JP2007251917A - Similarity calculation apparatus and method, and echo erasure apparatus and method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide similarity calculation techniques for highly accurately estimating similarity in a short time, and to provide echo erasure techniques utilizing the similarity calculation method. <P>SOLUTION: In the present invention, first of all, a similarity coefficient of a spectrum analyzed for each predetermined sample number (frame) and predetermined frequency interval (frequency value). Within a predetermined range in a frequency base direction and within a predetermined range in a time base direction, the sum of products of two signal spectrums and the sum of squares of amplitudes of respective signal spectrums (power spectrums) are determined. Then, the similarity coefficient is determined from the sum of the squares of the sum of products of the two signal spectrums and the squares (power spectral) of amplitudes of the respective signal spectrums. In the present invention, an output erasing echo by means of a method of gain calculation or adaptive filtering using the similarity coefficient is then obtained. <P>COPYRIGHT: (C)2007,JPO&amp;INPIT

Description

  The present invention relates to a similarity calculation device and a similarity calculation method for calculating the similarity between two signal spectra. The present invention also relates to an echo erasing device and an echo erasing method for erasing an acoustic echo using the similarity calculating device or the similarity calculating method.

FIG. 1 shows a functional configuration example of a conventional echo canceling apparatus. FIG. 2 shows a processing flow of the echo canceling apparatus 100. An output signal e (k) in which an echo generated when the reproduction signal x (k) goes from the speaker 1 to the microphone 2 is eliminated is generated. The conventional echo suppression processing based on short-time spectral amplitude (STSA) estimation uses the characteristics that human auditory characteristics are insensitive to the phase and the statistical characteristics of speech and echo. The amplitude component of the echo is subtracted in the frequency domain. Here, k is a number (sample point number) indicating discrete time at a predetermined interval. Sampling is replacing a value in a certain section of a variable with one representative value in order to convert an analog audio signal into a digital signal. For example, the sampling frequency is 16 kHz (16000 times per second). The signal given to the speaker 1 and the signal collected by the microphone 2 are analog signals. In the following description, since a digital signal is handled, it is necessary to perform conversion by a DA converter and an AD converter, respectively, but this is naturally not shown.
The echo canceller 100 includes a reproduction signal frequency analysis unit 101, a collected sound signal frequency analysis unit 102, a similarity coefficient calculation unit 103, a gain calculation unit 104, an integration unit 6, and a frequency synthesis unit 105. .

The frequency analysis unit 101 receives the reproduction signal x (k) as an input and outputs a reproduction signal spectrum _{Xω, l} (step S101). Here, the frequency value ω is the number of the frequency value of the spectrum obtained at a predetermined frequency interval, and the frame number l is the number of the frame subjected to frequency analysis. For example, 256 reproduction signals x (k-255),..., X (k) sampled at 16 kHz are set as one frame, and frequency analysis is performed while shifting half frames (here, 128 points), and the reproduction signal x (k) is obtained. In a frame unit, the frequency band up to 8 kHz is converted into a reproduction signal spectrum X _{ω, l} (ω = 1,..., 128) represented by 128 sample points.

と定義される。類似度係数ｒ_ω，ｌは、０〜１の間の実数値をとる。ここで、Ｅ_ｔ［］は時間領域での変数の期待値（平均値）を表す。例えば、Ｅ_ｔ［Ｘ^＊ _ω，ｌ・Ｙ_ω，ｌ］、Ｅ_ｔ［｜Ｘ_ω，ｌ｜^２］、Ｅ_ｔ［｜Ｙ_ω，ｌ｜^２］はそれぞれ、

により算出することができる。ここで、Ｘ^＊ _ω，ｌはＸ_ω，ｌの共役複素数、ｌ−１は１フレーム過去の値、αは時定数の重み係数（例えば０．３）を表す。
ゲイン計算部１０４は、類似度係数ｒ_ω，ｌと収音信号スペクトルＹ_ω，ｌを入力とし、ゲイン係数Ｇ_ω，ｌを出力する（ステップＳ１０４）。ゲイン係数Ｇ_ω，ｌは０〜１の実数値をとる。類似度が高いとき（類似度係数ｒ_ω，ｌが１に近いとき）には小さい値、類似度が低いとき（類似度係数ｒ_ω，ｌが０に近いとき）には大きい値をとる。積算部６は、収音信号スペクトルＹ_ω，ｌにゲイン係数Ｇ_ω，ｌを積算し、エコー消去信号スペクトルＥ_ω，ｌを得る（ステップＳ６）。周波数合成部１０５は、各周波数成分ωに対応するエコー消去信号スペクトルＥ_ω，ｌから、時間領域の信号ｅ（ｋ）を再合成して出力する（ステップＳ１０５）。
C. Beaugeant, V. Turbin, P. Scalart, A.Gilloire, “New optimal fi1tering approaches forhands-free telecommunication terminals,” SignalProcessing, vo1.64, no.1, pp.33-47, Jan. 1998. The frequency analysis unit 102 receives the collected sound signal y (k) as an input, and outputs a collected sound signal spectrum Y _{ω, l} (step S102). The similarity calculation unit 103 receives the reproduction signal spectrum X _{ω, l} and the collected sound signal spectrum Y _{ω, l} and outputs a similarity coefficient r _{ω, l} (step S103). A method using a coherence function has been proposed so far for estimating the similarity coefficient (Non-Patent Document 1). Reproduced signal spectrum X _{omega, l} sound collection signal spectrum Y _omega, similarity coefficients _l r _{omega, l} is

Is defined. The similarity coefficient _{rω, l} takes a real value between 0 _{and 1} . Here, E _t [] represents the expected value (average value) of the variable in the time domain. For ^{_{example, E t [X * ω,}} l · Y ω, l], E t [| X ω, l | 2], E t [| Y ω, l | 2] , respectively,

Can be calculated. _Here, ^{X *} ω, _l is X _omega, complex conjugate of _l, l-1 represents a 1-frame past values, alpha weighting factor is a time constant (e.g., 0.3).
The gain calculation unit 104 receives the similarity coefficient r _{ω, l} and the collected sound signal spectrum Y _{ω, l} and outputs the gain coefficient G _{ω, l} (step S104). The gain coefficient Gω _{, l} takes a real value from 0 to ₁ . When the similarity is high (when the similarity coefficient _{rω, l} is close to 1), the value is small, and when the similarity is low (when the similarity coefficient _{rω, l} is close to 0), the value is large. Integration unit 6, the sound collection signal spectrum Y _{omega, l} in the gain coefficient G _omega, integrates the _l, echo-canceled signal spectrum E _omega, obtain _l (step S6). The frequency synthesizer 105 re-synthesizes and outputs the time-domain signal e (k) from the echo cancellation signal spectrum E _{ω, l} corresponding to each frequency component ω (step S105).
C. Beaugeant, V. Turbin, P. Scalart, A. Gilloire, “New optimal fi1tering approaches forhands-free telecommunication terminals,” SignalProcessing, vo1.64, no.1, pp.33-47, Jan. 1998.

