JP2010011265A - Echo canceller apparatus and its method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an echo canceller apparatus having an improved follow-up ability for echo path variation. <P>SOLUTION: The apparatus includes: a measurement sound wave generator 8 which generates a measurement signal having a measurement frequency; a mixing section 1 which mixes the measurement signal with the received signal; a second linear adaptive filter 10 which generates a second echo replica signal by a second filter while changing a second filter coefficient, and minimizes power of a second differential signal from a second subtracter 11 to obtain the minimized second filter coefficient; a filter coefficient DB 13 for storing paired groups of first filter coefficient candidates and second filter coefficient candidates in advance; a filter coefficient selector 14 which searches for the second filter coefficient candidate corresponding to the minimized second filter coefficient from the filter coefficient DB 13 and sets as the first filter coefficient the first filter coefficient candidate contained in the group of the searched second filter coefficient candidate; and a first linear filter 6 which generates a first echo replica signal using the set first filter coefficient. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to an echo cancellation technique, and more particularly to an echo canceller apparatus using a filter coefficient learning method of an adaptive filter and a method thereof.

  In a loudspeaker device that transmits and receives audio through a transmission line, outputs the received far-end audio to a speaker, and transmits near-end audio received by a microphone, when acoustic coupling occurs in the space between the speaker and the microphone, A problem arises in that the acoustic output occurs when the speaker output wraps around the microphone input.

  In order to eliminate this acoustic echo, an echo canceller apparatus using a linear adaptive filter of Patent Document 1 (FIG. 4) is generally known.

  The far-end sound transmitted as a reception signal, which is a reception input, is amplified and output through a speaker, and is received by a microphone as an acoustic echo via an echo path. Since the near-end voice is also received by the microphone, the microphone input is the sum of the near-end voice and acoustic echo. Of these, since only near-end speech is desired to be transmitted and output, the problem is how to remove the acoustic echo mixed in the microphone input. This is an echo cancellation problem.

  Since the received signal that is the source of the acoustic echo is known, if the received signal is subjected to the same conversion as the transfer characteristic of the echo path, the same signal (echo replica) as the acoustic echo mixed in the microphone input is internally generated. Can be produced. By subtracting this from the microphone input, it should be possible to remove acoustic echo from the transmission output.

  The linear adaptive filter is a filter for simulating the transfer characteristic of the echo path, and the transfer characteristic is calculated and reflected in the received signal (hereinafter referred to as “reference signal” when called as an input to the filter). By doing so, an echo replica signal that simulates an acoustic echo is generated.

  The subtractor can generate approximate near-end speech by subtracting this echo replica signal from the microphone input, and if this is transmitted and output, the speech with reduced acoustic echo will reach the destination. become. Reflection of transfer characteristics, that is, generation of an echo replica signal is performed by convolution of a reference signal and a filter coefficient.

  The linear adaptive filter performs filter adaptation that minimizes the power of the difference signal (also referred to as residual signal or error signal) of the subtractor by selecting when there is no near-end speech in order to simulate the transfer characteristic of the echo path. .

  As this adaptation method, several adaptation algorithms such as an LMS algorithm and an NLMS algorithm are known. Since the adaptive algorithm is disclosed in Non-Patent Document 1, a typical LMS algorithm will be briefly described here.

  A filter update expression for obtaining a new filter coefficient for reducing the difference signal from the value of the difference signal at time t, the value of the reference signal at that time, the filter coefficient, and the parameter that determines the magnitude of the update is repeatedly calculated at each time. As a result, the power of the differential signal is gradually minimized. This minimization (hereinafter referred to as “adaptation” or “learning convergence”) process takes a certain amount of time, but the time can be shortened by increasing μ.

  However, as a result, it is known that the filter coefficient becomes oscillating and the accuracy of the echo replica signal is deteriorated. On the other hand, by taking μ small, the accuracy of the echo replica signal can be increased, but on the other hand, the time until the filter coefficient converges to an appropriate value becomes long, and the followability to the echo path fluctuation becomes worse.

  A state where there is no near-end voice and there is far-end voice is called “single talk”. On the other hand, a state in which near-end voice and far-end voice exist simultaneously is called “double talk”. In the single talk state, the microphone input is only the acoustic echo that the far-end speech has circulated. In this state, when the linear adaptive filter learns the filter coefficient so that the power of the differential signal is minimized and, in short, zero, the generated echo replica signal can be brought close to the acoustic echo. The learned filter coefficient simulates the transfer characteristic of an unknown echo path.

  The double talk detector determines the presence or absence of near-end speech so that the linear adaptive filter can perform such learning. As disclosed in the prior art in Patent Document 1, this determination can be made most simply by determining the ratio of the received input level of the received signal to the differential output level of the differential signal.

  When there is only far-end speech, the difference signal after echo cancellation becomes small, so this level ratio becomes large. On the other hand, when there is near-end speech, the differential signal after echo cancellation is not so small, so this level ratio R is relatively small. Therefore, when the level ratio R becomes smaller than the predetermined threshold, it can be determined that the state where the near-end voice is present (double talk state).

  If the double talk detector determines that it is in the double talk state, the linear adaptive filter temporarily stops learning. This is because, in the double talk state, near-end speech exists in the differential signal, so even if an adaptive algorithm that minimizes the differential signal power is used, the filter cannot correctly learn the echo path transfer characteristics.

  Note that since the learning of the linear adaptive filter proceeds in the non-double talk state, filter adaptation is performed even when neither the near-end speech nor the far-end speech is present. Originally, there is no meaning in learning if there is no far-end speech, but an adaptive algorithm such as LMS moves the filter coefficient in proportion to the size of the far-end speech, so there is no real harm.

  In addition, in this double talk determination method, the level ratio becomes small when there is no near-end speech and no far-end speech, and there is a possibility that it is erroneously determined as a double talk state. Similarly, no problem will occur even if learning is stopped. In Patent Document 1, the detection accuracy is further improved by dynamically changing the detection threshold before and after double talk detection. In addition to this, many double-talk detection methods have been proposed.

Conventionally, as disclosed in Patent Document 2, an attempt is made to improve followability to echo path fluctuations by dynamically selecting a filter having a plurality of adaptive filters and determining a transmission output. Has been proposed.
"Acoustic system and digital signal processing" published by IEICE Japanese Patent No. 3202256 Japanese Patent No. 4026693 JP 2007-336364 A Japanese Patent No. 2923964

  As described above, the conventional typical echo canceller apparatus can follow the fluctuation of the echo path in the non-double talk state, for example, the movement of the human body, by continuously learning the linear adaptive filter except for the double talk state. It has become.

  However, even if the echo path fluctuates in the double talk state, this cannot be followed, and in this case, there has been a problem that an unerased echo becomes large due to the fluctuation.

  Accordingly, the present invention has been made in view of the above problems, and an object of the present invention is to provide an echo canceller apparatus and a method thereof with improved followability to echo path fluctuations.

