JP3887247B2 - Signal separation device and method, signal separation program, and recording medium recording the program - Google Patents

Description

と表現できる。ここで＊は畳み込み、ｔは時間を示す。インパルス応答ｈ_qpは遅延や残響を表現するためにある程度の長さを持つ。
信号分離の目的は、分離のためのＦＩＲ(Finite Impulse Response)フィルタ係数ｗ_rqと
【数２】

を求めることにある。残響も含めて分離するためには、フィルタ係数ｗ_rqもある程度の長さを持つ必要がある。室内で定常状態になっている音源の音を止めた後にその音圧レベルが60dB下がるまでの時間を残響時間と呼ぶが、フィルタ長が残響時間に比べて短かすぎれば、フィルタ長を超える残響の部分の分離ができなくなり分離性能が低下する。逆に長いフィルタは、求めるべきフィルタ係数１個当たりの観測信号量の低下という代償を支払うため、フィルタ長が残響時間に比べて長すぎても分離性能が低下する。従って、分離のためのフィルタは、残響時間に応じた長さを持つのが望ましい。
【０００３】
特殊な状況以外では、信号源ｓ_p（ｔ）やインパルス応答ｈ_qpを正確に知ることはできず、センサでの観測信号ｘ_q（ｔ）からフィルタ係数ｗ_rqを求める必要がある。一般に源信号ｓ_p（ｔ）は独立であるため、独立成分分析(ICA:Independent Component Analysis)を用いて分離のためのフィルタ係数ｗ_rqを計算できる。ＩＣＡを用いた信号分離の手法には様々なものがあるが、残響に対処するためには周波数領域での手法が有効である。上記の畳み込み混合の問題が、各周波数毎の瞬時混合の問題に置き換えることができるからである。センサｑでの観測信号ｘ_q（ｔ）に短時間離散フーリエ変換を適用した結果をＸ_q(ｆ,ｍ)とする。ここでｆは周波数、ｍはフレーム番号である。各周波数毎に見ると、分離のための周波数特性Ｗ_rq（ｆ）と
【数３】

を、瞬時混合のＩＣＡを適用して求めることになる。全周波数にわたって周波数特性Ｗ_rq（ｆ）を求めたら、これに逆離散フーリエ変換を適用することで、分離のためのフィルタ係数ｗ_rqが最終的に求まる。
【０００４】
ＩＣＡによって作られる分離フィルタの周波数特性を解析すると、多くの場合は、除去したい信号の方向に適応的に死角を向けていることが分かる。つまり、複数のセンサで観測される信号の位相差が信号の到来方向に依存することを利用し、除去したい方向から来る信号のパワーが小さくなるように分離フィルタの周波数特性Ｗ_rq（ｆ）を調整している。ここで発生する位相差は、信号の周波数とセンサの間隔に依存する。センサ間隔が一定であれば、周波数が高くなるほど位相差が大きくなる。また、周波数が一定であれば、センサ間隔が広くなるほど位相差が大きくなる。十分な分離性能を達成するためには、ある程度の位相差が発生するのが望ましい。従って、低周波数の信号を分離するためには、センサ間隔を広くするのが良い。しかし、高周波数の信号に対してセンサ間隔を広くしすぎると、位相差が２πを超えて複数の方向に死角を向けることになり、除去したくない方向の信号まで除去することになる。
【０００５】
【発明が解決しようとする課題】
上記のように、分離フィルタの長さは残響時間に応じたものがふさわしい。しかし実際には、残響時間は周波数毎に異なっており、低周波数ほど長く、高周波数ほど短い。また、センサ間隔に関しては、低周波数には広いものが望ましく、高周波数には狭いものが望ましい。このように、周波数毎に適切なフィルタ長やセンサ間隔は異なる。
従来の技術では、全周波数帯域を１つの分離システムで分離するため、周波数が大きく異なる帯域に対しても同じフィルタ長やセンサ間隔で処理を行わなければならず、上記の要望に対応することができない。そこで本発明の目的は、上記の要望に鑑み、周波数帯域毎に適切なフィルタ長やセンサ間隔を適用できる信号分離装置を提供することにある。
【０００６】
【課題を解決するための手段】
上記目的を達成するために、本発明では、混合信号を分離信号に変換するためのデジタルフィルタの係数を計算する手段と、該デジタルフィルタを用いて混合信号を分離信号に変換する手段を備える分離システムが複数個存在する。そして、信号をいくつかの周波数帯域に分割するために、各分離システムは指定された周波数帯域のみを通過させる手段を持ち、全体としてはそれら複数の分離システムの出力を統合する手段を持つ。さらに、周波数帯域毎に高い分離性能を得るために、混合信号を分離信号に変換するデジタルフィルタの長さや、混合信号を観測する複数のセンサ間の相対位置が各分離システム毎に異なる。
また、本発明を２信号の分離に適用する場合においては、２つの信号源の方向を推定し、推定された方向と各分離システムが用いる２センサの間隔から、２信号が２センサそれぞれに到達する時間差を計算し、各分離システムが通過させる周波数帯域を決定する手段を備える。
【０００７】
【発明の実施の形態】
[全体構成]
図１は、本発明の信号分離装置の構成を示すブロック図である。
センサ（マイクロフォン）１は、空間内で混合された源信号ｓを収音し、混合信号（観測信号）ｘを出力する。観測信号ｘは複数の分離システム２に入力され、観測信号を分離信号ｙ’に変換する。それぞれの分離システムから出力される分離信号は統合部３で統合して統合分離信号、すなわち、復元された源信号ｙが生成される。
信号分離装置において、基本となる分離システム２が複数個存在する。各分離システムは、指定された周波数帯域の信号のみを分離して通過させる。また、通過させる周波数帯域の特性に合わせて、分離するためのデジタルフィルタの長さが異なったり、用いる複数のセンサの相対位置が異なったりしている。各分離システムの出力は、後段に位置する統合部３で統合され、全体として全周波数帯域の分離信号が得られる。
次に、信号分離方法を説明する。
