JP2006157499A - Acoustic echo canceller, hands-free telephone using the same, and acoustic echo cancellation method - Google Patents

Abstract

[PROBLEMS] When performing a learning process multiple times from sampling to the next sampling, if the reference signal of a portion with a small delay and the reference signal of a portion with a large delay are used at the same frequency, the transfer characteristic of an acoustic space changes. Therefore, even if it approaches the transfer characteristic based on the reference signal of the part with a lot of delay, a difference from the actually operating transfer characteristic occurs and it tends to be unstable.
An update frequency control unit 10 sets the delay of a reference signal to a magnitude of a delay of a reference signal storage unit 6 instead of using the reference signal for the delay and the delay part of the reference signal storage unit 6 at the same frequency. Accordingly, the filter coefficient of the filter coefficient accumulating unit 6 is updated by the filter coefficient updating unit, so that the transfer characteristic of the latest acoustic space can be reflected, and the transfer characteristic of the filter close to the transfer characteristic of the actual acoustic space can be obtained. Therefore, a stable acoustic echo canceller with a large echo cancellation amount can be obtained.
[Selection] Figure 1

Description

  The present invention relates to a stable acoustic echo canceller that has a rapid convergence and reflects the transfer characteristics of the latest acoustic section.

  The acoustic echo canceller is equipped with an adaptive algorithm, and performs an operation of removing the acoustic echo by estimating the transfer characteristic of the acoustic path between the speaker and the microphone and generating / subtracting a pseudo echo. A representative example of an adaptive algorithm implemented in an echo canceller is an LMS (Least Mean Square) method and an improved learning identification method.

  However, when the above algorithm is implemented by an acoustic echo canceller, there is a problem that the convergence speed is slow.

There are the following two methods for improving the convergence speed.
(1) A digital signal processor (DSP) having a large processing capability is used, and a learning process is performed a plurality of times from sampling to the next sampling so as to converge quickly.
(2) By controlling the tap length, it is possible to reduce the estimation of one acoustic path and the generation / subtraction of pseudo echoes, and even a DSP with a small amount of processing performs multiple estimation processes from sampling to the next sampling. By making it converge quickly.

  Among them, a technique that is devised so that the tap length is controlled so that estimation processing can be performed a plurality of times from sampling to the next sampling is disclosed.

  A conventional acoustic echo canceller that performs multiple estimation control will be described with reference to FIG.

  The number of reference signal data is the number of powers of 2 (eg, 4096), the number of filter coefficient taps is the number of powers of 2 less than the number of reference signal data (eg, 1024), and the number of reference signal data is filtered. It is assumed that the coefficient tap number is a power of 2 (4 times in this example). In this case, the length of the reference signal is divided into four equal parts, and the filter coefficient is updated a plurality of times using the respective reference signals divided into four equal parts. In this example, first, a filtering process for simulating reverberation characteristics is performed on the section 4 with an acoustic echo canceller (step 1), and an error between the simulated result and the signal due to the actual reverberation characteristics is calculated (step 2). The filter coefficient is updated from the error (step 3).

Next, the coefficient update is performed on the section 3, the coefficient update is performed on the section 2, and finally the coefficient update is performed on the section 1 (see, for example, Non-Patent Document 1).
Tsuyoshi Usagawa, and three others, "A New Adaptive Algorithm Forcused on the Convergence Characteristics by Coloered Input Signal: Variable Tap Length LMS", IEICE Transactions on Fundamentals of Electronics, Communications and Computer Sciences, Institute of Electronics, Information and Communication Engineers, 1992, VOL . E-75A, NO. 11, p. 1493-1499

  However, in the above-described conventional configuration, the reference signal with a small delay and the reference signal with a large delay are used at the same frequency. The transfer characteristic of the acoustic space is changing, and even if it approaches the transfer characteristic based on the reference signal of the part with a lot of delay, a difference from the actually operating transfer characteristic occurs, and it tends to be unstable.

  In particular, when the tap length or order of the filter is increased in order to cope with a long reverberation time, this difference becomes significant. As a specific example, when the echo canceller is executed at 8 kHz sampling, when the learning identification method of 4000 taps is executed and the above-described learning is performed a plurality of times, the tap with the longest delay is reduced from the actually input signal. Data from past 0.4 to 0.5 seconds will be used. In an environment where an acoustic echo canceller is used, the movement of a person for about 0.5 seconds affects the transfer characteristics from the speaker to the microphone.

  However, although not described in the literature, since the coefficient update of the acoustic echo canceller is performed based on a reference signal for a long time, it may be strong against the influence of noise and the like.

  Instead of using the reference signal with a small delay part and the reference signal with a large delay part at the same frequency, the frequency at which the reference signal is used for updating the filter coefficient is changed according to the magnitude of the delay.

  With this configuration, it is possible to update the filter coefficient so that the transfer characteristic corresponding to the filter coefficient is close to the transfer characteristic of the acoustic space.

  For this reason, the transfer characteristics of the latest acoustic space can be reflected, the transfer characteristics of the filter close to the transfer characteristics of the actual acoustic space can be obtained, the amount of echo cancellation is large, and a stable acoustic echo canceller can be obtained.

In the present invention,
An acoustic echo canceller that samples the sound from the speaker as a reference signal, updates the filter coefficient for the sampled reference signal, and cancels the echo of the sound input via the microphone by filtering based on the updated filter coefficient There,
A reference signal storage unit for storing an input reference signal;
A filter coefficient update unit that executes filter coefficient update processing for the reference signal accumulated by the reference signal accumulation unit;
It has a configuration of an update frequency control unit that controls the frequency of update processing by the filter coefficient update unit in accordance with the delay of the input reference signal.

