JP4041770B2 - Acoustic echo cancellation method, apparatus, program, and recording medium - Google Patents

Description

これら受話信号の信号ブロックを部分予測エコー生成部７３₁〜７３_Mにてそれぞれ各部分予測インパルス応答と畳み込んで、周波数領域の部分予測エコー信号が生成され、これが部分予測エコー信号の総和を加算部７４でとって逆ＦＦＴ変換部７５で逆高速離散フーリエ変換を行って時間領域の予測エコー信号ｙ（ｋ）が生成される。
【００１０】
予測エコー信号ベクトルｙ^（ｋ）とブロック化部７６よりの収音信号ベクトルｙ（ｋ）との差が減算部７７でとられ、その誤差信号ベクトルがＦＦＴ変換部７８で式（３）に示すように周波数領域に変換される。
Ｅ（ｋ）←ＦＦＴ（［０，…，０，ｙ ^T（ｋ）−ｙ^^T（ｋ）］^T）（３）このようにＬ′個のサンプル用誤差信号ベクトルの前にＬ′個の０を付けて２Ｌ′のサンプルとして周波数領域に変換する。
【００１１】
周波数領域受話信号Ｘ ₁（ｋ）は共役部７３_1aでその複素共役Ｘ ₁ ^*（ｋ）に変換され、周波数領域の誤差信号Ｅ（ｋ）と乗算部７３_1bで乗算され、その乗算結果ｄＷ ₁に対し、乗算部７３_1cでステップサイズμが乗算され、その結果μ ｄＷ ₁が更新部７３_1dでそれまでの周波数領域の部分予測インパルス応答Ｗ ₁（ｋ）に対し加算されて、これが更新される。他の遅延受話信号Ｘ ₂（ｋ），…，Ｘ _M（ｋ）と誤差信号Ｅ（ｋ）とが部分予測エコー生成部７３₂，…，７３_Mでそれぞれ同様に処理される。このように周波数領域での信号ベクトルＸ ₁（ｋ），…，Ｘ _M（ｋ），Ｅ（ｋ）をもちい、適応フィルタの各部分予測インパルス応答を次式で更新する。

ただしＸ _m ^*（ｋ）（ｍ＝１，…，Ｍ）はＸ _m（ｋ）の複素共役である。図９ではｄＷ _m（ｋ）＝Ｘ _m ^*（ｋ）Ｅ（ｋ）（ｍ＝１，…，Ｍ）を用いている。またμはフィルタ係数更新の大きさを決めているステップサイズである。図９中の部分予測エコー生成部７３₁〜７３_M内に示されるように、Ｍ個の部分予測エコー経路の部分インパルス応答の更新には、同一のステップサイズμが適用されていることに注意されたい。
このＭＤＦ法の処理遅延は従来法２のそれの１／Ｍになり、従来法２よりも演算量は多少増加するが適応フィルタの更新頻度はＭ倍になり収束速度が改善する。
【００１２】
【非特許文献１】
大賀、山崎、金田、「音響システムとディジタル処理」、電子情報通信学会、1997、pp.139-142
【非特許文献２】
E.R.Ferrara,“Fast Implementation of LMS adaptive filters,”IEEE Trans.Acoust., Speech,Signal Processing, vol.ASSP-28, pp.474-475(1980)
【非特許文献３】
J.S.Soo and K.K.Pang:“Multidelay Block Frequency Domain Adaptive Filter,”IEEE Trans.on ASSP, vol.ASSP-38, no.2, pp.373-376(1990)
【００１３】
【発明が解決しょうとする課題】
適応フィルタによりエコーを十分に消去するには、そのフィルタ長がエコー経路のインパルス応答長と同等である必要がある。残響時間の比較的長い部屋で拡声通話システムを確実に動作させるには、適応フィルタ長を長くとる必要がある。フィルタ長が長いほど適応フィルタの収束速度は遅くなるために、従来法３に対して、一層の収束速度向上が要求されている。
【００１４】
【課題を解決するための手段】
この発明によれば、受話信号を予測エコー経路により周波数領域でフィルタ処理して予測エコーを求め、その予測エコーと収音信号との誤差信号を小さくするように、予測エコー経路を構成する適応フィルタのフィルタ係数を周波数領域で制御する低演算量の音響エコー消去方法において、適応フィルタのタップ数（適応フィルタ長）をＭ（２以上の整数）分割し、Ｍ個の部分予測エコー経路を構成し、
特に部分予測エコー経路の部分インパルス応答の更新に対して、同一のステップサイズではなく、エコー経路のインパルス応答包絡線の減衰傾向に合わせて減衰するステップサイズを適用することを特徴とする。つまり一般に室内インパルス応答すなわちエコー経路のインパルス応答の包絡は図８に示したように時間とともにほぼ一定の傾向で減衰することが知られている。インパルス応答の変化量も同様の性質を持つ。この発明はこの性質を適応フィルタの係数更新に反映させたものである。
【００１５】
従ってインパルス応答の前半に位置し、係数変化の最も大きい部分予測インパルス応答Ｗ ₁（ｋ）の更新には大きめのステップサイズが設定される。またインパルス応答の後半に位置し、係数変化の小さい部分予測インパルス応答Ｗ _M（ｋ）の更新には小さいステップサイズが設定される。これにより、ＭＤＦ法に基づく周波数領域処理の適応フィルタに対して、同一の処理遅延、ほぼ同等の計算量で収束速度を向上させることが可能になる。
この発明方法は、スピーカＰ個（Ｐは２以上の整数）およびマイクロホン１個からなる拡声通話システムにもそのまま適用可能である。また、スピーカＰ個（Ｐは１以上の整数）およびマイクロホンＱ個（Ｑは２以上の整数）からなる拡声通話システムも、スピーカＰ個およびマイクロホン１個からなる拡声通話システムをＱ個並列に設置することで実現可能である。
【００１６】
【発明の実施の形態】
実施形態１
はじめにスピーカ１個マイクロホン１個からなる拡声通話システムにこの発明を適用した実施形態１を図１を参照して説明する。図１中の図９と対応する部分に同一参照番号を付けてある。以下では適応フィルタの全体長をＬ、その分割数をＭとする。
受話信号ｘ（ｋ）はブロック化部７１およびＦＦＴ変換部７２にて、下記ステップ１にしたがってブロック化されて周波数領域に変換される。受話信号は所定時間Ｌ′＝Ｌ／Ｍサンプルごとに長さ２Ｌ′のブロック信号として処理される。周波数領域の受話信号Ｘ ₁とその順次Ｌ′サンプル分遅延された信号Ｘ ₂〜Ｘ _Mが部分予測エコー生成部７３₁〜７３_Mにてフィルタ処理されて、周波数領域で部分予測エコーが生成される。加算部７４にてこれら部分予測エコーの総和がとられ、逆ＦＦＴ変換部７５を経て時間領域の予測エコーに変換される。この一連の処理が下記ステップ２に対応する。収音信号ｙ（ｋ）は、ブロック化部７６でＬ′サンプルごとにブロック化され、ベクトル減算部７７で予測エコー信号ベクトルとの差が求められたのち、ＦＦＴ変換部７８にて周波数領域に変換されて誤差信号Ｅ（ｋ）が求められる。この一連の処理が下記ステップ３に対応する。各部分予測エコー生成部７３_mでは、下記ステップ４、５にしたがって部分予測インパルス応答が更新される。
【００１７】
ステップ１
受話信号ｘ（ｋ）を、Ｌ′＝Ｌ／Ｍサンプルごとに長さ２Ｌ′の信号ベクトルにブロック化し、ＦＦＴを適用して式（５）に示すように周波数領域に変換する。
Ｘ ₁（ｋ）＝diag（FFT（［ｘ（k-2L′＋１），…，ｘ（ｋ）］^T)) （５）
ただし上式のｄｉａｇ（ ）は、周波数領域に変換した信号ベクトルを対角成分のみに各周波数成分を持つ行列（対角行列）に変換する。同時に、過去の信号ブロックを縦続接続された遅延部７９₁〜７９_M-1でそれぞれシフトする。遅延部７９₁，…，７９_M-1より式（２）と同様にそれぞれ順次Ｌ′ずつ多く遅延された信号ブロックＸ ₂（ｋ），…，Ｘ _M（ｋ）が出力される。