複数の分離システム２は、入力された混合信号ｘを分離信号ｙ’に変換するためのデジタルフィルタの係数を計算し（ステップ１）、計算されたデジタルフィルタ係数によりデジタルフィルタを用いて混合信号を分離信号に変換し（ステップ２）、指定された周波数帯域のみを通過させる（ステップ３）を有する。なお、（ステップ３）は、（ステップ１）の前、（ステップ２）の後（図２ 参照）、あるいは（ステップ１）と（ステップ２）の間（図３ 参照）のいずれでもよい。統合部３においては、複数の（ステップ１〜３）により生成された分離信号ｙ’を統合して復元された源信号ｙを生成する（ステップ４）。
【０００８】
[分離システム]
基本となる分離システム２の実施形態を２種類示す。
１つ目は、図２に示すように、一般の帯域フィルタ２−２を、ＩＣＡによる分離部と分離フィルタから構成されるＩＣＡによる分離部に直列接続した構成である。帯域フィルタ２−２は分離信号の分離前、分離後どちらにでも接続できる。
このような構成は、良く知られている帯域フィルタを流用できることが利点となるが、一方で、処理される信号が通過すべきデジタルフィルタの総和長が長くなり、観測信号から分離信号への遅延が大きくなるという欠点もある。
２つ目は、分離のためのフィルタに帯域制限の機能を同時に持たせる構成である。本構成では、１つ目の構成ほどフィルタ長が長くならないという利点がある。
【０００９】
図３を参照して分離フィルタに帯域制限の機能を持たせる分離システム、方法について説明する。
まず、ＩＣＡによるフィルタ係数生成部２−３は、ＩＣＡを用いて帯域制限を行わない普通の分離を行い、分離のためのフィルタ係数ｗ_rqを求める。次に、離散フーリエ変換部２−４は、これらに離散フーリエ変換を適用し、各フィルタの周波数特性Ｗ_rq（ｆ）を求める。そして、帯域制限処理部２−５は、指定された周波数帯域以外の通過を制限するために、指定された周波数帯域以外の周波数特性を０にして新たな周波数特性Ｗ_rq'（ｆ）を求める。最後に、逆フーリエ変換部２−６において、これらを逆フーリエ変換して求めた周波数特性を反映するフィルタ係数ｗ_rq'とし、分離フィルタ２−７は、これらを用いて実際に信号分離を行う。ただしこのままでは、遮断周波数付近で急峻な特性となり、現実的な長さのフィルタでは実現すべき周波数特性との誤差が大きくなる。これを緩和するには、得られたフィルタ係数にハニング(Hanning)窓などの窓を掛ければ良い。
【００１０】
[フィルタ長を変化させる例]
次に、分離フィルタの長さを各分離システム毎に異なるものにする実施例を示す。
残響時間は、信号が混合する環境（部屋の大きさ、壁の材質など）に依存し、前もって正確に知ることは難しいが、音の場合では、125Hzから250Hzにかけて最も長くなり、1000Hz以上では短めに安定するのが一般的である。従って、サンプリング周波数8000Hzの場合の例としては、500Hz以下を通過させる分離システムではフィルタ長を4096(512ms)とし、500Hz以上を通過させる分離システムではフィルタ長を2048(256ms)とする。分離フィルタはその長さの半分ほどの残響に対応できるため、この場合は、500Hz以下では256msまでの残響を500Hz以上では128msまでの残響を分離することができる。
【００１１】
[センサ間隔を変化させる例]
次に、センサ間隔を各分離システム毎に異なるものにする実施例として、間隔が28.3mmであるセンサを用いる分離システムと、間隔が141.5mmであるセンサを用いる分離システムを備えた音信号分離装置を考える。間隔ｄを141.5mmとしても複数の方向に死角を作らない周波数の範囲を考えると、その条件はｃを音速として２πｆｃ^-1ｄ＜πであり、結果としてｆ＜1201.4Hzである。従って、センサ間隔が141.4mmである分離システムでは1201Hz以下を通過させ、センサ間隔が28.3mmである分離システムでは1201Hz以上を通過させる。これにより、複数の方向に死角を作らない範囲で、できるだけ広い幅のセンサを利用することができる。
【００１２】
[信号源方向を推定してより適切な周波数範囲を決定]
しかし、信号源方向が推定できれば、より適切な周波数範囲を求めることが可能となる。本発明では、信号源の数が２である場合にその装置、方法を提供する。
図４を参照して、センサ間隔がｄである分離システムが扱うべき最大周波数を求める構成、方法、すなわち推定信号源方向による周波数帯域の決定部、方法を説明する。
まず、信号源方向推定部４−１は、ＭＵＳＩＣ法などの良く知られた信号源方向を推定する方法を用いて、観測信号から信号源方向θ₁、θ₂を推定する。時間差計算部４−２は、信号源方向θ₁、θ₂とセンサ間の間隔ｄをもとに次の計算を行う。信号の到来方向はセンサを結ぶ線に垂直となる方向を90°とすると、間隔ｄの２つのセンサに到達する時間差は、θ₁方向から来る信号の場合はｃ^-1ｄcosθ₁、θ₂方向から来る信号の場合はｃ^-1ｄcosθ₂となる。ただしｃは信号の速度である。次にこれらの差を周波数ｆにおける位相差として考えると、２πｆｃ^-1ｄcosθ₁および２πｆｃ^-1ｄcosθ₂となる。周波数範囲決定部４−３は、以下に示すように周波数帯域を決定する。これらの位相差の差（２πｆｃ^-1ｄcosθ₁−２πｆｃ^-1ｄcosθ₂）がπに近づくほど、ゲイン特性の差を大きくすることができるため、信号分離の性能を高めやすい。そして、位相差の差がπを超えて２πに近づくにつれて、ゲイン特性の差が小さくなり分離性能が除々に低下する。また、πを超えた場合には、θ₁とθ₂の間のゲイン特性が最大となる方向ができるため、残響を多く拾ってしまう恐れがある。従って本実施例では、位相差がπを大きく超えない範囲の周波数を扱うことにし、（２πｆｃ^-1ｄcosθ₁−２πｆｃ^-1ｄcosθ₂）≦απとする。ここでαは、位相差の差がπをどれだけ超えても良いかを設定するパラメータであり、１か１を少し超える値である。本条件を整理すると、扱う最大周波数はｆ_max＝αｃ／[２ｄ（cosθ₁−cosθ₂）]となる。