Alternatively, the acoustic echo that samples the sound from the speaker as a reference signal, updates the filter coefficient for the sampled reference signal, and cancels the echo of the voice input via the microphone by the filter processing based on the updated filter coefficient Cancellation method,
A reference signal accumulating step for accumulating an input reference signal;
A filter coefficient update step for executing a filter coefficient update process for the reference signal accumulated by the reference signal accumulation unit;
It has a configuration called an update frequency control step for controlling the frequency of update processing by the filter coefficient update unit in accordance with the magnitude of the delay of the input reference signal.

  As a result, instead of using the reference signal with a small delay part and the reference signal with a large delay part at the same frequency, the frequency at which the reference signal is used for updating the filter coefficient is changed according to the magnitude of the delay. Therefore, it is possible to reflect the transfer characteristics of the latest acoustic space, to obtain a filter transfer characteristic close to that of the actual acoustic space, and to obtain a stable acoustic echo canceller with a large amount of echo cancellation.

In addition, the reference signal storage unit
A plurality of divided storage units each having a section obtained by dividing the section of the reference signal storage unit into a plurality of parts, or divided into a plurality of equal parts,
The filter coefficient update unit may execute a filter coefficient update process on each reference signal accumulated by the division accumulation unit.

Alternatively, the reference signal accumulation step
Having a plurality of divided storage memories each having a section divided into a plurality of sections, or divided into a plurality of equal sections, which the storage memory processed in the reference signal storage process has,
In the filter coefficient update step, the filter coefficient update process may be executed for each reference signal stored in the divided storage memory.

  With this configuration, in the filter coefficient update algorithm using each section, processing can be performed using a finite number of filter coefficient update units corresponding to each length. Moreover, by processing the same length section, it can be processed by one filter coefficient updating unit, and the processing amount of the algorithm is reduced.

  Further, the length of the section included in the reference signal storage unit (storage memory) is a power of 2, and the number of powers of 2 different from the length of the section included in the reference signal storage unit (storage memory) The divided storage memory may be equally divided. With this configuration, the processing amount of the algorithm when realized by a DSP or the like is further reduced.

  Further, the sections of the divided storage unit (divided storage memory) may overlap each other. With this configuration, there is no restriction that the section length of the entire reference signal storage section (storage memory) needs to be an integral multiple of the section length of the filter coefficient storage section. The number of times of performing the estimation process a plurality of times before the next sampling can be increased.

  In addition, when the section of the reference signal storage unit (storage memory) is divided, the section of the divided signal storage unit (divided storage memory) is configured such that the sections of the divided storage unit (divided storage memory) overlap each other. May be the same length. This configuration reduces the amount of algorithm processing.

  Further, when dividing the section of the reference signal storage unit (storage memory), the divided reference signal storage unit (storage memory) is divided so that the sections of the divided storage unit (divided storage memory) overlap each other. You may make it be the length of the power of 2 of the same length. With this configuration, the processing amount of the algorithm when realized by a DSP or the like is further reduced.

  Further, the control unit may control the update frequency so that the frequency is small at a place where the delay is small and the frequency is small at a place where the delay is small. With this configuration, an acoustic echo canceller that follows the latest transfer characteristics can be realized. Also, considering that there are speakers around the microphone and environmental noise in the signal from the transmission signal input terminal, the acoustic echo canceller algorithm is shorter than the case of using only the latest section with a short filter coefficient. It is also difficult to be affected by people and environmental noise.

A filter coefficient storage unit;
A reverberation time estimation unit for estimating reverberation time based on the filter coefficient accumulated in the filter coefficient accumulation unit;
A filter length control unit that controls the update frequency control unit based on the calculation processing amount based on the order or stage number of the filter processing corresponding to the reverberation time estimated by the reverberation time estimation unit may be used.

Alternatively, a reverberation time estimation unit that estimates reverberation time based on filter coefficients stored in the filter coefficient memory;
A filter length control step for controlling the update frequency control step according to the calculation processing amount based on the order or stage number of the filter processing corresponding to the reverberation time estimated in the reverberation time estimation step may be adopted.

  With this configuration, it is possible to provide an acoustic echo canceller that operates in accordance with the reverberation time in an actually used environment.

The reverberation time estimation unit
A first section composed of a first start point and a first end point determined with reference to the position of the filter coefficient having the largest filter coefficient in the filter coefficient storage unit; and the filter coefficient of the filter coefficient storage unit A section start point end point calculation unit for obtaining a second start point and a second section composed of the second end point at which the filter coefficient determined based on the position of the filter coefficient having the maximum magnitude attenuates;
A reverberation time simple estimation unit that estimates reverberation time based on the filter coefficient magnitude of the first interval of the filter coefficient accumulation unit and the filter coefficient magnitude of the second interval of the filter coefficient accumulation unit may be used. .

Alternatively, the reverberation time estimation step
A first interval composed of a first start point and a first end point determined based on the position of the filter coefficient having the largest filter coefficient stored in the filter coefficient memory, and stored in the filter coefficient memory A section start point for obtaining a second start point at which the filter coefficient determined based on the position of the filter coefficient having the largest filter coefficient size is attenuated and a second section composed of the second end point End point calculation process,
A reverberation time simple estimation step of estimating the reverberation time based on the size of the filter coefficient in the first section of the filter coefficient memory and the size of the filter coefficient in the second section of the filter coefficient memory may be adopted.

  With this configuration, it is possible to determine whether the adaptive filter is functioning sufficiently in any filter processing section length, including a short filter processing section, and a filter processing section that can realize transfer characteristics with converged filter coefficients You can set the length.

  Further, the section start point end point calculation may be performed in which the length of the first section and the length of the second section of the filter coefficient storage unit (filter coefficient memory) are set to the same length. With this configuration, it is possible to compare the sizes of each of the sections without normalizing them. Normalization division is not necessary, and the amount of processing can be reduced.

