JP6502307B2 - Echo cancellation apparatus, method and program therefor - Google Patents

Description

  The present invention relates to a method of canceling echo due to acoustic coupling between a speaker and a microphone.

  When the speaker and the microphone are disposed in the same space, the reproduced sound from the speaker wraps around to the microphone, so that echo (echo) mixes in the collected signal collected by the microphone. An echo cancellation method for removing echoes from a collected signal has conventionally been devised (see Non-Patent Document 1).

  The conventional echo cancellation method will be described with reference to FIG. FIG. 1 shows a functional block diagram of an echo canceller which determines an estimate of the transfer characteristic of an acoustic transfer path in the frequency domain. Here, a signal frame for every N samples is configured with respect to the collected signal y (n), and echo cancellation processing is performed with one frame as a processing unit, and the effective length of the impulse response of the sound transmission path is L Assume it is a sample.

  First, the frequency domain conversion unit 10 inputs signals x (nN−L + 1), x (nN−L + 2),... To the speaker 3 from the (N + L−1) sample past to the present time n. , x (n) and subjecting this to discrete Fourier transform to obtain input signals X (0), X (1),..., X (N + L-1) in the frequency domain.

  Next, in each frequency number k (k = 0, 1, ..., N + L-1), the echo estimation unit 20 estimates the echo estimated value Y ^ (k) = H (k) X (k) in the frequency domain. Calculate H (k) is an estimated value of the transfer characteristic of the sound transfer path in the frequency domain.

  The time domain transform unit 30 performs inverse discrete Fourier transform on the echo estimated value Y ^ (k) in the frequency domain to convert the echo estimated value in the time domain y ^ (nN-L + 1), y ^ (nN-L +). 2), ..., y ^ (n) are obtained. Furthermore, the elements y ^ (n-N + 1), y ^ (n-N + 2), ..., y ^ (n) for the last N samples are extracted and output to the subtraction unit 40.

  The subtractor 40 outputs N samples of collected sound signals y (n-N + 1), y (n-N + 2),..., Y (n) and echo estimated values y ^ (n-N + 1) in the time domain. , y ^ (n-N + 2),..., y ^ (n), and the residual signal e (n-N + 1), e (n-N + 2),. Get). This residual signal is also output to the transmitting end 5 as an echo cancellation signal.

  The frequency domain transformation unit 50 adds a zero sequence of L points before the sequence of the residual signals e (n-N + 1), e (n-N + 2),. Thus, a residual signal E (k) in the frequency domain of N + L points is obtained.

  The transfer path characteristic update unit 60 updates the estimated value H (k) of the transfer characteristic of the sound transfer path in the frequency domain at each frequency number k according to the following equation.

Here, μ _k is a step size for adjusting the update amount, P (k) is a normalization coefficient depending on the magnitude of the power of the input signal, and ^* is a complex conjugate.

  The echo canceler 90 returns to the processing of the frequency domain transform unit 10 when discrete time for N samples has passed, and reacquires the input signal x (n) and the collected signal y (n), repeat.

C. Breining et al., "Acoustic echo control. An application of very-high-order adaptive filters", in IEEE Signal Processing Magazine, vol. 16, no. 4, pp. 42-69, Jul 1999.

  Many echo cancellation methods are designed for voice calls, and it is required to reduce the processing delay for echo cancellation in order to realize smooth calls. For this reason, it is necessary to estimate echo estimation values sequentially for a small number of samples of the collected signal and output an echo cancellation signal. In many cases, by first estimating the impulse response of the acoustic transmission path between the speaker and the microphone, and applying the input signal x (n) to the speaker 3 thereto, estimation of the echo estimation value is realized. The impulse response has a length of several tens to several hundreds ms, and estimates of all these sample values are made with reference to a few samples of the collected signal sequentially given and samples of the input signal to the speaker 3 , Will run. For this reason, the impulse response can be obtained by sequentially solving an indeterminate simultaneous equation in which the number of unknowns (the number of samples of the impulse response) exceeds the number of equations (the number of samples of the collected signal) Each sample will be estimated. For this reason, in the conventional echo cancellation method, it takes a certain time until the echo signal is practically sufficiently canceled. Also in the process shown in FIG. 1, the estimated value H (k) of the transfer characteristic of the sound transfer path in the frequency domain at each frequency number k is properly estimated, and is constant to sufficiently cancel the echo (echo). Requires time signal input.

  The present invention provides an echo canceler capable of generating a series of echo cancellation signals in which echoes are canceled more appropriately than in the past, and a method and program thereof, for a given sample sequence of collected sound signals. With the goal.

