JP4920511B2 - Multichannel echo canceller - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To always perform stable echo cancellation without sound quality deterioration during multichannel reproduction, and without depending on either double talk time or single talk time. <P>SOLUTION: A multichannel echo canceler according to the present invention includes an echo cancellation section which: receives loudspeaker input signals (sp1 , sp2) including a second acoustic signal to be respectively input to a plurality of loudspeakers (10, 20) provided in a first location and detection signals (m1, m2) detected by a plurality of microphones (11, 21) provided in the first location; separates a first acoustic signal and the second acoustic signal, which are included in the detection signals (m1, m2) by performing signal processing based on an independent component analysis; and cancels as an echo the second acoustic signal, which is included in the detection signals (m1, m2) by outputting only the separated first acoustic signal to a plurality of loudspeakers (30, 40) provided in a second location. <P>COPYRIGHT: (C)2008,JPO&amp;INPIT

Description

  The present invention relates to a multi-channel echo canceller, and more particularly to a multi-channel echo canceller used for a conference system, a hands-free telephone or the like.

  In recent years, multi-channel acoustic systems such as conference systems and hands-free telephones that mutually transmit acoustic signals, which are voices between speakers in remote locations, have been realized. When this acoustic system is realized between, for example, a first location and a second location, a plurality of microphones for detecting the voice of the speaker itself and a story existing at a remote location are provided in each of the first and second locations. And a plurality of speakers for listening to a person's voice. Each speaker at the first location is connected to each microphone at the second location, and each microphone at the first location is connected to each speaker at the second location. Thereby, for example, the speaker S1 existing in the first location can hear the voice of the speaker S2 existing in the second location through each speaker in the first location. Also, the voice of the speaker S1 can be heard by the speaker S2 through each microphone at the first location.

  However, such an acoustic system has a problem that the echo must be canceled. For example, when the speaker S2 utters a sound, the sound is amplified by each speaker at the first location through each microphone at the second location. Here, each microphone is provided in the first location. For this reason, the voice of the speaker S2 amplified by each speaker at the first location is detected by each microphone at the first location. As a result, the speaker S2 listens to the sound generated by himself / herself through each speaker in the second place in addition to the sound of the speaker S1. As described above, the sound of the user himself / herself that is amplified by the speaker for listening to the sound of the speaker existing at a distant place becomes an unnecessary echo for the speaker.

  Therefore, conventionally, a multichannel echo canceller using an adaptive filter has been proposed as a multichannel echo canceller for canceling such echo. FIG. 8 is a diagram showing a configuration of a multichannel echo canceller 9 using a conventional adaptive filter used in an acoustic system. The acoustic system shown in FIG. 8 shows a case where there are two channels. In the acoustic system shown in FIG. 8, the speaker S1 exists as a sound source (near end sound source) on the near end side, and the speaker S2 exists as a sound source (far end sound source) on the far end side. Suppose that On the near-end side, speakers 10 and 20 for amplifying the far-end acoustic signal composed of the speech of the far-end speaker S2 in stereo and the near-end acoustic signal composed of the speech of the near-end speaker S1 are provided. Microphones 11 and 21 for detection are provided. On the far end side, speakers 30 and 40 for amplifying the near end acoustic signal in stereo, and microphones 31 and 41 for detecting the far end acoustic signal are provided. In the acoustic system shown in FIG. 8, it is assumed that the multi-channel echo canceller 9 is provided only on the near end side as an example.

  In FIG. 8, the multi-channel echo canceller 9 includes adaptive filters 91 to 94, adders 95 and 97, and subtractors 96 and 98. The adaptive filter 91 estimates the transfer characteristic h11 (ω) from the speaker 10 to the microphone 11 based on the output signal from the subtractor 96. ω is a frequency. The adaptive filter 91 convolves the estimation result eh11 (ω) with the speaker input signal sp1 to be input to the speaker 10 and outputs the result. The adaptive filter 92 estimates the transfer characteristic h 21 (ω) from the speaker 20 to the microphone 11 based on the output signal from the subtractor 96. The adaptive filter 92 convolves the estimation result eh21 (ω) with the speaker input signal sp2 to be input to the speaker 20 and outputs the result. The adaptive filter 93 estimates the transfer characteristic h12 (ω) from the speaker 10 to the microphone 21 based on the output signal from the subtractor 98. The adaptive filter 93 convolves the estimation result eh12 (ω) with the speaker input signal sp1 to be input to the speaker 10 and outputs the result. The adaptive filter 94 estimates the transfer characteristic h22 (ω) from the speaker 20 to the microphone 21 based on the output signal from the subtractor 98. The adaptive filter 94 convolves the estimation result eh22 (ω) with the speaker input signal sp2 to be input to the speaker 20 and outputs the result.