  The conventional similarity calculation method based on coherence uses only the feature quantity of the time vector. In this method, assuming that the echo path is unchanged, the similarity is calculated from the long-time feature quantity of the time vector. Therefore, when the echo path fluctuates, it takes a certain time until the similarity can be accurately estimated. In practical use, the echo path sometimes fluctuates frequently, so that the degree of similarity cannot be calculated accurately. In addition, erroneous estimation of similarity was one of the causes of musical noise. An object of the present invention is to provide a similarity calculation technique that estimates a similarity with high accuracy in a short time, and to provide an echo suppression technique that suppresses the occurrence of musical noise.

First, a means for obtaining a similarity coefficient of a spectrum analyzed every predetermined number of samples (frame) and every predetermined frequency interval (frequency value) will be described. In the present invention, the sum of the product of two signal spectra and the sum of the square of the amplitude of each signal spectrum are determined within a predetermined range in the frequency axis direction and within a predetermined range in the time axis direction. Then, a similarity coefficient is obtained from the sum of the square of the sum of the products of the two signal spectra and the sum of the square of the amplitude of each signal spectrum (power spectrum).
In the present invention, using the similarity coefficient, an output in which echo is eliminated by a method such as gain calculation or an adaptive filter is obtained.

  In the present invention, by considering the feature quantity of the frequency vector, the feature quantity that can be used for similarity calculation can be supplemented even when the time vector is shortened. Therefore, even if the time vector is shortened, it is possible to prevent a decrease in the accuracy of similarity calculation. Further, when applied to the echo cancellation technique, even if the echo path changes frequently, it can be tracked in a short time.

[First Embodiment]
FIG. 3 shows a functional configuration example of the echo cancellation apparatus of the first embodiment. FIG. 4 shows an example of the processing flow of the echo canceller 200. The echo cancellation apparatus 200 includes a reproduction signal frequency analysis unit 101, a collected sound signal frequency analysis unit 102, a two-dimensional similarity calculation unit 203, a gain calculation unit 104, an integration unit 6, and a frequency synthesis unit 105. The The configuration is the same as that of the prior art except that the similarity calculation unit 103 is replaced with a two-dimensional similarity calculation unit 203. Below, the two-dimensional similarity calculation part 203 different from FIG. 1 is demonstrated.

The two-dimensional similarity calculation unit 203 includes a first calculation unit 2031 that calculates the expected value of the cross spectrum, a second calculation unit 2032 that calculates the sum of the squares of the amplitudes of the reproduction signal spectrum (power spectrum), and the collected sound signal. Third calculation means 2033 is provided for calculating the sum of the square of the spectrum amplitude (power spectrum). The two-dimensional similarity calculation unit 203 receives the reproduction signal spectrum X _{ω, l} and the sound collection signal spectrum Y _{ω, l} as inputs. Then, the first calculation unit 2031, the second calculation unit 2032, and the third calculation unit 2033 perform the expected value (average value) E _{t, f} [X ^* _{ω, l} · Y of the two-dimensional vector of time and frequency. _{ω, l} ], E _{t, f} [| X _{ω, l} | ² ], E _{t, f} [| Y _{ω, l} | ² ],

により算出する（ステップＳ２０３１、Ｓ２０３２、Ｓ２０３３）。ここで、ｍ、ｎ、Ｍ、Ｎはそれぞれ自然数、ω＋ｍはωからｍサンプルシフトした値、ｌ＋ｎはｌからｎフレームシフトした値である。例えば、Ｍを５、Ｎを１０とする。このように、時間軸方向だけでなく、周波数方向にも加算することで、サンプル数を増やすことができる。したがって、時間軸方向のサンプル数を少なく（言い換えると、短時間に）することができる。なお、上記のＥ_ｔ，ｆ［Ｘ^＊ _ω，ｌ・Ｙ_ω，ｌ］、Ｅ_ｔ，ｆ［｜Ｘ_ω，ｌ｜^２］、Ｅ_ｔ，ｆ［｜Ｙ_ω，ｌ｜^２］は、単なる和を求めており、期待値（平均値）を求める式とはなっていない。しかし、後述の類似度係数ｒ’_ω，ｌを求める式で、Ｅ_ｔ，ｆ［Ｘ^＊ _ω，ｌ・Ｙ_ω，ｌ］の二乗が分子、Ｅ_ｔ，ｆ［｜Ｘ_ω，ｌ｜^２］とＥ_ｔ，ｆ［｜Ｙ_ω，ｌ｜^２］の積が分母となるため、単なる和を求めても、類似度係数ｒ’_ω，ｌの値は同じである。

(Steps S2031, S2032, and S2033). Here, m, n, M, and N are natural numbers, ω + m is a value obtained by shifting m samples from ω, and l + n is a value obtained by shifting n frames from l. For example, M is 5 and N is 10. Thus, the number of samples can be increased by adding not only in the time axis direction but also in the frequency direction. Therefore, the number of samples in the time axis direction can be reduced (in other words, in a short time). The above E _{t, f} [X ^* _{ω, l} · Y _{ω, l} ], E _{t, f} [| X _{ω, l} | ² ], E _{t, f} [| Y _{ω, l} | ² ] are It is simply a sum, not an equation for the expected value (average value). However, in a formula for _calculating a similarity coefficient r ′ _{ω, l} described later, the square of E _{t, f} [X ^* _{ω, l} · Y _{ω, l} ] is a numerator, and E _{t, f} [| X _{ω, l} | ² ] And E _{t, f} [| Y _{ω, l} | ² ] is the denominator, and the value of the similarity coefficient r ′ _{ω, l} is the same even if a simple sum is obtained.

In the present embodiment, the calculation means for calculating the sum of the square of the amplitude of the reproduction signal spectrum (power spectrum) and the calculation means for calculating the sum of the square of the amplitude of the collected sound signal spectrum (power spectrum) are separated. . However, the input spectrum can be selected and the sum of the square of the amplitude of the reproduced signal spectrum (power spectrum) and the sum of the square of the amplitude of the collected signal spectrum (power spectrum) can be calculated by the same calculation means. Also good.
Next, the expected value (average value) of the two-dimensional vector of time and frequency E _{t, f} [X ^* _{ω, l} · _{Yω, l} ], E _{t, f} [| X _{ω, l} | ² ], E _{t , F} [| Y _{ω, l} | ² ], the similarity coefficient r ′ _{ω, l} between the reproduction signal spectrum X _{ω, l} and the collected sound signal spectrum Y _{ω, l} is obtained.

により求める（ステップＳ２０３５）。類似度係数ｒ’_ω，ｌは０〜１の間の実数値をとる。
このような方法により類似度係数を求めるので、時間ベクトルを短時間にした場合にも類似度計算に利用できる特徴量を補うことができる。したがって、時間ベクトルを短時間にしても類似度計算の精度の低下を防ぐことができる。また、エコー消去技術に適用した場合には、エコー経路が頻繁に変化する場合でも、短時間で追従できる。

(Step S2035). The similarity coefficient r′ω _{, l} takes a real value between 0 _{and 1} .
Since the similarity coefficient is obtained by such a method, it is possible to supplement the feature quantity that can be used for the similarity calculation even when the time vector is shortened. Therefore, even if the time vector is shortened, it is possible to prevent a decrease in the accuracy of similarity calculation. Further, when applied to the echo cancellation technique, even if the echo path changes frequently, it can be tracked in a short time.