【００１３】
本発明の信号分離装置は、ＣＰＵやメモリ等を有するコンピュータと利用者端末と記録媒体から構成することができる。
記録媒体は、ＣＤ−ＲＯＭ、磁気ディスク装置、半導体メモリ等の機械読み取り可能な記録媒体であり、ここに記録された信号分離プログラムは、コンピュータに読み取られ、コンピュータの動作を制御し、コンピュータ上に前述した各構成要素、すなわち、分離システム、統合部等を実現する。
【００１４】
【発明の効果】
従来技術および本発明を用いて、２音源を分離した際の分離性能の比較を図５に示す。
本結果を得るに際し、ＲＷＣＰ(Real World Computing partnership：新情報処理開発機構) 実環境音声・音響データベースから選んだ残響時間300msのインパルス応答に、ＡＳＪ(The Acoustical Society of Japan:日本音響学会)研究用音声コーパスから選んだ８秒の音声データを畳み込んで混合信号を作成した、横軸はフィルタ長とマイク間隔の組み合わせを示し、「2048」および「4096」はフィルタ長を、「28.3」と「141.5」はそれぞれマイク間隔をmmで示している。
左側の４組み合わせは従来技術によるものであり、フィルタ長やマイク間隔を周波数毎に変えることはできない。
右側の５組み合わせは本発明によるものであり、フィルタ長やマイク間隔を周波数帯域毎に変えることができる。
「項２」および「項３」は、それぞれ請求項３および請求項４の発明を利用したことを示す。
「項２」の場合は、基本となる分離システムを２つ備え、796Hz以下の周波数に対してはフィルタ長を4096とし、796Hz以上4000Hz以下の周波数に対してはフィルタ長を2048とした。
「項３」の場合も同様に基本となる分離システムを２つ備え、796Hz以下の周波数に対してはマイク間隔を141.5mmとし、796Hz以上4000Hz以下の周波数に対してはマイク間隔を28.3mmとした。
796Hzという境界は、請求項３および請求項４記載の発明の実施例においてα＝１、音速ｃ＝340m／s、センサ間隔ｄ＝141.5mm、推定信号源方向θ₁＝50°、θ₂＝150°として決定された。
比較のため、全ての組み合わせにおいて低域（796Hz以下）での分離性能と高域（796Hz以上4000Hz以下）での分離性能を示した。
従来技術の４組み合わせの結果を見ると、低域の分離性能を高めようとすると、高域の分離性能が低下してしまうことが分かる、一方、本発明では、高域の分離性能を保ったまま、低域の分離性能を高めることが可能となる。
例えば、「項２」と「28.3」の組み合わせでは、従来技術における「2048」と「28.3」の組み合わせの高域の分離性能を保ったまま、「4096」と「28.3」の組み合わせの低域の分離性能を達成できる。
このように、本発明によれば、扱う信号の周波数帯域毎に適切なフィルタ長とセンサ間隔を設定することが可能となり、全帯域に対して単一のフィルタ長とセンサ間隔しか設定できなかった従来技術よりも分離性能が向上していることが分かる。
また、請求項３と請求項４の発明を併用することで、さらに分離性能を向上させることも可能であることが分かる。
【図面の簡単な説明】
【図１】本発明の構成を示すブロック図。
【図２】帯域フィルタを分離部に直列接続した分離システム構成図。
【図３】分離フィルタに帯域制限の機能を持たせる分離システム構成、及び方法の説明図。
【図４】推定信号源方向による周波数帯域の決定部の構成、及び方法の説明図。
【図５】分離性能の比較を示す図。
【符号の説明】
１・・・センサ
２・・・分離システム
２−１・・ＩＣＡによる分離部、２−２・・・帯域フィルタ、２−３・・・ＩＣＡによるフィルタ係数生成部、２−４・・・離散フーリエ変換部、２−５・・・帯域制限処理部、２−６・・・逆離散フーリエ変換部、２−７・・・分離フィルタ
４−１・・・信号源方向推定部、４−２・・・時間差計算部、４−３・・・周波数範囲決定部[0001]
BACKGROUND OF THE INVENTION
The present invention belongs to the field of signal processing, and relates to a signal separation technique for restoring a source signal as accurately as possible from a mixture of a plurality of signals in space. According to the present technology, it is possible to accurately extract a target signal in an actual environment where various interference signals are generated. As an application example of the sound signal of the present invention, a speech recognizer with a high recognition rate can be configured even in a situation where the speaker and the microphone are separated and the microphone picks up sound other than the speaker's voice. can give.
[0002]
[Prior art]