In order to solve the above problems, according to one aspect of the present invention, the echo canceler has the number N of samples as a processing unit of echo cancellation processing more than the number of samples of the effective length of the assumed impulse response. The first frequency domain transform unit converts the input signal x in the time domain into the input signal X in the frequency domain every N samples, the index indicating the number of repetitions is p, and the input signal X and the frequency domain are large. Using the estimated value H _p of the transfer characteristic of the acoustic transfer path to obtain the echo estimated value Y ^ _p in the frequency domain, and the echo estimated value Y ^ _p as the echo estimated value y ^ _p in the time domain A time domain transformation unit to be converted, a subtraction unit to obtain a time domain residual signal e _p which is a difference between the time domain collected sound signal y and an echo estimated value y ^ _p , and a time domain residual signal e _p a second frequency domain converter for converting the residual signal E _p in the frequency domain, Using the force signal X and the residual signal E _p, and updates the estimated value H _p, and the transmission path characteristic update section for obtaining an estimate H _{p + 1} after updating, and the input signal X and the collected sound signal y Control to repeat the processing in the echo estimation unit, the time domain conversion unit, the subtraction unit, the second frequency domain conversion unit, and the transmission path characteristic update unit until the updated estimated value H _{p + 1} converges using Including the department.

In order to solve the above problems, according to another aspect of the present invention, the echo cancellation method has the number N of samples as a processing unit of echo cancellation processing more than the number of samples of the effective length of the assumed impulse response. The first frequency domain conversion step of converting the input signal x in the time domain to the input signal X in the frequency domain every N samples, the index indicating the number of repetitions is p, and the input signal X and the frequency by using the estimated value H _p of the transfer characteristic of the acoustic transmission paths in the region, and the echo estimation step of obtaining the echo estimate Y ^ _p in the frequency domain, of the echo estimate Y ^ _p time domain echo estimate y ^ _p and the time domain conversion step of converting into a subtraction step of obtaining a residual signal e _p in the time domain which is the difference between the collected signal y and the echo estimate y ^ _p in the time domain, the residual signal e _p in the time domain to convert the residual signal E _p in the frequency domain A second frequency domain transforming step, by using the input signal X and the residual signal E _p, and updates the estimated value H _p, and a transmission path characteristic update step of obtaining an estimate H _{p + 1} after updating, The echo estimation step, the time domain conversion step, the subtraction step, the second frequency domain conversion step, and the transfer path characteristic update step until the updated estimated value H _{p + 1} converges using the input signal X and the collected sound signal y Repeat the process in and.

  According to the present invention, it is possible to generate an echo cancellation signal sequence in which the echo is canceled more appropriately than in the past, for a given sample sequence of the collected sound signal.

The functional block diagram of the echo canceler which concerns on a prior art. FIG. 2 is a functional block diagram of an echo canceler according to the first embodiment. The figure which shows the example of the processing flow of the echo canceler which concerns on 1st embodiment. The figure for demonstrating the process example of the transfer path characteristic update part which concerns on 1st embodiment. The figure for demonstrating the process example of the transmission path characteristic update part which concerns on the modification of 1st embodiment.

  Hereinafter, embodiments of the present invention will be described. In the drawings used in the following description, the same reference numerals are given to constituent parts having the same functions and steps for performing the same processing, and redundant description will be omitted. In the following description, the symbol “^” or the like used in the text should be written directly above the preceding character, but due to the limitation of the text notation, it is written immediately after the character. In the formula, these symbols are described at their original positions. Moreover, the processing performed in each element unit of a vector or a matrix is applied to all elements of the vector or the matrix unless otherwise noted.

<Point of the first embodiment>
In the echo cancellation method according to this embodiment, the number of equations to be satisfied (the number N of samples of the sound collection signal) is larger than the number of unknowns (the number L of samples of the impulse response). That is, in the present embodiment, it is assumed that the delay of the output of the echo cancellation signal is slightly delayed. For example, in a voice recognition system, not a voice communication application, when a guidance voice or a notification sound prompting a speaker to utter a voice is reproduced from the speaker, the echo (echo) coming around for voice recording is erased, and the utterer It is assumed that only the voice of is given to the speech recognition system. In this case, in some cases, the delay time to output the echo cancellation signal may be slightly increased as compared with the voice communication application. In this embodiment, as described above, it is assumed that the output delay is slightly delayed compared to the voice communication application. That is, in the prior art of FIG. 1, the number N of samples of the collected sound signal handled in one processing unit corresponds to, for example, about 10 ms in voice communication, in this case about 100 ms or several 100 ms. It is assumed that the length can be set. Even in this case, although the prior art of FIG. 1 is applicable, when applied as it is, the problem that the echo is not sufficiently eliminated for a certain period of time still occurs. This is because, in the case of voice communication applications, the number of unknowns (number L of samples of impulse response) exceeds the number of equations to be satisfied (number N of samples of collected sound signal), and L> N underdetermined simultaneous equations There is a problem in the situation where it is not possible to uniquely determine the estimated value of the transfer characteristic. On the other hand, in the present embodiment, the problem is solved by solving a well-determined simultaneous equation of L <N. For example, the estimated value of the transfer characteristic can be uniquely determined by the least squares method. However, the estimated value of the transfer characteristic obtained by the prior art of FIG. 1 is not a solution of the least squares method. For this reason, there still remains the problem that the echo is not sufficiently eliminated for a certain period of time. In order to apply the least squares method in practice, it is necessary to solve the inverse of the correlation matrix for the input signal, but it is very difficult to solve the inverse of the matrix of very high dimensionality numerically and stably is there. For this reason, in the present embodiment, without directly solving the inverse matrix, the estimated value of the transfer characteristic is recursively realized by a method of approaching the least square solution.