[Second Embodiment]
The echo canceling apparatus 200 ′ of the second embodiment will also be described with reference to the functional configuration example of FIG. 3 and the processing flow example of FIG. The difference from the first embodiment is that an expected value (average value) of a two-dimensional vector of time and frequency E _{t, f} [X ^* _{ω, l} · Y _{ω, l} ], E _{t, f} [| X _{ω, l} In the calculation of | ² ], E _{t, f} [| Y _{ω, l} | ² ], the weighted two-dimensional window function W _{m, n} having the maximum value at the point (ω, l) is used.

  Therefore, the two-dimensional similarity calculation unit 203 ′ of the echo canceller 200 ′ has a weight coefficient calculation unit 2034. The weight coefficient calculation means 2034 is, for example,

により重み付け二次元窓関数を求める（ステップＳ２０３４）。ただし、Ｉは減衰率を決定する値（例えば０．５）である。
次に、第１の計算手段２０３１’、第２の計算手段２０３２’、第３の計算手段２０３３’が、時間と周波数の二次元ベクトルの期待値（平均値）Ｅ_ｔ，ｆ［Ｘ^＊ _ω，ｌ・Ｙ_ω，ｌ］、Ｅ_ｔ，ｆ［｜Ｘ_ω，ｌ｜^２］、Ｅ_ｔ，ｆ［｜Ｙ_ω，ｌ｜^２］を、

により算出する（ステップＳ２０３１’、Ｓ２０３２’、Ｓ２０３３’）。
その他の処理は、第１実施形態と同じである。
［第３実施形態］
図５に、第３実施形態のエコー消去装置の機能構成例を示す。また、図６にエコー消去装置２００ｂの処理フロー例を示す。エコー消去装置２００ｂは、再生信号用の周波数分析部１０１、収音信号用の周波数分析部１０２、二次元類似度計算部２０３ｂ、ゲイン計算部１０４、積算部６、および周波数合成部１０５から構成される。類似度計算部１０３が二次元類似度計算部２０３ｂに置き換わった以外は、従来技術と同じ構成である。以下に、図１と異なる二次元類似度計算部２０３ｂについて説明する。
なお、第３実施形態のエコー消去装置は、再生信号と収音信号のクロススペクトルを式（１）の代わりに式（２）で計算する点で、第１実施形態のエコー消去装置と異なる。

To obtain a weighted two-dimensional window function (step S2034). However, I is a value (for example, 0.5) which determines an attenuation factor.
Next, the first calculation unit 2031 ′, the second calculation unit 2032 ′, and the third calculation unit 2033 ′ use the expected value (average value) E _{t, f} [X ^* _ω] of the two-dimensional vector of time and frequency. _{, L} · Y _{ω, l} ], E _{t, f} [| X _{ω, l} | ² ], E _{t, f} [| Y _{ω, l} | ² ],

(Steps S2031 ′, S2032 ′, S2033 ′).
Other processes are the same as those in the first embodiment.
[Third Embodiment]
FIG. 5 shows an example of the functional configuration of the echo cancellation apparatus of the third embodiment. FIG. 6 shows a processing flow example of the echo canceller 200b. The echo canceller 200b includes a reproduction signal frequency analysis unit 101, a collected sound signal frequency analysis unit 102, a two-dimensional similarity calculation unit 203b, a gain calculation unit 104, an integration unit 6, and a frequency synthesis unit 105. The The configuration is the same as that of the prior art except that the similarity calculation unit 103 is replaced with a two-dimensional similarity calculation unit 203b. Below, the two-dimensional similarity calculation part 203b different from FIG. 1 is demonstrated.
Note that the echo canceller of the third embodiment is different from the echo canceller of the first embodiment in that the cross spectrum of the reproduction signal and the collected sound signal is calculated by Formula (2) instead of Formula (1).

により算出する（ステップＳ２０３１ｂ）。
第１パワースペクトル計算手段２０３２は、再生信号スペクトルＸ_ω，ｌの振幅の二乗の期待値（平均値）Ｅ_ｔ，ｆ［｜Ｘ_ω，ｌ｜^２］を

により算出する（ステップＳ２０３２）。
第２パワースペクトル計算手段２０３３は、収音信号スペクトルＹ_ω，ｌの振幅の二乗の期待値（平均値）Ｅ_ｔ，ｆ［｜Ｙ_ω，ｌ｜^２］を、

により算出する（ステップＳ２０３３）。 The two-dimensional similarity calculation unit 203b is a cross amplitude spectrum calculation unit 2031b that calculates the sum of the products of the amplitude spectrum of the reproduction signal and the collected sound signal, and a first that calculates the sum of the square of the amplitude of the reproduction signal spectrum (power spectrum) A power spectrum calculation unit 2032, a second power spectrum calculation unit 2033 that calculates the sum of the square of the amplitude of the collected sound signal spectrum (power spectrum), and a similarity calculation unit 2035 are provided. The two-dimensional similarity calculation unit 203b receives the reproduction signal spectrum _{Xω, l} and the collected sound signal spectrum _{Yω, l} as inputs.
The cross amplitude spectrum calculating means 2031b is an expected value (average value) E _{t, f} [| X _{ω, l} | · | of the product of the amplitude of the reproduction signal spectrum X _{ω, l and} the amplitude of the collected sound signal spectrum Y _{ω, l.} Y _{ω, l} |]

(Step S2031b).
The first power spectrum calculation means 2032 calculates the expected value (average value) E _{t, f} [| X _{ω, l} | ² ] of the square of the amplitude of the reproduction signal spectrum X _{ω, l} .

(Step S2032).
The second power spectrum calculation means 2033 calculates the expected value (average value) E _{t, f} [| Y _{ω, l} | ² ] of the square of the amplitude of the collected sound signal spectrum Y _{ω, l} ,

(Step S2033).

Here, m, n, M, and N are natural numbers, ω + m is a value obtained by shifting m samples from ω, and l + n is a value obtained by shifting n frames from l. For example, M is 5 and N is 10. Thus, the number of samples can be increased by adding not only in the time axis direction but also in the frequency direction. Therefore, the number of samples in the time axis direction can be reduced (in other words, in a short time).
E _{t, f} [| X _{ω, l} | · | Y _{ω, l} |], E _{t, f} [| X _{ω, l} | ² ], E _{t, f} [| Y _{ω, l} | ² ] Merely calculates the sum and is not an equation for calculating the expected value (average value). However, in a formula for _calculating a similarity coefficient r ′ _{ω, l} described later, the square of E _{t, f} [X _{ω, l} · Y _{ω, l} ] is a numerator, and E _{t, f} [| X _{ω, l} | ² ]. And E _{t, f} [| Y _{ω, l} | ² ] is the denominator. Therefore, even if a simple sum is obtained _, the value of the similarity coefficient r ′ _{ω, l} is the same.