【００１８】
ステップ２
各部分予測エコー生成部７３_m（ｍ＝１，２，…，Ｍ）において、受話信号Ｘ _m（ｋ）と部分予測インパルス応答Ｗ _m（ｋ）とを周波数成分ごとに積をとることで、受話信号ベクトルをフィルタ処理した周波数領域の部分予測エコーを得る。これら部分予測エコー信号ベクトルの総和を加算部７４でとって逆ＦＦＴ変換部７５で逆ＦＦＴを適用し、更にブロック整形部７５ａでＬ′サンプル時間ごとのＬ′サンプルの時間領域の予測エコー信号ベクトルｙ^（ｋ）とする。
ｙ^(k)＝[０ _{L ′} Ｉ _{L ′}]ＦＦＴ^-1(Ｘ ₁(k)Ｗ ₁(k)＋…＋Ｘ _M(ｋ)Ｗ _M(ｋ)) （６）
ただし、０ _{L ′}はＬ′×Ｌ′の零行列、Ｉ _{L ′}はＬ′×Ｌ′の単位行列である。
【００１９】
ステップ３
ブロック化された収音信号と予測エコーとの誤差信号ベクトルをＦＦＴ変換部７８で式（３）と同様にＬ′個の誤差サンプルの前に０をＬ′個付けて周波数領域の信号Ｅ（ｋ）に変換する。
Ｅ（ｋ）＝ＦＦＴ（［０，…，０，ｙ ^T（ｋ）−ｙ^^T（ｋ）］^T）（３）
ただし
ｙ（ｋ）＝［ｙ（ｋ−Ｌ′＋１）…ｙ（ｋ）］^T （７）
である。
【００２０】
ステップ４
部分予測エコー生成部７３₁，…，７３_mの各部分予測インパルス応答Ｗ ₁（ｋ），…，Ｗ _M（ｋ）を、それぞれ異なるステップサイズμ₁…μ_Mをもちいて周波数領域で式（８）に示すように更新する。ステップサイズμ_m（ｍ＝１、・・・、Ｍ１）は、ステップサイズ生成部７０で次式の演算により生成する。
μ₁＝μ₁
μ_m＝μ₁α^m-1（ｍ＝１、・・・、Ｍ１）
つまりステップサイズμ₁，…，μ_Mは一定の減衰率αで指数的に減衰するように設定生成する。減衰率αは室内インパルス応答包絡の減衰傾向すなわち残響時間から決められる。例えば適用室について実測して求める。またμ₁は０〜１の値に設定する。この例では周波数領域の誤差信号Ｅ（ｋ）に補正行列Ｓを乗算部７８２で乗算し、この乗算結果と各受話信号Ｘ _m（ｋ）の複素共役Ｘ _m ^*（ｋ）（ｍ＝１，…，Ｍ）とを乗算部７３_mbで乗算した行列にステップサイズμ_mが乗算部７３_mcで乗算される。
【００２１】

ここで、補正行列Ｓ（ｋ）は、遅延部７９₁，…，７９_M-1の出力信号を補正行列算出部７８１に入力して
Ｓ（ｋ）＝diag（［１／ｒ₁（ｋ）…１／ｒ_{L ′}（ｋ）］） （９）
ｒ_j（ｋ）＝βｒ_j（ｋ−Ｌ′）＋（１−β）Σ _m=1 ^Mμ_mＴ²（Ｘ _m（ｋ），ｊ） （１０）
により算出される対角行列である。ただしＴ（Ｘ _m（ｋ），ｊ）は行列Ｘ _m（ｋ）の（ｊ，ｊ）成分を抜き出す関数である。行列Ｓ（ｋ）の対角要素の分母のｒ_j（ｋ）は、周波数成分ごとに部分予測エコー生成部７３₁〜７３_Mの各入力受話信号パワーの重み付き総和を求めたものである。βは前回の短時間平均パワーの総和ｒ_j（ｋ−Ｌ′）と今回の短時間パワーとの短時間平均をとるための平滑化定数であり、０〜１の値をとる。音声のように有色性信号の場合、修正ベクトルｄＷ _m（ｋ）に行列Ｓ（ｋ）をかけることは受話信号の白色化処理に対応し、有色信号が入力されたときの適応フィルタの収束速度を向上させることが知られている。
図１に示した例では、周波数領域の修正ベクトルｄＷ _m（ｋ）＝Ｘ _m ^*（ｋ）Ｓ（ｋ）Ｅ（ｋ）（ｍ＝１，…，Ｍ）をもちいている。
【００２２】
ステップ５
これら周波数領域の各部分予測エコー経路のインパルス応答Ｗ ₁（ｋ＋Ｌ′），…，Ｗ _M（ｋ＋Ｌ′）について、周波数領域のベクトルと時間領域の部分予測エコー経路のインパルス応答とが１対１に対応するように次式（１２）により、フィルタ更新部７３_md内で整形する。
ｖ _m（ｋ＋Ｌ′）＝［Ｉ _L ０ _L］ＩＦＦＴ（Ｗ _m（ｋ＋Ｌ′））（１１）
Ｗ _m（ｋ＋Ｌ′）＝ＦＦＴ（［ｖ _m ^T（ｋ＋Ｌ′），０，…，０］^T）（１２）
ＦＦＴ［ ］内の０数はＬ′個である。
【００２３】
この実施形態１に示すようにこの発明においては所定時間（Ｌ′サンプル）ごとに、受話信号を分割し、その分割された受話信号（２Ｌ′サンプル）を周波数領域信号Ｘ（ｋ）に変換し、その信号Ｘ（ｋ）を部分予測エコー経路部でフィルタ処理して部分予測エコーを生成し、その連続する複数のＭ個の所定時間に生成されたＭ個の部分予測エコーの総和を時間領域の予測エコーに変換し、連続するＭ個の所定時間の各受話信号Ｘ ₁（ｋ）〜Ｘ _M（ｋ）と誤差信号Ｅ（ｋ）との乗算により周波数領域での修正ベクトルｄＷ ₁（ｋ）〜ｄＷ _M（ｋ）を生成し、これら修正ベクトルｄＷ ₁（ｋ）〜ｄＷ _M（ｋ）と複数（Ｍ個）の所定時間内の各所定時間ごとに異なるステップサイズμ₁〜μ_Mとを用いて、部分予測エコー経路部を周波数領域で更新する。
【００２４】
実施形態２
この発明の実施形態２は、図２に示すようにスピーカＰ個（Ｐは２以上の整数）とマイクロホン１個からなる拡声通話システムにこの発明を適用した場合である。以下の適応アルゴリズムでは、適応フィルタ長の全体長をＬ、その分割数をＭとするときオーバラップセーブ法をもちいてＬ′＝Ｌ／Ｍサンプルごとに長さ２Ｌ′のブロック信号を処理する。図１に示したエコー消去装置中の逆ＦＦＴ変換部７５、ブロック整形部７５ａ、ブロック化部７６、減算部７７、ＦＦＴ変換部７８、補正行列算出部７８１、及び乗算部７８２を除いた部分が各受話端子１_pにチャネル予測エコー生成部８_pとして接続され、これらチャネル予測エコー生成部８_pにＦＦＴ変換部７８よりの誤差信号Ｅ（ｋ）が入力される。ここでｐ
＝１，…，Ｐである。
各チャネル予測エコー生成部８_pは図３に示す構成となりこれに入力された受話信号ｘ_p（ｋ）を図１に示した場合と同様に処理する。
【００２５】
ステップ１
Ｐチャネルの受話信号ｘ_p（ｋ）（ｐ＝１，…，Ｐ）を、それぞれＬ′＝Ｌ／Ｍサンプルごとに長さ２Ｌ′の信号ベクトルにブロック化し、式（５）と同様にＦＦＴを適用して周波数領域に変換する。
Ｘ _p,1(ｋ)＝diag（ＦＦＴ（[ｘ_p(ｋ−２Ｌ′＋１)，…，ｘ_p(ｋ)]^T））
ただし上式のｄｉａｇ（ ）は、周波数領域に変換した信号ベクトルを対角成分に各周波数成分を持つ行列に変換しており、これ以降の説明の便宜をはかるためにもちいている。同時に、過去の信号ブロックを縦続接続された遅延部を１遅延部ずつそれぞれシフトして式（２）と同様にＸ _p,2（ｋ），…，Ｘ _p,M（ｋ）を得る。