First, signal separation is formulated. Assume that N signals are mixed and observed by M sensors. The present invention deals with a situation where the signal is attenuated / delayed depending on the distance from the signal source to the sensor, and reverberation occurs due to reflection of the signal by walls or the like. Such mixing, ## EQU1 ## with an impulse response h _qp from the signal source s _p (t) to the sensor x _q (t)

Can be expressed as Here, * indicates convolution and t indicates time. The impulse response h _qp has a certain length to express delay and reverberation.
The purpose of signal separation is the FIR (Finite Impulse Response) filter coefficient _wrq for separation and

Is to seek. In order to perform separation including reverberation, the filter coefficient w _rq needs to have a certain length. The time until the sound pressure level drops by 60 dB after stopping the sound of a sound source that is in a steady state in the room is called the reverberation time, but if the filter length is too short compared to the reverberation time, the reverberation exceeds the filter length. This part cannot be separated, and the separation performance deteriorates. Conversely, a long filter pays the price of a reduction in the amount of observation signal per filter coefficient to be obtained, so that the separation performance is degraded even if the filter length is too long compared to the reverberation time. Therefore, it is desirable that the separation filter has a length corresponding to the reverberation time.
[0003]
Outside special situations, it is not possible to know the signal source s _p (t) and the impulse response h _qp accurately, it is necessary from the observed signal x _q at the sensor (t) obtaining the filter coefficients w _rq. Generally the source signal s _p (t) is independent, independent component analysis (ICA: Independent Component Analysis) can be calculated filter coefficients w _rq for separation using. There are various signal separation methods using ICA, but a method in the frequency domain is effective in dealing with reverberation. This is because the problem of convolutional mixing can be replaced with the problem of instantaneous mixing for each frequency. The result of applying the short-time discrete Fourier transform to the observation signal x _q (t) at the sensor q is X _q (f, m). Here, f is a frequency and m is a frame number. Looking at each frequency, the frequency characteristics W _rq (f) for separation and

Is obtained by applying ICA of instantaneous mixing. When the frequency characteristic W _rq (f) is obtained over all frequencies, the inverse discrete Fourier transform is applied to the frequency characteristic W _rq (f) to finally obtain the filter coefficient w _rq for separation.