により求める（ステップＳ２０３５）。類似度係数ｒ’_ω，ｌは０〜１の間の実数値をとる。 In the present embodiment, the calculation means for calculating the sum of the square of the amplitude of the reproduction signal spectrum (power spectrum) and the calculation means for calculating the sum of the square of the amplitude of the collected sound signal spectrum (power spectrum) are separated. . However, the input spectrum can be selected and the sum of the square of the amplitude of the reproduced signal spectrum (power spectrum) and the sum of the square of the amplitude of the collected signal spectrum (power spectrum) can be calculated by the same calculation means. Also good.
Next, the similarity calculation means 2035 calculates the expected value (average value) E _{t, f} [| X _{ω, l} | · | Y _{ω, l} |], E _{t, f} [| Using X _{ω, l} | ² ] and E _{t, f} [| Y _{ω, l} | ² ], the similarity coefficient r ′ _ω between the reproduction signal spectrum X _{ω, l} and the collected sound signal spectrum Y _{ω, l l}

(Step S2035). The similarity coefficient r′ω _{, l} takes a real value between 0 _{and 1} .

  Thus, by calculating the similarity only from the amplitude spectrum, a correlation error due to a phase shift can be avoided, and the similarity can be calculated with high accuracy. In addition, since the similarity coefficient is obtained by such a method, it is possible to supplement the feature amount that can be used for similarity calculation even when the time vector is shortened. Therefore, even if the time vector is shortened, it is possible to prevent a decrease in the accuracy of similarity calculation. Furthermore, when applied to the echo cancellation technique, even if the echo path changes frequently, it can follow in a short time.

Hereinafter, an example of a functional configuration of the similarity calculation unit will be described with reference to FIG.
The reproduction signal spectrum X _{ω, l} output from the frequency analysis unit 101 is converted into X _{1 -M, 1 to} X _{1 + M, 1} , X _{2 -M, 1 to} X _{2 + M, 1} ,. _{ω-M, l ~X ω +} M, l, ..., X ωa-M, l ~X ωa + M, is frequency divided as _l. In addition, the collected sound signal spectrum Y _{ω, l} output from the frequency analysis unit 102 is converted into Y _{1-M, 1 to} Y _{1 + M, l} , Y _{2, l to} Y _{2 + M, l} ,. Y _{ω-M, l to} Y _{ω + M, l} ,..., Y _{ωa-M, l to} Y _{ωa + M, l} . Here, ωa is the maximum value of the frequency value number. For example, in each frame, when the spectrum is represented by 128 sample points in the frequency band up to 8 kHz (ω = 1,..., 128), ωa = 128.
Note that when ω <0 or ωa <ω, X _{ω, l} , Y _{ω, l} is outside the range of each frame, so the overlap division unit 2037 and the overlap division unit 2039 have X _{ω, l} = Y _{ω, l} = 0 is output.
Then, from X _{1-M, l to} X _{1 + M, l} and Y _{1- M, l} to Y _{1 + M, l} , E _{t, f} [| X _{1, l} | · | Y _{1, l} |], E _{t, f} [| X1 _{, l} | ² ] and _{Et, f} [| Y1 _{, l} | ² ] are obtained, and the similarity coefficient r'1 _{, l} is calculated therefrom. This is done by shifting the spectrum one by one for all ω, that is, for all X _{ω-M, 1 to} X _{ω + M, l} and Y _{ω-M, 1} to Y _{ω + M, l.} The degree coefficient r ′ _{ω, l} is calculated.

により重み付け二次元窓関数を求める（ステップＳ２０３４）。ただし、Ｉは減衰率を決定する値（例えば０．５）である。 [Fourth Embodiment]
The echo canceling apparatus 200b ′ of the fourth embodiment will also be described with reference to the functional configuration example of FIG. 5 and the processing flow example of FIG. The difference from the third embodiment is that an expected value (average value) E _{t, f} [X _{ω, l} · Y _{ω, l} ], E _{t, f} [| X _{ω, l} | ² ], E _{t, f} [| Y _{ω, l} | ² ] is to use a weighted two-dimensional window function W _{m, n} having a maximum value at the point (ω, l).
Therefore, the two-dimensional similarity calculation unit 203b ′ of the echo canceller 200b ′ has a weight coefficient calculation unit 2034. The weight coefficient calculation means 2034 is, for example,

To obtain a weighted two-dimensional window function (step S2034). However, I is a value (for example, 0.5) which determines an attenuation factor.

により算出する（ステップＳ２０３１ｂ’、Ｓ２０３２’、Ｓ２０３３’）。
その他の処理は、第３実施形態と同じである。 Next, the first calculation unit 2031b ′, the second calculation unit 2032 ′, and the third calculation unit 2033 ′ perform the expected value (average value) E _{t, f} [X _{ω + m, l + n} · Y _{ω + m, l + n} ], E _{t, f} [| X _{ω, l} | ² ], E _{t, f} [| Y _{ω, l} | ² ],

(Steps S2031b ′, S2032 ′, S2033 ′).
Other processes are the same as those in the third embodiment.

[Fifth Embodiment]
FIG. 8 shows a functional configuration example of the echo cancellation apparatus of the fifth embodiment. FIG. 9 shows a processing flow example of the echo canceller 200c. The echo canceller 200c includes a reproduction signal frequency analysis unit 101, a collected sound signal frequency analysis unit 102, a two-dimensional similarity calculation unit 203c, a gain calculation unit 104, an integration unit 6, and a frequency synthesis unit 105. The The configuration is the same as that of the prior art except that the similarity calculation unit 103 is replaced with a two-dimensional similarity calculation unit 203c. Below, the two-dimensional similarity calculation part 203c different from FIG. 1 is demonstrated.
Note that the echo canceller of the fifth embodiment is different from the echo canceller of the first embodiment in that the cross spectrum of the reproduction signal and the collected sound signal is calculated by Equation (3) instead of Equation (1).