  Many of the conventional echo cancellation methods are applied to voice communications, and under the requirement to reduce the delay time of the output signal (echo cancellation signal), solve underdetermined simultaneous equations whose solution is not uniquely determined at that time. Thus, the transfer characteristic of the sound transfer path between the speaker and the microphone was estimated. Therefore, it takes a certain time to estimate the transfer characteristic, and a time interval occurs in which the echo can not be sufficiently eliminated at the beginning of the process or immediately after the change of the transfer characteristic. In the present embodiment, the application (for example, a voice recognition system other than voice communication) that allows the number of equations to be satisfied (the number N of samples of the collected sound signal) to be larger than the number of unknowns (the number L of samples of impulse response). When applying the echo cancellation method to applications where the requirements for delay time of the output signal are relaxed), the problem of solving an overdetermined simultaneous equation in which the estimation of the transfer characteristics of the sound transfer path can uniquely determine the solution each time I focused on what I can handle as However, even though the requirements have changed, simply applying the conventional echo cancellation method as it is does not solve the overdetermined simultaneous equations, and the echo cancellation may still be insufficient. We have devised a new method for finding the least squares solution of simultaneous equations of. In this method, in order to avoid the numerical instability of directly calculating the inverse matrix of the very high-dimensional correlation matrix for the input signal which is usually required in the least squares method, the conventional system of underdetermined simultaneous equations is used. The evaluation process of the recursive residual signal using the sample of the same input signal and the collection signal which did not make sense in the case of solution was added. Thus, in the over-decision requirement, the evaluation of the newly added recursive residual signal acts as a numerically stable substitute for directly solving the inverse matrix calculation in the least squares method, It is possible to estimate the transfer characteristics of the sound transfer path accurately.

<Election cancellation device 100>
FIG. 2 shows a functional block diagram of the echo canceller 100 according to the first embodiment, and FIG. 3 shows its processing flow.

  Echo cancellation apparatus 100 includes a frequency domain transform unit 110, an echo estimation unit 120, a time domain transform unit 130, a subtractor 140, a frequency domain transform unit 150, a transfer path characteristic update unit 160, and a control unit 170.

The echo canceler 100 receives an input signal x (n) input via the receiving end 2 and a collected signal y (n) collected via the microphone 4. The speaker 3 and the microphone 4 are disposed in the same space. The input signal x (n) is reproduced by the speaker 3, and the reproduced sound wraps around to the microphone 4, so that echo (echo) mixes in the collected signal y (n) collected by the microphone 4. The echo canceler 100 removes the echo (echo) estimation value from the collected signal y (n), obtains an echo cancellation signal (error signal e _P (n)), and outputs the echo cancellation signal (error signal e _P (n)) to the transmitting end 5. The receiving end 2 and the transmitting end 5 are connected to, for example, a dialogue system or the like which performs speech while performing speech recognition and realizes dialogue with the user.

  The processing content of each part will be described below.

<Frequency domain converter 110>
The frequency domain conversion unit 110 receives the input signal x (n), and acquires the input signal x (n) to the speaker 3 from the past N + L-1 samples to the present time n every time N samples are acquired. The signals x (nN-L + 1), x (nN-L + 2),..., X (n) are acquired, and (L + N) time domain input signals x (nN-L + 1), Convert x (nN-L + 2), ..., x (n) into frequency domain input signals X (0), X (1), ..., X (N + L-1) (S110), and estimate echo It outputs to the part 120 and the transfer path characteristic update part 160. For example, the signal is converted to a frequency domain signal by discrete Fourier transform. In addition, let L be the number of samples of the assumed impulse response, and N be any integer larger than L. Assuming that k = 0, 1,..., N + L−1, X (0), X (1),..., X (N + L−1) are also expressed as X (k). The past input signals x (nN-L + 1), x (nN-L + 2), ..., x (n-1) are stored in the storage unit (not shown), and every N samples are acquired from the storage unit You just need to get it. Note that other data (acquired signals, signals obtained by calculation, and the like) may be stored in a storage unit (not shown) as necessary and acquired.