[0004]
Analysis of the frequency characteristics of the separation filter produced by ICA reveals that in many cases the blind spot is adaptively directed in the direction of the signal to be removed. In other words, the frequency characteristic W _rq (f) of the separation filter is set so that the power of the signal coming from the direction to be removed becomes small by utilizing the fact that the phase difference of the signals observed by the plurality of sensors depends on the arrival direction of the signal. It is adjusted. The phase difference generated here depends on the signal frequency and the sensor interval. If the sensor interval is constant, the phase difference increases as the frequency increases. If the frequency is constant, the phase difference increases as the sensor interval increases. In order to achieve sufficient separation performance, it is desirable that a certain phase difference occurs. Therefore, in order to separate low frequency signals, it is preferable to widen the sensor interval. However, if the sensor interval is too wide for a high-frequency signal, the phase difference exceeds 2π and the blind spots are directed in a plurality of directions, and signals in directions that are not desired to be removed are removed.
[0005]
[Problems to be solved by the invention]
As described above, the length of the separation filter is suitable according to the reverberation time. In practice, however, the reverberation time varies from frequency to frequency, and is longer at lower frequencies and shorter at higher frequencies. As for the sensor interval, a wide one is desirable for low frequencies and a narrow one is desirable for high frequencies. Thus, the appropriate filter length and sensor interval differ for each frequency.
In the conventional technology, since all frequency bands are separated by one separation system, it is necessary to perform processing with the same filter length and sensor interval even for bands with greatly different frequencies, and this can meet the above demands. Can not. Accordingly, an object of the present invention is to provide a signal separation device capable of applying an appropriate filter length and sensor interval for each frequency band in view of the above demand.
[0006]
[Means for Solving the Problems]
In order to achieve the above object, the present invention provides a separation means comprising means for calculating a coefficient of a digital filter for converting a mixed signal into a separated signal, and means for converting the mixed signal into a separated signal using the digital filter. There are multiple systems. In order to divide the signal into several frequency bands, each separation system has means for passing only the designated frequency band, and has means for integrating the outputs of the plurality of separation systems as a whole. Furthermore, in order to obtain a high separation performance for each frequency band, the length of the digital filter that converts the mixed signal into the separated signal and the relative position between the plurality of sensors that observe the mixed signal are different for each separation system.
When the present invention is applied to the separation of two signals, the directions of the two signal sources are estimated, and the two signals reach each of the two sensors from the estimated direction and the interval between the two sensors used by each separation system. Means for calculating a time difference to be determined and determining a frequency band to be passed by each separation system.
[0007]
DETAILED DESCRIPTION OF THE INVENTION
[overall structure]
FIG. 1 is a block diagram showing the configuration of the signal separation device of the present invention.
The sensor (microphone) 1 collects the source signal s mixed in the space and outputs a mixed signal (observation signal) x. The observation signal x is input to a plurality of separation systems 2 and converts the observation signal into a separation signal y ′. The separation signals output from the respective separation systems are integrated by the integration unit 3 to generate an integrated separation signal, that is, a restored source signal y.
In the signal separation device, there are a plurality of basic separation systems 2. Each separation system separates and passes only signals in a designated frequency band. Further, the length of the digital filter for separation differs according to the characteristics of the frequency band to pass through, or the relative positions of the plurality of sensors used differ. The output of each separation system is integrated by the integration unit 3 located in the subsequent stage, and a separation signal of the entire frequency band is obtained as a whole.
Next, a signal separation method will be described.
The plurality of separation systems 2 calculate digital filter coefficients for converting the input mixed signal x into the separated signal y ′ (step 1), and use the digital filter to calculate the mixed signal based on the calculated digital filter coefficients. The separated signal is converted (step 2), and only the designated frequency band is passed (step 3). (Step 3) may be either before (Step 1), after (Step 2) (see FIG. 2), or between (Step 1) and (Step 2) (see FIG. 3). In the integration unit 3, a source signal y restored by integrating a plurality of separated signals y ′ generated in (Steps 1 to 3) is generated (Step 4).
[0008]
[Separation system]
Two embodiments of the basic separation system 2 are shown.
First, as shown in FIG. 2, a general band-pass filter 2-2 is connected in series to an ICA separation unit composed of an ICA separation unit and a separation filter. The band-pass filter 2-2 can be connected either before or after separation of the separated signal.
Such a configuration has the advantage that a well-known bandpass filter can be used, but on the other hand, the total length of the digital filter through which the signal to be processed must pass increases, and the delay from the observed signal to the separated signal There is also a disadvantage that becomes larger.
The second is a configuration in which a filter for separation has a band limiting function at the same time. This configuration has the advantage that the filter length does not increase as much as the first configuration.
[0009]
With reference to FIG. 3, a separation system and method for providing a separation filter with a band limiting function will be described.