【００２６】
ステップ２
各部分予測エコー生成部において、受話信号と、部分予測インパルス応答とを周波数成分ごとに積をとることで、受話信号ベクトルをフィルタ処理し、周波数領域の部分予測エコーＸ _p,m（ｋ）Ｗ _p,m（ｋ）（ｐ＝１，…，Ｐ，ｍ＝１，…，Ｍ）を得る。これら部分予測エコーの和をとることで式（１３）に示すチャネルごとの予測エコーＹ^_p（ｋ）を求める。このＹ^_p（ｋ）がチャネル予測エコー生成部８_pの出力となる。
Ｙ^_p（ｋ）＝Ｘ _p,1（ｋ）Ｗ _p,1（ｋ）＋…＋Ｘ _p,M（ｋ）Ｗ _p,M（ｋ） （１３）
【００２７】
ステップ３
第１〜Ｐチャネルにおける周波数領域での予測エコー信号ベクトルの総和を加算部８ａでとり、その総和に対し、逆ＦＦＴ変換部７５、ブロック整形部７５ａにより式（６）と同様に逆ＦＦＴを適用しかつＬ′個のサンプルブロックに整形して、予測エコー信号ベクトルｙ^（ｋ）を求める。
ｙ^（ｋ）＝［０ _{L ′} Ｉ _{L ′}］ＦＦＴ^-1（Ｙ^₁（ｋ）＋…＋Ｙ^_p（ｋ）） （１４）
ただし、０ _{L ′}はＬ′×Ｌ′の零行列、Ｉ _{L ′}はＬ′×Ｌ′の単位行列である。
そして、収音信号と予測エコー信号との差である誤差信号ベクトルをＦＦＴ変換部７８で式（３）に示す周波数領域信号Ｅ（ｋ）に変換する。
Ｅ（ｋ）＝ＦＦＴ（［０，…，０，ｙ ^T（ｋ）−ｙ^^T（ｋ）］^T）（３）
このＦＦＴ変換はＬ′個のサンプルの誤差信号ベクトルの前にＬ′個の０を付けて行う。また、
ｙ（ｋ）＝［ｙ（ｋ−Ｌ′＋１）…ｙ（ｋ）］^T （７）
である。各チャネル予測エコー生成部８_pにおいて加算部７４で部分予測エコーの和を取ることなく、加算部８ａで、チャネル予測エコー生成部８₁〜８_Pより全部分予測エコーの総和をとってもよい。
【００２８】
ステップ４
第１〜Ｐチャネルにおける第１〜第Ｍ区間の各部分予測インパルス応答Ｗ _p,1（ｋ）…Ｗ _p,M（ｋ）（ｐ＝１，…，Ｐ）を区間ごとに異なるステップサイズμ₁…μ_Mをもちいて式（１５）に示すように周波数領域で更新する。
【００２９】