<Echo estimation unit 120>
The echo estimation unit 120 receives the input signal X (k) and the estimated value H _p (k) of the transfer characteristic of the acoustic transfer path, and uses these values to calculate an echo estimated value Y ^ _p (k in the frequency domain). ) Is output (S120) to the time domain conversion unit 130. For example, the echo estimated value ^ _p (k) is obtained by the following equation.
Y ^ _p (k) = H _p (k) X (k) (11)
As described above, in this embodiment, the process of evaluating the residual residual signal using the same input signal and the sample of the collected signal is added. p is an index indicating the number of repetitions when performing recursive evaluation processing, and it is assumed that p = 1, 2,.

<Time domain conversion unit 130>
The time domain transform unit 130 receives the echo estimated value Y ^ _p (k) in the frequency domain as an input, and estimates the echo estimated value Y ^ _p (0), Y ^ _p (1), ..., Y ^ _p (N) in the frequency domain. Convert + L-1) to the time domain echo estimate y ^ _p (nN-L + 1), y ^ _p (nN-L + 2), ..., y ^ _p (n) (S130) and subtract Output to section 140. For example, echo estimated values y ^ _p (n-N + 1), y ^ _p (n-N + 2), ..., y ^ _p (n) for the last N samples are extracted and output to the subtraction unit 140 May be The time domain conversion unit 130 may convert the echo estimation value in the frequency domain into an echo estimation value in the time domain by a conversion method corresponding to the inverse conversion of the conversion method of the frequency domain conversion unit 110. For example, the signal is converted to a time domain signal by inverse discrete Fourier transform.

<Subtractor 140>
The subtractor unit 140 receives the echo estimated values y ^ _p (nN-L + 1), y ^ _p (nN-L + 2), ..., y ^ _p (n) and the collected sound signal y (n). Do. Echo estimate _{y ^ p (nN-L +} 1), y ^ p (nN-L + 2), ..., y ^ p last N samples of the echo estimate y ^ _p of the (n) (n- N + 1), y ^ _p (n-N + 2), ..., y ^ _p (n), and the collected sound signal y (n-N + 1), y from N-1 samples past to the current time n The residual signal e _p (n-N + 1), e _p (n-N + 2), ..., e _p (n) which is the difference between (n-N + 2), ..., y (n) The signal is obtained (S140) and output to the frequency domain conversion unit 150. For example, e _p (i) = y (i) −y ^ _p (i) (i = n−N + 1, n−N + 2,..., N). In the case of p = P, the residual signal e _p (n−N + 1), e _p (n−N + 2),..., E _p (n) is also transmitted to the transmitting end 5 as the echo cancellation signal. Output.

<Frequency domain converter 150>
Frequency domain transform section 150, the residual signal e _p in the time domain (n-N + 1), e p (n-N + 2), ..., and enter the e _p (n), the frequency domain these values The residual signals E _p (0), E _p (1),..., E _p (N + L−1) are converted (S150), and output to the transfer path characteristic updating unit 160. The frequency domain transform unit 150 may transform the residual signal in the time domain into a residual signal in the frequency domain by a conversion scheme similar to that of the frequency domain transform unit 110. For example, a zero sequence of L points is added to a sequence of residual signals e _p (n-N + 1), e _p (n-N + 2), ..., e _p (n), and discrete Fourier transform , N + L point frequency domain residual signals E _p (k) are obtained.

<Transmission route characteristic update unit 160>
The transfer path characteristic updating unit 160 is configured to receive the input signals X (0), X (1),..., X (N + L-1) and the residual signal E _p (0), E _p (1) _,. The estimated value H _p (k) is updated using (N + L−1) as the input and these values are used (S 160), and the updated estimated value H _{p + 1} (k) is used as the echo estimation unit 120. Output to For example, the estimated value H _p (k) is updated by the following equation.