  Further, the average value of the square power of the filter coefficient may be used as the magnitude of the filter coefficient of the filter coefficient accumulation unit (filter coefficient memory). Or you may make it use the average value of the absolute value of a filter coefficient. Alternatively, the maximum absolute value of the filter coefficient may be used.

  With this configuration, when the average value of the square power of the filter coefficient is used, when implemented by a DSP, the DSP multiplication instruction can be effectively used to obtain a highly accurate filter coefficient size. Further, when the average value of the absolute values of the filter coefficients is used, it can be processed by a general-purpose CPU having no product-sum operation instruction. In addition, when the maximum absolute value of the filter coefficient is used, it can be processed by a general-purpose CPU that does not have a product-sum operation instruction, and it is not necessary to perform accumulation calculation. It is possible.

  Further, in a hands-free telephone characterized by using the acoustic echo canceller of the present invention, the speed of following a change in the room environment due to movement of a person is fast and stable, and the amount of echo cancellation is large. There is an effect that an easy hands-free telephone can be provided.

Example 1
Hereinafter, an acoustic echo canceller according to Embodiment 1 of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram of the first embodiment, and FIG. 2 is a diagram illustrating an operation of updating a filter coefficient.

  The received signal input from the received signal input terminal 1 is output to the received signal output terminal and simultaneously stored in the reference signal storage unit 5. The received signal is output as a sound wave from the speaker connected via the received signal output terminal 2 to the acoustic space. In addition, a signal obtained by picking up sound waves from the acoustic space with a microphone is input via the transmission signal input terminal 3 to obtain a transmission input signal. The filter coefficient accumulating unit 6 accumulates filter coefficients that are output from the speaker to the acoustic space by the operation of the filter coefficient updating unit 13 to be described later, and that simulate transfer characteristics when input by a microphone.

  The reference signal accumulated in the reference signal accumulation unit 5 and the filter coefficient accumulated in the filter coefficient accumulation unit are filtered by the first filter processing unit, and the sound wave inputted through the transmission signal input terminal 7 Is a simulated signal. Here, the first error calculation processing unit 8 outputs the output of the first filter processing unit from the signal from the transmission signal input terminal 3, and the echo generated in the acoustic space from the transmission signal output terminal 4 to the transmission partner. It transmits as a signal which excluded.

  Further, the signal from the transmission signal input terminal 3 is stored in the transmission input storage unit 9. The update frequency control unit 10 designates the reference signal section of the reference signal storage unit used for filter coefficient update and the data of the corresponding transmission input storage unit. The filter coefficient is updated using the filter coefficient of the filter coefficient accumulation unit 6 in this set. Further, as will be described later, the section of the reference signal storage unit is moved, and the filter coefficient update is controlled repeatedly.

  The second filter processing unit 11 performs filter processing using the reference signal in the section of the reference signal storage unit 5 specified by the update frequency control unit 10 and the filter coefficient of the filter coefficient storage unit 6. The second error calculation processing unit calculates an error between the output of the second filter processing unit 11 and the data of the transmission input storage unit 9 specified by the update frequency control unit 10, and sends the error to the filter coefficient update unit 13. And the filter coefficient is updated based on the error. Reference numeral 27 denotes the entire acoustic echo canceller.

  Here, if the filter coefficient is updated by the learning identification method, the update is performed by the following equation. An equation for calculating the error is shown in (Formula 1).

Here, d _k is a signal at time k in the corresponding section of the transmission input storage section, X _k is data in the corresponding section reference signal storage section, W _k is a filter coefficient in the corresponding section of the filter coefficient storage section, and ε _k is a result of the second error calculation processing unit. T represents the transpose of the matrix.

  An expression for updating the filter coefficient is shown in (Expression 2).

Here, W k _{+ 1} is the filter coefficient of the relevant section of the filter coefficient storing unit in the next sampling is time _k + 1, ‖X k ‖ ² norm of X _k, mu is a constant that controls the speed of advancing the learning .

  By creating a signal simulating reverberation characteristics in the reference signal storage unit 5, the filter coefficient storage unit 6 and the filter processing unit 7, and subtracting the transmission signal input terminal 3 from the signal from the first error calculation processing unit, A part of the function of the acoustic echo canceller, which is a function of generating an echo-free signal, is realized.

  Further, the result of the filter processing unit 2 performing the filtering process with the reference signal of the section of the divided storage unit that is a part of the reference signal storage unit 5 controlled by the update frequency control unit 10 and the filter coefficient of the filter coefficient storage unit An acoustic echo is calculated by calculating an error from the data of the transmission input accumulating unit 9 controlled by the update frequency control unit 10 by the second error calculation processing unit 12 and updating the filter coefficient by the filter coefficient updating unit 13. The remaining functions of the canceller function are realized. The basic function of the acoustic echo canceller is realized by combining the function of generating a signal without echo and the function of updating the filter coefficient.

  Here, the operation of the update frequency control unit 10 will be described with reference to FIG.

  First, the second filter process is performed using the section 1 of the divided accumulation unit which is a part of the reference signal accumulation unit 5 with a small delay and the filter coefficient of the filter coefficient accumulation unit 6 (step 1-1). Next, an error from the corresponding data in the transmission signal accumulation unit is calculated (step 1-2). Next, using the error, the filter coefficient is updated by the filter coefficient updating unit 13 (step 1-3).

  In the next step, the second filter processing is performed using the section 2 of the divided storage unit which is a part of the reference signal storage unit 5 and the filter coefficient of the filter coefficient storage unit 6 (step 2-1). Next, an error with the corresponding data in the transmission signal accumulation unit is calculated (step 2-2). Next, the filter coefficient is updated by the filter coefficient updating unit 13 using the error (step 2-3).

  Similarly, the filter coefficient can be updated by using the section 3 of the divided storage unit that is a part of the reference signal storage unit 5. In addition, the filter coefficient can be updated using the section 4 of the divided storage unit which is a part of the reference signal storage unit 5.