First, the filter coefficient generation unit 2-3 by ICA performs normal separation without band limitation using ICA, and obtains a filter coefficient _wrq for separation. Next, the discrete Fourier transform unit 2-4 applies the discrete Fourier transform to these to obtain the frequency characteristic W _rq (f) of each filter. Then, the band limitation processing unit 2-5 obtains a new frequency characteristic W _rq ′ (f) by setting the frequency characteristics other than the designated frequency band to 0 in order to restrict passage other than the designated frequency band. . Finally, in the inverse Fourier transform unit 2-6, these are used as filter coefficients _wrq 'reflecting the frequency characteristics obtained by inverse Fourier transform, and the separation filter 2-7 actually performs signal separation using these. . However, if it remains as it is, it becomes a steep characteristic near the cutoff frequency, and an error from the frequency characteristic to be realized becomes large with a filter having a realistic length. To alleviate this, the obtained filter coefficient may be multiplied by a window such as a Hanning window.
[0010]
[Example of changing the filter length]
Next, an embodiment in which the length of the separation filter is different for each separation system will be described.
The reverberation time depends on the environment in which the signal is mixed (room size, wall material, etc.), and it is difficult to know accurately in advance, but in the case of sound, it is the longest from 125Hz to 250Hz, and short at 1000Hz or more It is generally stable. Accordingly, as an example in the case of a sampling frequency of 8000 Hz, the filter length is 4096 (512 ms) in a separation system that passes 500 Hz or less, and the filter length is 2048 (256 ms) in a separation system that passes 500 Hz or more. Since the separation filter can cope with reverberation of about half of its length, in this case, reverberation up to 256 ms can be separated at 500 Hz or less and reverberation up to 128 ms at 500 Hz or more.
[0011]
[Example of changing the sensor interval]
Next, as an example in which the sensor interval is different for each separation system, a sound signal separation device including a separation system using a sensor with an interval of 28.3 mm and a separation system using a sensor with an interval of 141.5 mm think of. Considering a frequency range that does not create a blind spot in a plurality of directions even if the distance d is 141.5 mm, the condition is 2πfc ⁻¹ d <π where c is the speed of sound, and as a result, f <1201.4 Hz. Therefore, a separation system with a sensor interval of 141.4 mm passes 1201 Hz or less, and a separation system with a sensor interval of 28.3 mm passes 1201 Hz or more. As a result, a sensor having a width as wide as possible can be used within a range in which no blind spots are formed in a plurality of directions.
[0012]
[Estimate signal source direction to determine more appropriate frequency range]
However, if the signal source direction can be estimated, a more appropriate frequency range can be obtained. The present invention provides an apparatus and method when the number of signal sources is two.
With reference to FIG. 4, the configuration and method for obtaining the maximum frequency to be handled by the separation system whose sensor interval is d, that is, the determination unit and method of the frequency band based on the estimated signal source direction will be described.
First, the signal source direction estimation unit 4-1 estimates the signal source directions θ ₁ and θ ₂ from the observation signal by using a well-known method of estimating the signal source direction such as the MUSIC method. The time difference calculation unit 4-2 performs the following calculation based on the signal source directions θ ₁ and θ ₂ and the distance d between the sensors. Assuming that the direction of arrival of the signal is 90 ° in the direction perpendicular to the line connecting the sensors, the time difference to reach the two sensors at the interval d is c ⁻¹ dcos θ ₁ , θ ₂ direction for signals coming from the θ ₁ direction. In the case of the signal coming from, it is c ⁻¹ dcos θ ₂ . Where c is the speed of the signal. Next, when these differences are considered as phase differences at the frequency f, they are 2πfc ⁻¹ dcos θ ₁ and 2πfc ⁻¹ dcos θ ₂ . The frequency range determination unit 4-3 determines the frequency band as described below. As the difference between these phase differences (2πfc ⁻¹ dcos θ ₁ −2πfc ⁻¹ dcos θ ₂ ) approaches π, the gain characteristic difference can be increased, so that it is easy to improve the signal separation performance. As the difference in phase difference exceeds π and approaches 2π, the gain characteristic difference becomes smaller and the separation performance gradually decreases. Further, when π is exceeded, there is a possibility that the gain characteristic between θ ₁ and θ ₂ is maximized, so that a lot of reverberation may be picked up. Therefore, in this embodiment, the frequency in a range where the phase difference does not greatly exceed π is handled, and (2πfc ⁻¹ dcos θ ₁ −2πfc ⁻¹ dcos θ ₂ ) ≦ απ. Here, α is a parameter for setting how much the difference in phase difference may exceed π, and is a value slightly exceeding 1 or 1. When this condition is arranged, the maximum frequency to be handled is f _max = αc / [2d (cos θ ₁ −cos θ ₂ )].
[0013]
The signal separation device of the present invention can be composed of a computer having a CPU, a memory, etc., a user terminal, and a recording medium.
The recording medium is a machine-readable recording medium such as a CD-ROM, a magnetic disk device, or a semiconductor memory. The signal separation program recorded on the recording medium is read by a computer, and controls the operation of the computer. Each component mentioned above, ie, a separation system, an integration part, etc. are realized.
[0014]
【The invention's effect】
FIG. 5 shows a comparison of separation performance when two sound sources are separated using the prior art and the present invention.
