JP3916934B2 - Acoustic parameter encoding, decoding method, apparatus and program, acoustic signal encoding, decoding method, apparatus and program, acoustic signal transmitting apparatus, acoustic signal receiving apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To realize vector encoding/decoding having low quality degradation of LSP parameters in a silence interval or a steady state noise interval. SOLUTION: While conducting vector encoding/decoding of LSP parameters of moving average type voice, a vector of the spectrum equivalent to a steady noise interval or a vector from which beforehand obtained average value vector of a voice interval is subtracted, is stored as one vector C0 of a vector coding table 14A and a spectrum equivalent to a silence interval or a steady noise is outputted as one of the coded vectors.

Description

となる。ここで、w₀からw_mまでの重み係数の和を1あるいは、それに近い値であるとすると、y(n)は、C ₀+y _ave、すなわち、無音区間のLSPパラメータから求めたF、あるいはそれに近いベクトルを量子化ベクトルとして出力できるようになり、無音区間あるいは、定常雑音区間における符号化性能を高くすることができる。上述のような構成により、ベクトル符号帳１４Ａには１つの符号ベクトルとしてベクトルFの成分を含むベクトルが格納されることになる。このベクトルＦの成分を含む符号ベクトルとしては、量子化パラメータ生成部１５が平均ベクトルy _aveの成分を含む量子化ベクトルy(n)を生成する場合には、ベクトルFから平均ベクトルy _aveを減算したものを用い、平均ベクトルy _aveの成分を含まない量子化ベクトルy(n)を生成する場合は、ベクトルFそのものを用いる。
【００３０】
図２は、この発明の実施例を適用した復号化装置の構成例であり、符号帳２４と量子化パラメータ生成部２５とから構成されている。これら符号帳２４と量子化パラメータ生成部２５は図１の符号化装置における符号帳１４と量子化パラメータ生成部１５とそれぞれ同様に構成されている。図１の符号化装置より送られたパラメータ符号としてのインデックスIx(n), Iw(n)が入力され、インデックスIx(n)に対応する符号ベクトルx(n)がベクトル符号帳２４Ａより出力され、また、インデックスIw(n)に対応する重み係数セットw₀,w₁,...,w_mが、係数符号帳２４Ｂより出力される。ベクトル符号帳２４Ａからフレームごとに出力された符号ベクトルx(n)は直列接続されたバッファ部25B₁,...,25B_mに順次入力される。現フレームｎの符号ベクトルx(n)と、バッファ部25B₁,...,25B_mの1,...,mフレーム過去の符号ベクトルx(n-1),...,x(n-m)とに重み係数w₀,w₁,...,w_mを乗算器25A₀,25A₁,...,25A_mで乗算し、これら乗算結果を加算器２５Ｄで加算し、さらに、レジスタ２５Ｃに予め保持された音声信号全体のLSPパラメータの平均ベクトルy _aveを加算器２５Ｄに加えることにより得られた量子化ベクトルy(n)を復号LSPパラメータとして出力する。y _aveは、有声部の平均ベクトルあるいは、零ベクトルzとしておくことも可能である。
【００３１】
この発明ではこの復号化装置においても、図１に示した符号化装置と同じく、ベクトルC ₀をその一符号ベクトルとしてベクトル符号帳２４Ａに格納しておくことにより、音響信号の無音区間あるいは、定常雑音区間で求めたLSPパラメータベクトルFを出力することができる。
図１中の加算器１５Ｄ、図２中の加算器２５Ｄで平均ベクトルy _aveを加算しない（零ベクトルとする）場合は、ベクトル符号帳１４Ａ、２４Ａに無音区間あるいは定常雑音区間に相当するLSPパラメータベクトルFがベクトルC ₀の代わりに一つの符号ベクトルとして格納される。以下の説明においては各ベクトル符号帳１４Ａ、２４Ａに格納するLSPパラメータベクトルF又はベクトルC ₀を代表してベクトルC ₀として表記する。
【００３２】
図３に、図１中のベクトル符号帳１４Ａ、あるいは、図２中のベクトル符号帳２４Ａの構成例をベクトル符号帳４Ａとして示す。この例は１段のベクトル符号帳４１を用いた場合であり、ベクトル符号帳４１には、Ｎ個の符号ベクトルx ₁,...,x _Nがそのまま格納されており、入力されたインデックスIx(n)に応じて、Ｎ個の符号ベクトル中のいずれかが選択され出力される。この発明では、その中の１つの符号ベクトルxとして前記符号ベクトルC ₀が用いられている。ベクトル符号帳４１のＮ個の符号ベクトルは例えば従来と同様に学習により作られるが、この発明では、その内のベクトルC ₀に最も類似した（歪の小さい）１個のベクトルがC₀でおきかえられるか、あるいは単に追加される。
【００３３】
ベクトルC ₀を求めるにはいくつかの方法がある。１つとして、通常、無音区間あるいは定常雑音区間においては、入力音響信号のスペクトル包絡が平坦になるので、例えばｐ次のLSPパラメータベクトルFの場合、0〜πをp+1等分してπ/(1+p), 2π/(1+p), ..., π/(1+p)のように大きさがほぼ等間のｐ個の値をLSPパラメータベクトルとして使用してもよい。あるいは、無音区間、定常雑音区間の実際のLSPパラメータベクトルFからC ₀=F-y _aveにより求める。あるいは、白色雑音又はHoth雑音を入力してそのLSPパラメータを求め、パラメータベクトルFとして使用し、C ₀=F-y _aveを求めてもよい。なお音声信号全体のLSPパラメータの平均ベクトルy _aveは、一般にはベクトル符号帳４１の符号ベクトルxを学習する際に、全学習用ベクトルの平均ベクトルとして求めておく。
音響パラメータとしてp=10次のLSPパラメータを使用し、無音区間又は定常雑音区間のLSPパラメータを０〜πの間の値に正規化した１０次元のベクトルC ₀, y _ave 及びＦの例を次の表１に示す。
【００３４】
【表１】

ベクトルFはこの発明により符号帳に書き込まれる無音区間・定常雑音区間を代表するLSPパラメータの符号ベクトルの例である。このベクトルの要素の値はほぼ一定の間隔で増加しており、これは周波数スペクトラムがほぼ平坦なことを意味している。
実施例２
図４は、図１のLSPパラメータ符号器のベクトル符号帳１４Ａ、あるいは、図２のLSPパラメータ復号化装置のベクトル符号帳２４Ａの他の構成例を符号帳４Ａとして２段のベクトル符号帳を用いた場合である。１段目の符号帳４１にはＮ個のｐ次元符号ベクトルx ₁₁,...,x _1Nが格納されており、２段目の符号帳４２にはN'個のｐ次元符号ベクトルx ₂₁,...,x _2N'が格納されている。
【００３５】
まず、符号ベクトルを指定するインデックスIx(n)が入力されると、符号解析部４３でそのインデックスIx(n)を解析して、第１段目の符号ベクトルを指定するインデックスIx(n)₁と、第2段目の符号ベクトルを指定するインデックスIx(n)₂を得る。そして、各段のインデックスIx(n)₁，Ix(n)₂にそれぞれ対応するｉ番目及びi'番目の符号ベクトルx _1i, x _2i'を１段目符号帳４１、及び２段目符号帳４２より読み出し、加算部４４にて、両符号ベクトルを加算し、加算結果を符号ベクトルx(n)として出力する。
【００３６】
この２段構成のベクトル符号帳の場合、符号ベクトル探索は、まず１段目の符号帳４１のみを使って量子化歪が最小のものから順に所定数の候補符号ベクトルまで行う。この探索は、図１で示した係数符号帳１４Ｂの重み係数セットとの組み合わせで行う。次に、各候補の１段目符号ベクトルと２段目符号帳のそれぞれの符号ベクトルとの組み合わせについて量子化歪が最小となる符号ベクトルの組み合わせを探索する。
このように１段目符号帳４１を優先して符号ベクトルの探索を行う場合は、多段ベクトル符号帳４Ａの１段目符号帳４１内の１つの符号ベクトルとして前記符号ベクトルC ₀（又はF）を予め格納しておき、かつ、２段目の符号帳４２内の１つの符号ベクトルとして、零ベクトルzを予め格納しておく。これにより、符号帳４１から符号ベクトルC ₀が選択された場合に、符号帳４２から零ベクトルzが選択され、その結果、加算器４４から符号帳４Ａの出力として、無音区間や、定常雑音区間に相当する場合の符号ベクトルC ₀を出力できるような構成を実現している。零ベクトルzを格納しておかず、符号帳４１から符号ベクトルC ₀が選択された場合には符号帳４２からの選択加算を行わないように構成してもよい。
【００３７】
１段目符号帳４１の各符号ベクトルと、２段目符号帳の各符号ベクトルのすべての組み合わせについて探索を行う場合は、符号ベクトルC ₀と零ベクトルzは互いに別の符号帳であればどちらの符号帳に格納してもよい。符号ベクトルC ₀と零ベクトルzは無音区間や定常雑音区間では同時に選択される可能性が高いが、計算誤差その他の関係でこれらは必ずしも同時に選択されない場合もある。各段の符号帳において符号ベクトルC ₀や零ベクトルzは他の符号ベクトルの選択と同様に１つの符号ベクトルとして選択対象になる。
【００３８】
２段目符号帳４２に零ベクトルを格納しないでもよい。その場合、１段目符号帳４１からベクトルC ₀が選択されたときは、符号帳４２から符号ベクトルの選択を行わず、加算器４４から符号帳４１の符号C ₀をそのまま出力すればよい。
図４のように符号帳４Ａを多段の符号帳で構成することにより、選択可能な符号ベクトルの組み合わせ数だけ符号ベクトルを設けたことと実効的に同じであり、従って、図３のような１段の符号帳のみの場合と比べ、符号帳のサイズ（ここでは符号ベクトルの総数）を小さくすることができる利点がある。図４では２段のベクトル符号帳４１、４２で構成した場合を示したが、段数が３以上の場合は単に追加段数だけ符号帳を追加し、それぞれの段に対するインデックスによりそれぞれの段の符号帳から符号ベクトルを選択し、それらをベクトル合成するだけなので、容易に拡張可能である。
実施例３
図５は、図４の実施例のベクトル符号帳４Ａにおいて、１段目符号帳４１の各符号ベクトルに対し、予め決めたスケーリング係数を２段目符号帳４２から選択される符号ベクトルに対し乗算して、１段目符号帳４１からの符号ベクトルに加算して出力する場合である。スケーリング係数符号帳４５が設けられ、１段目符号帳４１のそれぞれの符号ベクトルx ₁₁,...,C ₀,...,x _1Nに対応して予め学習により決めた例えば0.5〜２程度のスケーリング係数s₁,...,s_Nが格納されており、１段目符号帳４１と共通のインデックスIx(n)₁によりアクセスされる。
【００３９】
まず、符号ベクトルを指定するインデックスIx(n)が入力されると、符号解析部４３でそのインデックスIx(n)を解析して、第１段目の符号ベクトルを指定するインデックスIx(n)₁と、第２段目の符号ベクトルを指定するインデックスIx(n)₂を得る。インデックスIx(n)₁に対応する符号ベクトルx _1iを１段目符号帳４１から読み出す。また、スケーリング係数符号帳４５から、そのインデックスIx(n)₁に対応したスケーリング係数s_iが読み出される。次に、インデックスIx(n)₂に対応する符号ベクトルx _2i'を２段目符号帳４２から読み出し、スケーリング係数s_iを乗算器４６において、２段目符号帳４２からの符号ベクトルx _2i'に乗算する。乗算により得られたベクトルと、１段目符号帳４１からの符号ベクトルx _1iを加算部４４で加算し、加算結果を符号帳４Ａからの符号ベクトルx(n)として出力する。
【００４０】
この実施例においても、符号ベクトルの探索は、まず、１段目符号帳４１のみを使用して量子化歪が最小のものから順に所定数の候補符号ベクトルを探索する。次に、各候補符号ベクトルと、２段目符号帳４２のそれぞれの符号ベクトルとの組み合わせについて量子化歪が最小となる組を探索する。この場合、スケーリング係数付き多段ベクトル符号帳４Ａに対して、１段目符号帳４１内の１つの符号ベクトルとして前記ベクトルC ₀を予め格納しておき、かつ、２段目の符号帳４２の１つの符号ベクトルとして、零ベクトルzを予め格納しておく。図４の場合と同様に、２つの符号帳４１、４２の符号ベクトル間の全組み合わせについて探索を行うのであれば、符号ベクトルC ₀と零ベクトルzは互いに別々の符号帳に格納するのであればどちらに格納してもよい。あるいは、前述の実施例と同様に零ベクトルzを符号帳に格納しなくてもよい。格納しない場合は、符号ベクトルC ₀が選択されたときには符号帳４２からの選択加算を行わない。
【００４１】
このようにして、無音区間や、定常雑音区間に相当する場合の符号ベクトルを出力することができる。符号ベクトルC ₀と零ベクトルzは無音区間や定常雑音区間では同時に選択される可能性が高いが計算誤差その他の関係でこれらは必ずしも同時に選択されない場合もある。各段の符号帳において、符号ベクトルC ₀や零ベクトルzは他の符号ベクトルの選択と同様に１つの符号ベクトルとして選択対象になる。図５の実施例のように、スケーリング係数符号帳４５を使用することにより、スケーリング係数の数Ｎだけ２段目符号帳を設けたことと実効的に同じであり、従って、より量子化歪の小さい符号化が実現できる利点がある。
実施例４
図６は、図１のパラメータ符号化装置のベクトル符号帳１４Ａ、あるいは、図２のパラメータ復号化装置のベクトル符号帳２４Ａを分割ベクトル符号帳４Ａとして構成しこの発明を適用した場合を示す。図６は２分割のベクトル符号帳で構成したが、分割数が３以上の場合も同様に拡張可能であるので、ここでは、分割数２の場合の実現について述べる。
【００４２】
この符号帳４Ａでは、Ｎ個の低次符号ベクトルx _L1,...,x _LNを格納した低次ベクトル符号帳４１_Lと、N'個の高次符号ベクトルx _H1,...,x _HN'を格納した高次ベクトル符号帳４１_Hを備える。出力符号ベクトルをx(n)とすると、低次及び高次ベクトル符号帳４１_L，４１_Hは、p次のうち、1〜k次までを低次、k+1〜p次までを高次として、それぞれの次元数のベクトルからなる符号帳を構成する。即ち、低次ベクトル符号帳４１_Lのｉ番目のベクトルは、
x _Li = (x_Li1, x_Li2, ..., x_Lik) (9)
で表され、高次ベクトル符号帳４１_Hのi'番目のベクトルは、
x _Hi' = (x_Hi'k+1, x_Hi'k+2, ..., x_Hi'p) (10)
で表される。入力されたインデックスIx(n)は解析部４３でIx(n)_LとIx(n)_Hに分けられ、これらIx(n)_LとIx(n)_Hに応じて各符号帳４１_L，４１_Hより、それぞれ低次と高次の分割ベクトルx _Li, x _Hi'が選択され、統合部４７でこれら分割ベクトルx _Li, x _Hi'が統合されて、出力符号ベクトルx(n)を生成する。即ち、統合部４７から出力される符号ベクトルをx(n)とすると、
x(n)＝(x_Li1, x_Li2, ..., x_Lik | x_Hi'k+1, x_Hi'k+2, ..., x_Hi'p)
となる。
【００４３】
この実施例では、低次ベクトルの符号帳４１_Lの１つのベクトルとして前記ベクトルC ₀の低次ベクトルC _0Lを格納し、かつ、高次ベクトルの符号帳４１_Hの１つのベクトルとして、前記ベクトルC ₀の高次ベクトルC _0Hを格納する。このようにして、無音区間や、定常雑音区間に相当する場合の符号ベクトルとして、
C ₀=(C _0L|C _0H)
を出力できる構成を実現している。しかも場合によっては、C _0Lと他の高次ベクトルの組合せ、あるいは他の低次ベクトルとC _0Hの組合せとして出力されることもある。図６のように分割ベクトル符号帳４１_L、４１_Hを設ければ、２つの分割ベクトルの組み合わせの数だけ符号ベクトルを設けたことと等価なので、それぞれの分割ベクトル符号帳のサイズを小さくすることができる利点がある。
実施例５
図７は、図１の音響パラメータ符号化装置のベクトル符号帳１４Ａ、あるいは、図２の音響パラメータ復号化装置のベクトル符号帳２４Ａの更に他の構成例を示し、符号帳４Ａを多段・分割ベクトル符号帳として構成した場合である。この符号帳４Ａは、図４の符号帳４Ａにおいて２段目の符号帳４２を図６と同様の２分割のベクトル符号帳で構成したものである。
【００４４】
１段目符号帳４１にはＮ個の符号ベクトルx ₁₁,...,x _1Nが格納されており、２段目低次符号帳４２_LにはN'個の分割ベクトルx _2L1,...,x _2LN'が格納されており、２段目高次符号帳４２_HにはN"個の分割ベクトルx _2H1,...,x _2HN"が格納されている。
入力されたインデックスIx(n)は符号解析部４３₁にて、１段目の符号ベクトルを指定するインデックスIx(n)₁と、２段目の符号ベクトルを指定するインデックスIx(n)₂とに解析される。１段目のインデックスIx(n)₁に対応するｉ番目の符号ベクトルx _1iを１段目符号帳４１より読み出す。また、２段目のインデックスIx(n)₂は解析部４３₂でIx(n)_2LとIx(n)_2Hに解析され、これらIx(n)_2L，Ix(n)_2Hにより２段目低次分割ベクトル符号帳４２_Lと、２段目高次分割ベクトル符号帳４２_Hのそれぞれi'番目及びi"番目の分割ベクトルx _2Li'及びx _2Hi"を選択し、これら選択された分割ベクトルが統合部４７でベクトル統合され、２段目の符号ベクトルx _2i'i"が生成される。加算部４４にて、１段目の符号ベクトルx _1iと２段目の統合ベクトルx _2i'i"が加算され、符号ベクトルx(n)として出力される。
【００４５】