  The present invention provides a loudspeaker output unit for loudly outputting a received signal having a speech frequency band to an acoustic space, a sound input unit for inputting sound from the acoustic space, and convolving a first filter coefficient with the received signal. A first filter that generates a first echo replica signal that is an echo replica of the received signal; a first bandpass filter that passes a signal in the speech frequency band from the input sound signal that is input; and the first bandpass filter A first subtractor that subtracts the first echo replica signal from the signal that has passed through the output unit and outputs a first differential signal; an echo canceller signal output unit that outputs the first differential signal as an echo cancelled echo canceller signal; A measurement sound wave generator for generating a measurement signal having an arbitrary measurement frequency, and a mixing unit for mixing the measurement signal with the reception signal A second filter that generates a second echo replica signal that is an echo replica of the measurement signal by convolving a second filter coefficient with the measurement signal; and a second bandpass that passes the signal of the measurement frequency from the input sound signal A filter unit, a second subtracter that subtracts the second echo replica signal from the signal that has passed through the second bandpass filter unit, and outputs a second differential signal, and the power of the second differential signal is minimized. A second filter update unit for correcting the second filter coefficient, a first filter coefficient candidate that is a candidate for the first filter coefficient, and a second filter coefficient candidate that is a candidate for the second filter coefficient are stored in advance as a set. The filter coefficient storage unit to search the second filter coefficient candidate corresponding to the current second filter coefficient from the filter coefficient storage unit, A filter coefficient selection unit that sets the first filter coefficient candidate included in the searched second filter coefficient candidate set as the first filter coefficient, and the first filter has the set first filter coefficient An echo canceller that generates the first echo replica signal using one filter coefficient.

  According to the present invention, the followability to the echo path fluctuation is improved even in the double talk state in the call voice band.

  Hereinafter, an embodiment of an echo canceller according to the present invention will be described with reference to the drawings.

(First embodiment)
First, the echo canceller apparatus according to the first embodiment, which is the most basic configuration of the present invention, will be described with reference to FIG. FIG. 1 shows a functional block configuration of an echo canceller apparatus according to this embodiment.

  As shown in FIG. 1, the echo canceller apparatus includes a mixer (referred to as “MIX” in the figure) 1 that mixes received voice and other voices at a predetermined ratio, a speaker 2 that outputs the output of the mixer 1 loudly, An echo path 3 generated by acoustic coupling, a microphone 4, a first band pass filter (referred to as “BPF1” in the figure) 5 that selectively passes sound in a frequency band (call band) used for voice calls, and echo of call band sounds From the speech band sound that has passed through a first linear filter (referred to as “LF1” in the figure) 6 and a first bandpass filter (referred to as “BPF1” in the figure) 5 that generate a replica (first echo replica signal) A first subtractor for obtaining an output of a first differential signal by subtracting an echo replica (second echo replica signal) generated by a linear filter (referred to as “LF1” in the figure) 6 Made, the first difference signal is transmitted outputted as the final proximity signal (echo canceller signal).

  If it is assumed that the filter coefficient of the first linear filter 6 simulates the echo path 3 of the speech band, an echo canceller for the loudspeaker device is configured by the above functional blocks.

  The echo canceller apparatus according to the present embodiment further includes functional blocks indicated by the units 8 to 12 in the drawing in the above configuration.

  As shown in FIG. 1, the echo canceller apparatus includes a measurement sound wave generator (referred to as “SGEN” in the figure) 8 that generates a measurement sound wave having a predetermined frequency, and a second band that selectively passes the measurement sound wave. A path filter (referred to as “BPF2” in the figure) 9, a second linear adaptive filter (referred to as “LAF2” in the figure) 10 for learning the echo path of this measured sound wave, a second subtractor 11, and a double of this measured sound wave A second double-talk detector (referred to as “DTD2” in the figure) 12 for detecting the talk state is further provided.

  For the second double-talk detector 12, the method disclosed in Patent Document 1 can be used. With this configuration, a measurement signal having a predetermined frequency generated by the measurement sound wave generator 8 is mixed with the reception signal by the mixer 1 and output from the speaker 2 as a measurement sound wave.

  The output measurement sound wave is received by the microphone 4 via the echo path 3 and passes through the second band-pass filter 9.

  The second linear adaptive filter 10 is a filter that simulates the transfer characteristic related to the measurement sound wave of the echo path 3, and acoustic reflection related to the measurement sound wave is calculated by reflecting this transfer characteristic in the measurement signal from the measurement sound wave generator 8. To generate a second echo replica signal.

  The second subtractor 11 generates a second differential signal by subtracting the second echo replica signal from the passing signal of the second bandpass filter 9.

  The second linear adaptive filter 10 learns the echo path of the measurement sound wave by performing adaptation that minimizes the power of the second differential signal while the second double talk detector 12 determines that it is in a non-double talk state. Filter adaptation is performed. For this adaptation, the LMS algorithm is used. As a result, the second linear adaptive filter 10 and the second subtractor 11 serve as an echo canceller for the measurement sound wave, but the output is not included in the transmission output.

  The echo canceller apparatus also has two filters 6 and 10, but the output of the filter 10 is not directly included in the transmission output, and the transmission output is determined only from the output of the filter 6.

  Further, the echo canceller apparatus has functional blocks indicated by the respective units 13 to 14 in FIG.

  As shown in FIG. 1, the echo canceller apparatus has at least a filter coefficient data candidate D = [W1, which is a set of filter coefficient W1 candidates for the first linear filter 6 and filter coefficient W2 candidates for the second linear adaptive filter 10. W2] is stored in the filter coefficient database (referred to as “FDB” in the figure) 13, and the filter coefficient W2 of the second linear adaptive filter 10 is updated every time the filter coefficient W2 is updated. A filter coefficient selector 14 that selects a filter coefficient data candidate D to be selected from the filter coefficient database 13 and sets a candidate for the filter coefficient W1 of the first linear filter 6 paired therewith to the first linear filter 6. . A method for selecting a suitable filter coefficient will be described later.

  The first linear filter 6 generates a first echo replica signal of the speech band from the received signal (reference signal) using the set filter coefficient W1.

  The first subtracter 7 generates an echo canceller signal in which acoustic echo is suppressed by subtracting the first echo replica signal from the speech band voice included in the microphone input that has passed through the first bandpass filter 5.

  Usually, even if the transfer characteristic of the echo path 3 is estimated at a very limited frequency such as a measurement sound wave, it cannot be expected that the characteristic accurately represents the transfer characteristic of the entire speech band. However, if the arrangement of the objects that reflect and absorb the sound that forms the acoustic coupling is the same, there is a certain correlation between the transfer characteristics measured at a specific limited frequency and the transfer characteristics of the entire speech band. It is established.

  Assuming this correlation, the echo canceller apparatus measures or simulates the filter coefficient W1 of the first linear filter 6 and the filter coefficient W2 of the second linear adaptive filter 10 under various expected conditions. A set of values under the same condition is stored in the filter coefficient database 13 as a filter coefficient data candidate D = [W1, W2].

  As a result, by measuring the transfer characteristic of the measured sound wave, that is, by obtaining the filter coefficient W2, the filter coefficient W1 of the first linear filter 6 suitable for the environment corresponding to this filter coefficient W2 is obtained as the filter coefficient data candidate D = [ W1, W2] can be selected and determined.

  In the above description, the filter coefficient data candidate D is a combination of at least the filter coefficient W1 of the first linear filter 6 and the filter coefficient W2 of the second linear adaptive filter 10.

  However, this assumes that the measurement frequency of the measurement signal does not change. If the measurement frequency of the measurement signal is different, the value of the learned filter coefficient W2 may be different even in the same environment. As will be described later, the echo canceller apparatus is configured such that the measurement sound wave generator 8 can change the measurement frequency of the measurement signal.