In order to obtain this result, RWCP (Real World Computing Partnership) is used for ASJ (The Acoustical Society of Japan) research in impulse response of 300 ms reverberation time selected from real environment speech and acoustic database. A mixed signal was created by convolving 8-second audio data selected from the audio corpus. The horizontal axis shows the combination of filter length and microphone interval, “2048” and “4096” indicate the filter length, “28.3” and “ “141.5” indicates the distance between the microphones in mm.
The four combinations on the left are according to the prior art, and the filter length and the microphone interval cannot be changed for each frequency.
The five combinations on the right side are according to the present invention, and the filter length and the microphone interval can be changed for each frequency band.
"Claim 2" and "Claim 3" indicate that the inventions of Claim 3 and Claim 4 were used, respectively.
In the case of “term 2”, two basic separation systems are provided, the filter length is 4096 for frequencies of 796 Hz or less, and the filter length is 2048 for frequencies of 796 Hz to 4000 Hz.
In the case of “Category 3” as well, two basic separation systems are provided, with a microphone spacing of 141.5 mm for frequencies below 796 Hz, and a microphone spacing of 28.3 mm for frequencies between 796 Hz and 4000 Hz. did.
The boundaries of 796 Hz are α = 1, sound velocity c = 340 m / s, sensor interval d = 141.5 mm, estimated signal source direction θ ₁ = 50 °, θ ₂ = in the embodiments of the inventions of claims 3 and 4. Determined as 150 °.
For comparison, the separation performance in the low range (796 Hz or less) and the separation performance in the high range (796 Hz to 4000 Hz) were shown for all combinations.
Looking at the results of the four combinations of the prior art, it can be seen that if the separation performance of the low band is increased, the separation performance of the high band is lowered. On the other hand, in the present invention, the separation performance of the high band is maintained. The low-frequency separation performance can be improved.
For example, in the combination of “term 2” and “28.3”, the low frequency range of the combination of “4096” and “28.3” is maintained while maintaining the high frequency separation performance of the combination of “2048” and “28.3” in the prior art. Separation performance can be achieved.
Thus, according to the present invention, it is possible to set an appropriate filter length and sensor interval for each frequency band of a signal to be handled, and only a single filter length and sensor interval can be set for the entire band. It can be seen that the separation performance is improved over the prior art.
Moreover, it turns out that it is also possible to improve isolation | separation performance further by using invention of Claim 3 and Claim 4 together.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a configuration of the present invention.
FIG. 2 is a configuration diagram of a separation system in which band-pass filters are connected in series to a separation unit.
FIG. 3 is an explanatory diagram of a separation system configuration and method for giving a band limiting function to a separation filter;
FIG. 4 is an explanatory diagram of a configuration and a method of a determination unit of a frequency band according to an estimated signal source direction.
FIG. 5 is a diagram showing a comparison of separation performance.
[Explanation of symbols]
DESCRIPTION OF SYMBOLS 1 ... Sensor 2 ... Separation system 2-1, ... Separator by ICA, 2-2 ... Band filter, 2-3 ... Filter coefficient generator by ICA, 2-4 ... Discrete Fourier transform unit, 2-5, band limitation processing unit, 2-6, inverse discrete Fourier transform unit, 2-7, separation filter 4-1, signal source direction estimation unit, 4-2 ... Time difference calculation unit, 4-3 ... Frequency range determination unit

Claims (8)

A signal separation device for separating a plurality of independent signals mixed in space from a mixed signal observed by a plurality of sensors,
Means for calculating a coefficient of a digital filter for converting a mixed signal into a separated signal using ICA (independent component analysis); and means for converting the mixed signal into a separated signal using a digital filter based on the coefficient of the digital filter. A plurality of separation systems having separation means by ICA having, and means for passing only a designated frequency band connected in series to the input unit of the separation means,
The length of the digital filter is different for each separation system,
Signal source direction estimation means for estimating the directions of the two signal sources from the observed mixed signal, and signals from the two signal sources reach the two sensors from the estimated direction and the interval between the two sensors used by each separation system. A time difference calculating means for calculating a time difference; and a frequency range determining means for determining a maximum frequency that each separation system passes based on a difference in phase difference determined from the time difference and the signal frequency,
Furthermore, a signal separation device comprising means for integrating the outputs of the plurality of separation systems. A signal separation device for separating a plurality of independent signals mixed in space from a mixed signal observed by a plurality of sensors,
Means for calculating a coefficient of a digital filter for converting a mixed signal into a separated signal using ICA (independent component analysis); and means for converting the mixed signal into a separated signal using a digital filter based on the coefficient of the digital filter. A plurality of separation systems having separation means by ICA having, and means for passing only a designated frequency band connected in series to the input unit of the separation means,
The relative position between the plurality of sensors is different for each separation system, the signal source direction estimation means for estimating the directions of the two signal sources from the observed mixed signal, the estimated direction and the two sensors used by each separation system Based on the interval, a time difference calculating means for calculating the time difference between the signals from the two signal sources reaching the two sensors, and the maximum frequency that each of the separation systems passes is determined based on the difference in phase difference determined from the time difference and the signal frequency. Comprising a frequency range determining means,