この実施例では、図４及び５の実施例と同様に、１段目の符号帳４１の１つの符号ベクトルとして前記ベクトルC ₀を格納し、かつ、２段目の分割ベクトル符号帳４２の低次分割ベクトル符号帳４２_L、高次分割ベクトル符号帳４２_Hのそれぞれ１つずつの分割ベクトルとして、分割零ベクトルz _L, z _Hを格納しておく。このようにすることにより、無音区間や、定常雑音区間に相当する場合の符号ベクトルを出力する構成を実現している。符号帳の段数は３以上でもよい。また分割ベクトル符号帳は任意の段に対して用いてもよく、また１段当りの分割ベクトル符号帳の数も２に限らない。分割する段数も１以上でもよい。更に、１段目符号帳４１と２段目符号帳４２_L、４２_H間のすべての組の符号ベクトルについて探索を行うのであれば、ベクトルC ₀及び分割零ベクトルz _L, z _Hは、互いに異なる段のどの符号帳に格納してもよい。あるいは、実施例２及び３と同様に分割零ベクトルを符号帳に格納しなくてもよい。格納しない場合は、ベクトルC ₀が選択されたときに符号帳４２_L, ４２_Hからの選択加算を行わない。
実施例６
図８は、図７の実施例のベクトル符号帳４Ａにおける分割ベクトル符号帳４２の低次符号帳４２_Lと高次符号帳４２_Hに対し、図５の実施例におけるスケーリング係数符号帳４５と同様のスケーリング係数符号帳４５_Lと４５_Hを設けた、スケーリング係数付き多段・分割ベクトル符号帳４Ａに、本発明を適用した例である。低次と高次の分割ベクトルにそれぞれ乗じるための係数として、低次スケーリング係数符号帳４５_Lと高次スケーリング係数符号帳４５_Hには、それぞれＮ個の値が例えば約0.5〜２程度の係数を格納しておく。
【００４６】
入力されたインデックスIx(n)は、解析部４３₁で、１段目の符号ベクトルを指定するインデックスIx(n)₁と、２段目の符号ベクトルを指定するインデックスIx(n)₂とに解析される。まず、インデックスIx(n)₁に対応する符号ベクトルx _1iを１段目符号帳４１から求める。また、インデックスIx(n)₁に対応して低次スケーリング係数符号帳４５_Lと高次スケーリング係数符号帳４５_Hとから、それぞれ低次スケーリング係数s_Liと高次スケーリング係数s_Hiが読み出される。次に、インデックスIx(n)₂は、解析部４３₂で、インデックスIx(n)_2LとIx(n)_2Hに解析され、これら、Ix(n)_2LとIx(n)_2Hにより２段目低次分割ベクトル符号帳４２_Lと２段目高次分割ベクトル４２_Hのそれぞれの分割ベクトルx _2Li', x _2Hi"を選択する。それら選択した分割ベクトルに対し、乗算器４６_L、４６_Hにおいて低次と高次のスケーリング係数s_Li, s_Hiを乗じて得られたベクトルを統合部４７で統合し、２段目の符号ベクトルx _2i'i"が生成される。加算部４４にて、１段目の符号ベクトルx _1iと２段目の統合ベクトルx _2i'i"を加算し、加算結果は符号ベクトルx(n)として出力される。
【００４７】
この実施例のスケーリング係数付き多段・分割ベクトル符号帳４Ａでは、１段目符号帳４１内の１つの符号ベクトルとして前記ベクトルC ₀を格納し、かつ、２段目の分割ベクトル符号帳の低次分割ベクトル符号帳４２_L、高次分割ベクトル符号帳４２_Hに分割ベクトルとして分割零ベクトルz _L, z _Hをそれぞれ格納する。こうすることにより、無音区間や定常雑音区間に相当する場合の符号ベクトルを出力する構成を実現している。符号帳の段数は３つ以上でもよい。その場合、２段目以降の２つ以上の段をそれぞれ分割ベクトル符号帳で構成してもよい。また、いずれの場合も１段当りの分割ベクトル符号帳の数に限らない。
実施例７
図９は、図１の音響パラメータ符号化装置のベクトル符号帳４Ａ、あるいは、図２の音響パラメータ復号化装置のベクトル符号帳２４Ａの更に別の構成例を示し、図７の実施例における１段目符号帳４１についても図６の実施例と同様の分割ベクトル符号帳により構成した場合である。この実施例では、１段目低次符号帳４１_LにはＮ個の低次分割ベクトルx _1L1,...,x _1LNが格納され、１段目高次符号帳４１_HにはN'個の高次分割ベクトルx _1H1,...,x _HN'が格納され、２段目低次符号帳４２_LにはN"個の低次分割ベクトルx _2L1,...,x _2LN"が格納され、２段目高次符号帳４２_HにはN'"個の高次分割ベクトルx _2H1,...,x _2HN'"が格納されている。
【００４８】
入力されたインデックスIx(n)は符号解析部４３にて、１段目のベクトルを指定するインデックスIx(n)₁と、２段目のベクトルを指定するインデックスIx(n)₂とに解析される。１段目のインデックスIx(n)₁に対応するベクトルを、１段目低次分割ベクトル符号帳４１_L、及び１段目高次分割ベクトル符号帳４１_Hのそれぞれｉ番目及びi'番目の分割ベクトルx _1Li, x _1Hi'を選択し、これらを統合部４７₁で統合して１段目の統合ベクトルx _1ii'を生成する。
また、２段目のインデックスIx(n)₂も1段目と同様に、２段目低次分割ベクトル符号帳４２_Lと、２段目高次分割ベクトル符号帳４２_Hのそれぞれi"番目及びi'"番目の分割ベクトルx _2Li", x _2Hi'"を選択し、これらを統合部４７₂で統合して２段目の統合ベクトルx _2i"i'"を生成する。加算部４４にて、１段目の統合ベクトルx _1ii'と２段目の統合ベクトルx _2i"i'"を加算し、加算結果を符号ベクトルx(n)として出力する。
【００４９】
この実施例では、１段目においては、図６の分割ベクトル符号帳構成と同様に、１段目の低次ベクトルの符号帳４１_Lの１つの符号ベクトルとして前記ベクトルC ₀の低次分割ベクトルC _0Lを格納し、かつ、１段目の高次ベクトルの符号帳４１_Hの１つの分割ベクトルとして、前記ベクトルC ₀の高次分割ベクトルC _0Hを格納し、かつ、２段目の分割ベクトル符号帳４２の低次分割ベクトル符号帳４２_L、２段目の高次分割ベクトル符号帳４２_Hのそれぞれ１つずつのベクトルとして、分割零ベクトルz _L, z _Hを格納する。この構成により無音区間や、定常雑音区間に相当する場合の符号ベクトルを出力できる構成を実現している。この場合も、多段の数は２に限らず、１段当りの分割ベクトル符号帳の数も２に限らない。
実施例８
図１０は、この発明を適用した音声信号送信装置および受信装置の構成を示すブロック図である。
【００５０】
音声信号101は、入力装置102によって電気的信号に変換されA/D変換装置103に出力される。A/D変換装置103は入力装置102から出力された（アナログ）信号をディジタル信号に変換し音声符号化装置104へ出力する。音声符号化装置104はA/D変換装置103から出力されたディジタル音声信号を後述する音声符号化方法を用いて符号化し、符号化情報をＲＦ変調装置105へ出力する。ＲＦ変調装置105は音声符号化装置104から出力された音声符号化情報を電波等の伝播媒体に載せて送出するための信号に変換し送信アンテナ106へ出力する。送信アンテナ106はＲＦ変調装置105から出力された出力信号を電波（ＲＦ信号）107として送信する。以上が音声信号送信装置の構成および動作である。
【００５１】
送信された電波（ＲＦ信号）108は受信アンテナ109によって受信され、ＲＦ復調装置110へ出力される。なお、図中の電波（ＲＦ信号）108は受信側から見た電波（ＲＦ信号）107のことであり、伝播路において信号の減衰や雑音の重畳がなければ電波（ＲＦ信号）107と全く同じものとなる。ＲＦ復調装置110は受信アンテナ109から出力されたＲＦ信号から音声符号化情報を復調して音声復号化装置111へ出力する。音声復号化装置111はＲＦ復調装置110から出力された音声符号化情報から後述する音声復号化方法を用いて音声信号を復号してD/A変換装置112へ出力する。D/A変換装置112は音声復号化装置111から出力されたディジタル音声信号をアナログの電気的信号に変換して出力装置113へ出力する。出力装置113は電気的信号を空気の振動に変換し音波114として人間の耳に聴こえるように出力する。以上が音声信号受信装置の構成および動作である。
【００５２】
上記のような音声信号送信装置および受信装置の少なくとも一方を備えることにより、移動通信システムにおける基地局装置および移動端末装置を構成することができる。
前記音声信号送信装置は、音声符号化装置104にその特徴を有する。図１１は音声符号化装置104の構成を示すブロック図である。
入力音声信号は図１０のA/D変換装置103から出力される信号であり、前処理部200に入力される。前処理部200ではＤＣ成分を取り除くハイパスフィルタ処理や後続する符号化処理の性能改善につながるような波形整形処理やプリエンファシス処理を行い、その処理された信号ＸinをLPC分析部201および加算器204およびパラメータ決定部212に出力する。LPC分析部201は、Ｘinについて線形予測分析を行い、分析結果（線形予測係数）をLPC量子化部202へ出力する。LPC量子化部202は、LSPパラメータ算出部１３と、パラメータ符号化部１０と、復号部１８と、パラメータ変換部１９とから構成されている。パラメータ符号化部１０は、図３〜９のいずれかの実施例によるこの発明のベクトル符号帳が適用された図１におけるパラメータ符号化部１０と同様の構成とされている。また、復号部１８も図３〜９のいずれかの符号帳が適用された図２の復号化装置と同様の構成とされている。
【００５３】
LPC分析部201から出力された線形予測係数（LPC）は、LSPパラメータ算出部１３でLSPパラメータに変換され、得られたLSPパラメータはパラメータ符号化部１０で図１を参照して説明したように符号化される。符号化して得た符号Ix(n), Iw(n)、即ち、量子化LPCを表す符号Ｌは多重化部２１３へ出力されると共に、それらの符号Ix(n), Iw(n)を復号部１８で復号して量子化されたLSPパラメータを得て、それをパラメータ変換部１９で再びLPCパラメータに変換し、その得られた量子化LPCパラメータを合成フィルタ203に与える。合成フィルタ203は、前記量子化LPCをフィルタ係数とし、加算器210から出力される駆動音源信号に対してフィルタ処理により音響信号を合成し、合成信号を加算器204へ出力する。
【００５４】
加算器204は前記Ｘinと前記合成信号との誤差信号εを算出し、聴覚重み付け部211へ出力する。聴覚重み付け部211は、加算器204から出力された誤差信号εに対して聴覚的な重み付けをおこない、聴覚重み付け領域での前記Ｘinに対する前記合成信号の歪みを算出し、パラメータ決定部212へ出力する。パラメータ決定部212は、聴覚重み付け部211から出力された前記符号化歪みが最小となるように、適応符号帳205と固定符号帳207と量子化利得生成部206とから生成されるべき信号を決定する。なお、聴覚重み付け部211から出力される符号化歪みの最小化だけでなく、前記Ｘinを用いた別の符号化歪み最小化方法を併用して前記３つの手段から生成されるべき信号を決定することにより、さらに符号化性能を改善することもできる。
【００５５】
適応符号帳205は、前記歪みが最小化された時の過去に加算器210によって出力された直前フレームn-1の音源信号をバッファリングしており、そのパラメータ決定部212から出力された適応ベクトル符号Ａによって特定される位置から音源ベクトルを切り出し、それを１フレーム長になるまで繰り返し繋いで所望の周期成分を含む適応ベクトルを生成し、乗算器208へ出力する。固定符号帳207には複数の１フレーム長の固定ベクトルが固定ベクトル符号に対応して格納されており、パラメータ決定部212から出力された固定ベクトル符号Ｆによって特定される形状を有する固定ベクトルを乗算器209へ出力する。
【００５６】
量子化利得生成部206は、パラメータ決定部212から出力された利得符号Ｇによって特定される適応ベクトルと固定ベクトルに対する量子化適応ベクトル利得g_Aと量子化固定ベクトル利得g_Fをそれぞれ乗算器208と209に与える。乗算器208は、量子化利得生成部206から出力された量子化適応ベクトル利得g_Aを、適応符号帳205から出力された適応ベクトルに乗じて、加算器210へ出力する。乗算器209は、量子化利得生成部206から出力された量子化固定ベクトル利得g_Fを、固定ベクトル符号帳207から出力された固定ベクトルに乗じて、加算器210へ出力する。
【００５７】
加算器210は、利得乗算後の適応ベクトルと固定ベクトルとをベクトル加算して合成フィルタ203および適応符号帳205へ出力する。最後に多重化部213は、LPC量子化部202から量子化LPCを表す符号Ｌを、パラメータ決定部212から適応ベクトルを表す適応ベクトル符号Ａおよび固定ベクトルを表す固定ベクトル符号Ｆおよび量子化利得を表す利得符号Ｇを、それぞれ入力し、これらの符号を多重化して符号化情報として伝送路へ出力する。
図１２は、図１０中の音声復号化装置111の構成を示すブロック図である。
【００５８】
図において、ＲＦ復調部110から出力された符号化情報は、多重化分離部1301によって多重化されている符号化情報を個々の符号Ｌ．Ａ，Ｆ，Ｇに分離される。分離されたLPC符号ＬはLPC復号化部1302に与えられ、分離された適応ベクトル符号Ａは適応符号帳1305に与えられ、分離された利得符号Ｇは量子化利得生成部1306に与えられ、分離された固定ベクトル符号Ｆは固定符号帳1307へ与えられる。LPC復号化部1302は図２と同様に構成された復号部1302Aと、パラメータ変換部1302Bとから構成されている。多重化分離部1301から与えられた符号Ｌ=(Ix(n), Iw(n))は復号部1302Aで、図２に示したように、LSPパラメータ領域で復号し、それをLPCに変換し、合成フィルタ1303に出力する。
【００５９】
適応符号帳1305は、多重化分離部1301から出力された適応ベクトル符号Ａで指定される位置から適応ベクトルを取り出して乗算器1308へ出力する。固定符号帳1307は、多重化分離部1301から出力された固定ベクトル符号Ｆで指定される固定ベクトルを生成し、乗算器1309へ出力する。量子化利得生成部1306は、多重化分離部1301から出力された利得符号Ｇで指定される適応ベクトル利得g_Aと固定ベクトル利得g_Fを復号し乗算器1308および1309へそれぞれ出力する。乗算器1308は、前記適応符号ベクトルに前記適応符号ベクトル利得g_Aを乗算して、加算器1310へ出力する。乗算器1309は、前記固定符号ベクトルに前記固定符号ベクトル利得g_Fを乗算して、加算器1310へ出力する。加算器1310は、加算器1308 及び1309から出力された利得乗算後の適応ベクトルと固定ベクトルの加算を行い、合成フィルタ1303へ出力する。合成フィルタ1303は、加算器1310から出力されたベクトルを駆動音源信号として、LPC復号化部1302によって復号されたフィルタ係数を用いて、フィルタ合成を行い、合成した信号を後処理部1304へ出力する。後処理部1304は、ホルマント強調やピッチ強調といったような音声の主観的な品質を改善する処理や、定常雑音の主観的品質を改善する処理などを施した上で、最終的な復号音声信号として出力する。
【００６０】
上述では音声信号のスペクトル包絡を表わす線形予測係数と等価なパラメータとしてLSPパラメータを用いたが、他のパラメータ、例えばαパラメータ、パーコール係数などを用いてもよい。これらを用いた場合も、無音区間や定常雑音区間においてはスペクトラム包絡が平坦になるため、この区間でのパラメータの計算は容易に行うことができ、例えばｐ次のαパラメータの場合は０次を1.0, 1〜p次を0.0とすればよい。その他の音響パラメータを使用する場合でも、ほぼ平坦なスペクトラム包絡を表すように決めた音響パラメータのベクトルであればよい。なおLSPパラメータは量子化効率が良い点から実用的である。
【００６１】
上述において、ベクトル符号帳として、多段構成する場合は、ベクトルC ₀を例えばC ₀＝C ₀₁＋C ₀₂と二つの合成ベクトルで表わし、C ₀₁、C ₀₂を互いに異なる段の符号帳に格納してもよい。
更にこの発明は音声信号の符号化、復号化のみならず音楽信号など一般の音響信号の符号化、復号化にも適用できる。
また、この発明の装置はコンピュータによりプログラムを実行させて音響信号の符号化及び復号化を行うこともできる。図１３は図３〜９のいずれかのこの発明による符号帳を使用した、図１及び２の音響パラメータ符号化装置及び復号化装置、更にその符号化方法及び復号化方法を適用した図１１及び１２の音響信号符号化装置及び復号化装置をコンピュータで実行する実施形態を示す。
【００６２】
この発明を実施するコンピュータは通信網に接続されたモデム410、音響信号の入出力を行う入出力インタフェース420、ディジタル音響信号あるいは音響信号符号を一時的に蓄積するバッファメモリ430、符号化及び復号化処理をそこで実行するランダムアクセスメモリ（ＲＡＭ）440、データの入出力及びプログラム実行の制御を行う中央処理装置（ＣＰＵ）450、符号化及び復号化プログラムが格納されたハードディスク460、記録媒体470Mを駆動する駆動装置470から構成され、これらは互いに共通のバス480で接続されている。
【００６３】
記録媒体470Mとしては、コンパクトディスクＣＤ，ディジタルビデオディスクＤＶＤ，磁気光学ディスクＭＯ，メモリカード、その他どのような種類の記録媒体を使用してもよい。ハートディスク460には、図１１及び１２の音響信号符号化装置及び復号化装置において実施する符号化方法及び復号化方法をコンピュータによる処理手順で表したプログラムが格納されている。そのプログラムは、サブルーチンとして図１及び２の音響パラメータ符号化及び復号化を実行するプログラムを含む。
【００６４】
入力音響信号を符号化する場合は、ＣＰＵ450はハードディスク460から音響信号符号化プログラムをＲＡＭ440に読み込み、入出力インタフェース420を介してバッファメモリ430に取り込んだ音響信号をフレームごとにＲＡＭ440内で符号化プログラムに従って処理を行うことにより符号化し、得られた符号を符号化音響信号データとして、例えばモデム410を介して通信網に送出する。あるいは一時的にハードディスク460に保存する。あるいは記録媒体駆動装置470により記録媒体470Mに書き込む。
【００６５】
入力符号化音響信号データを復号する場合は、ＣＰＵ450はハードディスク460から復号プログラムをＲＡＭ440に読み込む。音響符号データがモデム410を介して通信網からバッファメモリ430にダウンロードされるか、あるいは記録媒体470Mから駆動装置470によりバッファメモリ430に読み込まれる、ＣＰＵ440はフレームごとに音響符号データをＲＡＭ440において復号プログラムに従って処理し、得られた音響信号データを入出力インタフェース420から出力する。
【００６６】
【発明の効果】