  In this way, the filter coefficients can be updated independently for the sections of the divided accumulation units of the reference signal. In this figure, the frequency of the reference signal storage unit is changed not according to the same frequency but from the division 1 of the division accumulation unit with a small delay of the reference signal to the division 4 of the division accumulation unit with a large delay according to the magnitude of the delay. Thus, it is possible to control which section of the divided accumulation unit is used so that the learning of the acoustic echo canceller proceeds. Here, in FIG. 2, the embodiment is shown in the section of four divided accumulation units, but is not limited to the section of four divided accumulation units.

  Thereby, the learning of the acoustic echo canceller can be performed quickly and the accuracy can be improved by changing the frequency at which the reference signal is used for updating the filter coefficient in accordance with the magnitude of the delay.

  Further, in FIG. 2, the reference signal storage unit is divided into equal division storage units and matched with the filter coefficient storage unit, so that the filter coefficient update unit 13 using the respective division storage units can be used. The segment of the divided accumulation unit with the same length can be targeted, and the processing amount of the algorithm is reduced.

  Furthermore, by making the reference signal storage unit a power of 2 and making the number of equals a power of 2 smaller than that, the processing amount of the algorithm when realized by a DSP or the like is further reduced.

  In addition, in the acoustic echo canceller, considering that the transmission characteristic of the current time is expressed as the delay is small, by increasing the frequency of updating the filter coefficient using the section of the divided accumulation unit with a small delay, An acoustic echo canceller that follows the latest transfer characteristics can be realized. In addition, considering that there are speakers around the microphone and environmental noise in the signal from the transmission signal input terminal, the acoustic echo canceller algorithm uses only the latest segmented storage section with a short filter coefficient. In comparison, it is also less susceptible to speaker and environmental noise.

  A specific implementation example when implemented by an actual DSP will be described.

  Consider a case where an echo canceller is realized by a learning identification method at 8000 Hz sampling. The sampling interval is 125 μs from the sampling frequency. It is assumed that the DSP can perform filtering processing and filter coefficient updating processing corresponding to a total of 8192 taps per sample. Assume that 4096 samples are accumulated in the reference signal accumulation unit.

  Assume that a filter coefficient of 128 ms is required from the reverberation time of the room. Here, the number of taps corresponding to 128 ms is 1024 taps. In this case, since only a total of 1024 taps of hardware resources are used in the filtering process and the update process of the filter coefficient, the DSP processing capacity for the remaining 7168 taps is surplus. Then, the filter coefficient can be updated 7168 ÷ 1024 = 7 times, and the filter coefficient can be updated a total of 8 times / 1 sample.

  In the update frequency control unit 10, the first filter coefficient update is 0 to 1023 taps of the reference signal storage unit, the second filter coefficient update is 1024 to 2047 taps of the reference signal storage unit, and the third filter coefficient update is a reference signal. The fourth update of the filter coefficient from 2048 to 3071 taps of the storage unit controls the reference signal storage unit to be used like the reference signal storage unit of 3072 to 4095 taps of the reference signal storage unit. Since the reference signal accumulating unit accumulates 4096 samples, the data in the entire reference signal accumulating unit is used up by four updates. Therefore, in the remaining four times, the reference signal storage unit having 0 to 1023 taps to which a reference signal with a small delay is input is selected three more times, and 1024 to 2047 is selected one more time. The update frequency control unit performs control so that the update at 1023 is 4 times, the update at 1024 to 2047 is 2 times, the update at 2048 to 3071 is 1 time, and the update at 3072 to 4095 is 1 time.

  With the above configuration, it is possible to achieve high-speed convergence and high-accuracy estimation of the acoustic echo canceller. Particularly when used in a room with a short reverberation time, high-speed and reliable convergence is possible.

(Example 2)
Hereinafter, an acoustic echo canceller according to Embodiment 2 of the present invention will be described with reference to the drawings. FIG. 3 is a diagram showing the operation of updating the filter coefficient. The operation of the update frequency control unit 10 will be described with reference to FIG.

  First, the second filter process is performed using the section 1 of the divided accumulation unit which is a part of the reference signal accumulation unit 5 with a small delay and the filter coefficient of the filter coefficient accumulation unit 6 (step 1-1). Next, an error from the corresponding data in the transmission signal accumulation unit is calculated (step 1-2). Next, using the error, the filter coefficient is updated by the filter coefficient updating unit 13 (step 1-3).

  In the next step, the second filter processing is performed using the section 2 of the divided storage unit which is a part of the reference signal storage unit 5 and the filter coefficient of the filter coefficient storage unit 6 (step 2-1). Next, an error with the corresponding data in the transmission signal accumulation unit is calculated (step 2-2). Here, the difference from the second embodiment is that the section 1 of the divided storage section and the section 2 of the divided storage section overlap. Next, the filter coefficient is updated by the filter coefficient updating unit 13 using the error (step 2-3).

  Similarly, the filter coefficient can be updated by using the section 3 of the divided storage unit that is a part of the reference signal storage unit 5. In addition, the filter coefficient can be updated using the section 4 of the divided storage unit which is a part of the reference signal storage unit 5.

  In this way, the filter coefficients can be updated independently for the sections of the divided accumulation units of the reference signal. In this figure, the frequency of the reference signal storage unit is changed not according to the same frequency but from the division 1 of the division accumulation unit with a small delay of the reference signal to the division 4 of the division accumulation unit with a large delay according to the magnitude of the delay. Thus, it is possible to control which section of the divided accumulation unit is used so that the learning of the acoustic echo canceller proceeds. Here, in FIG. 3, the embodiment is shown in the section of four divided storage sections, but is not limited to the section of four divided storage sections.

  Thereby, the learning of the acoustic echo canceller can be performed quickly and the accuracy can be improved by utilizing the bias of delay.