  Therefore, the filter coefficient data candidate D must include the measurement frequency F of the measurement signal when W2 is learned in addition to W1 and W2. The measurement frequency F is information indicating which frequency is used among the L measurement frequency candidates set in advance. The element corresponding to the measurement frequency used is 1 and the other elements are 0. L-dimensional vector. As a result of including this F, the filter coefficient data becomes D = [W1, W2, F].

  The filter coefficient selector 14 also selects a filter coefficient data candidate D from which both F and W2 are suitable from the filter coefficient database 13 as described below.

When the filter length of the first linear filter 6 for generating the first echo replica signal in the speech band is N, the filter coefficient W1 is represented by an N-dimensional vector.

  If the filter length of the second linear adaptive filter 10 for learning the echo path with the measurement sound wave is M, the filter coefficient W2 is an M-dimensional vector.

  Further, the measurement frequency F when learning W2 is an L-dimensional vector as described above.

  The number of F elements having 1 is not limited to one, and a plurality of elements are 1 when a measurement sound wave obtained by synthesizing a plurality of frequencies is used. In addition, the arrangement of the frequencies corresponding to the elements of the vector F may be appropriate, but here, the ones having a low frequency to a high frequency are arranged in order.

  At this time, an element which does not have 1 adjacent to an element of F having 1 has a value larger than 0 and smaller than 1, for example, 0.5. It should be noted that this value is not given when the frequency of adjacent elements is far beyond a predetermined threshold. By doing in this way, since the vector F which used the group of the similar measurement frequency becomes similar, it is convenient when evaluating the adaptability mentioned later. Data obtained by combining these three types of vectors (W1, W2, F) is set as filter coefficient data candidate D (k) = [W1 (k), W2 (k), F (k)]. Remember. Here, the subscript k indicates the kth data stored in the filter coefficient database 13.

  A method for selecting the filter coefficient data candidate D to which the filter coefficient selector 14 matches is as follows.

  The filter coefficient of the current second linear adaptive filter 10 is W2c, the measurement frequency at that time is Fc, and the kth filter coefficient data candidate stored in the filter coefficient database 13 is D (k) = [W1 (k), W2 (k), F (k)].

  First, for all k, a value CF (k) obtained by normalizing the inner product of Fc and F (k) with a norm is obtained, and only filter coefficient data for which the value has obtained an arbitrary threshold value or more is picked up. This threshold value can be adjusted. However, if it is desired to obtain only a more exact match, it is possible to pick out only those with CF (k) = 1.0.

  Then, among the extracted filter coefficient data, a value CW (k) obtained by normalizing the inner product of W2c and W2 (k) with a norm is obtained, and the filter coefficient data for obtaining the maximum value when the value is equal to or greater than an arbitrary threshold value Select.

Finally, W1 (k) in the selected filter coefficient data is set as a new filter coefficient of the first linear filter 6. Here, CF (k) and CW (k) are expressed by the following equation (1).

  Note that when the filter coefficient data candidate D in which CF or CW is equal to or greater than an arbitrary threshold is not found, the echo canceller apparatus fails because an appropriate filter coefficient W1 cannot be selected. In order to avoid such a situation, filter coefficient data candidates of sufficient variations are stored in the filter coefficient database 13.

  When the measurement sound wave generator 8 does not change the measurement frequency of the measurement signal, the vector F described in the filter coefficient data candidate D is not necessary. At the same time, it is not necessary to pick out data by CF (k) described above. Therefore, in this case, only CW (k) is evaluated for all the filter coefficient data candidates D (k) stored in the filter coefficient database 13. Also, the echo canceller apparatus can set the measurement frequency to either fixed or variable depending on the parameter.

  Next, the handling will be described while considering the measurement sound wave.

  The measurement frequency of the measurement signal in the echo canceller apparatus may be in the human audible range (generally 20 Hz to 20000 Hz), but in that case, the measurement sound wave is output at a level lower than the level of the far-end voice included in the reception input. Good. In this way, the measurement sound wave is less likely to be annoying to the user.

  In addition, the measurement frequency of the measurement signal is a high frequency range where the human auditory level falls even in the audible range (depending on age, for example, 15000 Hz or more near the mosquito sound), or even outside the audible range (generally 20 Hz or less or 20000 Hz or more ) Is preferable because the measurement sound wave does not disturb the user.

  Although there is a proposal (for example, Japanese Patent Application Laid-Open No. 2007-336364) that estimates an echo delay amount using a non-audible sound wave, it does not select the coefficient W1 of the echo cancellation filter as in this embodiment.

  In addition, a method for pre-learning an adaptive filter for echo cancellation using teaching voice prior to a call has been proposed (for example, Japanese Patent No. 2923964). However, it is the coefficient W1 of the first linear filter 6 that is learned by the teaching speech in the case of an echo canceller. Therefore, in order for this pre-learning to succeed, the frequency component of the teaching voice needs to be similar to the call voice. On the other hand, in the echo canceller apparatus of the present embodiment, the measurement frequency component of the measurement signal does not need to be similar to the call voice in principle. Note that the echo canceller apparatus can arbitrarily set the measurement frequency by a parameter.

  However, even if the measurement sound wave and the call voice do not need to be similar, when the correlation between the echo path of the measurement sound wave and the echo path of the call band is used like an echo canceller device, the measurement is performed with a low frequency. There is a possibility that the relationship between the filter coefficient W2 of the sound wave and the filter coefficient W1 of the speech band covering a wider frequency is not one-to-one correspondence but one-to-many correspondence. That is, even if a large fluctuation does not appear in the echo path of the measurement sound wave, the echo path of the call band may be fluctuating. When such a measurement frequency is used, the apparatus cannot follow the fluctuation of the echo path that does not appear in W2. Therefore, how to select the measurement frequency is important.

  In order to improve this problem, the echo canceller apparatus can also make the measurement sound wave a mixed sound of a plurality of frequency components.

  In this case, the second band-pass filter 9 is a comb filter (comb filter) having center permeability with respect to each frequency constituting the measurement sound wave. In this way, the echo path variation can be measured with higher accuracy, and the correlation between W1 and W2 can be further stabilized. Note that the echo canceller apparatus can set how many measurement frequencies are used by parameters.

  In particular, by using measurement sound waves mixed with at least one frequency component from each of a low frequency range (for example, 20 Hz or less) and a high frequency range (for example, 15000 Hz or more, or 20000 Hz or more), the user's ears are not disturbed. In addition, the probability of measuring the echo path of the environment with better accuracy increases.

  Since the low frequency component of the measurement sound wave is close to the low frequency range of the call band and the high frequency component is close to the high frequency range of the call band, the measurement sound wave with the high and low positions sandwiching the call band up and down is used. As a result, the correlation between the transmission characteristic of the measurement sound wave and the transmission characteristic of the speech band can be stabilized more than when not.

Next, a method for dynamically selecting the measurement frequency according to the environment will be described. The echo canceller apparatus can select the measurement frequency from the following three viewpoints.

    (A) Select a frequency where no other sound source exists in the environment. That is, ensuring a single talk.

    (B) Select a frequency with strong echo. That is, ensuring sensitivity to echo path fluctuations.

(C) Select a frequency having a large echo path variation. That is, ensuring sensitivity to echo path fluctuations First, the above (a) will be described.

  It is important to select a frequency that does not have another sound source in the environment as the measurement frequency. If a measurement frequency in which no other sound source exists can be used, the measured sound wave is always in a single talk state even if the call band is in a double talk state. The filter coefficient W1 of the first linear filter 6 can be switched, and in principle, followability to echo path fluctuation can be maintained in full time. Therefore, the echo canceller apparatus dynamically searches for and selects such frequencies as follows.