  The adder 95 receives the output signal from the adaptive filter 91 and the output signal from the adaptive filter 92, and adds these output signals. The subtracter 96 receives the detection signal m1 detected by the microphone 11 and the output signal from the adder 95, and subtracts the output signal from the adder 95 from the detection signal m1. Thereby, the output signal y1 from the subtracter 96 becomes a signal in which the voice of the far-end speaker S2 which is an echo is canceled. The output signal y1 from the subtractor 96 is transmitted to the far end side and is amplified by the far end side speaker 30. The adder 97 receives the output signal from the adaptive filter 93 and the output signal from the adaptive filter 94 and adds these output signals. The subtracter 98 receives the detection signal m2 detected by the microphone 21 and the output signal from the adder 97, and subtracts the output signal from the adder 97 from the detection signal m2. Thereby, the output signal y2 from the subtractor 98 becomes a signal in which the voice of the far-end speaker S2 which is an echo is canceled. The output signal y2 from the subtracter 98 is transmitted to the far end side and is amplified by the far end side speaker 40.

  Here, for the estimation of transfer characteristics performed by the adaptive filters 91 to 94, a learning identification method (LMS) generally used as a learning method of the adaptive filter is used. Specifically, the adaptive filters 91 and 92 estimate the transfer characteristics so that the power of the output signal y1 from the subtractor 96 is minimized. The adaptive filters 93 and 94 estimate the transfer characteristics so that the power of the output signal y2 from the subtractor 98 is minimized.

  Hereinafter, problems of the conventional multi-channel echo canceller 9 will be described. In order to obtain the echo cancellation effect in FIG. 8, the correct transfer characteristics must be estimated in each of the adaptive filters 91 to 94. For example, in the case of the adaptive filter 91, the estimation result eh11 (ω) needs to coincide with the transfer characteristic h11 (ω). However, the conventional multi-channel echo canceller 9 cannot estimate a correct transfer characteristic unless only one of the speaker input signal sp1 and the speaker input signal sp2 is amplified. That is, unless the speaker 10 or the speaker 20 is operating in a monaural reproduction state, correct transfer characteristics cannot be estimated.

式（１）で表されたｓ２（ω）成分は、エコーである。よって、適応フィルタ９１および９２は、擬似エコーである加算器９５からの出力信号が式（１）で表されたｓ２（ω）成分と同じになるように、伝達特性を推定すればよい。加算器９５からの出力信号が式（１）で表されたｓ２（ω）成分と同じになれば、出力信号ｙ１のパワーが最小となり（つまり、ｓ１（ω）成分のみとなり）、エコーがキャンセルされる。 At the time of multi-channel reproduction (in this case, during stereo reproduction), usually both the speaker 10 and the speaker 20 operate, and a signal having a correlation is input to the speakers 10 and 20. For example, it is assumed that the voice of the speaker S2 is detected in stereo in the far-end microphones 31 and 41 shown in FIG. Further, the voice of the speaker S2 is s2 (ω), the transfer characteristic from the speaker S2 to the microphone 31 is a21 (ω), and the transfer characteristic from the speaker S2 to the microphone 41 is a22 (ω). At this time, the speaker input signal sp1 input to the speaker 10 is s2 (ω) · a21 (ω), and the speaker input signal sp2 input to the speaker 20 is s2 (ω) · a22 (ω). Since the speaker input signals sp1 and sp2 both include s2 (ω), they have a correlation. Further, the detection signal m1 (ω) detected by the microphone 11 is as shown in Expression (1).

The s2 (ω) component expressed by Equation (1) is an echo. Therefore, the adaptive filters 91 and 92 may estimate the transfer characteristics so that the output signal from the adder 95, which is a pseudo echo, is the same as the s2 (ω) component expressed by Expression (1). When the output signal from the adder 95 becomes the same as the s2 (ω) component expressed by the equation (1), the power of the output signal y1 becomes minimum (that is, only the s1 (ω) component), and the echo is canceled. Is done.

However, m1 (ω) expressed by the equation (1) includes a signal obtained by multiplying s2 (ω) by a predetermined transfer characteristic, and the speaker input signals sp1 and sp2 are also s2 (ω). And a product multiplied by a predetermined transfer characteristic. This means that the s2 (ω) component expressed by the equation (1) can be reproduced by using either the speaker input signal sp1 or the speaker input signal sp2. Therefore, a plurality of solutions (for example, Expression (2) or Expression (3)) exist in the transfer characteristic eh11 (ω) estimated by the adaptive filter 91 and the transfer characteristic eh21 (ω) estimated by the adaptive filter 92. Will do.

  As described above, the conventional multi-channel echo canceller 9 has a problem that, during multi-channel reproduction, the correct transfer characteristic cannot be estimated due to the indefiniteness of the solution, and the echo cancellation effect cannot be obtained stably.