  Furthermore, by overlapping the sections of the divided storage section, there is no restriction that the section length of the entire reference signal storage section needs to be an integral multiple of the section length of the filter coefficient storage section. The degree of freedom is increased with respect to the determination method, and the number of times of performing the estimation processing a plurality of times from sampling to the next sampling can be increased.

  Further, in FIG. 3, the divided sections of the divided storage units are made the same length and matched with the filter coefficient storage units, so that the algorithms of the filter coefficient update unit 13 using the sections of the respective divided storage units have the same length. This interval can be targeted, and the processing amount of the algorithm is reduced.

  Furthermore, by making the reference signal storage unit a power of 2 and making the number of equal parts a power of 2 smaller than that, the processing amount of the algorithm when realized by a DSP or the like is further reduced.

  In addition, in the acoustic echo canceller, considering that the transmission characteristic of the current time is expressed as the delay is small, by increasing the frequency of updating the filter coefficient using the section of the divided accumulation unit with a small delay, An acoustic echo canceller that follows the latest transfer characteristics can be realized. In addition, considering that there are speakers around the microphone and environmental noise in the signal from the transmission signal input terminal, the acoustic echo canceller algorithm uses only the latest segmented storage section with a short filter coefficient. In comparison, it is also less susceptible to speaker and environmental noise.

(Example 3)
Hereinafter, an acoustic echo canceller according to Embodiment 3 of the present invention will be described with reference to the drawings. FIG. 4 is a block diagram of the third embodiment, and FIG. 5 is a detailed block diagram of the reverberation time estimation unit 14 of the third embodiment.

  In addition, about the thing which has the structure similar to Example 1, the same code | symbol is attached | subjected and the description is abbreviate | omitted.

  4, the difference from the first embodiment is that the reverberation time estimation unit 14 estimates the reverberation time using the filter coefficients stored in the filter coefficient storage unit in FIG. 1, and the reverberation time of the reverberation time estimation unit 14. The filter length for setting the filter length of the first filter processing unit 7 and setting the processable processing amount for updating the filter coefficient a plurality of times from the processing amount based on the filter length in the update frequency control unit 10 It is the control unit 15.

  As a reverberation time estimation method in the reverberation time estimation unit 14, a method of measuring in advance before using the acoustic echo canceller, or a stationary noise is output from the speaker when initializing the use of the acoustic echo canceller, and the stationary noise is obtained. There is a method of measuring the time for which the signal input to the microphone is attenuated to a preset level after the output is stopped.

  Therefore, based on the estimated reverberation time, the filter length corresponding to the actual reverberation time is set in consideration of the attenuation amount to be canceled by the acoustic echo canceller.

  The reverberation time is defined as a period from when a sound of a certain level is output and becoming steady to when the sound output is stopped until the sound is attenuated by 60 dB. The reverberation time can be estimated by the reverberation time estimator 14 by the known method as described above.

  Next, the operation of the filter length control unit 15 will be described. Assuming that the ability to be attenuated by the acoustic echo canceller is 30 dB, about half of the reverberation time estimated by the reverberation time estimation unit 14 can be set as a necessary filter length.

  The reverberation time estimated as described above is used in place of the reverberation time known in advance by the update frequency control unit in the first embodiment, thereby operating in accordance with the reverberation time in the environment actually used. An acoustic echo canceller can be provided.

  Further, as another embodiment, a method for estimating reverberation time while operating an acoustic echo canceller is a block diagram of the reverberation time estimation unit 14 in FIG. 5 and a graph of filter coefficients when the learning identification method is actually operated. A description will be given based on FIGS.

  In this embodiment, the reverberation time estimation unit 14 determines the interval for obtaining the size of the filter coefficient for estimating the reverberation time, and the reverberation time based on the size of the filter coefficient in the interval. It is comprised from the reverberation time simple estimation part 17 which estimates. In this configuration, first, a reverberation time estimation method including the section start point / end point calculation unit 16 and the reverberation time simple estimation unit 17 will be described.

  The section start point / end point calculation unit 16 refers to the filter coefficient of the filter coefficient storage unit 6 and uses the first start point to the first end point in the vicinity of the tap position of the largest filter coefficient and the time axis. It is a processing unit for calculating the second start point and the second end point, which compares the subsequent filter coefficients from the second start point to the second end point.

  The reverberation time simple estimation unit 17 compares the size of the filter coefficient from the first start point to the first end point with the size of the filter coefficient from the second start point to the second end point, thereby determining the filter processing section. decide. In practice, when creating a DSP program for an acoustic echo canceller, a value convenient for the program is often selected as the filter length. Convenient values when creating a DSP program include powers of 2, 512, 1024, 2048, and the like, and an operation for selecting an optimum section length is performed.

  FIG. 6 and FIG. 7 show filter coefficients after the adaptive filter actually operates and converges. The horizontal axis represents the time corresponding to 8000 Hz sampling, and the vertical axis represents the filter coefficient of each tap when expressed by signed 16 bits. Hereinafter, since the filter coefficient represents an impulse response that is a transfer characteristic when the acoustic echo canceller is actually operated, the filter coefficient is used as an expression representing the impulse response.

  FIG. 6 shows a filter coefficient for a room with a reverberation time of 100 ms, and FIG. 7 shows a filter coefficient for a room with a reverberation time of 400 ms. It can be seen that not only the attenuation time but also the tap position of the maximum filter coefficient is different. It can also be seen that the filter coefficient does not become zero at the time of convergence due to the influence of environmental noise and the degree of misadjustment even if it converges.

  Based on the fact that the filter coefficient becomes a substantially constant value, it can be estimated from the graph of FIG. 6 that the optimum filter processing interval is about 1000 taps and the optimum filter processing interval is about 2500 taps from the graph of FIG.