  When the measurement sound wave generator 8 generates a measurement sound wave, a sound having the measurement frequency of the measurement signal exists in the environment. Here, when a sound including the same frequency as the measurement sound wave is generated by another sound source, the second double talk detector 12 detects the double talk state.

  Therefore, the measurement sound wave generator 8 monitors the detection state of the second double talk detector 12 along with application of a certain measurement frequency F.

  If the current measurement sound wave frequency F is frequently in a double talk state, the measurement sound wave generator 8 has another sound source having the same frequency component as at least one of the frequencies constituting F in the environment. The measurement frequency is changed so as to avoid such a frequency. This frequency is measured as a ratio of the double talk detection period to the measurement sound wave generation period.

  When the double-talk detection frequency becomes equal to or greater than an arbitrary threshold value, the measurement sound wave generator 8 searches for an effective measurement frequency F ′ different from the current F.

  The measurement sound wave generator 8 notifies the second band pass filter 9 of the frequency to be used so that it can operate even if the measurement frequency of the measurement signal is changed. A filter bank for the frequency can be formed dynamically. Specifically, the second band pass filter 9 is configured as a filter bank composed of a set of narrow band pass filters that pass predetermined frequencies different from each other, and the measurement sound wave generator 8 By combining filter bank outputs corresponding to one or a plurality of frequencies notified by the above, a comb filter (comb filter) is formed. Therefore, “filter bank output” refers to the output of each bandpass filter forming the filter bank, and “output of the second bandpass filter 9” refers to the output of the comb filter formed by combining them.

  One method for searching for an effective measurement frequency F ′ is to examine the output of the second doubletalk detector 12 while switching the frequencies one by one.

  Specifically, in the predetermined frequency candidate list B, the frequency excluding the current measurement frequency F is set as the search candidate frequency C, and this is output in a predetermined order. Similarly, the double-talk detection frequency for a predetermined period is measured. Since a predetermined period is required to obtain the frequency of one frequency, when a necessary number of frequencies having a frequency less than an arbitrary threshold are found, this is set as a new measurement frequency F ′, and the search is terminated here. To. In this way, F ′ can be obtained as early as possible. However, if the necessary number of frequencies cannot be found by the above processing, the frequencies less than the necessary number are supplemented in ascending order of frequency.

For example, the frequency candidate list is represented by B = [f ₁ , f ₂ ,. . . , F ₂₀ ] and the current measurement frequency is F = [f ₂ , f ₈ , f ₁₈ ], the search candidate frequency is C = [f ₁ , f ₃ ,. . . , F ₇ , f ₉ ,. . . , F ₁₇ , f ₁₉ , f ₂₀ ]. At this time, the number of necessary frequencies is 3 as can be seen from the number of elements of F. In order to search for three new frequencies, the frequency of double talk detection was measured in order for each candidate of C. As a result, frequencies below the threshold were found in the order of f ₁ , f ₅ , and f ₇ , and as a result, the new measured frequency was F ′ = [f ₁ , f ₅ , f ₇ ]. As can be seen from this example, once a search is performed, a total exchange of measurement frequencies usually occurs.

  Further, as another method for searching for an effective measurement frequency, the second bandpass filter 9 is configured to be a filter bank for all frequencies of search candidates, and output of each filter bank is stopped in a state where measurement sound wave output is stopped. There is a way to check the size of.

  The measurement sound wave generator 8 sets the frequency excluding the current measurement frequency F in the predetermined frequency candidate list B as the search candidate frequency C, notifies the second band pass filter 9 and passes these frequencies, respectively. The filter bank is dynamically configured to Then, with the measurement sound wave stopped, the time average power of each filter bank output is measured, and a necessary number of frequencies are selected in ascending order to obtain a new measurement frequency F ′.

For example, the frequency candidate list is represented by B = [f ₁ , f ₂ ,. . . , F ₂₀ ] and the current measurement frequency is F = [f ₂ , f ₈ , f ₁₈ ], the search candidate frequency is C = [f ₁ , f ₃ ,. . . , F ₇ , f ₉ ,. . . , F ₁₇ , f ₁₉ , f ₂₀ ]. At this time, the required number of measurement frequencies is three. Therefore, in order to search for three new frequencies, when the filter bank output is examined for each candidate of C, it is assumed that the time average power is small in the order of f ₁₉ , f ₅ , and f ₇ . As a result, the new measurement frequency is F ′ = [f ₅ , f ₇ , f ₁₉ ]. This method is also a total replacement of measurement frequencies.

  This method allows you to check the existence of another sound source that has the same frequency component as the measured sound wave at the same time for multiple frequencies by forcibly creating a far-end audio state during the stop of the measured sound wave and monitoring the filter bank output during that time. Is a method that can easily and quickly search for an effective measurement frequency.

  Apart from this, the echo canceller device can also operate as follows.

  The measurement sound wave generator 8 periodically repeats the generation and stop of the measurement sound wave, and the second double talk detector 12 has a predetermined period during which the measurement sound wave is stopped, specifically, a certain reverberation time from the stop of the measurement sound wave. After the elapse of (corresponding to the filter length of the second linear adaptive filter 10), the time average power of each filter bank output of the second bandpass filter 9 is monitored until the next measurement sound wave is generated. At this time, if the output of the filter bank of the second bandpass filter 9 is observed with a magnitude equal to or greater than an arbitrary threshold, another sound source exists at the frequency passing through the filter bank (in the double talk state). Yes). Further, when the second double talk detector 12 detects that at least one filter bank output exceeds an arbitrary threshold value, the second double talk detector 12 outputs a double talk state.

  The measurement sound wave generator 8 monitors the above-described detection state of the second double talk detector 12 together with the application of a certain measurement frequency F, and when the double talk detection frequency exceeds an arbitrary threshold value, Searches for another effective measurement frequency F ′.

  In the above-described double talk detection method, it is impossible to know the double talk state during measurement sound wave generation. Therefore, the measurement sound wave generator 8 increases the time density of the double talk detection timing by shortening the cycle of repeating the generation and stop of the measurement sound wave as much as possible. If the double talk state continues longer than the above-described repetition period, a simple but effective double talk detector can be obtained by this method. Note that the echo canceller apparatus can set which of the above-described double talk detection methods is used as a parameter.

  In addition, since the appropriate value of the above-mentioned threshold varies depending on the actual use environment, it is determined by appropriately adjusting when the apparatus is introduced. Therefore, the various threshold values in this embodiment are parameters that can be adjusted as appropriate.

  When the second double talk detector 12 is operated in this way, its output can indicate not only the presence / absence of double talk detection but also at which measurement frequency the double talk is occurring.

  Therefore, the measurement sound wave generator 8 can search for another frequency only for the measurement frequency in the double talk state. This is shown below.

  The measurement sound wave generator 8 sets the frequency excluding the current measurement frequency F in the predetermined frequency candidate list B as the search candidate frequency C, notifies the second band pass filter 9 and passes these frequencies, respectively. Configure the filter bank to Then, with the measurement sound wave stopped, the time average power of each filter bank output is measured, and the insufficient frequency is selected in ascending order to obtain a new measurement frequency F ′.