により算出する（ステップＳ２０３１ｃ）。すなわち、クロススペクトル計算手段２０３１ｃは、２つの信号のクロススペクトルの時間軸方向の所定の範囲における和の絶対値を求め、その絶対値について周波数軸方向の所定の範囲における和を求める。
第１パワースペクトル計算手段２０３２は、再生信号スペクトルＸ_ω，ｌの振幅の二乗の期待値（平均値）Ｅ_ｔ，ｆ［｜Ｘ_ω，ｌ｜^２］を

により算出する（ステップＳ２０３２）。
第２パワースペクトル計算手段２０３３は、収音信号スペクトルＹ_ω，ｌの振幅の二乗の期待値（平均値）Ｅ_ｔ，ｆ［｜Ｙ_ω，ｌ｜^２］を、

により算出する（ステップＳ２０３３）。 The two-dimensional similarity calculation unit 203c is a cross spectrum calculation unit 2031c that calculates the sum of the cross spectrums of the reproduction signal and the collected sound signal, and a first power spectrum calculation that calculates the sum of the squares (power spectra) of the amplitudes of the reproduction signal spectra. Means 2032, second power spectrum calculation means 2033 for calculating the sum of the squares of the amplitude (power spectrum) of the collected sound signal spectrum, and similarity calculation means 2035 are provided. The two-dimensional similarity calculation unit 203c receives the reproduction signal spectrum _{Xω, l} and the sound collection signal spectrum _{Yω, l} as inputs.
Cross spectrum calculation unit 2031c is reproduced signal spectrum X _{omega, l} sound collection signal spectrum Y _omega, the expected value of the cross-spectrum of the _l (mean _{value) E t, f [| X} ω, l | · | Y ω, l |]

(Step S2031c). That is, the cross spectrum calculation means 2031c calculates the absolute value of the sum of the cross spectra of the two signals in a predetermined range in the time axis direction, and calculates the sum of the absolute values in the predetermined range in the frequency axis direction.
The first power spectrum calculation means 2032 calculates the expected value (average value) E _{t, f} [| X _{ω, l} | ² ] of the square of the amplitude of the reproduction signal spectrum X _{ω, l} .

(Step S2032).
The second power spectrum calculation means 2033 calculates the expected value (average value) E _{t, f} [| Y _{ω, l} | ² ] of the square of the amplitude of the collected sound signal spectrum Y _{ω, l} ,

(Step S2033).

Here, m, n, M ₁ , M ₂ , N ₁ and N ₂ are natural numbers, ω + m is a value obtained by shifting m samples from ω, and l + n is a value obtained by shifting 1 frame by n frames. For example, M ₁ is 5, M ₂ is 5, N ₁ is 10, and N ₂ is 10. Thus, the number of samples can be increased by adding not only in the time axis direction but also in the frequency direction. Therefore, the number of samples in the time axis direction can be reduced (in other words, in a short time).
E _{t, f} [| X _{ω, l} | · | Y _{ω, l} |], E _{t, f} [| X _{ω, l} | ² ], E _{t, f} [| Y _{ω, l} | ² ] Merely calculates the sum and is not an equation for calculating the expected value (average value). However, in a formula for _calculating a similarity coefficient r ′ _{ω, l} described later, the square of E _{t, f} [X _{ω, l} · Y _{ω, l} ] is a numerator, and E _{t, f} [| X _{ω, l} | ² ]. And E _{t, f} [| Y _{ω, l} | ² ] is the denominator. Therefore, even if a simple sum is obtained _, the value of the similarity coefficient r ′ _{ω, l} is the same.

により求める（ステップＳ２０３５）。類似度係数ｒ’_ω，ｌは０〜１の間の実数値をとる。 In the present embodiment, the calculation means for calculating the sum of the square of the amplitude of the reproduction signal spectrum (power spectrum) and the calculation means for calculating the sum of the square of the amplitude of the collected sound signal spectrum (power spectrum) are separated. . However, the input spectrum can be selected and the sum of the square of the amplitude of the reproduced signal spectrum (power spectrum) and the sum of the square of the amplitude of the collected signal spectrum (power spectrum) can be calculated by the same calculation means. Also good.
Next, the similarity calculation means 2035 calculates the expected value (average value) E _{t, f} [| X _{ω, l} | · | Y _{ω, l} |], E _{t, f} [| Using X _{ω, l} | ² ] and E _{t, f} [| Y _{ω, l} | ² ], the similarity coefficient r ′ _ω between the reproduction signal spectrum X _{ω, l} and the collected sound signal spectrum Y _{ω, l l}

(Step S2035). The similarity coefficient r′ω _{, l} takes a real value between 0 _{and 1} .

  The first embodiment is a similarity calculation device based on a two-dimensional complex correlation, and the third embodiment is a similarity calculation device based on a two-dimensional amplitude correlation, whereas the fifth embodiment is a complex correlation in the time direction. By calculating the correlation in the frequency direction from the absolute value of, the correlation error due to the phase shift in the frequency direction can be avoided and the similarity can be calculated with higher accuracy than the two-dimensional complex correlation. Also, the feature amount increases by considering the phase in the time direction, and the similarity can be calculated with high accuracy even if the time vector is shortened by two-dimensional amplitude correlation.

により重み付け二次元窓関数を求める（ステップＳ２０３４）。ただし、Ｉは減衰率を決定する値（例えば０．５）である。 [Sixth Embodiment]
The echo canceling apparatus 200c ′ of the sixth embodiment will also be described with reference to the functional configuration example of FIG. 8 and the processing flow example of FIG. The difference from the fifth embodiment is that an expected value (average value) E _{t, f} [X _{ω, l} · Y _{ω, l} ], E _{t, f} [| X _{ω, l} | ² ], E _{t, f} [| Y _{ω, l} | ² ] is to use a weighted two-dimensional window function W _{m, n} having a maximum value at the point (ω, l).
Therefore, the two-dimensional similarity calculation unit 203c ′ of the echo canceller 200c ′ has a weight coefficient calculation unit 2034. The weight coefficient calculation means 2034 is, for example,

To obtain a weighted two-dimensional window function (step S2034). However, I is a value (for example, 0.5) which determines an attenuation factor.

により算出する（ステップＳ２０３１ｃ’、Ｓ２０３２’、Ｓ２０３３’）。
その他の処理は、第５実施形態と同じである。 Next, the first calculation unit 2031c ′, the second calculation unit 2032 ′, and the third calculation unit 2033 ′ perform the expected value (average value) E _{t, f} [X _{ω + m, l + n} · Y _{ω + m, l + n} ], E _{t, f} [| X _{ω, l} | ² ], E _{t, f} [| Y _{ω, l} | ² ],

(Steps S2031c ′, S2032 ′, S2033 ′).
Other processes are the same as those in the fifth embodiment.

[Seventh Embodiment]
FIG. 10 shows a functional configuration example of an echo canceller using the two-dimensional similarity calculation unit and the adaptive filter of the present invention. A conventional technique may be used for the adaptive filter 301. FIG. 11 shows a functional configuration example of the adaptive filter. Further, the processing flow of the echo canceller 300 is shown in FIG. The echo canceller 300 includes a reproduction signal frequency analysis unit 101, a collected sound signal frequency analysis unit 102, a two-dimensional similarity calculation unit 203, and an adaptive filter 301. The adaptive filter 301 includes a reproduction signal frequency analysis unit 101 ′, a collected sound signal frequency analysis unit 102 ′, an update amount calculation unit 302, an integration unit 6 ′, a filter update unit 303, a filter calculation unit 304, and an addition unit 7. And a frequency synthesizer 105 ′.