  As a method for determining whether to lengthen or shorten the filter processing interval from the length of the filter processing interval determined in advance as an initial value, a filter using a reverberation characteristic having an attenuation characteristic as shown in FIG. By comparing the size of the filter coefficient in the vicinity of the tap position where the coefficient is the maximum with the size of the filter coefficient after that, it is determined whether or not the currently set filter processing section length is appropriate.

  The size of the filter coefficient from the first start point to the first end point in the vicinity of the tap position of the largest filter coefficient in the filter coefficient storage unit 6, and the filter coefficient from the second start point to the second end point on the time axis. By comparing the size, it is determined whether the filter processing section is to be lengthened or shortened.

  When the filter coefficient magnitude from the first start point to the first end point of the filter coefficient accumulating unit 6 is A, and the filter coefficient magnitude from the second start point to the second end point is B, the filter processing is performed by the following equation. Judge whether to make the section longer or shorter.

  Judgment conditions are as follows. If α> B / A, the filter processing section is shortened because the filter processing section has sufficiently converged. If α <B / A <β, the filter processing section is used as it is because it is almost converged. If B / A> β, the filter processing section is lengthened because it has not converged yet. Here, α and β are coefficients for determination that are determined in advance.

  If the current filter processing section length is 1024 taps, the filter processing section length is changed to 512 taps to shorten the filter processing section length, and 2048 taps to increase the filter processing section length. It will be changed to the filter processing section length.

  The signal for actually operating the acoustic echo canceller does not operate with a signal such as white noise having an ideal frequency spectrum spread. Actually, for example, a signal whose frequency spectrum is biased, such as speech, is determined from the first start point to the first end point, and the second start point to the second end point is determined. I must. For this reason, an error has occurred even in the converged filter coefficient as compared with the learning result in the ideal state. In order to remove this error, it is necessary to reduce the error by using a certain section length.

  First, the operation of the section start point / end point calculation unit 15 will be described with reference to FIGS. 8 and 9 to determine how to determine the first start point, the first end point, the second start point, and the second end point.

  First, a method for setting a section determined from the first start point to the first end point will be described. Since it is necessary to be a section including the tap having the largest filter coefficient, the first start point is set to a tap position 30% before the filter processing section from the tap position having the largest filter coefficient. Further, the first end point is set to a tap position 30% behind the convolution section from the tap position having the largest filter coefficient.

  First, the current filter processing section of the first filter processing unit is set to 512 taps. In FIG. 8, the tap position with the largest filter coefficient is the position of 40 taps. The tap position before 153 taps, which is 30% of the 512 taps of the filter processing section, from the tap position with the largest filter coefficient is set as the second start point. In this case, since it comes before the top tap, it is set as the top tap position. Further, the tap position after 153 taps, which is 30% of the 512 taps of the filter processing section, from the tap position having the largest filter coefficient is set as the second end point.

  Similarly, in FIG. 9, the tap position with the largest filter coefficient is the position of 130 taps. The tap position before 153 taps, which is 30% of the 512 taps of the filter processing section, from the tap position with the largest filter coefficient is set as the second start point. Further, a tap position that is 153 taps after 30% of the 512 taps of the filter processing section from the tap position having the largest filter coefficient is set as the second end point.

  Next, the section setting method determined from the second start point to the second end point is as follows. The tap position next to the first end point is set as the second start point, and the end of the filter processing section is set as the second end point.

  In FIG. 8, the 194 tap at the tap position next to the first end point is set as the second start point, and the 512 tap which is the last tap position in the filter processing section is set as the second end point.

  In FIG. 9, the 284 tap at the tap position next to the first end point is set as the second start point, and the 512 tap which is the last tap position in the filter section is set as the second end point.

  The cumulative square sum of the filter coefficients in the section determined from the first start point to the first end point is normalized with the tap length of the corresponding section. Similarly, the cumulative sum of squares of the filter coefficients in the section determined from the second start point to the second end point is normalized with the tap length of the corresponding section.

  The section-normalized cumulative square sum determined from the first start point to the first end point and the section-normalized cumulative square sum determined from the second start point to the second end point. Is used to change the filter processing section longer, as it is, or shorter, based on the judgment criteria.

  In the above configuration, it is possible to determine whether the adaptive filter is sufficiently functioning in any filter processing section length including a short filter processing section, and filter processing that can realize transfer characteristics with converged filter coefficients The section length can be set.

  In this embodiment, in determining the second end point from the first start point, the first end point, and the second start point, a section length of 30% of the filter processing section length is set. By setting a value that matches the characteristics of the transfer function at, an optimal effect can be obtained.

  In the third embodiment, in order to calculate the size of the filter coefficient, the section length from the first start point to the first end point and the section from the second start point to the second end point of the filter coefficient storage unit 6 are calculated. By setting the lengths to the same length, the sizes can be compared with each other without performing a normalization operation with the section length. Normalization division is not necessary, and the amount of processing can be reduced.

  Further, in the third embodiment, as a method for calculating the magnitude of the filter coefficient, the cumulative square sum obtained by normalizing the section of the filter coefficient is obtained, but the absolute value of the filter coefficient averaged in the section is obtained as the magnitude of the filter coefficient. Even if the value is used, the same effect can be obtained. When the absolute value is used, the general-purpose CPU having no product-sum operation instruction can be processed in the above configuration.

  In the third embodiment, as a method for calculating the magnitude of the filter coefficient, the cumulative square sum of the filter coefficients is obtained. However, the maximum value of the filter coefficient in the section from the start point to the end point is used as the magnitude of the filter coefficient. However, the same effect can be obtained. When the maximum value is used, in the above configuration, it can be processed by a general-purpose CPU that does not have a product-sum operation instruction, and it is not necessary to perform accumulation calculation, so it is possible to eliminate overflow of operations due to accumulation. It is.

Example 4
Hereinafter, a hands-free telephone according to Embodiment 4 of the present invention will be described with reference to the drawings.

