means). .

Claims (12)

Simulating the acoustic characteristics of the echo path by an adaptive filter using the NLMS algorithm based on the digitized output audio signal, and generating a pseudo echo signal by multiplying the acoustic signal by the filter coefficient of the adaptive filter. An echo canceller device that performs an echo canceling process of subtracting the pseudo echo signal from the input audio signal and removing an echo signal component generated by the echo path,
The calculation of the filter coefficient is performed by a method for improving the adaptation speed until the initial pull-in period in which a predetermined time elapses from the start of the echo cancellation process, and after the elapse of the initial pull-in period, the adaptive accuracy is improved by a method. An echo canceller device characterized by switching to be executed. When calculating a learning coefficient which is a calculation parameter of the filter coefficient based on a reciprocal of a power ratio between an echo removal signal and a transmission signal, the learning coefficient is larger than a result of the calculation in the initial pull-in period. 2. The echo canceller device according to claim 1, wherein the learning coefficient is obtained by the calculation after the initial pull-in period has elapsed after replacing with a constant value. When calculating a learning coefficient, which is an operation parameter of the filter coefficient, based on the reciprocal of a power ratio between an echo removal signal and a transmission signal, in the initial pull-in period, the learning coefficient is calculated by using the power ratio The echo canceller device according to claim 1, wherein the learning coefficient is obtained by the calculation after the initial pull-in period has elapsed. A double-talk detecting means for detecting a double-talk state occurring between the transmission system and the reception system,
4. The echo canceller according to claim 1, wherein when a double talk state occurs during the initial pull-in period, the calculation of the filter coefficient is performed by a method that improves adaptation accuracy. 5. apparatus. Simulating the acoustic characteristics of the echo path by an adaptive filter using the NLMS algorithm based on the digitized output audio signal, and generating a pseudo echo signal by multiplying the acoustic signal by the filter coefficient of the adaptive filter. In the filter coefficient calculation method in the echo cancellation process of subtracting the pseudo echo signal from the input audio signal to remove an echo signal component generated by the echo path,
The calculation of the filter coefficient is performed by a method for improving the adaptation speed until the initial pull-in period in which a predetermined time elapses from the start of the echo cancellation process, and after the elapse of the initial pull-in period, the adaptive accuracy is improved by a method. A filter coefficient calculation method for echo cancellation processing, characterized by switching to be executed. When calculating a learning coefficient which is a calculation parameter of the filter coefficient based on a reciprocal of a power ratio between an echo removal signal and a transmission signal, the learning coefficient is larger than a result of the calculation in the initial pull-in period. 6. The filter coefficient calculating method according to claim 5, wherein the learning coefficient is obtained by the calculation after the initial pull-in period elapses, and the predetermined value is replaced by a constant value. When calculating a learning coefficient, which is an operation parameter of the filter coefficient, based on a reciprocal of a power ratio between an echo removal signal and a transmission signal, the learning coefficient is multiplied by the power ratio in the initial pull-in period. 6. The method according to claim 5, wherein the calculation is performed by the calculation after the elapse of the initial pull-in period. 6. The method according to claim 5, wherein when a double talk state occurs between a transmission system and a reception system during the initial pull-in period, the calculation of the filter coefficient is performed in a manner that improves adaptive accuracy. 8. The filter coefficient calculation method for echo cancellation processing according to any one of claims 1 to 7. Simulating the acoustic characteristics of the echo path by an adaptive filter using the NLMS algorithm based on the digitized output audio signal, and generating a pseudo echo signal by multiplying the acoustic signal by the filter coefficient of the adaptive filter. A computer program executed by a processor for implementing an echo cancellation processing function of subtracting the pseudo echo signal from the input audio signal and removing an echo signal component generated by the echo path,
The calculation of the filter coefficient is performed by a method of improving the adaptation speed until the initial pull-in period in which a predetermined time elapses from the start of the echo canceling process, and after the elapse of the initial pull-in period, the adaptive accuracy is improved by a method. A computer program characterized by being switched to be executed. When calculating a learning coefficient, which is a calculation parameter of the filter coefficient, based on the reciprocal of the power ratio between the echo removal signal and the transmission signal, the learning coefficient is larger than the result of the calculation in the initial pull-in period. The computer program according to claim 9, wherein the learning coefficient is calculated by a predetermined value, and after the elapse of the initial pull-in period, the learning coefficient is calculated. When calculating a learning coefficient, which is a calculation parameter of the filter coefficient, based on the reciprocal of a power ratio between an echo removal signal and a transmission signal, in the initial pull-in period, the learning coefficient is calculated by applying the power ratio to an equation of the calculation. 10. The computer program according to claim 9, wherein the learning coefficient is calculated by the calculation after the initial pull-in period has elapsed. Detect the double talk state that occurs between the transmission system and the reception system,
The computer according to any one of claims 9 to 11, wherein when a double talk state occurs during the initial pull-in period, the calculation of the filter coefficient is performed in a manner that improves adaptation accuracy. program.

Images