図１４の表１に、効果を表す例として、この発明により符号帳に無音区間のベクトルC ₀と零ベクトルzを埋め込んだ場合と、従来のように符号帳にベクトルC ₀を埋め込まない場合の音響パラメータ符号化装置の量子化性能について示す。表１において、縦軸は、ケプストラム歪みであり、これは、対数スペクトル歪みに相当するものであり、デシベル(dB)表示としている。ケプストラム歪みの小さいほど量子化性能が良い事を示す。また、歪みを計算する音声区間として、すべての区間での平均（Total)、無音区間・音声の定常区間以外の区間(Mode 0)、及び音声の定常区間( Mode 1)で、平均歪みを求めた。無音区間の存在するのは、Mode 0 であり、そこでの歪みは、提案符号帳の方が0.11dB低く、無音及び零ベクトルを挿入する効果が有ることがわかる。また、Totalでのケプストラム歪みも提案符号帳を使った場合に低くなっており、音声定常区間でも劣化がないことから、この発明による符号帳の有効性が明らかである。
【００６７】
このように、この発明では、現在のフレームの符号ベクトルと過去に出力された符号ベクトルの重み付き和、あるいは、それに予め求めておいた平均ベクトルを加えたベクトルにより線形予測係数と等価なパラメータを量子化する符号化において、ベクトル符号帳に格納されているベクトルとして、無音区間あるいは定常雑音区間に相当するパラメータベクトル、あるいは、そのパラメータベクトルから前記平均ベクトルを差し引いたベクトルを符号ベクトルとして選択し、その符号を出力することを可能にしているので、これらの区間での品質劣化の少ない符号化・復号化方法及びその装置を提供できる。
【図面の簡単な説明】
【図１】この発明による符号帳が適用された音響パラメータ符号化装置の機能構成を示すブロック図。
【図２】この発明による符号帳が適用された音響パラメータ復号化装置の機能構成を示すブロック図。
【図３】 LSPパラメータ符号化及び復号化のためのこの発明によるベクトル符号帳の構成例を示す図。
【図４】多段構成とした場合のこの発明によるベクトル符号帳の構成例を示す図。
【図５】分割ベクトル符号帳で構成した場合のこの発明によるベクトル符号帳の構成例を示す図。
【図６】多段ベクトル符号帳にスケーリング係数を適用した場合のこの発明によるベクトル符号帳の構成例を示す図。
【図７】２段目符号帳を分割ベクトル符号帳で構成した場合のこの発明によるベクトル符号帳の構成例を示す図。
【図８】図７の符号帳における２つの分割ベクトル符号帳に対しそれぞれスケーリング係数を適用した場合のベクトル符号帳の構成例を示す図。
【図９】図４の多段ベクトル符号帳の各段を分割ベクトル符号帳とした場合のベクトル符号帳の構成例を示す図。
【図１０】Ａはこの発明による符号化方法が適用された音声信号伝送装置の構成例を示すブロック図、Ｂはこの発明による復号化方法が適用された音声信号受信装置の構成例を示すブロック図。
【図１１】この発明による符号化方法が適用された音声信号符号化装置の機能構成を示す図。
【図１２】この発明による復号化方法が適用された音声信号復号化装置の機能構成を示す図。
【図１３】この発明による符号化装置及び復号化装置をコンピュータで実施する場合の構成例を示す図。
【図１４】この発明の効果を説明するためのグラフ。[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a low bit rate audio signal encoding / decoding method and apparatus in a mobile communication system or the Internet for encoding and transmitting an audio signal such as an audio signal or a music signal, and an audio parameter encoding / decoding applied to them. The present invention relates to a computerization method and apparatus, and a program for executing these methods on a computer.
[0002]
[Prior art]
In the fields of digital mobile communication and voice storage, voice coding apparatuses that compress voice information and encode it with high efficiency are used in order to effectively use radio waves and storage media. Such a speech coding apparatus uses a method using a model suitable for representing a speech signal so that a high-quality speech signal can be represented even at a low bit rate. For example, a CELP (Code Excited Linear Prediction) method is a widely used method at a bit rate of 4 kbit / s to 8 kbit / s. For the CELP technology, see M.R.Schroeder and B.S.Atal: "Code-Excited Linear Prediction (CELP): High-quality Speech at Very Low Bit Rates", Proc. ICASSP-85, 25.1.1, pp.937-940, 1985 ".
[0003]
The CELP speech coding method is based on a speech synthesis model corresponding to a human speech mechanism, and a speech signal is obtained from a filter expressed by a linear prediction coefficient representing vocal tract characteristics and an excitation signal that drives the filter. Is synthesized. Specifically, the digitized speech signal is divided into certain frame lengths (about 5 ms to 50 ms), the speech signal is linearly predicted for each frame, and the prediction residual (excitation signal) by linear prediction is known. The encoding is performed using an adaptive code vector and a fixed code vector consisting of the waveform. The adaptive code vector is stored in the adaptive codebook as a vector representing a drive excitation signal generated in the past, and is used to represent a periodic component of the speech signal. The fixed code vector is stored as a vector having a predetermined number of waveforms prepared in advance in the fixed codebook, and is mainly used to express aperiodic components that cannot be expressed in the adaptive codebook. As a vector stored in the fixed codebook, a vector composed of a random noise sequence, a vector expressed by a combination of several pulses, or the like is used.
[0004]
An algebraic fixed codebook is one of the typical fixed codebooks that express the fixed code vector by a combination of several pulses. Specific contents of the algebraic fixed codebook are shown in “ITU-T Recommendation G.729” and the like.
In conventional speech coding systems, speech linear prediction coefficients are converted into parameters such as partial autocorrelation (PARCOR) coefficients and line spectrum pairs (also called line spectrum pairs (LSP)), and further quantized. After being converted into a digital code, it is stored or transmitted. Details of these methods are described in, for example, Sadaaki Furui, “Digital Audio Processing” (Tokai University Press).
[0005]
In the encoding of the linear prediction coefficient, the LSP parameter encoding method includes a weighted vector obtained by multiplying the code vector output from the vector codebook in the past one or more frames by the weighting coefficient selected from the weighting coefficient codebook. Alternatively, the quantization parameter of the current frame is expressed by a vector obtained by adding the average vector of the LSP parameters of the entire speech signal obtained in advance to this vector, and the quantization parameter is obtained from the input speech. The code vector to be output by the vector codebook and the weighting coefficient set to be output by the weighting coefficient codebook are selected so that the distortion with respect to the LSP parameter, that is, the quantization distortion is minimized or sufficiently small, and these are selected as the LSP parameter. Output as a code.
[0006]
This is generally referred to as weighted vector quantization or moving average (MA) prediction vector quantization if the weight coefficient is considered as a prediction coefficient from the past.
On the decoding side, from the received vector code and weighting coefficient code, the code vector of the current frame and the past code vector are multiplied by the weighting coefficient, or the average vector of the LSP parameters of the entire speech signal obtained in advance is added. Output as a quantized vector of the current frame.
[0007]
The vector codebook that outputs the code vector of each frame includes a basic one-stage vector quantizer, a divided vector quantizer obtained by dividing the vector dimension, a multistage vector quantizer having two or more stages, or A multistage / divided vector quantizer, which combines a multistage and a divided vector quantizer, is possible.
[0008]
[Problems to be solved by the invention]
In the above conventional LSP parameter encoder / decoder, the number of frames is large in the silent section and the stationary noise section, and the encoding process and the decoding process are multistage configurations. Thus, it was impossible to output a vector so that the synthesized parameters changed smoothly. This is because the vector codebook used for encoding is usually obtained by learning, but in this learning, there is no sufficient amount of silent intervals or stationary noise intervals in the learning speech. If the vector corresponding to the interval is not necessarily reflected and learned, or if the number of bits given to the quantizer is small, these codebooks contain sufficient quantization vectors corresponding to the non-speech interval. Could not be designed.
[0009]
In such an LSP parameter encoder / decoder, the quantization performance in the non-speech interval cannot be sufficiently exhibited in the encoding at the time of actual communication, and the quality of the reproduced sound cannot be avoided. Such a problem has occurred not only in the encoding of acoustic parameters equivalent to the linear prediction coefficient representing the spectral envelope of the audio signal, but also in the same encoding for music signals.
The object of the present invention has been made in view of the above points, and corresponds to a silence interval and a stationary noise interval in encoding / decoding of an acoustic parameter equivalent to a linear prediction coefficient representing a spectrum envelope of a conventional acoustic signal. By making it easier to output vectors, an acoustic parameter encoding / decoding method and apparatus with little quality degradation in these sections, an acoustic signal encoding / decoding method and apparatus using these, and these methods are provided. The object is to provide a computer-implemented program.