  In FIG. 10, the line input terminal 29 is electrically connected to the reception signal input terminal 1 of the acoustic echo canceller 27 according to the second embodiment inside the housing 28. Similarly, the transmission signal output terminal 4 of the acoustic echo canceller 27 is electrically connected to the line signal output terminal 30.

  A speaker 25 and a microphone 26 are fixed inside the housing 28, and the speaker 25 is electrically connected to the reception signal output terminal 2 of the acoustic echo canceller 27. A microphone 26 is electrically connected to the transmission signal input terminal 3 of the acoustic echo canceller 27.

  In this embodiment, the speaker 25 is installed toward the upper surface of the housing 27 so that the sound from the line input terminal 29 can be reproduced. In addition, four microphones 26 are installed so as to have horizontal directivity in order to pick up the voices of the speakers around the entire casing 28.

  Also, the signal from the line signal input terminal 29 is connected to the reception signal terminal 21 of the acoustic echo canceller 27, and the sound is output from the speaker 25, so that the person around the hands-free telephone can hear the sound from the line. Can.

  The sound wave of the sound output from the speaker 25 in the housing 28 is a sound wave to which the reverberation of the installed room is added, and the sound wave generated from the periphery of the housing 28 is collected from the microphone 25, This is input to the transmission signal input terminal 3 of the acoustic echo canceller 27.

  The acoustic echo canceller 27 removes the component output from the speaker 25 and added with room reverberation from the signal collected by the microphone 26. Therefore, only sound waves generated from the periphery of the housing 28 are output to the line signal output terminal 30 connected to the transmission signal output terminal 2 of the acoustic echo canceller 27. As a result, only the voices of people around the housing 28 can be output to the line signal output terminal 30, and a hands-free telephone can be realized.

  The configuration using the acoustic echo canceller according to the fourth embodiment provides a hands-free telephone that is fast and stable following a change in the room environment due to movement of a person, and that has a large echo cancellation amount and is easy to talk. be able to.

  Throughout the above embodiment, the following operations and effects are obtained.

  In the present invention, when the filter coefficient is updated a plurality of times from sampling to the next sampling, the filter coefficient is frequently updated using the reference signal with a small delay, and the reference signal with a large delay is used. The filter coefficient is updated little, and the filter coefficient is updated.

  For this reason, the transfer characteristics of the latest acoustic space can be reflected, the transfer characteristics of the filter close to the transfer characteristics of the actual acoustic space can be obtained, the amount of echo cancellation is large, and a stable acoustic echo canceller can be obtained. In addition, by updating the filter coefficient using the reference signal with a portion with a large delay, there is an effect that it is possible to provide a highly accurate echo canceller and a hands-free telephone having characteristics that are strong against the influence of noise and the like.

  As described above, since the acoustic echo canceller according to the present invention has a large echo cancellation amount and enables a stable echo cancellation function, it can be applied to a hands-free telephone or the like as an application of the acoustic echo canceller.

1 is a block diagram of an acoustic echo canceller in Embodiment 1 of the present invention. The figure which shows the operation | movement of the filter coefficient update in Example 1 of this invention. The figure which shows the operation | movement of the filter coefficient update in Example 2 of this invention. Block diagram of acoustic echo canceller in Embodiment 3 of the present invention Block diagram of reverberation time estimation unit in Embodiment 3 of the present invention Graph showing an example of impulse response in a room with short reverberation time Graph showing an example of impulse response in a room with a long reverberation time (A) is a graph which shows the example of the impulse response of the room where reverberation time is short, (b) is a figure which shows the area which determines the magnitude | size of the filter coefficient of Example 2. FIG. (A) is a graph which shows the example of the impulse response of a room with a long reverberation time, (b) is a figure which shows the area which determines the magnitude | size of the filter coefficient of Example 2. FIG. The perspective view of the hands-free telephone in Example 4 of the present invention The figure which shows the operation | movement of the filter coefficient update in the conventional acoustic echo canceller

Explanation of symbols

5 Reference Signal Storage Unit 6 Filter Coefficient Storage Unit 7 First Filter Processing Unit 8 First Error Calculation Processing Unit 9 Transmission Input Storage Unit 10 Update Frequency Control Unit 11 Second Filter Processing Unit 12 Second Error Calculation Processing Section 13 Filter coefficient update section 14 Reverberation time estimation section 15 Filter length control section 16 Section start point end point calculation section 17 Reverberation time simple estimation section 25 Speaker 26 Microphone 27 Acoustic echo canceller 28 Case 29 Line signal input terminal 30 Line signal output terminal

Claims (29)