For example, the frequency candidate list is represented by B = [f ₁ , f ₂ ,. . . , F ₂₀ ] and the current measurement frequency is F = [f ₂ , f ₈ , f ₁₈ ], the search candidate frequency is C = [f ₁ , f ₃ ,. . . , F ₇ , f ₉ ,. . . , F ₁₇ , f ₁₉ , f ₂₀ ]. Here, if another sound source exists at frequencies f ₈ and f ₁₈ included in F, the shortage is two frequencies that substitute for f ₈ and f ₁₈ . Therefore, when the filter bank output is examined for each candidate of C, it is assumed that the time average power is small in the order of f ₅ and f ₇ . As a result, the new measurement frequency is F ′ = [f ₂ , f ₅ , f ₇ ].

  The monitoring of another sound source and the change of the measurement frequency described above assume a state where the echo canceller is operating as an echo canceller. As a result, the device temporarily loses its ability to follow echo path fluctuations during frequency search, but by using the filter bank output method described above, it is possible to reduce the period during which the echo path fluctuations cannot be followed. become.

  On the other hand, assuming that the echo canceller apparatus is not in such an operating state and is in a standby state before the start of a voice call, for example, the measurement sound wave generator 8 selects the measurement frequency as described above during that time. It is also possible to do only.

  Specifically, the filter bank of the second band-pass filter 9 is dynamically configured to pass the respective frequencies for all the frequencies in the predetermined frequency candidate list B, and the generation of the measurement sound wave is stopped. Measure the time average power of the filter bank output. Then, a required number of frequencies are selected in order of increasing time-average power and set as the measurement frequency F.

  By doing this, before the start of the call, it is possible to select in advance the measurement frequency that does not have another sound source in the environment, suppress the frequency of change of the measurement frequency during the call, and the period during which the echo path fluctuation cannot be followed Further reduction is possible. The echo canceller can select this function with a parameter.

  Next, the above (b) will be described.

  If a measurement frequency at which acoustic echoes are less likely to occur in the operating environment is used, W2 sensitivity to echo path fluctuations cannot be increased. Therefore, in the echo canceller apparatus, the measurement sound wave generator 8 can dynamically search for a frequency having a high echo intensity and select a measurement frequency.

  The echo intensity is determined by monitoring the time average power of each filter bank output of the second band pass filter 9 during the generation of the measurement sound wave. The measurement sound wave generator 8 monitors the output of each filter bank and, when detecting a frequency that does not exceed an arbitrary threshold value for a predetermined period, searches for a different frequency.

  Specifically, in the predetermined frequency candidate list B, the frequency excluding the current measurement frequency F is set as the search candidate frequency C, and the filter bank of the second bandpass filter 9 is configured to pass these frequencies. To do.

  Thereafter, the time average power of each filter bank output is measured in a state where a measurement sound wave is generated by synthesizing all frequency components of the search candidate frequency C with equal power.

  Then, the missing frequency is selected in descending order and set as a new measurement frequency F ′. The echo canceller can select this function with a parameter.

  In addition, as in the case of detecting another sound source as described above, this frequency search can be executed before the start of a call.

  Next, the above (c) will be described.

  In the operating environment, the measurement frequency at which the filter coefficient W2 of the second linear adaptive filter 10 does not relatively change is considered to be a frequency that is insensitive to echo path fluctuations. Therefore, in the echo canceller apparatus, the measurement sound wave generator 8 can select a measurement frequency by dynamically searching for a frequency whose amount of time change of the filter coefficient W2 is relatively larger than other frequencies. .

  Here, the amount of change of the filter coefficient W2 is defined as the absolute value of the M elements of the time difference vector ΔW2 = W2 (t + 1) −W2 (t) of W2 during the generation of the measurement sound wave. . The measurement sound wave generator 8 monitors M elements of ΔW2, and searches for another measurement frequency when the amount of change does not exceed an arbitrary threshold value for a predetermined period.

  The search is performed as follows. In the predetermined frequency candidate list B, the frequency excluding the current measurement frequency F is set as the search candidate frequency C, and the filter bank of the second bandpass filter 9 is dynamically configured so that these frequencies are passed in order. Similarly, in the state where measurement sound waves including only the frequency components are generated in order, the amount of change in W2 with respect to each frequency is measured. Then, a necessary number of frequencies are selected in the descending order of the value and set as a new measurement frequency F ′. The echo canceller can select this function with a parameter. Further, as in the case of the detection of another sound source as described above, this frequency search can be executed before the start of a call.

  Further, in the echo canceller apparatus, it is possible to use a free frequency of the call band as the measurement frequency in order to bring the echo path of the measurement frequency close to the echo path of the call band.

  As in the case of the second bandpass filter 9, the first bandpass filter 5 is configured by a filter bank, and the measurement sound wave generator 8 calculates the time average power of the filter bank output from the past to the present time for a predetermined period. Store as P. Here, P is a vector whose number of elements is the number of filter banks of the first bandpass filter 5.

  As in the case of detection of another sound source described above, the measurement sound wave generator 8 that has detected that the double-talk detection frequency has reached an arbitrary threshold or more during operation at a certain measurement frequency F is stored in the time-average power stored. A necessary number of vacant frequencies are selected from the data P in ascending order of time average power, and set as a new measurement frequency F ′. The echo canceller can select this function with a parameter. In addition, as in the case of detecting another sound source as described above, this frequency search can be executed before the start of a call.

  When the frequency of the call band is the measurement frequency, the measurement sound wave is a sound that can be heard by the user. Therefore, the measurement sound wave generator 8 outputs more than when using the non-audible range so that the sound volume of the measurement sound wave is sufficiently smaller than the sound volume of the far-end sound received and input so that the sound is less harsh. A sound wave is generated at a lower level.

  In the echo canceller apparatus, the filter lengths of the first linear filter 6 and the second linear adaptive filter 10 can be set to different lengths.

  For example, by making the filter length of the second linear adaptive filter 10 shorter than the filter length of the first linear filter 6, the first linear filter 6 can be used as an adaptive filter as in the conventional typical echo canceller. The calculation cost required for learning the filter can be reduced as compared with the case where the learning is performed with a long length.

(Second Embodiment)
Next, an echo canceller apparatus according to a second embodiment of the present invention will be described with reference to FIG. FIG. 2 shows a functional block configuration of the echo canceller apparatus of the present embodiment. The echo canceller apparatus of this embodiment has a configuration in which a learning mode function capable of learning and storing the filter coefficient data is added to the first embodiment.

  With this configuration, the failure of the apparatus can be prevented when the filter coefficient data candidate D in which the CF or CW is equal to or greater than an arbitrary threshold is not found.

  Each part 1-5 in FIG. 2 and 7-14 is the same functional block as 1st Embodiment. However, in the first embodiment, the first linear filter 6 becomes the first linear adaptive filter (referred to as “LAF1” in the figure) 15, and a new first double talk detector (in the figure, in the figure). A filter coefficient register (referred to as “FREG” in the figure) 17 and a filter coefficient register 16 (referred to as “DTD1”) are added.

  As in the first embodiment, as shown in FIG. 2, a mixer 1 that mixes and outputs a far-end voice and other voices, a speaker 2 that outputs a loud output from the mixer 1, and an echo path 3 generated by acoustic coupling. , A microphone 4, a first band pass filter 5 that selectively allows voice in the call band to pass, and a first subtractor 7.

  Further, as shown in FIG. 2, the first linear adaptive filter 15 for learning the echo path of the speech band using an adaptive algorithm, not the simple linear filter, and the first double talk detection for detecting the double talk state of the speech band A container 16.