[0010]
[Means for Solving the Problems]
The present invention is originally intended for encoding / decoding acoustic parameters equivalent to linear prediction coefficients representing the spectral envelope of acoustic signals, that is, parameters such as LSP parameters, α parameters, and Percoll parameters (hereinafter simply referred to as acoustic parameters). One feature is that an acoustic parameter vector representing a substantially flat spectrum envelope corresponding to a silent section or a stationary noise section, which is not obtained by learning, is added to the codebook so that it can be selected. In the conventional technique, a vector including an acoustic parameter vector component representing a substantially flat spectrum envelope is calculated in advance and stored as one vector of a vector codebook, and a multistage vector quantization configuration, In the divided vector quantization configuration, the code vector is output.
[0011]
An acoustic parameter encoding method according to the present invention includes:
(a) calculating an acoustic parameter equivalent to a linear prediction coefficient representing a spectral envelope characteristic of an acoustic signal for each frame of a certain time length;
(b) From a vector codebook that stores a plurality of code vectors corresponding to indexes representing them, code vectors output in at least one previous past frame and code vectors selected in the current frame On the other hand, a weighted vector is generated by multiplying and adding each of the weighting coefficients of a set selected from the coefficient codebook that stores one or more sets of weighting coefficients corresponding to the indexes representing them. Obtaining a vector including a component as a candidate of a quantized acoustic parameter for the acoustic parameter of the current frame;
(c) A set of code vectors of the vector codebook and a set of weighting coefficients of the coefficient codebook is determined using a criterion so that distortion of the quantized acoustic parameter candidates with respect to the calculated acoustic parameter is minimized. Determining and outputting an index representing a set of the determined code vector and weight coefficient as a quantization code of the acoustic parameter;
Including
The vector codebook includes a vector including an acoustic parameter vector component representing a substantially flat spectrum envelope as one of stored code vectors.
[0012]
An acoustic parameter decoding method according to the present invention includes:
(a) a vector codebook storing a plurality of code vectors of acoustic parameters equivalent to linear prediction coefficients representing the spectral envelope characteristics of an acoustic signal in correspondence with indexes representing them, and one or more sets of weighting coefficients Outputting a code vector corresponding to an index represented by a code input for each frame and a set of weighting coefficients from a coefficient codebook stored corresponding to an index representing the set;
(b) The weight coefficient of the set output to the code vector output from the vector codebook in the nearest past frame and the code vector output from the vector codebook in the current frame, respectively. Multiplying and adding to generate a weighted vector and outputting a vector containing the components of the weighted vector as a decoded quantized vector of the current frame;
Including
The vector codebook includes a vector including an acoustic parameter vector component representing a substantially flat spectrum envelope as one of stored code vectors.
[0013]
The acoustic parameter encoding device according to the present invention is:
A parameter calculating means for analyzing the input acoustic signal for each frame and calculating an acoustic parameter equivalent to a linear prediction coefficient representing the spectral envelope characteristics of the acoustic signal;
A vector codebook that stores a plurality of code vectors corresponding to indexes representing them;
A coefficient codebook that stores one or more sets of weighting coefficients in correspondence with indexes representing those sets;
The code vector for the current frame output from the vector code book and the code vector output from the nearest past at least one frame are multiplied by each of the set of weight coefficients selected from the coefficient code book. A quantization parameter generating means for generating a weighted vector by adding together and outputting a vector including the generated component of the weighted vector as a candidate of a quantized acoustic parameter for the acoustic parameter of the current frame;
A distortion calculation unit that calculates distortion of the quantized acoustic parameter with respect to the acoustic parameter calculated by the parameter calculation unit;
Indexes representing the code vectors of the vector codebook and the set of coefficient codebooks are determined using a standard so that the distortion is minimized, and the set of the determined code vector and the set of weighting coefficients is respectively represented Codebook search control unit for outputting as a code of the acoustic parameters,
Including
The vector codebook is configured to include a vector including an acoustic parameter vector component representing a substantially flat spectrum envelope as one code vector.
[0014]
The acoustic parameter decoding device according to the present invention is:
A vector codebook storing a plurality of code vectors of acoustic parameters equivalent to linear prediction coefficients representing the spectral envelope characteristics of the acoustic signal in correspondence with indexes representing them;
A coefficient codebook storing one or more sets of weighting coefficients in correspondence with indexes representing them;
One code vector is output from the vector codebook according to an index represented by a code input for each frame, a set weight coefficient is output from the coefficient codebook, and the code vector output in the current frame; A weighted vector is generated by multiplying the code vector output in the nearest past frame by the weight coefficient of the set output in the current frame, and adding them together, and the components of the generated weight vector Quantization parameter generating means for outputting a vector including a decoded quantized acoustic parameter of the current frame,
And is configured to include
In the vector codebook, a vector including an acoustic parameter vector component representing a substantially flat spectrum envelope is stored as one of the code vectors.
[0015]
An acoustic signal encoding apparatus for encoding an input acoustic signal according to the present invention includes:
Means for encoding the spectral characteristics of the input acoustic signal using the acoustic parameter encoding method;
An adaptive codebook holding an adaptive code vector representing a periodic component of the input acoustic signal;
A fixed codebook storing a plurality of fixed vectors; and
A sound source vector generated based on the adaptive code vector from the adaptive code book and the fixed vector from the fixed code book is input as an excitation signal, and a synthesized acoustic signal is used using a filter coefficient based on the quantized acoustic parameter. Filter means for synthesizing
An adaptive code vector and a fixed vector selected from the fixed codebook and the adaptive codebook are determined so that distortion of the synthesized acoustic signal with respect to the input acoustic signal is reduced, and the determined adaptive code vector and fixed vector are respectively determined. Means for outputting corresponding adaptive and fixed codes;
And is configured to include.
[0016]
An acoustic signal decoding apparatus that decodes an input code according to the present invention and outputs an acoustic signal,
Means for decoding an acoustic parameter equivalent to a linear prediction coefficient representing a spectral envelope characteristic from an input code using the acoustic parameter decoding method;
A fixed codebook storing a plurality of fixed vectors; and
An adaptive codebook holding an adaptive code vector indicating the periodic components of the synthesized acoustic signal;
Means for taking out the corresponding fixed vector from the fixed codebook by the input adaptive code and fixed code, taking out the corresponding adaptive code vector from the adaptive codebook, and combining them to generate an excitation vector;
Filter means for setting a filter coefficient based on the acoustic parameter and reproducing an acoustic signal by the excitation vector;
And is configured to include.
[0017]
An acoustic signal encoding method for encoding an input acoustic signal according to the present invention includes:
(A) encoding the spectral characteristics of the input acoustic signal using the acoustic parameter encoding method;
(B) A sound source vector generated based on an adaptive code vector from an adaptive code book that holds an adaptive code vector indicating a periodic component of the input acoustic signal and a fixed vector from a fixed code book that stores a plurality of fixed vectors And generating a synthesized acoustic signal by synthesizing and filtering with a filter coefficient based on the quantized acoustic parameter,
(C) The adaptive code vector and fixed vector to be selected from the fixed codebook and the adaptive codebook are determined so that the distortion of the synthesized sound signal with respect to the input sound signal is reduced, and the determined adaptive code vector and the fixed codebook are fixed. Outputting an adaptive code and a fixed code corresponding to each vector,
Including.
[0018]
An acoustic signal decoding method for decoding an input code according to the present invention and outputting an acoustic signal is as follows:
(A) using the acoustic parameter decoding method, decoding an acoustic parameter equivalent to a linear prediction coefficient representing a spectral envelope characteristic from an input code;
(B) Based on the input adaptive code and fixed code, the adaptive code vector is extracted from the adaptive code book holding the adaptive code vector representing the periodic component of the input acoustic signal, and the fixed code book corresponding to the plurality of fixed vectors is stored. Extracting a fixed vector and combining the adaptive code vector and the fixed vector to generate an excitation vector;
(C) reconstructing a synthesized acoustic signal by synthesizing and filtering the excitation vector using a filter coefficient based on the acoustic parameter;
Including.
[0019]
The present invention described above can be provided in the form of a computer-executable program.
According to the present invention, in a weighted vector quantizer (or MA prediction vector quantizer), a vector including an acoustic parameter vector component representing a substantially flat spectrum envelope is obtained in advance as a code vector of a vector codebook. Therefore, the quantization vector corresponding to the acoustic parameter corresponding to the corresponding silent section or stationary noise section can be output.
[0020]
According to another embodiment of the present invention, when a multistage vector codebook is used as the configuration of the vector codebook of the acoustic parameter encoding apparatus / decoding apparatus, By storing a vector containing acoustic parameter vector components representing a substantially flat spectrum envelope, and by storing a zero vector in the codebook of the other stage, the corresponding silent section or stationary noise section A quantization vector corresponding to the acoustic parameter corresponding to can be output.
The zero vector is not necessarily stored. When the zero vector is not stored, when a vector including the acoustic parameter vector component representing the substantially flat spectrum envelope is selected from the codebook of one stage, the acoustic parameter vector component representing the substantially flat spectrum envelope is selected. Only the vector including the symbol may be output as the code vector candidate of the current frame.
[0021]
When the vector codebook is composed of divided vector codebooks, a plurality of divided vectors obtained by dividing the vector dimension including the acoustic parameter vector components representing a substantially flat spectrum envelope are used. By dividing and storing one by one in the divided vector codebook, each divided vector codebook is selected in the search of each divided vector codebook, and the vector obtained by integrating the divided vectors is assigned to the corresponding silent section or stationary noise section. A quantization vector corresponding to the corresponding acoustic parameter can be output.
[0022]
Furthermore, the vector quantizer has a multistage / divided vector quantization configuration, and by combining the technology of the multistage vector quantization configuration and the divided vector quantization configuration, acoustic parameters corresponding to the corresponding silent section or stationary noise section Can be output as a quantization vector corresponding to.
When the codebook has a multistage configuration, a scaling coefficient for each of the second and subsequent codebooks is provided as a scaling coefficient codebook corresponding to each code vector of the first stage codebook. By reading the scaling coefficient corresponding to the code vector selected in the code book from each scaling coefficient code book and multiplying the code vector selected from the second stage code book, encoding with smaller quantization distortion can be realized. .
[0023]
As described above, it is possible to provide an acoustic parameter encoding / decoding method and apparatus with little quality deterioration in the section, which is an object of the present invention.
In the acoustic signal encoding device according to the present invention, any one of the parameter encoding devices is used in an acoustic parameter region equivalent to the linear prediction coefficient in quantizing the linear prediction coefficient. According to this configuration, the same effect as any of the above can be obtained.
In the acoustic signal decoding device according to the present invention, any one of the parameter decoding devices is used in the acoustic parameter region equivalent to the linear prediction coefficient in decoding the linear prediction coefficient. According to this configuration, the same effect as any of the above can be obtained.
[0024]
DETAILED DESCRIPTION OF THE INVENTION
Example 1
Next, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing a configuration example of an acoustic parameter encoding apparatus according to an embodiment to which the linear prediction parameter encoding method according to the present invention is applied. This encoding apparatus includes a linear prediction analysis unit 12, an LSP parameter calculation unit 13, and a code book 14, a quantization parameter generation unit 15, a distortion calculation unit 16, and a code book search control unit 17 that constitute the parameter encoding unit 10. It consists of and. In the figure, a series of digitized audio signal samples, for example, are input from an input terminal T1. The linear prediction analysis unit 12 performs linear prediction analysis on the audio signal samples for each frame stored in the internal buffer, and calculates a set of linear prediction coefficients. In the following description, the underlined symbol represents a vector. Now, assuming that the order of the linear prediction analysis is p-order, the LSP parameter calculation unit 13 calculates an equivalent p-order LSP (line spectrum pair) parameter from the p-order linear prediction coefficient. Details of these treatment methods are described in the book of Furui described above. The p LSP parameters
f(n) = (f₁(n), f₂(n), ..., f_p(n)) Express as a vector as in (1). Here, the integer n indicates the number n of a certain frame, and the frame at that time is called a frame n.
[0025]
The codebook 14 is provided with a vector codebook 14A in which N code vectors representing the LSP parameter vector obtained by learning are stored, and a coefficient codebook 14B in which K weight coefficient sets are stored. The code vector corresponding to the index Ix (n) specifying the code vector and the index Iw (n) specifying the weight coefficient codex(n) and the weight coefficient set (w₀, w₁, ..., w_m) Is output. The quantization parameter generation unit 15 includes m buffer units 15B connected in series.₁, ..., 15B_m, M + 1 multiplier 15A₀, 15A₁, ..., 15A_m, A register 15C and a vector adder 15D. Code vector of current frame n selected as one of candidates from vector codebook 14Ax(n) and buffer 15B₁~ 15B_mCode vectors established for past frames n-1, ..., n-m stored inx(n-1), ...,xMultiplier 15A for each (n-m) code vector₀, ..., 15A_mThe weighting factor w of the set selected in₀, ..., w_mAre multiplied by the adder 15D, and the average vector of the LSP parameters of the entire audio signal obtained in advance is added.y _aveIs added from the register 15C to the adder 15D. In this way, a candidate for a quantization vector, that is, a candidate for an LSP parameter, from the adder 15D.y(n) is generated. Mean vectory _aveFor example, an average vector of the voiced part may be used, or a zero vector may be used as described later.
[0026]
Code vector selected from the vector codebook 14A for the current frame nx(n)
x(n) = (x₁(n), x₂(n), ..., x_p(n)) (2)
Similarly, the code vector determined one frame before isx(n-1) The code vector determined two frames beforex(n-2), the code vector determined m frames agoxIf (n-m) is set, quantization vector candidates for the current frame
y(n) = (y₁(n), y₂(n), ..., y_p(n)) (3) is
y(n) = w₀・x(n) + Σ_{j = 1} ^mw_j・x(n-j) +y _ave             It is expressed by (4). Here, the larger the value of m, the better the quantization efficiency, but the influence when a code error occurs extends to after m frames. Since m needs to be traced back to the past, m is appropriately selected as necessary. For voice communication, in the case of about 20 ms per frame, the value of m is 6 or less, and may be a value of 1 to 3. This m is also called a moving average prediction order.
[0027]
Quantization vector candidates obtained in this wayy(n) is sent to the distortion calculation unit 16 and the LSP parameter calculated by the LPS parameter calculation unit 13fCompute the quantization distortion for (n). The distortion d is defined by, for example, the following weighted Euclidean distance.
d = Σ_{i = 1} ^pr_i(f_i(n) −y_i(n))²                          (Five)
Where r_i, I = 1, ..., p are LSP parametersfWhen the weighting coefficient obtained from (n) is set so as to emphasize the vicinity of the formant frequency of the spectrum, the performance is good.
[0028]
The codebook search unit 17 sequentially changes the set of indexes Ix (n) and Iw (n) to be given to the codebook 14, and calculates the distortion d of Expression (5) as described above for each set of indexes. By repeating, the code vector of the vector code book 14A in the code book 14 and the weight coefficient set of the coefficient code book 14B are searched for those that minimize or sufficiently reduce the distortion d, which is the output of the distortion calculation unit 16. Then, these indices Ix (n) and Iw (n) are transmitted from the terminal T2 as the sign of the input LSP parameter. The codes Ix (n) and Iw (n) sent from the terminal T2 are sent to the decoder through the transmission path or stored in the storage device.
[0029]
Output code vector of the current framexWhen (n) is determined, the buffer unit 15B_jCode vector of past frame (n-j)x(n-j), j = 1, ..., m-1 are sequentially the next buffer unit 15B_{j + 1}The code vector of the current frame nx(n) is the buffer unit 15B₁Is input.