The processing flow (steps S101, S102, S203) of the reproduction signal frequency analysis unit 101, the collected sound signal frequency analysis unit 102, and the two-dimensional similarity calculation unit 203 is the same as in the first embodiment. Hereinafter, the processing flow (step S301) of the adaptive filter 301 will be described.
The reproduction signal frequency analysis unit 101 ′ receives the reproduction signal x (k) and outputs the reproduction signal spectrum X _{ω, l} (step S101 ′). The collected sound signal frequency analyzing unit 102 ′ receives the collected sound signal y (k) and outputs the collected sound signal spectrum Y _{ω, l} (step S102 ′). The update amount calculation unit 302 receives the reproduction signal spectrum X _{ω, l} and the echo cancellation signal spectrum E _{ω, l} as inputs, and calculates a filter update amount ΔH ^ _{ω, l} (step S302). As a calculation method of the filter update amount ΔH ^ _{ω, l} , for example, the FLMS (Fast Least-Mean Square) algorithm (Ferrara E.Jr., Widrow B., “Acoustics Speech and Signal Processing”, IEEE Transactions on Signal Processing, vol. .29, Issue 3, Jun 1981pp.679-683.). The filter update amount ΔH ^ _{ω, l} corresponds to the second term on the right side of the equation (15) in the document.

Integrating unit 6 ', the filter update amount [Delta] H ^ _omega, similarity coefficient r to _{l' omega,} integrates the _l, filter accumulated update value _{r 'ω, l · ΔH ^} ω, and outputs the _l (step S6') . The accumulating unit 6 ′ can suppress performance degradation of the echo canceller due to disturbance such as double talk (a state in which the signal of the sender is mixed in the echo signal) noise. The filter integration update amount may be another integration of the filter update amount ΔH ^ _{ω, l} and the similarity coefficient _{r′ω, l} .
The filter update unit 303 receives the filter integrated update amount _{r′ω, l} · ΔH ^ _{ω, l} and outputs the filter coefficient H ^ _{ω, l} (step S303). Here, the filter coefficient H ^ _{ω, l} is
H ^ _{ω, l} = H ^ _{ω, l-1} + _{r'ω, l} · ΔH ^ _{ω, l}
Ask for. Note that l−1 is a value indicating the previous frame. The similarity coefficient r ′ _{ω, l} is a value close to 1 when only the echo signal is collected by the microphone 2 (in the state of received single talk). In addition, when there is a lot of double talk or noise, the value is close to zero. The similarity coefficient r ′ _{ω, l} is added to the filter update amount ΔH ^ _{ω, l} , whereby the update amount of the adaptive filter is controlled.

The filter calculation unit 304 receives the reproduction signal spectrum _{Xω, l} and the filter coefficient H ^ _{ω, l} and outputs a pseudo echo signal spectrum D ^ _{ω, l} (step S304). The pseudo echo signal spectrum D ^ _{ω, l} is
D ^ _{ω, l} = H ^ _{ω, l}・_{Xω, l}
Ask for.
Adding unit 7 is collected sound signal spectrum Y _omega, the pseudo echo signal from the _l spectra D ^ _omega, subtracts _l, echo-canceled signal spectrum E _omega, obtain _l (step S7). The frequency synthesizer 105 ′ re-synthesizes and outputs the time domain signal e (k) from the echo cancellation signal spectrum E _{ω, l} corresponding to each frequency component ω (step S105 ′).

  Since the similarity coefficient is obtained by such a method, the feature quantity that can be used for similarity calculation can be supplemented even when the time vector is shortened. Therefore, even if the time vector is shortened, it is possible to prevent a decrease in the accuracy of similarity calculation. Further, when applied to the echo cancellation technique, even if the echo path changes frequently, it can be tracked in a short time. In place of the two-dimensional similarity calculation unit 203, the two-dimensional similarity calculation unit 203 ′ of the second embodiment, the two-dimensional similarity calculation unit 203b of the third embodiment, and the two-dimensional similarity calculation of the fourth embodiment. The unit 203b ′, the two-dimensional similarity calculation unit 203c of the fifth embodiment, and the two-dimensional similarity calculation unit 203c ′ of the sixth embodiment may be used.

他の実施形態における二次元類似度計算部においても同様である。
ここで、第１実施形態〜第６実施形態において、Ｍ_１、Ｍ_２はサンプリング周波数に依存する整数であり、例えば、サンプリング周波数が１６ｋＨｚの場合、Ｍ_１は２から１０の間の値、Ｍ_２も２から１０の値に設定するとよい。特に、Ｍ_１＝５、Ｍ_２＝５の付近でＭ_１とＭ_２を設定するのが望ましい。サンプリング周波数が２倍になれば、Ｍ_１とＭ_２の値も２倍にするとよい。一方、Ｎ_１、Ｎ_２は使用環境における残響時間及び雑音に係る時定数に依存する自然数で、例えば、Ｎ_１＝１０、Ｎ_２＝０とする。残響時間が長い場合には、例えば、Ｎ_１＝１００、Ｎ_２＝０とする。 [Modifications, etc.]
The similarity calculation method of the present invention may be applied to devices other than the echo canceller. In the two-dimensional similarity calculation unit according to the first to fourth embodiments, generally, a predetermined range −N _{1 to} N ₂ in the time axis direction and a predetermined range −M _{1 to} M _{2 in} the frequency axis direction. The sum of spectral products may be calculated over a range. That is, for example, the first calculation unit 2031 of the two-dimensional similarity calculation unit 203 according to the first embodiment sets N ₁ , N ₂ , M ₁ , and M ₂ as natural numbers, and replaces the equation (1) with the following equation: The sum of cross spectra may be calculated with

The same applies to the two-dimensional similarity calculator in other embodiments.
Here, in the first to sixth embodiments, M ₁ and M ₂ are integers depending on the sampling frequency. For example, when the sampling frequency is 16 kHz, M ₁ is a value between 2 and 10, M ₂ may also be set to a value from 2 to 10. In particular, it is desirable to set M ₁ and M ₂ in the vicinity of M ₁ = 5 and M ₂ = 5. If the sampling frequency is doubled, the values of M ₁ and M ₂ are preferably doubled. On the other hand, N ₁ and N ₂ are natural numbers depending on reverberation time and noise time constant in the usage environment, and for example, N ₁ = 10 and N ₂ = 0. When the reverberation time is long, for example, N ₁ = 100 and N ₂ = 0.

Each step of the similarity calculation method and echo cancellation method according to the present invention may be described as a computer-executable program, and the present invention may be executed by the computer. The program may be recorded on a computer-readable recording medium, and the program may be read and executed.
For example, as shown in FIG. 13, each unit is connected to a bus 50, and a similarity calculation program is stored in the memory 52, or an echo cancellation program is stored in the memory 53 from a CD-ROM, a hard disk, or the like, or via a communication line. Installed. When the CPU 51 executes this similarity calculation program or echo cancellation program, the similarity calculation method or echo cancellation method can be executed. The storage unit 55 is used for temporarily storing data.