An acoustic echo that samples sound from a speaker as a reference signal, updates a filter coefficient for the sampled reference signal, and cancels the echo of the voice input through a microphone by a filter process based on the updated filter coefficient Canceller,
A reference signal storage unit for storing the input reference signal;
A filter coefficient update unit that performs an update process of the filter coefficient for the reference signal accumulated by the reference signal accumulation unit;
An update frequency control unit that controls the frequency of the update processing by the filter coefficient update unit according to the magnitude of the delay of the input reference signal.
An acoustic echo canceller characterized by that. The reference signal storage unit
Each having a section obtained by dividing the section of the reference signal storage unit into a plurality of divided storage units;
2. The acoustic echo canceller according to claim 1, wherein the filter coefficient updating unit performs the filter coefficient updating process on each reference signal accumulated by the division accumulation unit. The divided accumulation unit
The acoustic echo canceller according to claim 2, further comprising a section obtained by equally dividing the section of the reference signal storage unit. The section of the reference signal storage unit is a power of 2.
The divided accumulation unit has a section equally divided by the power of 2 different from the length of the reference signal accumulation unit,
The acoustic echo canceller according to claim 3. The acoustic echo canceller according to claim 2, wherein sections of the divided accumulation unit overlap each other. 6. The acoustic echo canceller according to claim 5, wherein all the lengths of the sections of the divided accumulation unit are set to be the same. The acoustic echo canceller according to claim 6, wherein a section of the divided accumulation unit is set to a power of two. The update frequency control unit
The frequency of the update process by the filter coefficient update unit is controlled so that the frequency of the reference signal to which the sound from the speaker is input is high when the delay is small and the frequency is low when the delay is large. The acoustic echo canceller according to any one of claims 1 to 7. A filter coefficient storage unit;
A reverberation time estimation unit that estimates a reverberation time based on the filter coefficient accumulated in the filter coefficient accumulation unit;
A filter length control unit that controls the update frequency control unit according to an arithmetic processing amount based on the order or stage number of the filter processing corresponding to the reverberation time estimated by the reverberation time estimation unit,
The acoustic echo canceller according to any one of claims 1 to 8. The reverberation time estimation unit includes:
A first section composed of a first start point and a first end point determined based on the position of the filter coefficient having the largest filter coefficient in the filter coefficient storage section; and the filter of the filter coefficient storage section A section start point end point calculation unit for obtaining a second start point and a second section composed of the second end point at which the filter coefficient determined based on the position of the filter coefficient having the maximum coefficient magnitude attenuates;
A reverberation time simple estimation unit configured to estimate a reverberation time based on a filter coefficient magnitude of a first interval of the filter coefficient accumulation unit and a filter coefficient magnitude of a second interval of the filter coefficient accumulation unit. The acoustic echo canceller according to claim 9. 11. The acoustic echo canceller according to claim 10, further comprising: the section start point / end point calculation unit in which the length of the first section and the length of the second section of the filter coefficient storage unit are set to the same length. 12. The acoustic echo canceller according to claim 10 or 11, wherein an average value of square power of filter coefficients in a section is used as a filter coefficient size of the filter coefficient storage unit. The acoustic echo canceller according to claim 10 or 11, wherein an average value of absolute values of filter coefficients in a section is used as a filter coefficient size of the filter coefficient storage unit. The acoustic echo canceller according to claim 10 or 11, wherein the maximum value of the absolute value of the filter coefficient in the section is used as the size of the filter coefficient of the filter coefficient storage unit. A hands-free telephone using the acoustic echo canceller according to any one of claims 1 to 14. An acoustic echo that samples sound from a speaker as a reference signal, updates a filter coefficient for the sampled reference signal, and cancels the echo of the voice input via a microphone by filtering based on the updated filter coefficient Cancellation method,
A reference signal accumulation step for accumulating the input reference signal;
A filter coefficient update step for executing the update process of the filter coefficient for the reference signal stored by the reference signal storage unit;
An update frequency control step for controlling the frequency of the update process by the filter coefficient update unit according to the magnitude of the delay of the input reference signal.
An acoustic echo canceling method characterized by the above. The reference signal accumulation step includes
Each having a section obtained by dividing the section of the storage memory to be processed in the reference signal storage process into a plurality of divided storage memories;
17. The acoustic echo canceling method according to claim 16, wherein in the filter coefficient updating step, the filter coefficient updating process is executed for each reference signal accumulated in the divided accumulation memory. 18. The acoustic echo canceling method according to claim 17, further comprising a divided accumulation memory that equally divides a section of the accumulation memory processed in the reference signal accumulation step. The section of the storage memory to be processed in the reference signal storage step is a power of 2.
19. The acoustic echo cancellation method according to claim 18, wherein the divided storage memory is equally divided by a power of 2 different from the length of the storage memory. 18. The acoustic echo canceling method according to claim 17, wherein the section of the divided storage memory is composed of overlapping sections of the storage memory. 21. The acoustic echo cancellation method according to claim 20, wherein the sections of the divided storage memory are set to the same length. The acoustic echo canceling method according to claim 21, wherein the length of the section of the divided storage memory is set to a power of two. The update frequency control step includes
The frequency of the update processing by the filter coefficient updating step is controlled so that the frequency of the reference signal to which the sound from the speaker is input is high when the delay is small and the frequency is low when the delay is large. The acoustic echo canceling method according to any one of claims 16 to 22. A reverberation time estimation unit for estimating reverberation time based on the filter coefficients stored in the filter coefficient memory;
A filter length control step for controlling the update frequency control step according to an arithmetic processing amount based on the order or stage number of the filter processing corresponding to the reverberation time estimated in the reverberation time estimation step,
The acoustic echo cancellation method according to any one of claims 16 to 23, wherein: The reverberation time estimation step includes:
A first section composed of a first start point and a first end point determined based on the position of the filter coefficient having the largest filter coefficient stored in the filter coefficient memory, and the filter coefficient memory; A second start point at which the filter coefficient determined based on the position of the filter coefficient having the maximum filter coefficient stored in the reference is attenuated is obtained. Section start point end point calculation process,
And a reverberation time simple estimation step of estimating a reverberation time based on a filter coefficient magnitude in a first section of the filter coefficient memory and a filter coefficient magnitude in a second section of the filter coefficient memory. The acoustic echo cancellation method according to claim 24. 26. The acoustic echo cancellation method according to claim 25, further comprising: the section start point end point calculation step in which the length of the first section and the length of the second section of the filter coefficient memory are set to the same length. 27. The acoustic echo cancellation method according to claim 25 or 26, wherein an average value of square powers of filter coefficients in a section is used as a size of the filter coefficient stored in the filter coefficient memory. 27. The acoustic echo canceling method according to claim 25 or 26, wherein an average value of absolute values of the filter coefficients in the section is used as a size of the filter coefficient stored in the filter coefficient memory. 27. The acoustic echo canceling method according to claim 25 or 26, wherein the maximum value of the absolute value of the filter coefficient in the section is used as the size of the filter coefficient stored in the filter coefficient memory.

Images