A feature of the present invention is that, as one code vector stored in the vector codebook 14A used in the weighted vector quantization of the LSP parameter as described above or the coding by moving average vector quantization, Mean vectory _aveLSP parameter vector corresponding to silence or stationary noiseFOry _aveIf is not zero, its LSP parameter vectorFFromy _aveVector minusC ₀Is to store. That is,y _aveIf is not zero, the LSP parameter vector corresponding to the silence interval or stationary noise interval is
F= (F₁, F₂, ..., F_p)
Then, the code vector to be stored in the vector codebook 14A of FIG.C ₀Is
C ₀=F−y _ave
Calculate as In coding by moving average prediction in a silent section or a stationary noise section, as a code vector, continuously over m framesC ₀Is selected, the quantization vectory(n) is

It becomes. Where w₀To w_mIf the sum of the weighting factors is 1 or close to it,y(n) isC ₀+y _aveThat is, calculated from the LSP parameter of the silent sectionFAlternatively, a vector close to it can be output as a quantization vector, and the encoding performance in a silent section or a stationary noise section can be improved. With the configuration as described above, the vector codebook 14A has a vector as one code vector.FWill be stored. As a code vector including the component of this vector F, the quantization parameter generation unit 15 uses an average vector.y _aveQuantization vector containing components ofyWhen generating (n), a vectorFMean vectory _aveIs the average vectory _aveQuantization vector that does not containyWhen generating (n), vectorFUse itself.
[0030]
FIG. 2 shows an example of the configuration of a decoding apparatus to which the embodiment of the present invention is applied, and includes a codebook 24 and a quantization parameter generation unit 25. The code book 24 and the quantization parameter generation unit 25 are configured in the same manner as the code book 14 and the quantization parameter generation unit 15 in the encoding apparatus of FIG. The index vectors Ix (n) and Iw (n) as the parameter codes sent from the encoding device in FIG. 1 are input, and the code vector corresponding to the index Ix (n)x(n) is output from the vector codebook 24A, and the weight coefficient set w corresponding to the index Iw (n)₀, w₁, ..., w_mIs output from the coefficient codebook 24B. Code vector output for each frame from the vector codebook 24Ax(n) is the buffer unit 25B connected in series.₁, ..., 25B_mAre sequentially input. Code vector of current frame nx(n) and buffer unit 25B₁, ..., 25B_m1, ..., m frame past sign vectorx(n-1), ...,x(n-m) and weighting factor w₀, w₁, ..., w_mThe multiplier 25A₀, 25A₁, ..., 25A_mAnd the multiplication results are added by the adder 25D, and the average vector of the LSP parameters of the entire audio signal held in the register 25C in advance.y _aveQuantized vector obtained by adding to the adder 25Dy(n) is output as a decryption LSP parameter.y _aveIs the mean vector of the voiced part or the zero vectorzIt is also possible to leave
[0031]
In the present invention, in this decoding apparatus as well as the encoding apparatus shown in FIG.C ₀Is stored as one code vector in the vector codebook 24A, so that the LSP parameter vector obtained in the silent section or stationary noise section of the acoustic signal is stored.FCan be output.
The adder 15D in FIG. 1 and the adder 25D in FIG.y _aveIs not added (zero vector), the LSP parameter vector corresponding to the silent section or the stationary noise section is added to the vector codebooks 14A and 24A.FIs a vectorC ₀Is stored as a single code vector. In the following description, LSP parameter vectors stored in the respective vector codebooks 14A and 24AFOr vectorC ₀On behalf of vectorC ₀It describes as.
[0032]
FIG. 3 shows a configuration example of the vector codebook 14A in FIG. 1 or the vector codebook 24A in FIG. 2 as a vector codebook 4A. In this example, a single-stage vector codebook 41 is used, and the vector codebook 41 includes N code vectors.x ₁, ...,x _NAre stored as they are, and one of the N code vectors is selected and output according to the input index Ix (n). In the present invention, one of the code vectorsxAs the sign vectorC ₀Is used. The N code vectors of the vector codebook 41 are created by learning, for example, as in the prior art.C ₀One vector (small distortion) most similar to is C₀Replaced by or simply added.
[0033]
vectorC ₀There are several ways to find out. For example, since the spectral envelope of the input acoustic signal is usually flat in a silent section or a stationary noise section, for example, a p-order LSP parameter vectorFIn the case of 0 to π, p + 1 is divided equally and the sizes are approximately equal, such as π / (1 + p), 2π / (1 + p), ..., π / (1 + p) May be used as the LSP parameter vector. Alternatively, the actual LSP parameter vector in the silent section and stationary noise sectionFFromC ₀=F-y _aveAsk for. Alternatively, input white noise or hot noise to find the LSP parameter, and use the parameter vectorFUse asC ₀=F-y _aveYou may ask for. The average vector of LSP parameters for the entire audio signaly _aveIs generally a code vector of the vector codebook 41xIs learned as an average vector of all learning vectors.
A 10-dimensional vector using the L = 10th order LSP parameter as the acoustic parameter and normalizing the LSP parameter in the silent section or stationary noise section to a value between 0 and π.C ₀,y _ave as well asFAn example of this is shown in Table 1 below.
[0034]
[Table 1]

vectorFIs an example of a code vector of an LSP parameter that represents a silent section and a stationary noise section written in the codebook according to the present invention. The value of this vector element increases at almost constant intervals, which means that the frequency spectrum is substantially flat.
Example 2
FIG. 4 uses a two-stage vector codebook as the codebook 4A as another example of the vector codebook 14A of the LSP parameter encoder of FIG. 1 or the vector codebook 24A of the LSP parameter decoder of FIG. This is the case. The first-stage codebook 41 has N p-dimensional code vectorsx ₁₁, ...,x _1NAre stored in the second-stage codebook 42.x _{twenty one}, ...,x _{2N '}Is stored.
[0035]
First, when an index Ix (n) that designates a code vector is input, the code analyzer 43 analyzes the index Ix (n), and an index Ix (n) that designates the first-stage code vector.₁And an index Ix (n) that specifies the second-stage code vector₂Get. And each stage index Ix (n)₁, Ix (n)₂I-th and i'-th code vectors corresponding respectively tox _1i,x _{2i '}Are read out from the first-stage codebook 41 and the second-stage codebook 42, the adder 44 adds both code vectors, and the addition result is the code vector.xOutput as (n).
[0036]
In the case of this two-stage vector codebook, first, the code vector search is performed using only the first-stage codebook 41 to the predetermined number of candidate code vectors in order from the one with the smallest quantization distortion. This search is performed in combination with the weighting coefficient set of the coefficient codebook 14B shown in FIG. Next, a combination of code vectors that minimizes the quantization distortion is searched for a combination of each candidate first-stage code vector and each code vector of the second-stage codebook.
When the code vector search is performed with priority given to the first-stage codebook 41 in this way, the code vector as one code vector in the first-stage codebook 41 of the multistage vector codebook 4A.C ₀(OrF) And a zero vector as one code vector in the second-stage codebook 42zIs stored in advance. As a result, the code vector from the codebook 41C ₀Is selected from codebook 42zAs a result, as an output from the adder 44 to the codebook 4A, a code vector corresponding to a silent section or a stationary noise sectionC ₀Is realized. Zero vectorzStored in the codebook 41C ₀It may be configured not to perform selective addition from the codebook 42 when is selected.
[0037]
When searching for all combinations of each code vector of the first stage codebook 41 and each code vector of the second stage codebook, the code vectorC ₀And zero vectorzMay be stored in either codebook as long as they are different codebooks. Sign vectorC ₀And zero vectorzAre likely to be selected at the same time in a silent section or a stationary noise section, but these may not always be selected at the same time due to calculation errors and other relationships. Code vector in each stage codebookC ₀Or zero vectorzIs selected as one code vector in the same manner as other code vectors.
[0038]
The zero vector may not be stored in the second stage codebook 42. In that case, the vector from the first stage codebook 41C ₀Is selected, the code vector is not selected from the code book 42, and the code of the code book 41 is added from the adder 44.C ₀Can be output as is.
By configuring the codebook 4A with a multistage codebook as shown in FIG. 4, it is effectively the same as providing code vectors for the number of selectable code vector combinations. There is an advantage that the size of the code book (here, the total number of code vectors) can be reduced compared to the case of only the stage code book. Although FIG. 4 shows the case of the two-stage vector codebooks 41 and 42, when the number of stages is three or more, the codebook is simply added by the number of additional stages, and the codebook of each stage is determined by the index for each stage Since the code vectors are simply selected from these and the vectors are synthesized, they can be easily expanded.
Example 3
FIG. 5 is a diagram illustrating multiplication of the code vector selected from the second-stage codebook 42 by a predetermined scaling coefficient for each code vector of the first-stage codebook 41 in the vector codebook 4A of the embodiment of FIG. In this case, the code vector from the first-stage codebook 41 is added and output. A scaling coefficient codebook 45 is provided, and each code vector of the first-stage codebook 41x ₁₁, ...,C ₀, ...,x _1NFor example, a scaling factor s of about 0.5 to 2 determined in advance by learning.₁, ..., s_NIs stored, and the index Ix (n) common to the first-stage codebook 41 is stored.₁Accessed by.
[0039]
First, when an index Ix (n) that designates a code vector is input, the code analyzer 43 analyzes the index Ix (n), and an index Ix (n) that designates the first-stage code vector.₁And an index Ix (n) that specifies the second-stage code vector₂Get. Index Ix (n)₁Sign vector corresponding tox _1iAre read from the first-stage codebook 41. Also, from the scaling coefficient codebook 45, its index Ix (n)₁Scaling factor corresponding to_iIs read out. Next, index Ix (n)₂Sign vector corresponding tox _{2i '}Is read from the second-stage codebook 42 and the scaling coefficient s_iIn the multiplier 46, the code vector from the second-stage codebook 42x _{2i '}Multiply by. Vector obtained by multiplication and code vector from first stage codebook 41x _1iAre added by the adder 44, and the addition result is a code vector from the codebook 4A.xOutput as (n).
[0040]
Also in this embodiment, the code vectors are searched for a predetermined number of candidate code vectors in order from the smallest quantization distortion using only the first-stage codebook 41. Next, a set that minimizes the quantization distortion is searched for a combination of each candidate code vector and each code vector of the second-stage codebook 42. In this case, with respect to the multistage vector codebook 4A with a scaling coefficient, the vector is used as one code vector in the first stage codebook 41.C ₀Are stored in advance, and a zero vector is used as one code vector of the codebook 42 in the second stage.zIs stored in advance. As in the case of FIG. 4, if the search is performed for all combinations between the code vectors of the two codebooks 41 and 42, the code vectorC ₀And zero vectorzMay be stored in either one of them if they are stored in different codebooks. Alternatively, as in the previous embodiment, the zero vectorzMay not be stored in the codebook. Code vector if not storedC ₀When is selected, the selective addition from the code book 42 is not performed.
[0041]
In this way, it is possible to output a code vector corresponding to a silent section or a stationary noise section. Sign vectorC ₀And zero vectorzAre likely to be selected at the same time in a silent section or a stationary noise section, but these may not always be selected at the same time due to calculation errors and other relationships. In each stage codebook, code vectorC ₀Or zero vectorzIs selected as one code vector in the same manner as other code vectors. The use of the scaling coefficient codebook 45 as in the embodiment of FIG. 5 is effectively the same as providing the second-stage codebook by the number N of scaling coefficients. There is an advantage that a small encoding can be realized.
Example 4
FIG. 6 shows a case where the vector codebook 14A of the parameter encoding device of FIG. 1 or the vector codebook 24A of the parameter decoding device of FIG. 2 is configured as a divided vector codebook 4A and the present invention is applied. Although FIG. 6 is composed of a two-divided vector codebook, it can be similarly expanded when the number of divisions is 3 or more, so here, the implementation when the number of divisions is 2 will be described.
[0042]
In this codebook 4A, N low-order code vectorsx _L1, ...,x _LNLow-order vector codebook 41 storing_LAnd N 'higher-order code vectorsx _H1, ...,x _{HN '}High-order vector codebook 41_HIs provided. Output code vectorxIf (n), the low-order and high-order vector codebook 41_L, 41_HConstitutes a codebook composed of vectors of respective dimensionalities, with the 1st to kth order being the lower order and the k + 1 to pth order being the higher order among the pth order. That is, the low-order vector codebook 41_LI-th vector is
x _Li = (x_Li1, x_Li2, ..., x_Lik(9)
The high-order vector codebook 41_HThe i'th vector of
x _{Hi '} = (x_{Hi'k + 1}, x_{Hi'k + 2}, ..., x_Hi'p) (Ten)
It is represented by The input index Ix (n) is converted into Ix (n) by the analysis unit 43._LAnd Ix (n)_HThese Ix (n)_LAnd Ix (n)_HEach codebook 41 according to_L, 41_HLower and higher order split vectors, respectivelyx _Li,x _{Hi '}Is selected, and the integration unit 47 selects these divided vectors.x _Li,x _{Hi '}Are integrated into the output code vectorxGenerate (n). That is, the code vector output from the integration unit 47 isx(n)
x(n) = (x_Li1, x_Li2, ..., x_Lik | x_{Hi'k + 1}, x_{Hi'k + 2}, ..., x_Hi'p)
It becomes.
[0043]
In this embodiment, a low-order vector codebook 41 is used._LVector as a vector ofC ₀Low-order vectorC _0LAnd a high-order vector codebook 41_HAs a vector, the vectorC ₀Higher order vectorC _0HIs stored. In this way, as a code vector when it corresponds to a silent section or a stationary noise section,
C ₀= (C _0L|C _0H)
Is realized. In some cases,C _0LAnd other higher-order vectors or other lower-order vectorsC _0HMay be output as a combination. The divided vector codebook 41 as shown in FIG._L, 41_HIs equivalent to providing as many code vectors as the number of combinations of two divided vectors, and there is an advantage that the size of each divided vector codebook can be reduced.
Example 5
FIG. 7 shows still another configuration example of the vector codebook 14A of the acoustic parameter encoding device of FIG. 1 or the vector codebook 24A of the acoustic parameter decoding device of FIG. This is a case where it is configured as a codebook. This codebook 4A is configured by configuring the second-stage codebook 42 in the codebook 4A of FIG. 4 with a two-part vector codebook similar to FIG.
[0044]
The first codebook 41 has N code vectors.x ₁₁, ...,x _1NAre stored, and the second-stage low-order codebook 42 is stored._LContains N 'partition vectorsx _2L1, ...,x _{2LN '}Are stored, and the second-stage higher-order codebook 42 is stored._HContains N "split vectorsx _2H1, ...,x _{2HN "}Is stored.
The input index Ix (n) is the code analysis unit 43.₁At index Ix (n) that specifies the first-stage code vector₁And index Ix (n) that specifies the second-stage code vector₂And analyzed. First level index Ix (n)₁I-th code vector corresponding tox _1iAre read from the first-stage codebook 41. The second index Ix (n)₂The analysis unit 43₂At Ix (n)_2LAnd Ix (n)_2HAnd these Ix (n)_2L, Ix (n)_2HBy the second stage low-order divided vector codebook 42_LAnd the second-stage higher-order divided vector codebook 42_HI'th and i "th partition vectors, respectivelyx _{2Li '}as well asx _{2Hi "}These divided vectors are vector-integrated by the integration unit 47, and the second stage code vectorx _{2i'i "}Is generated. In addition unit 44, the first-stage code vectorx _1iAnd second-stage integration vectorx _{2i'i "}Are added to the sign vectorxOutput as (n).