Here, μ _k is a step size for adjusting the update amount, P (k) is a normalization coefficient depending on the magnitude of the power of the input signal, and ^* is a complex conjugate. Incidentally, in obtaining the estimated value H _{p + 1} (k) in the frequency domain, estimate H _{p + 1} estimates were converted time domain _{(k) h p + 1 (} 0), h p + 1 (1) ,..., H _{p + 1} (N + L−1) may be constrained in the time domain in the update of the transfer characteristic of the sound transfer path such that the elements other than the first half sample become zero. For example, (i) the updated estimated value H _{p + 1} (k) may be directly converted to the time domain, elements other than the first half samples may be replaced with zeros, and converted again to the frequency domain signal ii) Only the update term part of the second term of the right side of the equation (12) is converted to the time domain, the elements other than the first half sample are replaced with zeros, converted to the frequency domain signal again, and updated by the equation (12) You may By replacing the noise generated in the latter sample shown in FIG. 4A with zero (see FIG. 4B), it is possible to estimate the transfer characteristic more stably.

<Control unit 170>
The control unit 170 performs processes S120, S130, S140, S150, and S160 in the echo estimation unit 120, the time domain conversion unit 130, the subtraction unit 140, the frequency domain conversion unit 150, and the transfer path characteristic update unit 160 until the predetermined condition is satisfied. Control to repeat (S170). In the processes S120 and S160 in the echo estimation unit 120 and the transfer path characteristic update unit 160, the input signals X (0), X (1),..., X (N + L-1) are used regardless of the number of repetitions p. In processing S140 in the subtraction unit 140, the collected sound signals y (n), y (n-1),..., Y (n-N + 1) are used regardless of the number of repetitions p.

The predetermined condition may be any condition that can check whether the estimated value H _{p + 1} (k) has converged. For example, (i) whether or not the process has been repeated a predetermined number of times (if the process is repeated a predetermined number of times or more, it is assumed that the estimated value H _{p + 1} (k) converges), (ii) Whether or not the processing is repeated for a predetermined time (if the processing is repeated for a predetermined time or more, the estimated value H _{p + 1} (k) is assumed to converge), (iii) the estimated value H _{p +} It can be checked whether convergence has occurred by determining whether the difference between ₁ (k) and the estimated value H _p (k) is smaller than a predetermined threshold. For example, (i) when the process is repeated a predetermined number of times, (ii) when the process is repeated for a predetermined time, (iii) between the estimated value H _{p + 1} (k) and the estimated value H _p (k) When the difference is smaller than a predetermined threshold value, it is determined that the estimated value H _{p + 1} (k) has converged.

  In the case of (i), the control unit 170 counts the number of times p of the processes S120, S130, S140, S150, and S160, and repeats the above process until the predetermined number of times P is exceeded. Send

  In the case of (ii), the control unit 170 measures an elapsed time after performing the first process (for example, echo estimation process S120), and continues until a predetermined time (for example, discrete time for N samples) elapses. Control signals are transmitted to the respective units so as to repeat the above-described processing.

In the case of (iii), the control unit 170 obtains the estimated value H _{p + 1} (k) and the estimated value H _p (k), and determines whether the difference is larger than a predetermined threshold value. In such a case, control signals are transmitted to each unit so as to repeat the above-described processing.

Note that these conditions may be combined. For example, when the process is repeated a predetermined number of times of the condition (i), or the difference between the estimated value H _{p + 1} (k) and the estimated value H _p (k) is smaller than a predetermined threshold Then, it is determined that the estimated value H _{p + 1} (k) has converged.

Controller 170, the subtraction unit 140, until the condition is satisfied residual signal _{e p (n-N + 1} ), e p (n-N + 2), ..., e p (n) is the frequency domain converter Control is performed so that the signal is output only to 150 and is output to the frequency domain conversion unit 150 and the transmitting end 5 when the condition is satisfied. In the present embodiment, an example in which only the condition (i) is applied is shown. Therefore, if p = 1, 2,..., P, and p = 1, 2,..., P−1, the residual signal e _P (n−N + 1), e _P (n−N + 2) ,..., E _P (n) are output only to the frequency domain transform unit 150, and when p = P, they are output to the frequency domain transform unit 150 and the transmitting end 5.

When the discrete time for N samples has passed, the echo canceler 100 returns to the process S110, acquires the input signal x (n) and the collected sound signal y (n) again, and repeats the processes S110 to S170. In addition, whenever discrete time for N samples elapses, a past estimated value (for example, an estimated value H obtained last in one frame ago) is used as the value of the initial value H ₁ (k) of the estimated value H _p (k). _P (k)) may be inherited and used as it is, or may be reset to zero.

The difference from the prior art in FIG. 1 is that the process S170 is added, and the estimated value H _p (k) of the transfer characteristic of the updated frequency domain is reapplied to the same input signal and the collected sound signal. The difference is evaluated and the update is repeated recursively. By this repetition, the estimated value H _p (k) of the transfer characteristic can be brought close to the least square solution without performing the inverse matrix operation of the correlation matrix of the input signal. Therefore, according to this embodiment, in the case of L <N, it is possible to output a signal whose echo has been sufficiently canceled immediately after the start of the process.