  The first calculation means 2031 (FIG. 3) in the first embodiment, the first calculation means 2031 ′ (FIG. 3) in the second embodiment, and the cross amplitude spectrum calculation means 2031b (FIG. 5) in the third embodiment. The cross spectrum calculation means 2031c (FIG. 8) in the fourth embodiment and the cross spectrum calculation means 2031c ′ (FIG. 8) in the fifth embodiment are common in that they obtain the sum of the products of the spectra of the two signals. Yes. The product-sum calculation means in the claims and the specification means these calculation means.

The figure which shows the function structural example of the conventional echo cancellation apparatus. The figure which shows the processing flow of the conventional echo cancellation apparatus. The figure which shows the function structural example of the echo cancellation apparatus of 1st Embodiment, 2nd Embodiment. The figure which shows the processing flow of the echo cancellation apparatus of 1st Embodiment and 2nd Embodiment. The figure which shows the function structural example of the echo cancellation apparatus of 3rd Embodiment and 4th Embodiment. The figure which shows the processing flow of the echo cancellation apparatus of 3rd Embodiment and 4th Embodiment. The figure which shows the function structural example of the two-dimensional similarity calculation part 203b. The figure which shows the function structural example of the echo cancellation apparatus of 5th Embodiment and 6th Embodiment. The figure which shows the processing flow of the echo cancellation apparatus of 5th Embodiment and 6th Embodiment. The figure which shows the function structural example of the echo cancellation apparatus using the two-dimensional similarity calculation part and adaptive filter of this invention. The figure which shows the function structural example of an adaptive filter. The figure which shows the processing flow of the echo cancellation apparatus of 7th Embodiment. The figure which shows the structural example in case the similarity calculation program by this invention and the echo cancellation method are performed by computer.

Claims (24)

A similarity calculation device for calculating a similarity between two signal spectra,
Each signal spectrum is a spectrum analyzed at a predetermined number of samples (hereinafter referred to as “frame”) and a predetermined frequency interval (hereinafter referred to as “frequency value”).
Product sum calculation means for calculating the sum of products of two signal spectra within a predetermined range in the frequency axis direction and within a predetermined range in the time axis direction, and within a predetermined range in the frequency axis direction and a predetermined range in the time axis direction A power spectrum calculation means for obtaining a sum of squares of the amplitudes of each signal spectrum, a similarity coefficient calculation means for obtaining a similarity coefficient from the output of the product-sum calculation means and the output of the power spectrum calculation means apparatus. The similarity calculation apparatus according to claim 1,
The product sum calculation means is a cross spectrum calculation means for obtaining a sum of cross spectra of two signals within a predetermined range in the frequency axis direction and within a predetermined range in the time axis direction. apparatus. The similarity calculation apparatus according to claim 1,
The product sum calculation means is a cross amplitude spectrum calculation means for obtaining a sum of amplitude products of two signal spectra within a predetermined range in the frequency axis direction and within a predetermined range in the time axis direction. Similarity calculation device. The similarity calculation apparatus according to claim 1,
The product-sum calculation unit is a cross-spectrum calculation unit that calculates an absolute value of a sum of a cross spectrum of two signals in a predetermined range in the time axis direction and calculates a sum of the absolute value in a predetermined range in the frequency axis direction. The similarity calculation apparatus characterized by the above. The similarity calculation device according to claim 2,
When X _{ω, l} and Y _{ω, l} are signal spectra for the ω-th frequency value of the l-th frame and N ₁ , N ₂ , M ₁ , M ₂ are natural numbers,
The cross spectrum calculation means includes:

Is used to find the sum of the cross spectrum of the two signals,
The power spectrum calculation means includes

Also

Is used to find the sum of the product of the signal spectrum,
The similarity coefficient calculation means calculates the similarity coefficient r ′ _{ω, l} .

A similarity calculation device characterized by: The similarity calculation device according to claim 3,
When X _{ω, l} and Y _{ω, l} are signal spectra for the ω-th frequency value of the l-th frame and N ₁ , N ₂ , M ₁ , M ₂ are natural numbers,
The cross amplitude spectrum calculating means includes:

Is used to find the sum of the products of the amplitudes of the two signal spectra,
The power spectrum calculation means includes

Also

Is used to find the sum of the product of the signal spectrum,
The similarity coefficient calculation means calculates the similarity coefficient r ′ _{ω, l} .

A similarity calculation device characterized by: The similarity calculation device according to claim 4,
When X _{ω, l} and Y _{ω, l} are signal spectra for the ω-th frequency value of the l-th frame and N ₁ , N ₂ , M ₁ , M ₂ are natural numbers,
The cross spectrum calculation means includes:

Is used to find the sum of the cross spectrum of the two signals,
The power spectrum calculation means includes

Also

Is used to find the sum of the product of the signal spectrum,
The similarity coefficient calculation means calculates the similarity coefficient r ′ _{ω, l} .

A similarity calculation device characterized by: The similarity calculation device according to claim 2,
When X _{ω, l} and Y _{ω, l} are signal spectra for the ω-th frequency value of the l-th frame, N ₁ , N ₂ , M ₁ , M ₂ are natural numbers, and W _{m, n} is a weight,
The cross spectrum calculation means includes:

Is used to find the sum of the cross spectrum of the two signals,
The power spectrum calculation means includes

Also

Is used to find the sum of the product of the signal spectrum,
The similarity coefficient calculation means calculates the similarity coefficient r ′ _{ω, l} .

A similarity calculation device characterized by: The similarity calculation device according to claim 3,
When X _{ω, l} and Y _{ω, l} are signal spectra for the ω-th frequency value of the l-th frame, N ₁ , N ₂ , M ₁ , M ₂ are natural numbers, and W _{m, n} is a weight,
The cross amplitude spectrum calculating means includes:

Is used to find the sum of the products of the amplitudes of the two signal spectra,
The power spectrum calculation means includes

Also

Is used to find the sum of the product of the signal spectrum,
The similarity coefficient calculation means calculates the similarity coefficient r ′ _{ω, l} .

A similarity calculation device characterized by: The similarity calculation device according to claim 4,
When X _{ω, l} and Y _{ω, l} are signal spectra for the ω-th frequency value of the l-th frame, N ₁ , N ₂ , M ₁ , M ₂ are natural numbers, and W _{m, n} is a weight,
The cross spectrum calculation means includes:

Is used to find the sum of the cross spectrum of the two signals,
The power spectrum calculation means includes

Also

Is used to find the sum of the product of the signal spectrum,
The similarity coefficient calculation means calculates the similarity coefficient r ′ _{ω, l} .