[0045]
In this embodiment, as in the embodiments of FIGS. 4 and 5, the vector is used as one code vector of the first-stage codebook 41.C ₀And the low-order divided vector codebook 42 of the divided vector codebook 42 in the second stage_L, High-order division vector codebook 42_HAs a split vector for each of, split zero vectorz _L,z _HIs stored. In this way, a configuration for outputting a code vector in a case corresponding to a silent section or a stationary noise section is realized. The number of codebook stages may be three or more. The divided vector codebook may be used for an arbitrary stage, and the number of divided vector codebooks per stage is not limited to two. The number of stages to be divided may be one or more. Further, the first-stage codebook 41 and the second-stage codebook 42_L, 42_HIf you want to search for all pairs of code vectorsC ₀And split zero vectorz _L,z _HMay be stored in any codebook in different stages. Alternatively, the divided zero vector may not be stored in the codebook as in the second and third embodiments. If not, vectorC ₀Codebook 42 when is selected_L, 42_HSelective addition from is not performed.
Example 6
FIG. 8 shows a low-order codebook 42 of the divided vector codebook 42 in the vector codebook 4A of the embodiment of FIG._LAnd high-order codebook 42_HIn contrast, the scaling coefficient codebook 45 similar to the scaling coefficient codebook 45 in the embodiment of FIG._LAnd 45_HThis is an example in which the present invention is applied to a multistage / divided vector codebook 4A with a scaling coefficient. A low-order scaling coefficient codebook 45 is used as a coefficient for multiplying the low-order and high-order divided vectors respectively._LAnd higher-order scaling coefficient codebook 45_HEach stores a coefficient having N values of about 0.5 to 2, for example.
[0046]
The input index Ix (n) is obtained from the analysis unit 43.₁And index Ix (n) that specifies the first-stage code vector₁And index Ix (n) that specifies the second-stage code vector₂And analyzed. First, index Ix (n)₁Sign vector corresponding tox _1iIs obtained from the first-stage codebook 41. Also, index Ix (n)₁Corresponding to the low-order scaling coefficient codebook 45_LAnd higher-order scaling coefficient codebook 45_HAnd low-order scaling factor s_LiAnd higher-order scaling factor s_HiIs read out. Next, index Ix (n)₂The analysis unit 43₂And index Ix (n)_2LAnd Ix (n)_2HAnd these, Ix (n)_2LAnd Ix (n)_2HBy the second stage low-order divided vector codebook 42_LAnd the second-stage higher-order division vector 42_HEach split vector ofx _{2Li '},x _{2Hi "}Select. The multiplier 46 is used for the selected divided vectors._L, 46_HLow and high order scaling factors at_Li, s_HiThe vectors obtained by multiplying by are integrated by the integration unit 47 and the second stage code vectorx _{2i'i "}Is generated. In addition unit 44, the first-stage code vectorx _1iAnd second-stage integration vectorx _{2i'i "}And the addition result is the sign vectorxOutput as (n).
[0047]
In the multistage / divided vector codebook 4A with scaling coefficients of this embodiment, the vector is used as one code vector in the first stage codebook 41.C ₀And the low-order divided vector codebook 42 of the second-stage divided vector codebook_L, High-order division vector codebook 42_HSplit zero vector as split vectorz _L,z _HIs stored respectively. In this way, a configuration for outputting a code vector corresponding to a silent section or a stationary noise section is realized. The number of codebook stages may be three or more. In that case, two or more stages after the second stage may be configured by a divided vector codebook. In any case, the number of divided vector codebooks per stage is not limited.
Example 7
FIG. 9 shows still another configuration example of the vector codebook 4A of the acoustic parameter encoding device of FIG. 1 or the vector codebook 24A of the acoustic parameter decoding device of FIG. The eye codebook 41 is also configured by a divided vector codebook similar to the embodiment of FIG. In this embodiment, the first-stage low-order codebook 41_LContains N low-order split vectorsx _1L1, ...,x _1LNAre stored, and the first-stage higher-order codebook 41 is stored._HContains N 'higher-order split vectorsx _1H1, ...,x _{HN '}Are stored, and the second-stage low-order codebook 42 is stored._LContains N "low-order split vectorsx _2L1, ...,x _{2LN "}Are stored, and the second-stage higher-order codebook 42 is stored._HContains N '"higher-order split vectorsx _2H1, ...,x _{2HN '"}Is stored.
[0048]
The input index Ix (n) is an index Ix (n) that designates the first-stage vector in the code analysis unit 43.₁And index Ix (n) that specifies the second vector₂And analyzed. First level index Ix (n)₁, The first-stage low-order divided vector codebook 41_L, And first-stage higher-order divided vector codebook 41_HI-th and i'-th partition vectors, respectivelyx _1Li,x _{1Hi '}Are selected and these are integrated.₁Integration vector of the first stage by integrating withx _{1ii '}Is generated.
The second index Ix (n)₂Similarly to the first stage, the second stage low-order divided vector codebook 42_LAnd the second-stage higher-order divided vector codebook 42_HI "th and i '" th partition vectors, respectivelyx _{2Li "},x _{2Hi '"}Are selected and these are integrated.₂Integration vector of the second stage by integrating withx _{2i "i '"}Is generated. In the addition unit 44, the first-stage integrated vectorx _{1ii '}And second-stage integration vectorx _{2i "i '"}And add the result to the sign vectorxOutput as (n).
[0049]
In this embodiment, in the first stage, as in the divided vector codebook configuration of FIG._LThe vector as one sign vector ofC ₀Low-order splitting vectorC _0LAnd the codebook 41 of the first-order higher-order vector_HAs one split vector ofC ₀Higher-order splitting vectorC _0HAnd the low-order divided vector codebook 42 of the divided vector codebook 42 in the second stage_LSecond-stage higher-order divided vector codebook 42_HAs a vector of one eachz _L,z _HIs stored. With this configuration, a configuration capable of outputting a code vector in a case corresponding to a silent interval or a stationary noise interval is realized. Also in this case, the number of multistages is not limited to two, and the number of divided vector codebooks per stage is not limited to two.
Example 8
FIG. 10 is a block diagram showing the configuration of an audio signal transmitting apparatus and receiving apparatus to which the present invention is applied.
[0050]
The audio signal 101 is converted into an electrical signal by the input device 102 and output to the A / D conversion device 103. The A / D conversion device 103 converts the (analog) signal output from the input device 102 into a digital signal and outputs it to the speech encoding device 104. The audio encoding device 104 encodes the digital audio signal output from the A / D conversion device 103 using an audio encoding method described later, and outputs the encoded information to the RF modulation device 105. The RF modulation device 105 converts the speech coding information output from the speech coding device 104 into a signal for transmission on a propagation medium such as a radio wave and outputs the signal to the transmission antenna 106. The transmission antenna 106 transmits the output signal output from the RF modulation device 105 as a radio wave (RF signal) 107. The above is the configuration and operation of the audio signal transmitting apparatus.
[0051]
The transmitted radio wave (RF signal) 108 is received by the receiving antenna 109 and output to the RF demodulator 110. The radio wave (RF signal) 108 in the figure is the radio wave (RF signal) 107 viewed from the receiving side, and is exactly the same as the radio wave (RF signal) 107 unless there is signal attenuation or noise superposition in the propagation path. It will be a thing. The RF demodulator 110 demodulates speech coding information from the RF signal output from the receiving antenna 109 and outputs the demodulated speech information to the speech decoder 111. The voice decoding device 111 decodes a voice signal from the voice coding information output from the RF demodulation device 110 using a voice decoding method described later and outputs the decoded voice signal to the D / A conversion device 112. The D / A converter 112 converts the digital audio signal output from the audio decoder 111 into an analog electrical signal and outputs it to the output device 113. The output device 113 converts an electrical signal into air vibration and outputs it as a sound wave 114 so that it can be heard by the human ear. The above is the configuration and operation of the audio signal receiving apparatus.
[0052]
By including at least one of the above-described audio signal transmitting apparatus and receiving apparatus, a base station apparatus and a mobile terminal apparatus in a mobile communication system can be configured.
The voice signal transmitting apparatus is characterized by the voice encoding apparatus 104. FIG. 11 is a block diagram showing a configuration of speech encoding apparatus 104.
The input audio signal is a signal output from the A / D conversion device 103 in FIG. 10 and is input to the preprocessing unit 200. The pre-processing unit 200 performs waveform shaping processing and pre-emphasis processing that leads to performance improvement of high-pass filter processing for removing DC components and subsequent encoding processing, and the processed signal Xin is processed by an LPC analysis unit 201 and an adder 204. And output to the parameter determination unit 212. The LPC analysis unit 201 performs linear prediction analysis on Xin and outputs the analysis result (linear prediction coefficient) to the LPC quantization unit 202. The LPC quantization unit 202 includes an LSP parameter calculation unit 13, a parameter encoding unit 10, a decoding unit 18, and a parameter conversion unit 19. The parameter encoding unit 10 has the same configuration as the parameter encoding unit 10 in FIG. 1 to which the vector codebook of the present invention according to any one of the embodiments of FIGS. The decoding unit 18 has the same configuration as that of the decoding device in FIG. 2 to which any of the codebooks in FIGS.
[0053]
The linear prediction coefficient (LPC) output from the LPC analysis unit 201 is converted into an LSP parameter by the LSP parameter calculation unit 13, and the obtained LSP parameter is explained by the parameter encoding unit 10 with reference to FIG. Encoded. The codes Ix (n) and Iw (n) obtained by encoding, that is, the code L representing the quantized LPC is output to the multiplexing unit 213 and the codes Ix (n) and Iw (n) are decoded. An LSP parameter decoded and quantized by the unit 18 is obtained, converted into an LPC parameter again by the parameter converting unit 19, and the obtained quantized LPC parameter is given to the synthesis filter 203. The synthesis filter 203 uses the quantized LPC as a filter coefficient, synthesizes an acoustic signal by filter processing with respect to the driving sound source signal output from the adder 210, and outputs the synthesized signal to the adder 204.
[0054]
The adder 204 calculates an error signal ε between the Xin and the combined signal and outputs the error signal ε to the auditory weighting unit 211. The auditory weighting unit 211 performs auditory weighting on the error signal ε output from the adder 204, calculates the distortion of the synthesized signal with respect to the Xin in the auditory weighting region, and outputs the distortion to the parameter determination unit 212. . The parameter determination unit 212 determines a signal to be generated from the adaptive codebook 205, the fixed codebook 207, and the quantization gain generation unit 206 so that the coding distortion output from the auditory weighting unit 211 is minimized. To do. In addition to minimizing the coding distortion output from the auditory weighting unit 211, a signal to be generated from the three means is determined by using another coding distortion minimizing method using the Xin. As a result, the encoding performance can be further improved.
[0055]
The adaptive codebook 205 buffers the excitation signal of the immediately preceding frame n−1 output by the adder 210 in the past when the distortion is minimized, and the adaptive vector output from the parameter determination unit 212 A sound source vector is cut out from the position specified by the code A, and is repeatedly connected until it becomes one frame length to generate an adaptive vector including a desired periodic component, which is output to the multiplier 208. Fixed codebook 207 stores a plurality of one-frame-length fixed vectors corresponding to the fixed vector code, and multiplies the fixed vector having a shape specified by fixed vector code F output from parameter determination unit 212. Output to the device 209.
[0056]
The quantization gain generation unit 206 is a quantization adaptive vector gain g for the adaptive vector and the fixed vector specified by the gain code G output from the parameter determination unit 212._AAnd quantized fixed vector gain g_FAre fed to multipliers 208 and 209, respectively. Multiplier 208 performs quantization adaptive vector gain g output from quantization gain generation section 206._AIs multiplied by the adaptive vector output from the adaptive codebook 205 and output to the adder 210. The multiplier 209 receives the quantized fixed vector gain g output from the quantized gain generator 206._FIs multiplied by the fixed vector output from the fixed vector codebook 207 and output to the adder 210.
[0057]
Adder 210 adds the adaptive vector after gain multiplication and the fixed vector, and outputs the result to synthesis filter 203 and adaptive codebook 205. Finally, the multiplexing unit 213 receives the code L representing the quantized LPC from the LPC quantizing unit 202, the adaptive vector code A representing the adaptive vector, the fixed vector code F representing the fixed vector, and the quantization gain from the parameter determining unit 212. Representing gain codes G are respectively input, these codes are multiplexed, and output to the transmission line as encoded information.
FIG. 12 is a block diagram showing a configuration of speech decoding apparatus 111 in FIG.
[0058]
In the figure, the encoded information output from the RF demodulator 110 includes the encoded information multiplexed by the demultiplexer 1301 as individual codes L.P. Separated into A, F, and G. The separated LPC code L is given to the LPC decoding unit 1302, the separated adaptive vector code A is given to the adaptive codebook 1305, and the separated gain code G is given to the quantization gain generating unit 1306, The fixed vector code F is given to the fixed codebook 1307. The LPC decoding unit 1302 includes a decoding unit 1302A configured in the same manner as in FIG. 2, and a parameter conversion unit 1302B. The code L = (Ix (n), Iw (n)) given from the demultiplexing unit 1301 is decoded by the decoding unit 1302A in the LSP parameter region as shown in FIG. 2, and converted into LPC. And output to the synthesis filter 1303.
[0059]
The adaptive codebook 1305 extracts an adaptive vector from the position specified by the adaptive vector code A output from the demultiplexing unit 1301 and outputs it to the multiplier 1308. Fixed codebook 1307 generates a fixed vector designated by fixed vector code F output from demultiplexing section 1301, and outputs the fixed vector to multiplier 1309. The quantization gain generation unit 1306 is an adaptive vector gain g designated by the gain code G output from the demultiplexing unit 1301._AAnd fixed vector gain g_FAre output to multipliers 1308 and 1309, respectively. The multiplier 1308 adds the adaptive code vector gain g to the adaptive code vector._AAnd output to the adder 1310. The multiplier 1309 adds the fixed code vector gain g to the fixed code vector._FAnd output to the adder 1310. Adder 1310 adds the gain-multiplied adaptive vector and fixed vector output from adders 1308 and 1309 and outputs the result to synthesis filter 1303. Synthesis filter 1303 uses the vector output from adder 1310 as a driving excitation signal, performs filter synthesis using the filter coefficients decoded by LPC decoding unit 1302, and outputs the synthesized signal to post-processing unit 1304 . The post-processing unit 1304 performs processing for improving the subjective quality of speech such as formant emphasis and pitch emphasis, processing for improving the subjective quality of stationary noise, and the like as a final decoded speech signal. Output.
[0060]
In the above description, the LSP parameter is used as a parameter equivalent to the linear prediction coefficient representing the spectral envelope of the speech signal. However, other parameters such as an α parameter and a Percoll coefficient may be used. Even when these are used, the spectrum envelope is flat in the silent section and the stationary noise section, so that the parameter calculation in this section can be easily performed. 1.0, 1 ~ p order should be 0.0. Even when other acoustic parameters are used, any acoustic parameter vector determined so as to represent a substantially flat spectrum envelope may be used. The LSP parameter is practical from the viewpoint of good quantization efficiency.