<Effect>
With the above configuration, the echo cancellation method performs echo cancellation with one sound processing signal having a number N of samples greater than the number L of samples of the impulse response assumed to have transfer characteristics between the speaker 3 and the microphone 4. The least squares solution of the transfer characteristic can be approximately obtained without directly calculating the inverse of the correlation matrix for the input signal. As a result, it is not necessary to require input of a signal for a certain period of time to estimate transfer characteristics, and it is possible to output a signal whose echo has been sufficiently canceled immediately after the start of processing. In the present embodiment, the echo estimation (S120) and the transfer characteristic estimation (S160) are performed in the frequency domain, but these processes may be performed in the time domain. However, in this case, the amount of calculation increases. Also, in this case, the frequency domain transform unit 110, the time domain transform unit 130, and the frequency domain transform unit 150 may be deleted. The present invention can also be applied to echo cancellation in the case where the input signal to the speaker and the sound pickup signal from the microphone have multiple channels, and similar effects can be obtained.

<Modification>
If the impulse response length of the sound transmission path can be effectively regarded as shorter than N, the present invention can be applied and the effect can be obtained even in the case of L> N. That is, at the stage of sound reproduction or sound collection, a delay occurs before the input signal x (n) is actually reproduced from the speaker 3 due to signal buffer processing or the like, or the sound is collected by the microphone 4. When a delay occurs before the collected signal y (n) is actually acquired for echo cancellation processing, if this delay can not be known in advance, the value of L including its delay time In some cases, it may be necessary to set N larger than N. Even in that case, the present invention is applicable. Hereinafter, the application method will be described focusing on parts different from the first embodiment.

<Transmission route characteristic update unit 160>
The transfer path characteristic updating unit 160 is configured to receive the input signals X (0), X (1),..., X (N + L-1) and the residual signal E _p (0), E _p (1) _,. When the estimated value H _p (k) is updated using (N + L−1) as an input and these values are used (S 160), the updated estimated value H ′ _{p + 1} (k) In the domain, only the effective length L 'samples of the impulse response of the sound transmission path are constrained so as to take non-zero values. For example, once the updated estimated value H ′ _{p + 1} (k) is estimated in the time domain h ′ _{p + 1} (0), h ′ _{p + 1} (1),..., H ′ _{p + 1} (N In the first half L samples, the section including the maximum peak value is cut out by the length of the effective length L ′ of the impulse response of the sound transmission path. That is, the maximum peak value of the estimated values h ′ _{p + 1} (0), h ′ _{p + 1} (1),..., H ′ _{p + 1} (N + L−1) in the time domain is detected, and the peaks The sample value for L 'samples is cut out from the sample value of the value or from the sample value several samples before the peak value. After replacing elements other than the cut out interval with zero, estimated values h _{p + 1} (0), h _{p + 1} (1),..., H _{p + 1} (N + L-1) of the time domain after replacement Is again converted to the estimated value H _{p + 1} (k) in the frequency domain to constrain the transfer characteristic of the acoustic transfer path in the time domain. Note that N is set so as to satisfy L>N> L ′. In this modification, the problem is solved by solving a system of equations in which L '<N is substantially determined. As in the first embodiment, the estimated value of the transfer characteristic is recursively brought closer to the least squares solution. be able to. Further, among the first half L samples shown in FIG. 5A, noise occurring in areas other than the section including the largest peak value is eliminated by replacing it with zero (see FIG. 5B), and more stable transfer characteristics can be estimated. . Such restraint in the time domain that narrows the effective range of the first half L sample can be flexibly coped with the fluctuation of the transfer characteristic if it is not applied at the initial stage of repetition of S120 to S160. For example, assuming that S120 to S160 are repeated ten times (P = 10), the first five times do not constrain the effective range of the first half L sample, and the estimated sound is used in the second five times of the repetition It can be implemented to specify a section in which the impulse response of the transmission path takes the maximum peak value, and apply a constraint that makes the other than that section zero.

<Effect>
With such a configuration, the same effect as that of the first embodiment can be obtained when L> N.

<Other Modifications>
The present invention is not limited to the above embodiments and modifications. For example, the various processes described above may be performed not only in chronological order according to the description, but also in parallel or individually depending on the processing capability of the apparatus that executes the process or the necessity. In addition, changes can be made as appropriate without departing from the spirit of the present invention.

<Program and Recording Medium>
In addition, various processing functions in each device described in the above-described embodiment and modification may be realized by a computer. In that case, the processing content of the function that each device should have is described by a program. By executing this program on a computer, various processing functions in each of the above-described devices are realized on the computer.