  For the first double-talk detector 16, the method disclosed in Patent Document 1 can be used.

  The first linear adaptive filter 15 generates a first echo replica signal of speech band voice using the received signal as a reference signal.

  The first subtracter 7 obtains a first differential signal as a transmission output by subtracting the first echo replica signal generated by the first linear adaptive filter 15 from the speech band sound that has passed through the first bandpass filter 5. .

  Further, the first linear adaptive filter 15 performs filter adaptation that minimizes the power of the first differential signal by the LMS algorithm in a period in which the first double talk detector 16 does not detect double talk. As a result, an echo canceller using the linear adaptive filter as illustrated in FIG. 2 is configured.

  As in the first embodiment, as shown in FIG. 2, a measurement sound wave generator 8 that generates a measurement sound wave having a predetermined frequency, a second bandpass filter 9 that selectively passes the measurement sound wave, A second linear adaptive filter 10 that learns the echo path of the measurement sound wave, a second subtractor 11, and a second double talk detector 12 that detects the double talk state of the measurement sound wave are included.

  A measurement sound wave of a predetermined frequency generated by the measurement sound wave generator 8 is mixed with a reception signal by the mixer 1 and output from the speaker 2. The output measurement sound wave is received by the microphone 4 via the echo path 3 and passes through the second band-pass filter 9.

  The second linear adaptive filter 10 is a filter that simulates the transfer characteristic related to the measurement sound wave of the echo path 3, and acoustic reflection related to the measurement sound wave is calculated by reflecting this transfer characteristic in the measurement signal from the measurement sound wave generator 8. To generate a second echo replica signal.

  The second subtractor 11 generates a second differential signal by subtracting the second echo replica signal from the passing signal of the second bandpass filter 9.

  The second linear adaptive filter 10 is a filter that learns the echo path of this measurement sound wave by performing adaptation that minimizes the power of the second differential signal while the double talk detector 12 determines that it is in a non-double talk state. Adapt.

  As shown in FIG. 2, a filter coefficient data candidate D for storing a measurement frequency of the measurement signal, a filter coefficient W2 of the second linear adaptive filter 10 and a filter coefficient W1 of the first linear adaptive filter 15 is stored. Based on the filter coefficient database 13, the filter coefficient W 2 of the second linear adaptive filter 10 and the measurement frequency, a filter coefficient selector 14 that selects a predetermined filter coefficient W 1 from the filter coefficient database 13 and sets it in the first linear adaptive filter 15. Have

  Further, as shown in FIG. 2, the filter coefficient register 17 for storing the filter coefficient W1 of the first linear adaptive filter 15, the filter coefficient W2 of the second linear adaptive filter 10 and the measurement frequency F in the filter coefficient database 13 as a set. Have

  In the echo canceller apparatus, there are cases where sufficient filter coefficient data candidates D are not accumulated in the filter coefficient database 13.

  That is, when the filter coefficient selector 14 selects suitable filter coefficient data from the data candidates in the filter coefficient database 13 based on the filter coefficient W2 of the second linear adaptive filter 10 and the measurement frequency F, the measurement frequency F and the filter There is a case where a matching factor of the coefficient W2 is not found.

  At that time, the filter coefficient selector 14 instructs the first linear adaptive filter 15 to learn the echo path and the filter coefficient register 17 to start the operation, and enters the learning mode.

  In the learning mode, the filter coefficient register 17 starts to monitor the change of the filter coefficient W1 of the first linear adaptive filter 15, and when the amount of change converges below an arbitrary threshold, the filter coefficient W1 at this time, The filter coefficient W2 of the second linear adaptive filter 10 and the measurement frequency F are paired and registered in the filter coefficient database 13 as data candidates.

  When this registration is completed, the learning mode by the filter coefficient register 17 is stopped, and the filter coefficient selector 14 starts operating again.

  According to the present embodiment, the echo canceller apparatus described above, even if there is no sufficient accumulation of filter coefficient data candidates in advance, or even in an unknown environment where the accumulated filter coefficient data is not valid. By performing the learning mode, filter coefficient data that is valid in the environment is gradually accumulated during operation, and after accumulation, operation as in the first embodiment makes it possible to change the echo path even if the speech bandwidth is double talk. Can follow.

  As in the first embodiment, the echo canceller can also dynamically search for the measurement frequency based on information such as the presence / absence of another sound source, the echo intensity, and the time change amount of the filter coefficient W2.

  When the echo canceller performs a frequency search while operating as an echo canceller, the filter coefficient selector 14 is stopped and the first linear adaptive filter 15 is operated as an adaptive filter to change the echo path of the speech band. Follow. As a result, the present apparatus does not lose the ability to follow the echo path fluctuation during the frequency search as in the first embodiment.

  Further, in the echo canceller, the measurement sound wave generator 8 periodically operates the first linear adaptive filter 15 as an adaptive filter, the change of the filter coefficient W1 during that period, and the second linear adaptive currently operating at this time. It is also possible to compare the change of the filter coefficient W2 of the filter 10 and evaluate whether or not the two are synchronized.

When a filter coefficient is represented by a vector W, a vector formed by arranging the norm (length) for a predetermined period in a time series is defined as Q. Q is a vector representing the time change of W. The degree of synchronization CS representing the degree of synchronization is defined by the following equation (2) as a value obtained by normalizing the inner product of the time change vectors Q1 and Q2 obtained for the filter coefficients W1 and W2, respectively, with a norm. When the measurement sound wave generator 8 detects that the CS is less than an arbitrary threshold value, the measurement sound wave generator 8 searches for an effective measurement frequency F ′ different from the current measurement frequency F.

  The search is performed as follows.

  In the predetermined frequency candidate list B, the frequency excluding the current measurement frequency F is set as the search candidate frequency C, and the filter bank of the second bandpass filter 9 is dynamically configured so that these frequencies are passed in order. Then, in a state where measurement sound waves including only the frequency components are generated in order, the degree of synchronization CS for each frequency is measured.

  Then, a necessary number of frequencies are selected in the descending order of the value and set as a new measurement frequency F ′. The echo canceller can select this function with a parameter. Further, as in the case of the detection of another sound source as described above, this frequency search can be executed before the start of a call.

(Third embodiment)
Next, an echo canceller apparatus according to a third embodiment of the present invention will be described with reference to FIG. FIG. 3 shows a functional block configuration of the echo canceller apparatus of this embodiment.

  The echo canceller apparatus of the present embodiment is newly added to the second embodiment as an echo cancellation capability evaluator (referred to as “ECEV” in the figure) 18 and a filter coefficient evaluator (referred to as “FEVEV” in the figure) 19. The parts 1 to 17 in FIG. 3 are the same as those of the second embodiment.

  In this embodiment, the echo cancellation capability evaluator 18 evaluates the echo cancellation capability, and when the echo cancellation capability is less than a predetermined standard, the filter coefficient selection failure by the filter coefficient selector 14 is detected, and the filter coefficient registration is performed. The learning mode by the device 17 is entered.

The echo cancellation capability is evaluated using the echo cancellation capability amount E defined by the following equation (3) as a logarithmic value of the microphone input power measured in the single talk state divided by the differential output power of the first differential signal. To do.

  The echo cancellation capability evaluator 18 calculates the moving average power while the first double talk detector 16 outputs the non-double talk state for each of the microphone input and the first differential signal, and the echo cancellation capability is calculated from the calculated value. The quantity E is determined and monitored.