Furthermore, a signal separation device comprising means for integrating the outputs of the plurality of separation systems. A signal separation device for separating a plurality of independent signals mixed in space from a mixed signal observed by a plurality of sensors,
Means for calculating coefficients of a digital filter for converting a mixed signal into a separated signal using ICA (independent component analysis); and discrete Fourier transform means for obtaining a frequency characteristic of each filter by performing discrete Fourier transform on the coefficients of the digital filter. Band limiting processing means for obtaining a new frequency characteristic by restricting passage other than the specified frequency from the frequency characteristics of each filter; and an inverse discrete Fourier transform means for obtaining a coefficient of a digital filter reflecting the new frequency characteristic; A plurality of separation systems having means for converting a mixed signal into a separated signal using a digital filter according to a coefficient of the digital filter reflecting a new frequency characteristic,
The length of the digital filter is different for each separation system,
Signal source direction estimation means for estimating the directions of the two signal sources from the observed mixed signal, and signals from the two signal sources reach the two sensors from the estimated direction and the interval between the two sensors used by each separation system. A time difference calculating means for calculating a time difference; and a frequency range determining means for determining a maximum frequency that each separation system passes based on a difference in phase difference determined from the time difference and the signal frequency,
Furthermore, a signal separation device comprising means for integrating the outputs of the plurality of separation systems. A signal separation device for separating a plurality of independent signals mixed in space from a mixed signal observed by a plurality of sensors,
Means for calculating coefficients of a digital filter for converting a mixed signal into a separated signal using ICA (independent component analysis); and discrete Fourier transform means for obtaining a frequency characteristic of each filter by performing discrete Fourier transform on the coefficients of the digital filter. Band limiting processing means for obtaining a new frequency characteristic by restricting passage other than the specified frequency from the frequency characteristics of each filter; and an inverse discrete Fourier transform means for obtaining a coefficient of a digital filter reflecting the new frequency characteristic; A plurality of separation systems having means for converting a mixed signal into a separated signal using a digital filter according to a coefficient of the digital filter reflecting a new frequency characteristic,
The relative position between the plurality of sensors is different for each separation system, the signal source direction estimation means for estimating the directions of the two signal sources from the observed mixed signal, the estimated direction and the two sensors used by each separation system Based on the interval, a time difference calculating means for calculating the time difference between the signals from the two signal sources reaching the two sensors, and the maximum frequency that each of the separation systems passes is determined based on the difference in phase difference determined from the time difference and the signal frequency. Comprising a frequency range determining means,
Furthermore, a signal separation device comprising means for integrating the outputs of the plurality of separation systems. A signal separation method for separating a plurality of independent signals mixed in space from a mixed signal observed by a plurality of sensors,
Calculating a coefficient of a digital filter for converting a mixed signal into a separated signal using ICA (independent component analysis); converting a mixed signal into a separated signal using a digital filter according to the coefficient of the digital filter; a plurality of sets of steps of passing only the frequency band specified prior to separation of the separated signals,
For each set of steps, the length of the digital filter is varied to estimate the direction of the two signal sources from the observed mixed signal, and from the estimated direction and the distance between the two sensors used by each separation system, Calculating a time difference at which a signal from a signal source reaches each of the two sensors, and determining a maximum frequency to pass based on a difference in phase difference determined from the time difference and the frequency of the signal,
And a step of integrating the signals generated in the plurality of sets of steps. A signal separation method for separating a plurality of independent signals mixed in space from a mixed signal observed by a plurality of sensors,
Calculating a coefficient of a digital filter for converting a mixed signal into a separated signal using ICA (independent component analysis); obtaining a frequency characteristic of each filter by performing a discrete Fourier transform on the coefficient of the digital filter; A step of obtaining a new frequency characteristic by restricting passage of a frequency other than the specified frequency from the frequency characteristic of the filter, a step of obtaining a coefficient of a digital filter reflecting the new frequency characteristic, and a digital filter reflecting the new frequency characteristic A plurality of sets of steps for converting a mixed signal into a separated signal using a digital filter with a coefficient of
The length of the digital filter differs for each set of steps,
Estimating the direction of the two signal sources from the observed mixed signal, and calculating the time difference at which the signals from the two signal sources reach each of the two sensors from the estimated direction and the interval between the two sensors used by each separation system. And determining a maximum frequency to be passed based on a difference in phase difference determined from a time difference and a signal frequency,
And a step of integrating the signals generated in the plurality of sets of steps. As a signal separating apparatus according to any of claims 1 to 4, the signal separation program for causing a computer to function. A recording medium readable by a computer on which the signal separation program according to claim 7 is recorded.

Images