  The program describing the processing content can be recorded in a computer readable recording medium. As the computer readable recording medium, any medium such as a magnetic recording device, an optical disc, a magneto-optical recording medium, a semiconductor memory, etc. may be used.

  Further, this program is distributed, for example, by selling, transferring, lending, etc. a portable recording medium such as a DVD, a CD-ROM or the like in which the program is recorded. Furthermore, the program may be stored in a storage device of a server computer, and the program may be distributed by transferring the program from the server computer to another computer via a network.

  For example, a computer that executes such a program first temporarily stores a program recorded on a portable recording medium or a program transferred from a server computer in its own storage unit. Then, at the time of execution of the process, the computer reads the program stored in its storage unit and executes the process according to the read program. In another embodiment of the program, the computer may read the program directly from the portable recording medium and execute processing in accordance with the program. Furthermore, each time a program is transferred from this server computer to this computer, processing according to the received program may be executed sequentially. In addition, a configuration in which the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes processing functions only by executing instructions and acquiring results from the server computer without transferring the program to the computer It may be Note that the program includes information provided for processing by a computer that conforms to the program (such as data that is not a direct command to the computer but has a property defining the processing of the computer).

  In addition, although each device is configured by executing a predetermined program on a computer, at least a part of the processing content may be realized as hardware.

Claims (5)