ただしＸ _p,m ^*（ｋ）（ｐ＝１，…，Ｐ、ｍ＝１，…，Ｍ）はＸ _p,m（ｋ）の複素共役である。
第ｐチャネルの第ｍ区間の部分予測インパルス応答の周波数領域における修正ベクトルはｄＷ _p,m（ｋ）＝Ｘ _p,m ^*（ｋ）Ｅ（ｋ）（ｍ＝１，…，Ｍ）である。
【００３０】
ステップ５
第１〜Ｐチャネルの第１〜第Ｍ区間の各部分予測エコー経路の部分インパルス応答Ｗ _p,1（ｋ＋Ｌ′），…，Ｗ _p,M（ｋ＋Ｌ′）について、周波数領域のベクトルと時間領域の部分予測エコー経路インパルス応答が１対１に対応するように式（１１），（１２）と同様な次式により整形する。
ｖ _p,m（ｋ＋Ｌ′）＝［Ｉ _L ０ _L］IFFT（Ｗ _p,m（ｋ＋Ｌ′））
Ｗ _p,m（ｋ＋Ｌ′）＝FFT（［ｖ _p,m ^T（ｋ＋Ｌ′），０，…，０］^T）
【００３１】
実施形態３
実施形態３は図４に示すようにスピーカＰ個（Ｐは１以上の整数）、マイクロホンＱ個（Ｑは２以上の整数）からなる拡声通話システムにこの発明を適用した場合である。図２に示したエコー消去装置がＱ個、９₁…９_Qとして設けられ、各エコー消去装置９_q（ｑ＝１，…，Ｑ）はＰ個の受話端子１₁〜１_Pからの各チャネルの受話信号ｘ₁（ｋ）〜ｘ_p（ｋ）と１個のマイクロホン３_qからの収音信号ｙ_q（ｋ）が入力され、その収音信号ｙ_q（ｋ）に対しエコー消去を行って送話端子４_qに出力する。つまり、この実施形態３では図２に示したスピーカＰ個マイクロホン１個からなる拡声通話システムにおけるエコー消去装置が各マイクロホンごとに並列に設置されていることになる。
【００３２】
実験例
この発明方法の性能を検証するために、スピーカ１個マイクロホン１個からなる拡声通話システムを想定して、数値シミュレーションを行った。この数値シミュレーションでは、サンプリング周波数を８ｋＨｚに設定し、音響エコー経路２３として残響時間３００ｍｓの部屋で実測した室内伝達関数を１６００タップに打ち切って音響エコーを生成した。適応フィルタについて、タップ数Ｌ＝１０２４、分割数Ｍ＝４とした。
【００３３】
従来法として白色化処理を含むＭＤＦ法をもちい、そのステップサイズをμ＝０．３に設定した。またこの発明方法として実施形態１をもちい、そのステップサイズをμ₁＝０．３、μ₂＝０．３×０．６、μ₃＝０．３×０．６²、μ₄＝０．３×０．６³に設定した。
入力信号として白色雑音をもちいた場合について、エコー経路推定値の相対誤差（Misalignment)の変化を図５に示す。相対誤差が−２０ｄＢに達する時間で比較すると、各部分予測エコー経路について更新時ステップサイズを残響特性に応じて別々に設定することにより、適応フィルタの推定速度が約３０％向上していることが分かる。
【００３４】
実施形態１において、補正行列Ｓ（ｋ）による誤差信号Ｅ（ｋ）に対する補正を行わなくてもよい。つまり図１において、補正行列算出部７８１、乗算部７８２を省略してもよい。また図３において、図１と同様に補正行列算出部７８１、乗算部７８２を設けて、誤差信号Ｅ（ｋ）に対し、各チャネル対応の補正を行ってもよい。
図１、図２に示した各音響エコー消去装置をコンピュータにより機能させてもよい。この場合は前述したこの発明による音響エコー消去方法の各過程をコンピュータにより実行させるためのプログラムをＣＤ−ＲＯＭ、磁気ディスクなどの記録媒体から又は通信回線を介してコンピュータにダウンロードし、そのコンピュータにそのプログラムを実行させればよい。
【００３５】
【発明の効果】
以上述べたようにこの発明によれば拡声通話システムのエコー消去において、エコー経路のインパルス応答を所定時間区間ごとに部分エコー経路インパルス応答に分割し、部分エコー経路インパルス応答の予測を周波数領域経由のブロック信号処理とし、その際、部分エコー経路ごとに異なるステップサイズで部分予測エコー経路インパルス応答を更新することにより、音響エコー経路インパルス応答の推定速度向上と低演算量化を実現することができる。
【図面の簡単な説明】
【図１】この発明の実施形態１の機能構成例を示す図。
【図２】この発明の実施形態２の機能構成例を示す図。
【図３】図２中のチャネル予測エコー生成部８_pの機能構成例を示す図。
【図４】この発明の実施形態３の機能構成例を示す図。
【図５】この発明方法と従来法によるエコー経路インパルス応答予測の収束挙動の数値シミュレーションの例を示す図。
【図６】従来の時間領域でのみ処理するエコー消去装置の機能構成を示す図。
【図７】従来の周波数領域を経由する適応フィルタ処理を用いるエコー消去装置の機能構成を示す図。
【図８】インパルス応答とその分割例を示す図。
【図９】従来のＭＤＦ法を適用した適応フィルタ処理を用いるエコー消去装置の機能構成を示す図。[0001]
BACKGROUND OF THE INVENTION
The present invention is applied to a loudspeaker communication system, and relates to a method, an apparatus, a program, and a recording medium for canceling acoustic echo that is an obstacle to a call and sometimes causes howling.
[0002]
[Prior art]
In the loudspeaking call system, the received voice is loudspeaked from a speaker and picked up by a microphone to generate an acoustic echo, and the processing thereof becomes a problem. When the loop gain of a closed loop formed including a ground-based loudspeaker communication system is greater than 1, acoustic echo causes howling and makes the call impossible. Even when the loop gain is smaller than 1, the acoustic echo causes an adverse effect such as a call failure or discomfort. In order to realize a more natural call environment, an acoustic echo canceller (echo canceller) that cancels acoustic wraparound from the speaker to the microphone is required.
[0003]
FIG. 6 shows an acoustic echo canceller (canceller) composed of a 1-channel reproduction system and a 1-channel sound collection system. The received signal from the receiving terminal 1 is reproduced as an acoustic signal by the speaker 2, and this acoustic signal goes around the microphone 3 through the acoustic echo path 23.
The received signal is x (k) (k is a discrete time), the echo signal collected by the microphone 3 is y (k), and the acoustic echo path 23 from the speaker (reproducer) 2 to the microphone (sound collector) 3 Is an impulse response h (k), and its length is L. The echo signal and the received signal have the following relationship.
y (k) = Σ_{i = 0} ^L-1h (i) x (ki)
In addition, impulse response and input signal
h= [H (0) ... h (L-1)]^T
x(K) = [x (k)... X (k−L + 1)]^T
When the vectorization is performed as follows, the convolution of the received signal and the impulse response is briefly described as follows. here[ ]^TRepresents the transpose of a vector.
y (k) =h ^T x(K)
[0004]
This acoustic echo is erased by an echo canceling unit 5 connected between the reception terminal 1 and the transmission terminal 4. The received signal x (k) is input to the adaptive filter 51 for generating a predicted echo signal to generate a predicted echo signal. The difference between the collected sound signal y (k) from the microphone 3 and the predicted echo signal is the subtractor 52. And an error signal e (k) is produced. Based on the error signal e (k) and the past received signal, the filter coefficient of the filter 51 for generating the predicted echo signal is updated so that the error between the collected sound signal and the predicted echo signal is reduced.
[0005]
"Conventional method 1"
First, a case will be described in which the filter coefficient is updated using an NLMS (Normalized Least Mean Square) algorithm (see Non-Patent Document 1). The number of taps of the response filter 51 for predictive echo generation is L, and the filter coefficient vector isw(K) The predicted echo signal y ^ (k) is obtained by convolving the received signal with the filter coefficient.
y ^ (k) =w ^T(K)x(K)
Correction vector using difference e (k) = y (k) −y ^ (k) between collected sound signal y (k) and predicted echo signal
dw(K) = e (k)x(K) /x ^T(K)x(K)
And the coefficient of the adaptive filter is updated by the following equation.
w(K + 1) =w(K) + μ dw(K)
However, μ is a step size set to stabilize the estimation.
In the NLMS method, since the convolution calculation for predictive echo generation and the correction of the adaptive filter are performed every sample in the time domain as described above, the calculation amount is very large although there is no processing delay. It is also known that the convergence speed, that is, the time until the estimated echo path by the adaptive filter 51 converges to the true echo path is slow.
[0006]
"Conventional method 2"
An adaptive algorithm that significantly reduces the amount of computation is described in E.I. R. Proposed by Ferrara (see Non-Patent Document 2). In this algorithm, the modification of the adaptive filter is changed from processing for each sample to block processing for each L sample. Then, convolution signal processing for predictive echo signal generation is blocked and performed via the frequency domain by fast discrete Fourier transform (FFT).
In this algorithm, the correction vector is calculated by convolution of the error signal and the received signal. This calculation is as follows when the adaptive filter is modified at time k.
dw(K) = Σ_{i = 0} ^L-1e (ki)x(Ki)
The convolution processing for correcting the adaptive filter can also be efficiently performed using FFT, and the total amount of calculation can be greatly reduced. The functional configuration is shown in FIG.
The received signal x (k) is divided into blocks each having 2L samples while being shifted by L samples by the blocking unit 61, and the signal of each block is frequency domain signal by the FFT transforming unit 62 by fast Fourier transform (FFT).XAnd its complex conjugateX ^*Is generated by the conjugate generation unit 63a, and the residual signal which is similarly set to the frequency domain for each block by the FFT conversion unit 68.EAre multiplied by the multiplication unit 63b, and the correction vector d in the frequency domainWIs multiplied by the step size μ by the multiplier 63c, and the frequency domain adaptive filter coefficient is multiplied by the coefficient updater 63d.WΜdWIs added to the filter coefficientWIs updated. This filter coefficientWAnd frequency domain received signalXAre multiplied by the multiplication unit 63e, and the multiplication resultWXIs subjected to inverse fast discrete Fourier transform by the inverse FFT transform unit 64 and converted into a time domain signal, which is taken out as a block of L samples by the block shaping unit 65, and the predicted echo signal for each blockyAn echo signal obtained by blocking ^ (k) by the blocking unit 66 for each L sample.yThe error signal for each block is subtracted from (k) by the subtraction unit 67.e(K) is obtained.
As described above, in the method performed in the frequency domain, it is necessary to block the signal every L samples (block length: 2L), which causes a delay of at least L samples. In addition, the adaptive filter is updated every L samples, and the convergence speed is not particularly improved.
[0007]
"Conventional method 3"
In the adaptive algorithm of the conventional method 2, the problem that the processing delay is large and the frequency of updating the adaptive filter is low is described in J. S. This is solved by the algorithm proposed by Soo (see Non-Patent Document 3). This algorithm introduces the concept of a multi-delay filter (hereinafter abbreviated as MDF) in order to reduce the processing delay.
[0008]
In the frequency domain signal processing, convolution processing is realized by an overlap-save method. The MDF method takes advantage of the fact that this convolution process can be divided into overlapping save processes between smaller blocks. If the MDF is applied with the division number 4, as shown in FIG. 8, the impulse response is divided into four on the time axis, and the partial impulse response and the received signal are convoluted to generate a partial prediction echo signal. Corresponding to obtaining a predicted echo signal.
[0009]
If the tap length of the adaptive filter is set to L and the number of divisions is set to M (where L is divisible by M), the MDF method can obtain a predicted echo for each L ′ = L / M sample. A functional configuration of the MDF method is shown in FIG. In the MDF method, the received signal x (k) is blocked into a signal vector having a length of 2L ′ by the blocking unit 71 for each L ′ = L / M sample using the overlap save method. Is a signal in the frequency domain by the FFT converter 72.X ₁Converted to (k).
X ₁(K) ← FFT ([x (k−2L ′ + 1),..., X (k)]^T(1)
The past M−1 signal blocks (signal vectors) are connected in cascade by a delay unit 79.₁~ 79_M-1Are sequentially shifted by one delay unit for each L ′ sample,₁~ 79_M-1FromX _L(K) ~X _M(K) is output as shown in equation (2).