[0061]
In the above, when the multi-stage configuration is used as the vector codebook,C ₀For exampleC ₀=C ₀₁+C ₀₂And two composite vectors,C ₀₁,C ₀₂May be stored in different codebooks.
Furthermore, the present invention can be applied not only to encoding and decoding of audio signals but also to encoding and decoding of general acoustic signals such as music signals.
In addition, the apparatus of the present invention can execute a program by a computer to encode and decode an acoustic signal. FIG. 13 shows the acoustic parameter encoding apparatus and decoding apparatus of FIGS. 1 and 2 using the code book according to the present invention of any of FIGS. 3 to 9, and FIG. 11 and FIG. 11 to which the encoding method and decoding method are applied. 12 shows an embodiment in which twelve acoustic signal encoding apparatuses and decoding apparatuses are executed by a computer.
[0062]
A computer embodying the present invention includes a modem 410 connected to a communication network, an input / output interface 420 for inputting and outputting audio signals, a buffer memory 430 for temporarily storing digital audio signals or audio signal codes, and encoding and decoding Drives random access memory (RAM) 440 for executing processing, central processing unit (CPU) 450 for controlling data input / output and program execution, hard disk 460 storing encoding and decoding programs, and recording medium 470M These are connected to each other by a common bus 480.
[0063]
As the recording medium 470M, a compact disk CD, a digital video disk DVD, a magneto-optical disk MO, a memory card, or any other type of recording medium may be used. The heart disk 460 stores a program representing the encoding method and the decoding method implemented in the acoustic signal encoding device and the decoding device of FIGS. The program includes a program for executing the acoustic parameter encoding and decoding shown in FIGS. 1 and 2 as a subroutine.
[0064]
When encoding the input acoustic signal, the CPU 450 reads the acoustic signal encoding program from the hard disk 460 into the RAM 440, and the acoustic signal fetched into the buffer memory 430 via the input / output interface 420 is encoded in the RAM 440 for each frame. The obtained code is transmitted as encoded sound signal data to the communication network via the modem 410, for example. Alternatively, it is temporarily stored in the hard disk 460. Alternatively, the data is written on the recording medium 470M by the recording medium driving device 470.
[0065]
When decoding input encoded acoustic signal data, the CPU 450 reads a decoding program from the hard disk 460 into the RAM 440. The acoustic code data is downloaded from the communication network via the modem 410 to the buffer memory 430 or read from the recording medium 470M into the buffer memory 430 by the driving device 470. The CPU 440 decodes the acoustic code data in the RAM 440 for each frame. The obtained acoustic signal data is output from the input / output interface 420.
[0066]
【The invention's effect】
As an example showing the effect, Table 1 in FIG.C ₀And zero vectorzAnd embedded in the codebook as beforeC ₀The quantization performance of the acoustic parameter encoding apparatus when no symbol is embedded will be described. In Table 1, the vertical axis represents cepstrum distortion, which corresponds to logarithmic spectral distortion and is expressed in decibels (dB). The smaller the cepstrum distortion, the better the quantization performance. In addition, the average distortion in all sections, total silence, sections other than the silent section / speech steady section (Mode 0), and speech steady section (Mode 1) are obtained as the speech sections for which distortion is calculated. It was. The silent section exists in Mode 0. The distortion in the proposed codebook is 0.11 dB lower, and it can be seen that there is an effect of inserting silence and zero vectors. Further, the cepstrum distortion in Total is low when the proposed codebook is used, and there is no deterioration even in the steady speech section, so the effectiveness of the codebook according to the present invention is clear.
[0067]
As described above, according to the present invention, a parameter equivalent to the linear prediction coefficient is obtained by a weighted sum of the code vector of the current frame and the code vector output in the past, or a vector obtained by adding a previously obtained average vector. In the encoding to be quantized, as a vector stored in the vector codebook, a parameter vector corresponding to a silent interval or a stationary noise interval, or a vector obtained by subtracting the average vector from the parameter vector is selected as a code vector, Since it is possible to output the code, it is possible to provide an encoding / decoding method and apparatus with little quality deterioration in these sections.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a functional configuration of an acoustic parameter encoding apparatus to which a codebook according to the present invention is applied.
FIG. 2 is a block diagram showing a functional configuration of an acoustic parameter decoding apparatus to which a codebook according to the present invention is applied.
FIG. 3 is a diagram showing a configuration example of a vector codebook according to the present invention for LSP parameter encoding and decoding.
FIG. 4 is a diagram showing a configuration example of a vector codebook according to the present invention in a multi-stage configuration.
FIG. 5 is a diagram showing a configuration example of a vector codebook according to the present invention when configured with a divided vector codebook.
FIG. 6 is a diagram showing a configuration example of a vector codebook according to the present invention when a scaling coefficient is applied to a multistage vector codebook.
FIG. 7 is a diagram showing a configuration example of a vector codebook according to the present invention when the second-stage codebook is configured by a divided vector codebook.
8 is a diagram showing a configuration example of a vector codebook when scaling coefficients are applied to two divided vector codebooks in the codebook of FIG.
FIG. 9 is a diagram showing a configuration example of a vector codebook when each stage of the multistage vector codebook in FIG. 4 is a divided vector codebook.
FIG. 10A is a block diagram showing a configuration example of a speech signal transmission apparatus to which the encoding method according to the present invention is applied, and B is a block diagram showing a configuration example of a speech signal receiving apparatus to which the decoding method according to the present invention is applied. Figure.
FIG. 11 is a diagram showing a functional configuration of a speech signal encoding apparatus to which the encoding method according to the present invention is applied.
FIG. 12 is a diagram showing a functional configuration of an audio signal decoding apparatus to which a decoding method according to the present invention is applied.
FIG. 13 is a diagram showing a configuration example when the encoding device and the decoding device according to the present invention are implemented by a computer.
FIG. 14 is a graph for explaining the effect of the present invention.

Claims (29)

An acoustic parameter encoding method,
(a) calculating an acoustic parameter equivalent to a linear prediction coefficient representing a spectral envelope characteristic of an acoustic signal for each frame of a certain time length;
(b) From a vector codebook that stores a plurality of code vectors corresponding to indexes representing them, code vectors output in at least one previous past frame and code vectors selected in the current frame On the other hand, a weighted vector is generated by multiplying and adding each of the weighting coefficients of a set selected from the coefficient codebook that stores one or more sets of weighting coefficients corresponding to the indexes representing them. Obtaining a vector including a component as a candidate of a quantized acoustic parameter for the acoustic parameter of the current frame;
(c) A set of code vectors of the vector codebook and a set of weighting coefficients of the coefficient codebook is determined using a criterion so that distortion of the quantized acoustic parameter candidates with respect to the calculated acoustic parameter is minimized. Determining and outputting an index representing a set of the determined code vector and weight coefficient as a quantization code of the acoustic parameter;
Including
The vector codebook is composed of a plurality of codebooks each storing a plurality of vectors corresponding to indexes representing them, and the codebook of one stage of the plurality of codebooks is substantially flat. A vector containing acoustic parameter vector components representing a simple spectrum envelope is stored as one vector,
In the step (b) , when a code vector other than the vector including the component of the acoustic parameter vector representing the substantially flat spectrum envelope is selected from the one-stage codebook, each vector from the plurality of codebooks Are selected and output as the selected code vector of the current frame, and a vector including the component of the acoustic parameter vector representing the substantially flat spectrum envelope is selected from the codebook of the one stage. A step including outputting a vector including an acoustic parameter vector component representing the substantially flat spectrum envelope as the selected code vector of the current frame. 2. The encoding method according to claim 1 , wherein at least one of the plurality of codebooks is divided and stored as a plurality of divided vectors obtained by dividing the dimension of the code vector into a plurality of division vectors. A code book, and an integration unit that integrates the divided vectors output from the plurality of divided vector code books and outputs the integrated vector as an output vector of the code book at that stage. The encoding method according to claim 1, wherein said step. (b) When (c) Jointly search for a predetermined number of code vectors with the least distortion by a code vector selected from the one-stage codebook, and then each of the predetermined number of code vectors and the remaining The method includes the step of obtaining the distortion for all combinations with code vectors selected one by one from the codebook of each stage, and determining a set of code vectors having the smallest distortion. An acoustic parameter encoding method,
(a) calculating an acoustic parameter equivalent to a linear prediction coefficient representing a spectral envelope characteristic of the acoustic signal for each frame of a certain time length;
(b) From a vector codebook that stores a plurality of code vectors corresponding to indexes representing them, code vectors output in at least one previous frame and code vectors selected in the current frame On the other hand, a weighted vector is generated by multiplying and adding each of the weighting coefficients of a set selected from the coefficient codebook that stores one or more sets of weighting coefficients corresponding to the indexes representing them. Obtaining a vector including a component as a candidate of a quantized acoustic parameter for the acoustic parameter of the current frame;
(c) Using a criterion so that distortion of the quantized acoustic parameter candidate with respect to the calculated acoustic parameter is minimized, a set of the code vector of the vector codebook and the weight coefficient of the coefficient codebook is determined. , the step of an index that represents the set of those determined code vector and the weighting factor determines and outputs a quantization code of the acoustic parameters,
Including
The vector codebook includes a plurality of codebooks each storing a plurality of codevectors and scaling coefficient codebooks provided for the second and subsequent codebooks. Stores a predetermined scaling coefficient corresponding to each code vector of the first-stage codebook,
In one stage of the multi-stage codebook, a vector containing acoustic parameter vector components representing the substantially flat spectrum envelope is stored .
Above Symbol step (b),
When a code vector other than the vector including the acoustic parameter vector component representing the substantially flat spectrum envelope is selected from the one-stage codebook, the second and subsequent stages for the code vector selected in the first stage Reads the corresponding scaling coefficient from the scaling codebook for the codebook of, multiplies the code vector selected from the codebook of the second and subsequent stages, outputs the multiplication result as a vector of each stage, and outputs the respective stages The vector and the first-stage vector are added together, and the addition result is output as a code vector from the vector codebook. An acoustic parameter vector representing the substantially flat spectrum envelope is output from the one-stage codebook. When a vector containing components is selected, the acoustic parameters representing the above-mentioned almost flat spectrum envelope are selected. Step of outputting the vector containing the components of Tabekutoru as a code vector from the vector codebook,
including. 5. The encoding method according to claim 4 , wherein at least one codebook of the second and subsequent stages among the plurality of codebooks is distributed and stored as a plurality of divided vectors obtained by dividing the code vector dimension into a plurality of dimensions. A plurality of divided vector codebooks,
The scaling coefficient codebook corresponding to the codebook of at least one stage includes a plurality of division vector scaling coefficient codebooks provided for the plurality of division vector codebooks, and each of the division vector scaling coefficients Each code vector of the code book stores a predetermined division vector scaling coefficient corresponding to each code vector of the first stage code book,
Step (b) above is
When a code vector other than a vector including an acoustic parameter vector component representing the substantially flat spectrum envelope is selected from the one-stage codebook, each selected from the plurality of divided vector codebooks of the at least one stage The divided vectors are read from the divided vector scaling coefficient codebooks corresponding to the index of the vector selected in the first-stage codebook, respectively multiplied , and multiplied. Integrating the divided vectors obtained in this way and outputting them as output vectors of the codebook at that stage ,
Including the. 6. The encoding method according to claim 1, wherein a vector including an acoustic parameter vector component representing the substantially flat spectrum envelope is obtained in advance from a parameter vector equivalent to the linear prediction coefficient. This is a vector generated by subtracting an average vector of parameters equivalent to the above linear prediction coefficient of the entire acoustic signal. 7. The encoding method according to claim 1 , wherein the parameter equivalent to the linear prediction coefficient is an LSP parameter. An acoustic parameter decoding method,
(a) a vector codebook storing a plurality of code vectors of acoustic parameters equivalent to linear prediction coefficients representing the spectral envelope characteristics of an acoustic signal in correspondence with indexes representing them, and one or more sets of weighting coefficients Outputting a code vector corresponding to an index represented by a code input for each frame and a set of weighting coefficients from a coefficient codebook stored corresponding to an index representing the set;
(b) The weight coefficient of the set output to the code vector output from the vector codebook in the nearest past frame and the code vector output from the vector codebook in the current frame, respectively. Multiplying and adding to generate a weighted vector and outputting a vector containing the components of the weighted vector as a decoded quantized vector of the current frame;
Including
The vector codebook is composed of a plurality of codebooks each storing a plurality of vectors corresponding to indexes representing them, and the codebook of one stage of the plurality of codebooks is substantially flat. Vector containing the components of the acoustic parameter vector representing the correct spectral envelope
Toru is stored as a vector,
When the code vector other than the vector including the component of the acoustic parameter vector representing the substantially flat spectrum envelope is output from the one-stage code book, the step (b) Are selected and output as the selected code vector of the current frame, and a vector including the component of the acoustic parameter vector representing the substantially flat spectrum envelope is selected from the codebook of the one stage. A step of outputting a vector including a component of an acoustic parameter vector representing the substantially flat spectrum envelope as a code vector of the current frame. 9. The decoding method according to claim 8 , wherein at least one of the plurality of codebooks is divided and stored as a plurality of divided vectors obtained by dividing the dimension of the code vector into a plurality of division vectors. A code book, and an integration unit that integrates the divided vectors output from the plurality of divided vector code books and outputs the integrated vectors as the output vector of the code book at that stage. An acoustic parameter decoding method,
(a) a vector codebook storing a plurality of code vectors of acoustic parameters equivalent to linear prediction coefficients representing the spectral envelope characteristics of the acoustic signal in correspondence with indexes representing them, and one or more sets of weighting coefficients Outputting a code vector corresponding to an index represented by a code input for each frame and a set of weighting coefficients from a coefficient codebook stored corresponding to an index representing the set;
(b) The set weight coefficients of the output sets for the code vector output from the vector codebook in the nearest past frame and the code vector output from the vector codebook in the current frame, respectively. Multiplying and adding to generate a weighted vector and outputting a vector containing the components of the weighted vector as a decoded quantized vector of the current frame;
Including
The vector codebook includes a plurality of codebooks each storing a plurality of codevectors and scaling coefficient codebooks provided for the second and subsequent codebooks. Stores a predetermined scaling coefficient corresponding to each code vector of the first-stage codebook,
A vector including the acoustic parameter vector component representing the substantially flat spectrum envelope is stored in one of the plurality of codebooks, and one codebook is stored in each of the remaining codebooks. Zero vectors are stored one by one,
Step (b) above is
When a code vector other than the vector including the acoustic parameter vector component representing the substantially flat spectrum envelope is output from the one-stage codebook, the second and subsequent stages with respect to the code vector selected in the first stage Read out the corresponding scaling factor from the scaling codebook for the codebook of, multiply the code vector selected from the codebook of the second and subsequent stages, and output the multiplication result as a vector of each stage ,
Vector addition of the output vector of each stage and the vector of the first stage, and output the addition result as a code vector from the vector codebook ,
When a vector including an acoustic parameter vector component representing the substantially flat spectrum envelope is selected from the one-stage codebook, a vector including the acoustic parameter vector component representing the substantially flat spectrum envelope is selected as the vector code. Outputting as a code vector from the book ,
including. 11. The decoding method according to claim 10, wherein among the plurality of codebooks, at least one codebook of the second and subsequent stages is distributed and stored as a plurality of divided vectors obtained by dividing the code vector dimension into a plurality of dimensions. A plurality of divided vector codebooks,
The scaling coefficient codebook corresponding to the codebook of the at least one stage includes a plurality of division vector scaling coefficient codebooks provided for the plurality of division vector codebooks, and each of the division vector scaling coefficient codes The book stores a plurality of division vector scaling coefficients corresponding to the respective code vectors of the first-stage codebook,