  Then, when it is detected that the echo canceling ability amount E is below an arbitrary threshold value, a stop command is sent to the filter coefficient selector 14 and an operation command is sent to the filter coefficient register 17.

  Upon receiving the stop command, the filter coefficient selector 14 causes the first linear adaptive filter 15 to start an echo path learning mode.

  At the same time, the filter coefficient register 17 monitors the amount of change of the filter coefficient W1 of the first linear adaptive filter 15 that has started the learning mode, and all the N elements of the difference vector ΔW1 = W1 (t + 1) −W1 (t) are monitored. When the state where the absolute value is less than an arbitrary threshold continues for a predetermined period, it is determined that the learning has converged, the filter coefficient W1 at this time, the filter coefficient W2 of the second linear adaptive filter 10, the measurement frequency F, Are registered in the filter coefficient database 13.

  At this time, if there is a filter coefficient data candidate D in which F and W2 coincide with the filter coefficient data to be registered as the filter coefficient data candidate D in the filter coefficient database 13, the W1 is overwritten with the new value of W1. . Otherwise, new filter coefficient data is generated and stored as filter coefficient data candidate D in the filter coefficient database 13.

  When this registration is completed, the filter coefficient register 17 stops and the filter coefficient selector 14 starts operating again to enter the normal operation mode.

  According to the present embodiment, the echo canceller apparatus performs the above-described learning process even if an unknown environmental fluctuation occurs such that the filter coefficient W1 that should be compatible as long as F and W2 are seen cannot completely cancel the echo. By doing so, filter coefficient data that is valid in the environment is accumulated during operation, and after the accumulation, the operation is performed as in the first embodiment, so that even if the speech band voice is double talk, the echo path fluctuation can be followed.

(Fourth embodiment)
Next, an echo canceller apparatus according to a fourth embodiment of the present invention will be described with reference to FIG. FIG. 4 shows a functional block configuration of the echo canceller apparatus of this embodiment.

  The echo canceller apparatus has a configuration in which a filter coefficient evaluator (referred to as “FVEV” in the figure) 19 is newly added to the third embodiment. As shown in FIG. 4, each part 1-18 is the same as the structure of 3rd Embodiment.

  In the present embodiment, the filter coefficient evaluator 19 monitors changes in the filter coefficient W2 of the second linear adaptive filter 10. Then, when even one of the M elements of the difference vector ΔW2 = W2 (t + 1) −W2 (t) has an absolute value equal to or greater than an arbitrary threshold value, it is determined that the echo path has fluctuated rapidly, and the filter A stop command is sent to the coefficient selector 14 and an operation command is sent to the filter coefficient register 17.

  Upon receipt of the stop command, the filter coefficient selector 14 causes the first linear adaptive filter 15 to start an echo path learning mode.

  At the same time, the filter coefficient register 17 monitors the amount of change of the filter coefficient W1 of the first linear adaptive filter 15 that has started the learning mode, and all the N elements of the difference vector ΔW1 = W1 (t + 1) −W1 (t) are monitored. When the state where the absolute value is less than an arbitrary threshold continues for a predetermined period, it is determined that the learning has converged. At this time, the filter coefficient W1, the filter coefficient W2 of the second linear adaptive filter 10, and the measurement frequency F are paired and registered in the filter coefficient database 13 as candidates.

  At this time, if the filter coefficient data candidate D whose F and W2 match with the filter coefficient data to be registered as a data candidate exists in the filter coefficient database 13, the W1 is overwritten with the new value of W1. Otherwise, a new filter coefficient data candidate D is generated and stored in the filter coefficient database 13.

  When this registration is completed, the filter coefficient register 17 stops and the filter coefficient selector 14 starts operating again to enter the normal operation mode.

  According to the present embodiment, the echo canceller apparatus evaluates whether or not the filter coefficient data stored in the filter coefficient database 13 is valid in the environment after the change when a sudden environmental change occurs. By learning the latest filter coefficient data immediately and then operating as in the first embodiment, it is possible to follow the echo path fluctuations even if the speech band voice is double talk.

(Example of change)
The present invention. The present invention is not limited to the configurations described in the above embodiments, and various modifications can be made without departing from the gist thereof.

(1) Modification 1
In each of the above embodiments, a linear echo that can be simulated by convolving a received signal with a linear filter has been considered as an acoustic echo. Therefore, an embodiment using a linear filter such as the first linear filter 6 or the first linear adaptive filter 15 has been shown as a filter for removing echo in the speech band. However, the acoustic echo may include a non-linear echo component due to distortion caused by the speaker or vibration of the speaker transmitted to the microphone through the housing. If this nonlinear echo component exists, it cannot be simulated by the linear adaptive filter, and this component remains unerased in the differential signal. On the other hand, it is known that this nonlinear echo component can be simulated by using a nonlinear filter.

  Therefore, as a modified example, the first linear filter 6 and the first linear adaptive filter 15 for removing the echo of the speech band can be configured by a nonlinear filter or a nonlinear adaptive filter.

  The transfer characteristics of the non-linear echo do not fluctuate greatly unless the performance of the speaker or the installation method of the microphone and the speaker in the housing change. For this reason, the correlation between the filter coefficient W2 by the measurement sound wave and the filter coefficient of the nonlinear filter largely depends on the linear echo component. Therefore, even if a non-linear filter is used, it is considered that the filter coefficient selector 14 can select an appropriate filter coefficient.

  As a non-linear filter, “Volterra Kernel” is known.

  Also, “adaptive Volterra Filter” is known as a nonlinear adaptive filter.

(2) Modification example 2
The mixer 1, the first bandpass filter 5, the first linear filter 6, the first subtractor 7, the measurement sound wave generator 8, the second bandpass filter 9, and the second linear adaptive filter of the echo canceller apparatus of each of the above embodiments. 10, the functions of the respective parts of the second subtractor 11, the second double talk detector 12, and the filter coefficient selector 14 can also be realized by a program stored in a computer.

It is the figure which showed the functional block structure of the echo canceller apparatus of 1st Embodiment which concerns on this invention. It is the figure which showed the functional block structure of the echo canceller apparatus of 2nd Embodiment. It is the figure which showed the functional block structure of the echo canceller apparatus of 3rd Embodiment. It is the figure which showed the functional block structure of the echo canceller apparatus of 4th Embodiment.