These received signal blocks are converted into partial prediction echo generators 73.₁~ 73_MThe partial prediction echo signal in the frequency domain is generated by convolution with each partial prediction impulse response at, and the sum of the partial prediction echo signals is taken by the adder 74, and the inverse fast discrete Fourier transform is performed by the inverse FFT transformer 75. Perform time domain predicted echo signaly(K) is generated.
[0010]
Predicted echo signal vectory^ (K) and collected sound signal vector from the block forming unit 76yThe difference from (k) is taken by the subtracting unit 77, and the error signal vector is transformed into the frequency domain by the FFT transforming unit 78 as shown in the equation (3).
E(K) ← FFT ([0, ..., 0,y ^T(K)-y^^T(K)]^T) (3) In this way, L ′ zeros are added in front of the L ′ sample error signal vectors to convert them into the frequency domain as 2L ′ samples.
[0011]
Frequency domain received signalX ₁(K) is the conjugate part 73._1aWith its complex conjugateX ₁ ^*Error signal in the frequency domain converted to (k)E(K) and multiplication unit 73_1bAnd the multiplication result dW ₁On the other hand, the multiplier 73_1cIs multiplied by the step size μ, resulting in μ dW ₁Is the update unit 73_1dUntil then, partially predicted impulse response in the frequency domainW ₁It is added to (k) and updated. Other delayed received signalX ₂(K), ...,X _M(K) and error signalE(K) is a partial prediction echo generation unit 73.₂, ..., 73_MAre processed in the same manner. Thus, the signal vector in the frequency domainX ₁(K), ...,X _M(K),EUsing (k), each partial prediction impulse response of the adaptive filter is updated by the following equation.

However,X _m ^*(K) (m = 1, ..., M) isX _mIt is a complex conjugate of (k). In FIG. 9, dW _m(K) =X _m ^*(K)E(K) (m = 1,..., M) is used. Further, μ is a step size that determines the magnitude of the filter coefficient update. Partial predictive echo generator 73 in FIG.₁~ 73_MNote that the same step size μ is applied to update the partial impulse responses of the M partial predicted echo paths, as shown in FIG.
The processing delay of the MDF method is 1 / M of that of the conventional method 2, and the amount of calculation is slightly increased as compared with the conventional method 2, but the update frequency of the adaptive filter is M times and the convergence speed is improved.
[0012]
[Non-Patent Document 1]
Oga, Yamazaki, Kanada, "Acoustic system and digital processing", IEICE, 1997, pp.139-142
[Non-Patent Document 2]
E.R.Ferrara, “Fast Implementation of LMS adaptive filters,” IEEE Trans.Acoust., Speech, Signal Processing, vol.ASSP-28, pp.474-475 (1980)
[Non-Patent Document 3]
J.S.Soo and K.K.Pang: “Multidelay Block Frequency Domain Adaptive Filter,” IEEE Trans.on ASSP, vol.ASSP-38, no.2, pp.373-376 (1990)
[0013]
[Problems to be solved by the invention]
In order to sufficiently cancel the echo by the adaptive filter, the filter length needs to be equal to the impulse response length of the echo path. In order to reliably operate the loudspeaker call system in a room having a relatively long reverberation time, it is necessary to increase the adaptive filter length. Since the convergence speed of the adaptive filter becomes slower as the filter length is longer, further improvement of the convergence speed is required for the conventional method 3.
[0014]
[Means for Solving the Problems]
According to the present invention, an adaptive filter that configures a predicted echo path so as to obtain a predicted echo by filtering the received signal in the frequency domain using the predicted echo path and reduce an error signal between the predicted echo and the collected sound signal. In a low-computation acoustic echo cancellation method for controlling the filter coefficient of the filter in the frequency domain, the number of taps (adaptive filter length) of the adaptive filter is divided into M (integers of 2 or more) to construct M partial prediction echo paths ,
In particular, for updating the partial impulse response of the partially predicted echo path, a step size that attenuates in accordance with the attenuation tendency of the impulse response envelope of the echo path is applied instead of the same step size. In other words, it is generally known that the envelope of the indoor impulse response, that is, the echo response of the echo path attenuates with a substantially constant tendency with time as shown in FIG. The amount of change in the impulse response has similar properties. The present invention reflects this property in the coefficient update of the adaptive filter.
[0015]
Therefore, it is located in the first half of the impulse response, and the partially predicted impulse response with the largest coefficient changeW ₁A larger step size is set for updating (k). In addition, it is located in the second half of the impulse response and has a small coefficient change.W _MA small step size is set for updating (k). This makes it possible to improve the convergence speed with the same processing delay and almost the same amount of calculation as compared with the frequency domain processing adaptive filter based on the MDF method.
The method of the present invention can be applied as it is to a loudspeaker call system comprising P speakers (P is an integer of 2 or more) and one microphone. In addition, a loudspeaker call system consisting of P speakers (P is an integer of 1 or more) and Q microphones (Q is an integer of 2 or more) is also provided with Q loudspeaker call systems consisting of P speakers and one microphone in parallel. This is possible.
[0016]
DETAILED DESCRIPTION OF THE INVENTION
Embodiment 1
First, a first embodiment in which the present invention is applied to a loudspeaker call system including one speaker and one microphone will be described with reference to FIG. 1 corresponding to those in FIG. 9 are denoted by the same reference numerals. In the following, the overall length of the adaptive filter is L, and the number of divisions is M.
The received signal x (k) is blocked in the block forming unit 71 and the FFT transforming unit 72 according to the following step 1 and converted into the frequency domain. The reception signal is processed as a block signal having a length of 2L ′ every predetermined time L ′ = L / M samples. Received signal in frequency domainX ₁And its delayed signal by L 'samplesX ₂~X _MIs a partial prediction echo generator 73.₁~ 73_MAnd a partial prediction echo is generated in the frequency domain. The summation of these partial prediction echoes is taken by the adder 74 and is converted into a prediction echo in the time domain via the inverse FFT converter 75. This series of processing corresponds to Step 2 below. The collected sound signal y (k) is blocked for each L ′ sample by the blocking unit 76, and after the difference from the predicted echo signal vector is obtained by the vector subtracting unit 77, the FFT conversion unit 78 puts it in the frequency domain. Converted error signalE(K) is required. This series of processing corresponds to Step 3 below. Each partial prediction echo generator 73_mThen, the partial prediction impulse response is updated according to steps 4 and 5 below.
[0017]
Step 1
The received signal x (k) is blocked into a signal vector having a length of 2L ′ for each L ′ = L / M sample, and is converted into the frequency domain as shown in Equation (5) by applying FFT.
X ₁(K) = diag (FFT ([x (k−2L ′ + 1),..., X (k)]^T)) (5)
However, diag () in the above equation converts the signal vector converted into the frequency domain into a matrix (diagonal matrix) having each frequency component only in the diagonal component. At the same time, a delay unit 79 in which past signal blocks are connected in cascade₁~ 79_M-1Shift each. Delay unit 79₁, ..., 79_M-1More signal blocks that are sequentially delayed by L ′ in the same manner as in equation (2)X ₂(K), ...,X _M(K) is output.