Step (b) above is
When a code vector other than the vector including the component of the acoustic parameter vector representing the substantially flat spectrum envelope is output from the one-stage codebook, each selected from the plurality of divided vector codebooks of the at least one stage The divided vectors are read from the divided vector scaling coefficient codebooks corresponding to the index of the vector selected in the first-stage codebook, respectively multiplied , and multiplied. Integrating the divided vectors obtained in this way and outputting them as output vectors of the codebook at that stage ,
Including the. 12. The decoding method according to claim 8, wherein a vector including a component of a parameter vector equivalent to the linear prediction coefficient is obtained in advance from a parameter vector equivalent to the linear prediction coefficient. This is a vector generated by subtracting an average vector of parameters equivalent to the linear prediction coefficient of the entire acoustic signal. In the decoding method according to any one of claims 12 to claim 8, it said linear predictive coefficients equivalent to the parameters is LSP parameters. An acoustic parameter encoding device,
A parameter calculating means for analyzing the input acoustic signal for each frame and calculating an acoustic parameter equivalent to a linear prediction coefficient representing the spectral envelope characteristics of the acoustic signal;
A vector codebook that stores a plurality of code vectors corresponding to indexes representing them;
A coefficient codebook that stores one or more sets of weighting coefficients in correspondence with indexes representing those sets;
The code vector for the current frame output from the vector code book and the code vector output from the nearest past at least one frame are multiplied by each of the set of weight coefficients selected from the coefficient code book. A quantization parameter generating means for generating a weighted vector by adding together and outputting a vector including the generated component of the weighted vector as a candidate of a quantized acoustic parameter for the acoustic parameter of the current frame;
A distortion calculation unit that calculates distortion of the quantized acoustic parameter with respect to the acoustic parameter calculated by the parameter calculation unit;
Indexes representing the code vectors of the vector codebook and the set of coefficient codebooks, and the representative indexes of the determined code vector and the set of weighting coefficients, using a standard that reduces the distortion Codebook search control unit for outputting as a code of the acoustic parameters,
Including
The vector codebook is composed of a plurality of codebooks each storing a plurality of vectors corresponding to indexes representing them, and the codebook of one stage of the plurality of codebooks is substantially flat. A vector containing acoustic parameter vector components representing a simple spectrum envelope is stored as one vector,
The quantization parameter generating means, if you select the code vector other than the vector containing the components of the acoustic parameter vector representing the substantially flat-spectrum envelope from the codebook of said one stage, the codebook of said plurality of stages Select each vector, add the vectors, and output as the selected code vector of the current frame, and from the one-stage codebook, a vector including the component of the acoustic parameter vector representing the substantially flat spectrum envelope If selected, means for outputting a vector containing a component of the acoustic parameter vector representing the substantially flat spectrum envelope as the selected code vector of the current frame. 15. The encoding device according to claim 14 , wherein at least one of the plurality of codebooks corresponds to an index representative of each of them as a plurality of divided vectors obtained by dividing the dimension of the code vector into a plurality. A plurality of divided vector codebooks distributed and stored, and an integration unit that integrates the divided vectors output from the plurality of divided vector codebooks and outputs them as output vectors of the codebook at that stage. An acoustic parameter encoding device,
Parameter calculating means for analyzing the input acoustic signal for each frame and calculating an acoustic parameter equivalent to a linear prediction coefficient representing the spectral envelope characteristics of the acoustic signal;
A vector codebook storing a plurality of code vectors corresponding to indexes representing them, and
A coefficient codebook storing one or more sets of weighting coefficients in correspondence with indexes representing those sets;
The code vector for the current frame output from the vector code book and the code vector output from the nearest past at least one frame are multiplied by each of the set of weight coefficients selected from the coefficient code book. A quantization parameter generating means for generating a weighted vector by adding together and outputting a vector including the generated component of the weighted vector as a candidate of a quantized acoustic parameter for the acoustic parameter of the current frame;
A distortion calculation unit that calculates distortion of the quantized acoustic parameter with respect to the acoustic parameter calculated by the parameter calculation unit;
Indexes representing the code vectors of the vector codebook and the set of coefficient codebooks, and the representative indexes of the determined code vector and the set of weighting coefficients, using a criterion that reduces the distortion Codebook search control unit for outputting as a code of the acoustic parameters,
Including
The above vector codebook is
A plurality of codebooks each storing a plurality of code vectors corresponding to indexes representing them;
Scaling coefficient codebook provided for each codebook in the second and subsequent stages, in which scaling coefficients determined in advance corresponding to the respective code vectors of the first stage codebook are stored in correspondence with indexes representing them Including
A vector including the acoustic parameter vector component representing the substantially flat spectrum envelope is stored in one codebook of the multistage codebook, and a zero vector is stored in the remaining codebook. Stored,
further,
When a code vector other than the vector including the acoustic parameter vector component representing the substantially flat spectrum envelope is selected from the one-stage codebook, the second and subsequent stages for the code vector selected in the first stage Read out the corresponding scaling factor from the scaling codebook for the codebook of, multiply the code vector selected from the codebook of the second and subsequent stages, and output the multiplication result as a vector of each stage ,
When a vector including the component of the acoustic parameter vector representing the substantially flat spectrum envelope is selected from the one-stage codebook, the vector including the component of the acoustic parameter vector representing the substantially flat spectrum envelope is selected as the current frame. Means for outputting as the selected code vector of
The including. 17. The encoding device according to claim 16 , wherein among the plurality of codebooks, at least one codebook of the second and subsequent stages is distributed and stored as a plurality of divided vectors obtained by dividing the dimension of the code vector into a plurality. A plurality of divided vector codebooks,
The scaling factor codebook corresponding to the at least one stage codebook is:
A plurality of division vector scaling coefficients codebook in which a plurality of division vector scaling coefficients are stored corresponding to the first-stage code vectors in correspondence with the plurality of division vector codebooks;
When a code vector other than a vector including a component of the acoustic parameter vector representing the substantially flat spectrum envelope is selected from the one-stage codebook, output from each of the plurality of divided vector codebooks of the at least one stage. is the relative split vector, multiplied respectively by reading the split vector for scaling coefficient corresponding to the index of the selected vector codebook of the first stage from each of the split vector for the scaling factor codebook,
Means for integrating the multiplication results and outputting them as an output vector of the codebook at that stage;
Including. An acoustic parameter decoding device,
A vector codebook storing a plurality of code vectors of acoustic parameters equivalent to linear prediction coefficients representing the spectral envelope characteristics of the acoustic signal in correspondence with indexes representing them;
A coefficient codebook storing one or more sets of weighting coefficients in correspondence with indexes representing them;
One code vector is output from the vector codebook according to an index represented by a code input for each frame, a set weight coefficient is output from the coefficient codebook, and the code vector output in the current frame; A weighted vector is generated by multiplying the code vector output in the nearest past frame by the weight coefficient of the set output in the current frame, and adding them together, and the components of the generated weight vector Quantization parameter generating means for outputting a vector including a decoded quantized acoustic parameter of the current frame,
Including
The vector codebook is composed of a plurality of codebooks each storing a plurality of vectors corresponding to indexes representing them, and the codebook of one stage of the plurality of codebooks is substantially flat. A vector containing acoustic parameter vector components representing a simple spectrum envelope is stored as one vector,
When the quantization parameter generation means selects a code vector other than the vector including the acoustic parameter vector component representing the substantially flat spectrum envelope from the one-stage codebook, Select vectors, add them and output as the selected code vector of the current frame, and select the vector containing the acoustic parameter vector component representing the substantially flat spectrum envelope from the one-stage codebook In this case, there is included means for outputting a vector including a component of the acoustic parameter vector representing the substantially flat spectrum envelope as the selected code vector of the current frame. 19. The decoding apparatus according to claim 18 , wherein at least one of the plurality of codebooks corresponds to an index representative of each of the plurality of divided vectors obtained by dividing the dimension of the code vector into a plurality of divided codebooks. A plurality of divided vector codebooks distributed and stored, and an integration unit that integrates the divided vectors output from the plurality of divided vector codebooks and outputs them as output vectors of the codebook at that stage. An acoustic parameter decoding device,
A vector codebook storing a plurality of code vectors of acoustic parameters equivalent to linear prediction coefficients representing the spectral envelope characteristics of the acoustic signal in correspondence with indexes representing them;
A coefficient codebook storing one or more sets of weighting coefficients in correspondence with indexes representing them;
One code vector is output from the vector codebook according to an index represented by a code input for each frame, a set weight coefficient is output from the coefficient codebook, and the code vector output in the current frame; A weighted vector is generated by multiplying the code vector output in the nearest past frame by the weight coefficient of the set output in the current frame, and adding them, and the components of the generated weight vector Quantization parameter generating means for outputting a vector including a decoded quantized acoustic parameter of the current frame,
Including
The above vector codebook is
A plurality of codebooks each storing a plurality of code vectors corresponding to indexes representing them;
Scaling coefficient codebook provided for each codebook in the second stage and thereafter, in which scaling coefficients determined in advance corresponding to the respective code vectors in the first stage codebook are stored in correspondence with indexes representing them Including
A vector including the acoustic parameter vector component representing the substantially flat spectrum envelope is stored in one codebook of the multistage codebook, and a zero vector is stored in the remaining codebook. Stored,
further,
When a code vector other than the vector including the acoustic parameter vector component representing the substantially flat spectrum envelope is selected from the codebook of the first stage, the second and subsequent stages with respect to the code vector selected in the first stage The corresponding scaling factor is read from the scaling codebook for the codebook of No. 2 and multiplied by the code vector selected from the codebook of the second and subsequent stages, and the multiplication result is output as a vector of each stage ,
A vector addition is performed on the output vector of each stage and the vector of the first stage, and the addition result is output as a code vector from the vector codebook, and the substantially flat spectrum envelope is output from the codebook of the one stage. When a vector including the component of the acoustic parameter vector representing is selected, a vector including the component of the acoustic parameter vector representing the substantially flat spectrum envelope is output as the selected code vector of the current frame. 21. The decoding apparatus according to claim 20 , wherein at least one codebook of the second and subsequent stages among the plurality of stages of codebooks is distributed and stored as a plurality of divided vectors obtained by dividing the code vector dimension into a plurality of dimensions. A plurality of divided vector codebooks,
The scaling factor codebook corresponding to the at least one stage codebook is:
A plurality of division vector scaling coefficients codebook in which a plurality of division vector scaling coefficients are stored corresponding to the first-stage code vectors in correspondence with the plurality of division vector codebooks;
When a code vector other than a vector including a component of the acoustic parameter vector representing the substantially flat spectrum envelope is selected from the one-stage codebook, output from each of the plurality of divided vector codebooks of the at least one stage. is the relative split vector, multiplied respectively by reading the split vector for scaling coefficient corresponding to the index of the selected vector codebook of the first stage from each of the split vector for the scaling factor codebook,
Means for integrating the multiplication results and outputting them as an output vector of the codebook at that stage ;
Including. An acoustic signal encoding device that encodes an input acoustic signal,
Means for encoding the spectral characteristics of the input acoustic signal using the acoustic parameter encoding method according to claim 1;
An adaptive codebook holding an adaptive code vector representing a periodic component of the input acoustic signal;
A fixed codebook storing a plurality of fixed vectors; and
A sound source vector generated based on the adaptive code vector from the adaptive code book and the fixed vector from the fixed code book is input as an excitation signal, and a synthesized acoustic signal is used using a filter coefficient based on the quantized acoustic parameter. Filter means for synthesizing
An adaptive code vector and a fixed vector selected from the fixed codebook and the adaptive codebook are determined so that distortion of the synthesized acoustic signal with respect to the input acoustic signal is reduced, and the determined adaptive code vector and fixed vector are respectively determined. Means for outputting corresponding adaptive and fixed codes;
Including. An acoustic signal decoding apparatus that decodes an input code and outputs an acoustic signal,
Means for decoding an acoustic parameter equivalent to a linear prediction coefficient representing a spectral envelope characteristic from an input code using the acoustic parameter decoding method according to claim 8 ;
A fixed codebook storing a plurality of fixed vectors; and
An adaptive codebook holding an adaptive code vector representing the periodic components of the synthesized acoustic signal;
Means for taking out the corresponding fixed vector from the fixed codebook by the input adaptive code and fixed code, taking out the corresponding adaptive code vector from the adaptive codebook, and combining them to generate an excitation vector;
Filter means for setting a filter coefficient based on the acoustic parameter and reproducing an acoustic signal by the excitation vector;
Including. An acoustic signal encoding method for encoding an input acoustic signal,
(A) using the acoustic parameter encoding method according to claim 1, encoding spectral characteristics of an input acoustic signal;
(B) Generated based on an adaptive code vector from an adaptive code book that holds an adaptive code vector representing a periodic component of the input acoustic signal and a fixed vector from a fixed code book that stores a plurality of fixed vectors Using a sound source vector as an excitation signal and performing a synthetic filter process with a filter coefficient based on the quantized acoustic parameter to generate a synthetic acoustic signal;
(C) The adaptive code vector and fixed vector to be selected from the fixed codebook and the adaptive codebook are determined so that the distortion of the synthesized sound signal with respect to the input sound signal is reduced, and the determined adaptive code vector and the fixed codebook are fixed. Outputting an adaptive code and a fixed code corresponding to each vector,
Including. An acoustic signal decoding method for decoding an input code and outputting an acoustic signal,
(A) decoding an acoustic parameter equivalent to a linear prediction coefficient representing a spectral envelope characteristic from an input code using the acoustic parameter decoding method according to claim 8 ;
(B) The corresponding adaptive code vector is extracted from the adaptive code book in which the adaptive code vector indicating the periodic component of the input acoustic signal is held by the adaptive code and the fixed code in the input code, and a plurality of fixed vectors are stored. Extracting a corresponding fixed vector from the fixed codebook, and vector-combining the adaptive code vector and the fixed vector to generate an excitation vector;
(C) reconstructing a synthesized acoustic signal by synthesizing and filtering the excitation vector using a filter coefficient based on the acoustic parameter;
Including. The program which implements the acoustic parameter encoding method in any one of Claims 1-7 with a computer. Program implementing the acoustic parameter decoding method according a computer in any one of claims 8-13. An acoustic input device for converting an acoustic signal into an electrical signal;
An A / D converter for converting a signal output from the acoustic input signal device into a digital signal;
The acoustic signal encoding device according to claim 14 , which encodes a digital signal output from the A / D conversion device;
An RF modulation device that performs modulation processing and the like on the encoded information output from the acoustic signal encoding device;
A transmitting antenna that converts a signal output from the RF modulation device into a radio wave and transmits the signal;
And an acoustic signal transmitting device. A receiving antenna for receiving radio waves,
An RF demodulator for demodulating a signal received by the receiving antenna;
The acoustic signal decoding device according to claim 18, which performs decoding processing of information obtained by the RF demodulation device;
A D / A converter for D / A converting the digital acoustic signal decoded by the acoustic signal decoding device, and an acoustic signal output device for converting an electrical signal output from the D / A converter to an acoustic signal ,
And an acoustic signal receiving device.

Images