Explanation of symbols

1 Mixer (MIX)
2 Speaker 3 Echo path 4 Microphone 5 First band pass filter (BPF1)
6 First linear filter 7 First subtractor 8 Measurement sound wave generator (SGEN)
9 Second bandpass filter (BPF2)
10 Second linear adaptive filter (LAF2)
11 Second subtractor 12 Second double talk detector (DTD2)
13 Filter coefficient database (FDB)
14 Filter coefficient selector (FSEL)

Claims (14)

A loudspeaker output unit for loudly outputting a received signal having a speech frequency band to an acoustic space;
A sound input unit for inputting sound from the acoustic space;
A first filter that generates a first echo replica signal that is an echo replica of the received signal by convolving a first filter coefficient with the received signal;
A first band-pass filter unit that passes the signal in the call frequency band in the input sound signal input;
A first subtractor that subtracts the first echo replica signal from the signal that has passed through the first bandpass filter unit and outputs a first differential signal;
An echo canceller signal output section for outputting the first differential signal as an echo cancelled echo canceller signal;
A measurement sound wave generator for generating a measurement signal having an arbitrary measurement frequency;
A mixing section for mixing the measurement signal with the reception signal;
A second filter that generates a second echo replica signal that is an echo replica of the measurement signal by convolving a second filter coefficient with the measurement signal;
A second band-pass filter section that passes the signal of the measurement frequency in the input sound signal;
A second subtracter that subtracts the second echo replica signal from the signal that has passed through the second bandpass filter unit and outputs a second differential signal;
A second filter updating unit for correcting the second filter coefficient so as to minimize the power of the second differential signal from the second subtractor;
A filter coefficient storage unit that pre-stores a first filter coefficient candidate that is a candidate for the first filter coefficient and a second filter coefficient candidate that is a candidate for the second filter coefficient;
A second filter coefficient candidate corresponding to the current second filter coefficient is searched from the filter coefficient storage unit, and the first filter coefficient candidate included in the searched second filter coefficient candidate group is the first filter coefficient candidate. An echo canceller apparatus comprising: a filter coefficient selection unit set as one filter coefficient, wherein the first filter generates the first echo replica signal using the set first filter coefficient. A learning mode for learning a set of the first filter coefficient candidate and the second filter coefficient candidate is performed, and the first filter is changed by the first filter while changing the first filter coefficient in the learning mode. Regenerating an echo replica signal, minimizing the power of the first differential signal from the first subtractor, and obtaining a first filter coefficient at the time of the minimization, a first filter updating unit;
In the learning mode, a combination of the minimized first filter coefficient and the second filter coefficient obtained by the second filter update unit is used as a set of the first filter coefficient candidate and the second filter coefficient candidate. The echo canceller according to claim 1, further comprising a filter coefficient registration unit that is stored in the filter coefficient storage unit as a set.   An evaluation unit that obtains an echo cancellation capability amount from the input sound signal and the first differential signal is further provided, and when the echo cancellation capability amount is lower than an arbitrary first threshold, the first filter update unit enters the learning mode. The echo canceller according to claim 2, wherein the echo canceller is installed. Further comprising an evaluation unit for obtaining a fluctuation amount of the transfer characteristic of the echo path in the acoustic space;
The echo canceller according to claim 2, wherein when the fluctuation amount is higher than an arbitrary second threshold, the first filter update unit enters the learning mode.   The echo canceller according to claim 1, wherein the measurement frequency is a non-audible frequency of 20 Hz or less or 15000 Hz or more. The measurement sound wave generator is
With the generation of the measurement signal stopped, with respect to the input sound signal, the time average power of an arbitrary period for each frequency in an arbitrary frequency candidate is measured, and the frequency candidates are counted in ascending order of the time average power. The echo canceller according to claim 1, wherein the measurement frequency is selected. The measurement sound wave generator is
Generating a sound wave obtained by synthesizing each frequency of a plurality of frequency candidates of the measurement frequency, and measuring an echo intensity in an arbitrary period for each frequency of the sound wave,
The echo canceller according to claim 1, wherein the frequency that is higher than an arbitrary third threshold value among the echo intensities is selected as the measurement frequency. The measurement sound wave generator is
A sound wave having one frequency among a plurality of frequency candidates of the measurement frequency is generated in order, a variation amount of the second filter coefficient within an arbitrary period for each sound wave is measured, and the variation amount is arbitrary The echo canceller according to claim 1, wherein the frequency higher than a fourth threshold is selected as the measurement frequency. The measurement sound wave generator is
Measuring the power for each frequency component of the input sound in the call frequency band, and selecting the frequency at which the time-average power of each period of power is lower than an arbitrary fifth threshold as the measurement frequency The echo canceller apparatus according to claim 1, wherein The measurement sound wave generator is
The echo canceller according to claim 1, wherein the level of the received signal within an arbitrary period is measured and the measured signal is output at a level smaller than the received signal.   The echo canceller according to claim 1, wherein a filter length of the second filter is shorter than a filter length of the first filter. The measurement sound wave generator is
A sound wave having one frequency among the frequency candidates of the measurement frequency is generated, and a degree of synchronization in which a change in the first filter coefficient and a change in the second filter coefficient with respect to the sound wave are synchronized is measured. The echo canceller according to claim 2, wherein the measurement frequencies are selected in descending order. A loudspeaker output unit for loudly outputting a received signal having a speech frequency band to an acoustic space;
A sound input unit for inputting sound from the acoustic space;
In an echo canceller method in an echo canceller apparatus comprising:
A first filter step of generating a first echo replica signal that is an echo replica of the received signal by convolving a first filter coefficient with the received signal;
A first bandpass filter step of passing a signal in the call frequency band in the input sound signal input;
A first subtraction step of subtracting the first echo replica signal from the signal passed through the first bandpass filter step to output a first differential signal;
An echo canceller signal output step of outputting the first differential signal as an echo cancelled echo canceller signal;
A measurement sound wave generation step for generating a measurement signal having an arbitrary measurement frequency;
A mixing step of mixing the measurement signal with the received signal;
A second filter step of generating a second echo replica signal that is an echo replica of the measurement signal by convolving a second filter coefficient with the measurement signal;
A second bandpass filter step of passing the signal of the measurement frequency in the input sound signal;
A second subtracting step of subtracting the second echo replica signal from the signal passed through the second bandpass filter step to output a second differential signal;
A second filter updating step for modifying the second filter coefficient so as to minimize the power of the second differential signal;
Search for a second filter coefficient candidate corresponding to the current second filter coefficient from a filter coefficient storage unit that is stored in advance by combining the first filter coefficient candidate of the first filter coefficient and the second filter coefficient candidate, A filter coefficient selecting step for setting the first filter coefficient candidate included in the searched second filter coefficient candidate set as the first filter coefficient, wherein the first filter step is the set An echo canceller method, wherein the first echo replica signal is generated using a first filter coefficient. A loudspeaker output unit for loudly outputting a received signal having a speech frequency band to an acoustic space;
A sound input unit for inputting sound from the acoustic space;
In an echo canceller program in an echo canceller apparatus comprising:
A first filter function for generating a first echo replica signal that is an echo replica of the received signal by convolving a first filter coefficient with the received signal;
A first band pass filter function for passing a signal in the call frequency band from the input sound signal input;
A first subtraction function for subtracting the first echo replica signal from a signal that has passed through the first bandpass filter function and outputting a first differential signal;
An echo canceller signal output function for outputting the first differential signal as an echo cancelled echo canceller signal;
A measurement sound wave generation function for generating a measurement signal having an arbitrary measurement frequency;
A mixing function for mixing the measurement signal with the received signal;
A second filter function for generating a second echo replica signal that is an echo replica of the measurement signal by convolving a second filter coefficient with the measurement signal;
A second band pass filter function for passing the signal of the measurement frequency from the input sound signal;
A second subtraction function for subtracting the second echo replica signal from the signal that has passed through the second bandpass filter function and outputting a second differential signal;
A second filter update function for correcting the second filter coefficient so that the power of the second differential signal is minimized;
Search for a second filter coefficient candidate corresponding to the current second filter coefficient from a filter coefficient storage unit that is stored in advance by combining the first filter coefficient candidate of the first filter coefficient and the second filter coefficient candidate, A filter coefficient selection function for setting the first filter coefficient candidate included in the searched second filter coefficient candidate set as the first filter coefficient is realized by a computer.
The echo canceller program, wherein the first filter function generates the first echo replica signal using the set first filter coefficient.

Images