[0018]
Step 2
Each partial prediction echo generator 73_m(M = 1, 2,..., M), the received signalX _m(K) and partially predicted impulse responseW _mBy multiplying (k) for each frequency component, a partial prediction echo in the frequency domain obtained by filtering the received signal vector is obtained. The sum of these partial predicted echo signal vectors is taken by the adder 74, and the inverse FFT is applied by the inverse FFT transform unit 75. Further, the block shaping unit 75a predicts the echo signal vector in the time domain of L 'samples for each L' sample time.yLet ^ (k).
y^ (k) = [0 _{L ′} I _{L ′}] FFT^-1(X ₁(k)W ₁(k) + ... +X _M(k)W _M(k)) (6)
However,0 _{L ′}Is the L ′ × L ′ zero matrix,I _{L ′}Is an L ′ × L ′ unit matrix.
[0019]
Step 3
The error signal vector of the collected sound signal and the predicted echo that has been blocked is added to the frequency domain signal by adding L ′ to the L ′ error samples before the L ′ error samples in the FFT transform unit 78 as in the equation (3).EConvert to (k).
E(K) = FFT ([0, ..., 0,y ^T(K)-y^^T(K)]^T(3)
However,
y(K) = [y (k−L ′ + 1)... Y (k)]^T      (7)
It is.
[0020]
Step 4
Partial prediction echo generator 73₁, ..., 73_mEach partial predicted impulse responseW ₁(K), ...,W _M(K) for different step sizes μ₁... μ_MTo update in the frequency domain as shown in equation (8). Step size μ_m(M = 1,..., M1) is generated by the step size generation unit 70 by the following equation.
μ₁= Μ₁
μ_m= Μ₁α^m-1(M = 1,..., M1)
In other words, step size μ₁, ..., μ_MIs generated to be exponentially attenuated at a constant attenuation rate α. The attenuation rate α is determined from the decay tendency of the room impulse response envelope, that is, the reverberation time. For example, it is obtained by actually measuring the applicable room. Μ₁Is set to a value between 0 and 1. In this example, the frequency domain error signalE(K) correction matrixSIs multiplied by a multiplier 782, and the multiplication result and each received signal are multiplied.X _mComplex conjugate of (k)X _m ^*(K) (m = 1,..., M) and multiplication unit 73_mbThe step size μ to the matrix multiplied by_mIs the multiplication unit 73_mcMultiplied by
[0021]

Where the correction matrixS(K) is the delay unit 79.₁, ..., 79_M-1Is input to the correction matrix calculator 781.
S(K) = diag ([1 / r₁(K) ... 1 / r_{L ′}(K)]) (9)
r_j(K) = βr_j(K−L ′) + (1−β)Σ _{m = 1} ^Mμ_mT²(X _m(K), j) (10)
Is a diagonal matrix calculated by T (X _m(K), j) are matricesX _mThis is a function for extracting the (j, j) component of (k). line; queue; procession; paradeSR of the denominator of the diagonal element of (k)_j(K) is a partial prediction echo generation unit 73 for each frequency component.₁~ 73_MThe weighted sum of each input received signal power is obtained. β is the sum of previous short-term average powers r_jThis is a smoothing constant for taking a short time average of (k−L ′) and the current short time power, and takes a value of 0 to 1. In the case of a colored signal such as speech, the correction vector dW _m(K) matrixSApplying (k) corresponds to whitening processing of the received signal, and is known to improve the convergence speed of the adaptive filter when a colored signal is input.
In the example shown in FIG. 1, the correction vector d in the frequency domainW _m(K) =X _m ^*(K)S(K)E(K) (m = 1,..., M) is used.
[0022]
Step 5
Impulse response of each partial prediction echo path in these frequency domainsW ₁(K + L '), ...,W _MFor (k + L ′), the filter update unit 73 is expressed by the following equation (12) so that the frequency domain vector and the impulse response of the partial prediction echo path in the time domain correspond one-to-one._mdFormat in.
v _m(K + L ′) = [I _L 0 _L] IFFT (W _m(K + L ′)) (11)
W _m(K + L ′) = FFT ([v _m ^T(K + L ′), 0,..., 0]^T(12)
The number of 0s in the FFT [] is L ′.
[0023]
As shown in the first embodiment, in the present invention, the reception signal is divided every predetermined time (L ′ samples), and the divided reception signals (2L ′ samples) are divided into frequency domain signals.X(K) to convert the signalX(K) is filtered by the partial prediction echo path section to generate a partial prediction echo, and the sum of the M partial prediction echoes generated at a plurality of consecutive M predetermined times is used as a time domain prediction echo. Convert each received signal for M consecutive times that are convertedX ₁(K) ~X _M(K) and error signalECorrection vector d in the frequency domain by multiplication with (k)W ₁(K) to dW _M(K) to generate these correction vectors dW ₁(K) to dW _MDifferent step sizes μ for each predetermined time within (k) and a plurality (M) of predetermined times₁~ Μ_MAnd update the partial prediction echo path section in the frequency domain.
[0024]
Embodiment 2
Embodiment 2 of the present invention is a case where the present invention is applied to a loudspeaker call system comprising P speakers (P is an integer of 2 or more) and one microphone as shown in FIG. In the following adaptive algorithm, when the total length of the adaptive filter length is L and the number of divisions is M, a block signal having a length of 2L ′ is processed for each L ′ = L / M sample using the overlap save method. 1 except for the inverse FFT conversion unit 75, block shaping unit 75a, blocking unit 76, subtraction unit 77, FFT conversion unit 78, correction matrix calculation unit 781, and multiplication unit 782 in the echo canceller shown in FIG. Each receiving terminal 1_pChannel prediction echo generator 8_pThese channel prediction echo generators 8 are connected as_pError signal from the FFT converter 78E(K) is input. Where p
= 1,..., P.
Each channel prediction echo generator 8_pIs configured as shown in FIG. 3, and the received signal x input thereto_p(K) is processed similarly to the case shown in FIG.
[0025]
Step 1
P channel received signal x_p(K) Block (p = 1,..., P) into signal vectors of length 2L ′ for each L ′ = L / M sample, and apply FFT to the frequency domain in the same manner as in equation (5). Convert.
X _{p, 1}(k) = diag (FFT ([x_p(k-2L '+ 1), ..., x_p(k)]^T))
However, diag () in the above equation converts the signal vector converted into the frequency domain into a matrix having each frequency component as a diagonal component, and is used for convenience of the following description. At the same time, the delay unit in which the past signal blocks are connected in cascade is shifted by one delay unit, respectively, as in the equation (2).X _{p, 2}(K), ...,X _{p, M}(K) is obtained.