Assuming that the number N of samples to be processed as the echo cancellation processing is larger than the number of samples of the effective length of the assumed impulse response, the input signal x in the time domain is converted to the input signal X in the frequency domain every N samples. A first frequency domain converter to convert;
An index indicating the number of repetition times and p, by using the estimated value H _p of the transfer characteristic of the acoustic transmission path of the input signal X and the frequency domain, and echo estimator for determining the echo estimate Y ^ _p in the frequency domain ,
And the time domain converter for converting the echo estimate Y ^ _p to echo estimate y ^ _p in the time domain,
A subtraction unit for obtaining the residual signal e _p in the time domain which is the difference between the collected signal y and the echo estimate y ^ _p in the time domain,
A second frequency domain converter for converting the residual signal e _p in the time domain to the residual signal E _p in the frequency domain,
A transfer path characteristic updating unit that updates the estimated value H _p using the input signal X and the residual signal E _p to obtain the updated estimated value H _{p + 1} ;
The echo estimation unit, the time domain conversion unit, the subtraction unit, and the second frequency domain conversion until the updated estimated value H _{p + 1} converges using the input signal X and the collected sound signal y And a control unit that performs control so as to repeat processing in the unit and the transfer path characteristic update unit,
Echo canceler. The echo canceler according to claim 1, wherein
Let L be the number of samples including the expected delay and the effective length of the impulse response, and N be any integer larger than L, and the first frequency domain transform unit generates (L + N) pieces for every N samples. , X (n) of the time domain input signal x (nN−L + 1), (nN−L + 2),..., X (n) to the frequency domain input signal X (0), X (1),. Convert to + L-1),
The echo estimation unit estimates estimated values H _p (0), H _p (1) of the transfer characteristics of the acoustic transmission path with the input signals X (0), X (1),..., X (N + L-1). ,..., H _p (N + L−1), and the echo estimated values Y ^ _p (0), Y ^ _p (1),..., Y ^ _p (N + L−1) in the frequency domain Ask for
The time domain transformation unit is configured to estimate the echo estimated values Y ^ _p (0), Y ^ _p (1), ..., Y ^ _p (N + L-1) as the time domain echo estimated value y ^ _p (nN- L + 1), y ^ _p (nN-L + 2), ..., y ^ _p (n),
The subtractor unit is configured to calculate a collected signal y (n−N + 1), y (n−N + 2),..., Y (n) and an echo estimated value y _p (n−N + 1), in the time domain. _{y ^ p (n-n +} 2), ..., the residual signal is a difference between _{y ^ p (n) e p} (n-n + 1), e p (n-n + 2), ..., e _{Find p} (n),
The second frequency domain transform unit is configured to receive the residual signal e _p (n−N + 1), e _p (n−N + 2),..., E _p (n) as a frequency domain residual signal in the time domain. Convert to E _p (0), E _p (1), ..., E _p (N + L-1),
The transfer path characteristic update unit is configured to update the input signal X (0), X (1),..., X (N + L-1) and the residual signal E _p (0), E _p (1),. The estimated values H _p (0), H _p (1), ..., H _p (N + L-1) are updated using E _p (N + L-1), and the updated estimated value H We obtain _{p + 1} (0), H _{p + 1} (1), ..., H _{p + 1} (N + L-1) ,
The controller controls the input signal X (0), X (1),..., X (N + L-1) and the collected sound signal y (n), y (n-1),. Until the updated estimated values H _{p + 1} (0), H _{p + 1} (1),..., H _{p + 1} (N + L−1) converge using The echo estimation unit, the time domain conversion unit, the subtraction unit, the second frequency domain conversion unit, and the transmission path characteristic update unit are controlled to be repeated.
Echo canceler. The echo canceler according to claim 1, wherein
Let L be the number of samples including the expected delay and the effective length of the impulse response, L 'be the number of samples of the effective length of the impulse response, and N be any integer greater than L' and less than L. The first frequency domain transform unit frequency-divides (L + N) time domain input signals x (nN−L + 1), (nN−L + 2),..., X (n) every N samples. Convert the input signal X (0), X (1), ..., X (N + L-1) of the domain
The echo estimation unit estimates estimated values H _p (0), H _p (1) of the transfer characteristics of the acoustic transmission path with the input signals X (0), X (1),..., X (N + L-1). ,..., H _p (N + L−1), and the echo estimated values Y ^ _p (0), Y ^ _p (1),..., Y ^ _p (N + L−1) in the frequency domain Ask for
The time domain transformation unit is configured to estimate the echo estimated values Y ^ _p (0), Y ^ _p (1), ..., Y ^ _p (N + L-1) as the time domain echo estimated value y ^ _p (nN- L + 1), y ^ _p (nN-L + 2), ..., y ^ _p (n),
The subtractor unit is configured to calculate a collected signal y (n−N + 1), y (n−N + 2),..., Y (n) and an echo estimated value y _p (n−N + 1), in the time domain. _{y ^ p (n-n +} 2), ..., the residual signal is a difference between _{y ^ p (n) e p} (n-n + 1), e p (n-n + 2), ..., e _{Find p} (n),
The second frequency domain transform unit is configured to receive the residual signal e _p (n−N + 1), e _p (n−N + 2),..., E _p (n) as a frequency domain residual signal in the time domain. Convert to E _p (0), E _p (1), ..., E _p (N + L-1),
The transfer path characteristic update unit is configured to update the input signal X (0), X (1),..., X (N + L-1) and the residual signal E _p (0), E _p (1),. When updating the estimated values H _p (0), H _p (1), ..., H _p (N + L-1) using E _p (N + L-1), the estimated values after the update H _{p + 1} (0), H _{p + 1} (1),..., H _{p + 1} (N + L−1) are estimated values in the time domain h _{p + 1} (nN−L + 1), h _{p + 1} (nN−L + 2),..., H _{p + 1} (n), the converted estimated values h _{p + 1} (nN−L + 1), h _{p + 1} (nN−L + 2),..., H _{p + 1} (n) is the updated estimated value H _{p + 1} (0), H _p so that the value becomes zero in the interval other than the interval of L ′ sample including the peak value Constrain the calculation of ₊₁ (1), ..., H _{p +1} (N + L-1),
The controller controls the input signal X (0), X (1),..., X (N + L-1) and the collected sound signal y (n), y (n-1),. Until the estimated values H _{p + 1} (0), H _{p + 1} (1),..., H _{p + 1} (N + L−1) converge using Control is performed so as to repeat processing in the time domain conversion unit, the subtraction unit, the second frequency domain conversion unit, and the transfer route characteristic update unit.
Echo canceler. Assuming that the number N of samples to be processed as the echo cancellation processing is larger than the number of samples of the effective length of the assumed impulse response, the input signal x in the time domain is converted to the input signal X in the frequency domain every N samples. A first frequency domain conversion step to convert;
An index indicating the number of repetition times and p, by using the estimated value H _p of the transfer characteristic of the acoustic transmission path of the input signal X and the frequency domain, and echo estimating step of obtaining the echo estimate Y ^ _p in the frequency domain ,
And the time domain conversion step of converting the echo estimate Y ^ _p to echo estimate y ^ _p in the time domain,
A subtraction step for obtaining a residual signal e _p in the time domain which is a difference between the collected sound signal y in the time domain and the echo estimated value y ^ _p ;
A second frequency domain transforming step of transforming the residual signal e _p in the time domain to the residual signal E _p in the frequency domain,
Updating the estimated value H _p using the input signal X and the residual signal E _p to obtain an updated estimated value H _{p + 1} ;
The echo estimation step, the time domain conversion step, the subtraction step, and the second frequency domain conversion until the updated estimated value H _{p + 1} converges using the input signal X and the collected sound signal y Repeating the process in the step and the transfer route characteristic update step;
Echo cancellation method.   A program for causing a computer to function as the echo canceler according to any one of claims 1 to 3.

Images