A similarity calculation device characterized by: A first frequency analysis unit for frequency analysis of the reproduction signal to obtain a reproduction signal spectrum;
A second frequency analysis unit that performs frequency analysis of the collected sound signal to obtain a collected sound signal spectrum;
The similarity calculation device according to any one of claims 1 to 10, which calculates a similarity coefficient between the reproduction signal spectrum and the collected sound signal spectrum;
A gain calculation unit for obtaining a gain coefficient from the collected sound signal spectrum for the similarity coefficient;
An integration unit that integrates the gain coefficient into the collected sound signal spectrum and obtains an echo cancellation spectrum;
An echo canceller comprising a frequency synthesizer that obtains a time-domain signal from the echo cancel spectrum. A first frequency analysis unit for frequency analysis of the reproduction signal to obtain a reproduction signal spectrum;
A second frequency analysis unit that performs frequency analysis of the collected sound signal to obtain a collected sound signal spectrum;
The similarity calculation device according to any one of claims 1 to 10, which calculates a similarity coefficient between the reproduction signal spectrum and the collected sound signal spectrum;
An echo cancellation apparatus comprising: an adaptive filter unit that obtains an output signal obtained by canceling a component of the reproduction signal in the sound pickup signal, using the reproduction signal, the sound pickup signal, and the similarity coefficient. A similarity calculation method for calculating a similarity between two signal spectra,
Each signal spectrum is a spectrum analyzed at a predetermined number of samples (hereinafter referred to as “frame”) and a predetermined frequency interval (hereinafter referred to as “frequency value”).
A product-sum calculation step for obtaining a sum of products of two signal spectra within a predetermined range in the frequency axis direction and within a predetermined range in the time axis direction; and within a predetermined range in the frequency axis direction and a predetermined range in the time axis direction A power spectrum calculation step for obtaining a sum of squares of amplitudes of respective signal spectra, a similarity coefficient calculation step for obtaining a similarity coefficient from the output of the product-sum calculation step and the output of the power spectrum calculation step. Method. The similarity calculation method according to claim 13,
The product sum calculation step is a cross spectrum calculation step for obtaining a sum of cross spectra of two signals within a predetermined range in the frequency axis direction and within a predetermined range in the time axis direction. Method. The similarity calculation method according to claim 13,
The product sum calculation step is a cross amplitude spectrum calculation step for obtaining a sum of products of amplitudes of two signal spectra within a predetermined range in the frequency axis direction and within a predetermined range in the time axis direction. Similarity calculation method. The similarity calculation method according to claim 13,
The product-sum calculation step is a cross-spectrum calculation step for obtaining an absolute value of a sum of a cross spectrum of two signals in a predetermined range in the time axis direction and calculating a sum of the absolute values in a predetermined range in the frequency axis direction. A similarity calculation method characterized by the above. The similarity calculation method according to claim 14, wherein
When X _{ω, l} and Y _{ω, l} are signal spectra for the ω-th frequency value of the l-th frame and N ₁ , N ₂ , M ₁ , M ₂ are natural numbers,
The cross spectrum calculation step includes:

Is used to find the sum of the cross spectrum of the two signals,
The power spectrum calculation step includes:

Also

Is used to find the sum of the product of the signal spectrum,
In the similarity coefficient calculation step, the similarity coefficient r ′ _{ω, l} is calculated.

A similarity calculation method characterized by being obtained by The similarity calculation method according to claim 15, comprising:
When X _{ω, l} and Y _{ω, l} are signal spectra for the ω-th frequency value of the l-th frame and N ₁ , N ₂ , M ₁ , M ₂ are natural numbers,
The cross amplitude spectrum calculation step includes:

Is used to find the sum of the products of the amplitudes of the two signal spectra,
The power spectrum calculation step includes:

Also

Is used to find the sum of the product of the signal spectrum,
In the similarity coefficient calculation step, the similarity coefficient r ′ _{ω, l} is calculated.

A similarity calculation method characterized by being obtained by The similarity calculation method according to claim 16, comprising:
When X _{ω, l} and Y _{ω, l} are signal spectra for the ω-th frequency value of the l-th frame and N ₁ , N ₂ , M ₁ , M ₂ are natural numbers,
The cross spectrum calculation step includes:

Is used to find the sum of the cross spectrum of the two signals,
The power spectrum calculation step includes:

Also

Is used to find the sum of the product of the signal spectrum,
In the similarity coefficient calculation step, the similarity coefficient r ′ _{ω, l} is calculated.

A similarity calculation method characterized by being obtained by The similarity calculation method according to claim 14, wherein
When X _{ω, l} and Y _{ω, l} are signal spectra for the ω-th frequency value of the l-th frame, N ₁ , N ₂ , M ₁ , M ₂ are natural numbers, and W _{m, n} is a weight,
The cross spectrum calculation step includes:

Is used to find the sum of the cross spectrum of the two signals,
The power spectrum calculation step includes:

Also

Is used to find the sum of the product of the signal spectrum,
In the similarity coefficient calculation step, the similarity coefficient r ′ _{ω, l} is calculated.

A similarity calculation method characterized by being obtained by The similarity calculation method according to claim 15, comprising:
When X _{ω, l} and Y _{ω, l} are signal spectra for the ω-th frequency value of the l-th frame, N ₁ , N ₂ , M ₁ , M ₂ are natural numbers, and W _{m, n} is a weight,
The cross amplitude spectrum calculation step includes:

Is used to find the sum of the products of the amplitudes of the two signal spectra,
The power spectrum calculation step includes:

Also

Is used to find the sum of the product of the signal spectrum,
In the similarity coefficient calculation step, the similarity coefficient r ′ _{ω, l} is calculated.

A similarity calculation method characterized by being obtained by The similarity calculation method according to claim 16, comprising:
When X _{ω, l} and Y _{ω, l} are signal spectra for the ω-th frequency value of the l-th frame, N ₁ , N ₂ , M ₁ , M ₂ are natural numbers, and W _{m, n} is a weight,
The cross spectrum calculation step includes:

Is used to find the sum of the cross spectrum of the two signals,
The power spectrum calculation step includes:

Also

Is used to find the sum of the product of the signal spectrum,
In the similarity coefficient calculation step, the similarity coefficient r ′ _{ω, l} is calculated.

A similarity calculation method characterized by being obtained by A first frequency analysis step of obtaining a reproduction signal spectrum by frequency analysis of the reproduction signal;
A second frequency analysis step of performing frequency analysis of the collected sound signal to obtain a collected sound signal spectrum;
The similarity calculation method according to any one of claims 13 to 22, wherein a similarity coefficient between the reproduced signal spectrum and the collected sound signal spectrum is calculated.
A gain calculating step for obtaining a gain coefficient from the collected sound signal spectrum from the similarity coefficient;
Integrating the gain factor to the collected sound signal spectrum to obtain an echo cancellation spectrum;
An echo cancellation method comprising: a frequency synthesis step of obtaining a time domain signal from the echo cancellation spectrum. A first frequency analysis step of obtaining a reproduction signal spectrum by frequency analysis of the reproduction signal;
A second frequency analysis step of performing frequency analysis of the collected sound signal to obtain a collected sound signal spectrum;
The similarity calculation method according to any one of claims 13 to 22, wherein a similarity coefficient between the reproduced signal spectrum and the collected sound signal spectrum is calculated.
An echo canceling method comprising: an adaptive filter step of obtaining an output signal from which a component of the reproduced signal in the collected sound signal is eliminated, using the reproduced signal, the collected sound signal, and the similarity coefficient.

Images