[0026]
Step 2
Each partial prediction echo generation unit multiplies the received signal and the partial predicted impulse response for each frequency component to filter the received signal vector, thereby generating a partial prediction echo in the frequency domain.X _{p, m}(K)W _{p, m}(K) (p = 1,..., P, m = 1,..., M) is obtained. By calculating the sum of these partial predicted echoes, the predicted echo for each channel shown in Equation (13)Y^_p(K) is obtained. thisY^_p(K) is a channel prediction echo generation unit 8._pOutput.
Y^_p(K) =X _{p, 1}(K)W _{p, 1}(K) + ... +X _{p, M}(K)W _{p, M}(K) (13)
[0027]
Step 3
The sum of the predicted echo signal vectors in the frequency domain in the first to P channels is taken by the adder 8a, and the inverse FFT is applied to the sum by the inverse FFT transform unit 75 and the block shaping unit 75a in the same manner as in equation (6). And a prediction echo signal vector shaped into L 'sample blocks.y^ (K) is obtained.
y^ (K) = [0 _{L ′} I _{L ′}] FFT^-1(Y^₁(K) + ... +Y^_p(K)) (14)
However,0 _{L ′}Is the L ′ × L ′ zero matrix,I _{L ′}Is an L ′ × L ′ unit matrix.
Then, an error signal vector, which is the difference between the collected sound signal and the predicted echo signal, is converted into a frequency domain signal represented by Expression (3) by the FFT transform unit 78.EConvert to (k).
E(K) = FFT ([0, ..., 0,y ^T(K)-y^^T(K)]^T(3)
The FFT conversion is performed by adding L ′ 0s before the error signal vector of L ′ samples. Also,
y(K) = [y (k−L ′ + 1)... Y (k)]^T          (7)
It is. Each channel prediction echo generator 8_pWithout adding the sum of the partial prediction echoes in the addition unit 74, the channel prediction echo generation unit 8 in the addition unit 8a.₁~ 8_PThe sum of all partial prediction echoes may be taken.
[0028]
Step 4
Each partially predicted impulse response of the first to Mth sections in the first to P channelsW _{p, 1}(K) ...W _{p, M}(K) (p = 1,..., P) is a step size μ that is different for each section.₁... μ_MTo update in the frequency domain as shown in equation (15).
[0029]

However,X _{p, m} ^*(K) (p = 1,..., P, m = 1,..., M) isX _{p, m}It is a complex conjugate of (k).
The correction vector in the frequency domain of the partial prediction impulse response of the m-th section of the p-th channel is dW _{p, m}(K) =X _{p, m} ^*(K)E(K) (m = 1,..., M).
[0030]
Step 5
Partial impulse response of each partial predicted echo path in the first to Mth sections of the first to P channelsW _{p, 1}(K + L '), ...,W _{p, M}(K + L ′) is shaped by the following equations similar to equations (11) and (12) so that the vector in the frequency domain and the partial predicted echo path impulse response in the time domain have a one-to-one correspondence.
v _{p, m}(K + L ′) = [I _L 0 _L] IFFT (W _{p, m}(K + L '))
W _{p, m}(K + L ′) = FFT ([v _{p, m} ^T(K + L ′), 0,..., 0]^T)
[0031]
Embodiment 3
Embodiment 3 is a case where the present invention is applied to a loudspeaker call system comprising P speakers (P is an integer of 1 or more) and Q microphones (Q is an integer of 2 or more) as shown in FIG. The number of echo cancellers shown in FIG.₁... 9_QEach echo canceling device 9 is provided as_q(Q = 1,..., Q) are P receiving terminals 1₁~ 1_PReceived signal x of each channel from₁(K) to x_p(K) and one microphone 3_qSound pickup signal y_q(K) is input and the collected sound signal y_qEcho canceling for (k) and sending terminal 4_qOutput to. That is, in the third embodiment, the echo canceller in the loudspeaker call system including one speaker P microphone shown in FIG. 2 is installed in parallel for each microphone.
[0032]
Experimental example
In order to verify the performance of the method of the present invention, a numerical simulation was performed assuming a loudspeaker communication system consisting of one speaker and one microphone. In this numerical simulation, an acoustic echo was generated by setting the sampling frequency to 8 kHz and cutting the room transfer function measured in a room with a reverberation time of 300 ms as the acoustic echo path 23 to 1600 taps. For the adaptive filter, the tap number L = 1024 and the division number M = 4.
[0033]
The MDF method including whitening treatment was used as a conventional method, and the step size was set to μ = 0.3. Further, as the method of the present invention, Embodiment 1 is used, and the step size is set to μ.₁= 0.3, μ₂= 0.3 x 0.6, μ_Three= 0.3 × 0.6², Μ_Four= 0.3 × 0.6^ThreeSet to.
FIG. 5 shows changes in the relative error (Misalignment) of the echo path estimation value when white noise is used as the input signal. Comparing with the time when the relative error reaches −20 dB, the estimated speed of the adaptive filter is improved by about 30% by setting the update step size separately for each partially predicted echo path according to the reverberation characteristics. I understand.
[0034]
In Embodiment 1, the correction matrixSError signal due to (k)ECorrection for (k) may not be performed. That is, in FIG. 1, the correction matrix calculation unit 781 and the multiplication unit 782 may be omitted. In FIG. 3, a correction matrix calculation unit 781 and a multiplication unit 782 are provided as in FIG.ECorrection for each channel may be performed on (k).
Each acoustic echo canceller shown in FIGS. 1 and 2 may be made to function by a computer. In this case, a program for causing the computer to execute the above-described acoustic echo canceling method according to the present invention is downloaded to a computer from a recording medium such as a CD-ROM or a magnetic disk or via a communication line. Just run the program.
[0035]
【The invention's effect】
As described above, according to the present invention, in the echo cancellation of the loudspeaker communication system, the echo response of the echo path is divided into partial echo path impulse responses every predetermined time interval, and the prediction of the partial echo path impulse response is performed via the frequency domain. By performing block signal processing and updating the partial predicted echo path impulse response with a different step size for each partial echo path, it is possible to improve the estimated speed of the acoustic echo path impulse response and reduce the amount of computation.
[Brief description of the drawings]
FIG. 1 is a diagram illustrating an example of a functional configuration according to a first embodiment of the invention.
FIG. 2 is a diagram showing a functional configuration example of a second embodiment of the present invention.
3 is a channel prediction echo generator 8 in FIG. 2;_pFIG.
FIG. 4 is a diagram showing an example functional configuration of a third embodiment of the invention.
FIG. 5 is a diagram showing an example of numerical simulation of convergence behavior of echo path impulse response prediction according to the method of the present invention and the conventional method.
FIG. 6 is a diagram showing a functional configuration of an echo canceling apparatus that performs processing only in a conventional time domain.
FIG. 7 is a diagram showing a functional configuration of an echo cancellation apparatus using adaptive filter processing that passes through a conventional frequency domain.
FIG. 8 is a diagram showing an impulse response and an example of its division.
FIG. 9 is a diagram showing a functional configuration of an echo canceller using adaptive filter processing to which a conventional MDF method is applied.

Claims (8)

The received signal from the receiving terminal is divided every predetermined time and converted to the frequency domain,
In partial prediction echo path unit, it generates a partial predicted echo to filter the received signal by taking the product of the received signal and the partial predicted impulse response of the frequency domain for each predetermined time,
A total of partial prediction echoes generated at a plurality of predetermined times is reconverted into the time domain for each predetermined time to generate a prediction echo,
Dividing the predicted echo in the time domain into blocks for each predetermined time;
The acoustic echo is eliminated by subtracting the predicted echo that is blocked in the time domain from the collected sound signal that is blocked at each predetermined time ,
An error signal before correction that is an error signal between the sound pickup signal and the predicted echo for each predetermined time is converted into a frequency domain,
Filter coefficient update of received signal power input to each partial prediction echo path section obtained previously
The current received signal power weighted sum by weighted addition of the short-time average of the new step size weighted sum and the path filter coefficient update step size weighted sum of the received signal power input to each partial predicted echo path section obtained this time For each frequency component corresponding to the reciprocal of the short-time average of the received signal power weighted sum and the above-mentioned error signal before correction in the frequency domain, respectively, as an error signal,
Multiplying each received signal of the plurality of predetermined times and the error signal for each component in the frequency domain to generate a correction vector,
An acoustic echo cancellation method, wherein the partial predicted impulse response of each partial predicted echo path unit is updated in the frequency domain using a different step size and a corresponding correction vector at each predetermined time. The received signal from the receiving terminal is divided into blocks every predetermined time,
Each divided block signal is converted to the frequency domain,
Delaying the frequency domain block signal, and generating M-1 (M is an integer of 2 or more) delay block signals whose delay amount is sequentially increased by the predetermined time,
Filtering the non-delayed frequency domain block signal and the M-1 delayed block signals according to frequency domain first to Mth partial predicted echo path impulse responses, respectively, to generate M partial predicted echoes;
These M partial prediction echoes are added to generate a frequency domain prediction echo,
Convert the predicted echo in the frequency domain to the time domain to generate a predicted echo,
Dividing the time domain predicted echo into blocks for each predetermined time;
Subtracting the predicted echo that has been blocked in the time domain from the collected sound signal that has been blocked at each predetermined time to generate an error signal before correction ,
Converts the uncorrected error signal frequency domain,
Short-time average of filter coefficient update step size weighted sum of received signal power input to each partial predicted echo path unit obtained last time, and that path of received signal power input to each partial predicted echo path unit determined this time The filter coefficient update step size weighted sum is weighted and added to obtain a short-time average of the current received signal power weighted sum. The product for each corresponding frequency component is obtained to correct the pre-correction error signal, and this is used as the frequency domain error signal for obtaining the correction vector,
Multiplying the non-delayed frequency domain block signal, the M-1 delayed block signals, and the frequency domain error signal for each corresponding frequency component to generate first to Mth modified vectors;
The first to Mth modified vectors are respectively multiplied by the first to Mth step sizes, and the first to Mth partial predicted echo path impulse responses are respectively updated with the multiplied modified vectors.
The acoustic echo canceling method according to claim 1, wherein the first to M-th step sizes are larger as the A-th (A = 1,..., M) number A is smaller. The number of the receiving terminals is P (P is an integer of 2 or more). For each received signal at each receiving terminal, the block signal is converted into the frequency domain, the M-1 delayed block signals are generated, and the M Generation of partial prediction echoes, generation of prediction echoes in the frequency domain, generation of error signals before correction , generation of the frequency domain error signals, generation of M correction vectors, first to Mth partial prediction echoes Update the path impulse response respectively
3. The acoustic echo canceling method according to claim 2 , wherein P prediction echoes in the frequency domain are added and converted to a time domain to obtain a prediction echo in the time domain. A second blocking unit that divides the reception signal from the reception terminal every predetermined time;
A second frequency domain transforming unit that transforms the divided received signal into a frequency domain;
A partial prediction echo path unit that generates a partial prediction echo by filtering the reception signal by taking a product of the reception signal in the frequency domain and a partial prediction impulse response every predetermined time; An inverse FFT transform unit that re-transforms the sum of partial prediction echoes generated at a plurality of predetermined times into the time domain for each predetermined time to generate a prediction echo;
A block shaping unit that divides the prediction echo in the time domain into blocks every predetermined time;
A first block-forming unit that receives a sound collection signal and divides the sound collection signal every predetermined time;
A subtractor that eliminates the acoustic echo by subtracting the predicted echo that is blocked in the time domain from the collected sound signal that is blocked every predetermined time;
A first frequency domain transforming unit that transforms an error signal before correction, which is an error signal between the collected sound signal at every predetermined time and the predicted echo, into a frequency domain;
Short-time average of filter coefficient update step size weighted sum of received signal power input to each partial predicted echo path unit obtained last time, and that path of received signal power input to each partial predicted echo path unit determined this time The filter coefficient update step size weighted sum is weighted and added to obtain a short-time average of the received signal power weighted sum, and the inverse of the short-time average of the received signal power weighted sum and the error signal before correction in the frequency domain. A correction matrix calculation unit for calculating an error signal by calculating a product for each corresponding frequency component;
A multiplier that multiplies each received signal of the plurality of predetermined times and the error signal for each component in the frequency domain to generate a correction vector;
An acoustic echo cancellation comprising: an updating unit that updates the partial predicted impulse response of each partial predicted echo path unit in the frequency domain using a different step size and a corresponding correction vector at each predetermined time apparatus. A second blocking unit that divides the reception signal from the reception terminal into blocks every predetermined time;
A second frequency domain transforming unit that transforms each of the divided block signals into a frequency domain;
A delay unit that delays the frequency domain block signal and generates M-1 (M is an integer of 2 or more) delay block signals, the delay amount of which is sequentially increased by the predetermined time;
The non-delayed frequency domain block signal and the M−1 delayed block signals are respectively filtered by first to Mth partial predicted echo path impulse responses in the frequency domain to generate M partial predicted echoes. 1 multiplication unit,
An adder for adding these M partial prediction echoes to generate a frequency domain prediction echo; an inverse FFT conversion unit for converting the frequency domain prediction echoes into the time domain and generating a prediction echo;
A block shaping unit that divides the predicted echo in the time domain into blocks every predetermined time;
A subtraction unit that generates a pre-correction error signal by subtracting the prediction echo blocked in the time domain from the sound pickup signal blocked every predetermined time;
A first frequency domain converter that converts the pre-correction error signal into a frequency domain;
Short-time average of filter coefficient update step size weighted sum of received signal power input to each partial predicted echo path unit obtained last time, and that path of received signal power input to each partial predicted echo path unit determined this time The filter coefficient update step size weighted sum is weighted and added to obtain a short-time average of the current received signal power weighted sum, and the inverse of the short-time average of the received signal power weighted sum and the uncorrected error signal in the frequency domain. A correction matrix calculation unit that calculates a product for each corresponding frequency component to correct the pre-correction error signal, and calculates this as a frequency domain error signal for determining the correction vector;
A second multiplier for multiplying the non-delayed frequency domain block signal, the M-1 delayed block signals, and the frequency domain error signal for each corresponding frequency component to generate first to Mth correction vectors;
A third multiplier for multiplying the first to Mth modified vectors by the first to Mth step sizes, respectively;
An acoustic echo canceller , comprising: an update unit that updates the first to Mth partial predicted echo path impulse responses with the corrected vector thus multiplied . The acoustic echo canceller according to claim 4 is connected in parallel to P (P is an integer of 1 or more) reception terminals in parallel with Q (Q is an integer of 2 or more), and these Q acoustic echo cancellers are connected to the Q acoustic echo cancellers. The acoustic echo canceling apparatus is characterized in that different sound pickup signals are inputted. Acoustic echo cancellation program for executing each process of the acoustic echo cancellation method according to any one of claims 1 to 3 on a computer. A computer-readable recording medium on which the acoustic echo cancellation program according to claim 7 is recorded.

Images