Therefore, conventionally, a technique (for example, Patent Document 1) has been proposed in which the level of the signal level of each channel is determined to select one channel for estimation processing. In addition, a technique for estimating a correct transfer characteristic by adding an additional signal to the speaker input signal sp1 and the speaker input signal sp2 (for example, Patent Document 2) has been proposed. Conventionally, by adopting these techniques, countermeasures against indefiniteness of solutions in the conventional multi-channel echo canceller 9 have been taken.
Japanese Patent No. 3407392 Japanese Patent No. 3073976

  However, in the technique disclosed in Patent Document 1, when the signal level difference between channels is small, the magnitude of the signal level of each channel cannot be correctly determined, and the correct transfer characteristic cannot be estimated. For this reason, the technique disclosed in Patent Document 1 cannot always perform echo cancellation stably. Further, in the technique disclosed in Patent Document 2, an additional signal is added to the speaker input signal sp1 and the speaker input signal sp2 in order to estimate a correct transfer characteristic. For this reason, in the speaker, there is a problem that the additional signal is also amplified in addition to the voice of the speaker, and the sound quality deteriorates due to the additional signal. As described above, in the techniques disclosed in Patent Documents 1 and 2 proposed as measures against the indefiniteness of the solution, stable echo cancellation cannot always be performed, or sound quality deterioration occurs.

  Therefore, the present invention can always perform stable echo cancellation without causing deterioration in sound quality during multi-channel playback, and can perform stable echo cancellation regardless of double talk or single talk. An object is to provide an echo canceller.

  The multi-channel echo canceller according to the present invention solves the above problems, and the multi-channel echo canceller according to the present invention is located at the first location detected by a plurality of microphones provided at the first location. First acoustic signal from one or more sound sources present and second sound from one or more sound sources present at the second location detected by a plurality of microphones provided at the second location. A multi-channel echo canceller for use in an acoustic system for transmitting signals to and from each other by using a plurality of speakers provided at each of the first and second locations, In addition to the first acoustic signal, the plurality of microphones provided at one location further includes a second sound that is amplified by a plurality of speakers provided at the first location. The multi-channel echo canceller is provided at the first location and the speaker input signal including the second acoustic signal to be input to each of the plurality of speakers provided at the first location. The detection signals of the plurality of microphones are input, signal processing based on independent component analysis is performed to separate the first acoustic signal and the second acoustic signal included in each detection signal, and the separated first acoustic signal An echo cancellation unit is provided that cancels the second acoustic signal included in each detection signal as an echo by outputting only the signal to a plurality of speakers provided at the second location.

  By performing signal processing based on independent component analysis, even if a second acoustic signal having a correlation is included in each speaker input signal, the first acoustic signal and the second acoustic signal included in each detection signal are included. The signal can be separated. Thereby, it is possible to always perform stable echo cancellation without causing deterioration of sound quality while solving the problem of indefiniteness of the solution during multi-channel reproduction. Furthermore, stable echo cancellation can be performed regardless of double talk or single talk.

  The first location corresponds to, for example, a location on the near end side in an embodiment described later. The first acoustic signal corresponds to a near-end acoustic signal in an embodiment described later. The second location corresponds to, for example, a far-end location in an embodiment described later. The second acoustic signal corresponds to a far-end acoustic signal in an embodiment described later.

  More preferably, the echo cancellation unit receives each speaker input signal and each detection signal as input, performs signal processing based on independent component analysis, and includes a first acoustic signal and a second acoustic signal included in each detection signal, A plurality of loudspeakers provided in the second place, the signals having a low correlation included in the first acoustic signal are separated by the number of detection signals, and only the separated signals having a low correlation are provided in the second location. A sound source separation unit that outputs to

  More preferably, the echo cancellation unit is provided corresponding to each of the plurality of microphones provided in the first place, and receives the detection signal of the corresponding microphone and each speaker input signal as input, and performs signal processing based on independent component analysis The first acoustic signal and the second acoustic signal included in the detection signal of the corresponding microphone are separated, and only the separated first acoustic signal is provided at the second location. It is preferable to have a plurality of sound source separation units that output to a speaker.

  More preferably, a separation matrix for separating the first acoustic signal and the second acoustic signal included in each detection signal is preset in the echo cancellation unit, and the separation matrix is the first location. A plurality of first matrix elements relating to respective transfer characteristics from a plurality of speakers provided in the first location to a plurality of microphones provided in the first location, the plurality of first matrices being learned according to independent component analysis The echo cancellation unit includes an element, and multiplies the separation matrix by an input vector constituted by each speaker input signal and each detection signal, and subtracts a second acoustic signal included in each detection signal from each detection signal. Thus, the first acoustic signal and the second acoustic signal included in each detection signal may be separated.

  More preferably, the separation matrix further includes a plurality of second matrix elements for each transfer characteristic from one or more sound sources present at the second location to a plurality of microphones provided at the second location, Of the second matrix elements, the matrix elements other than the diagonal of the separation matrix may be zero.

  More preferably, the separation matrix uses a plurality of second signals for separating signals having low correlation included in the second acoustic signal in each speaker input signal by the number of speaker input signals using each detection signal. The second matrix elements may be all zero.

  More preferably, the separation matrix includes a plurality of second matrix elements relating to each transfer characteristic from one or more sound sources present at the second location to a plurality of microphones provided at the second location, and each detection signal. And a plurality of third matrix elements for separating signals having a low correlation included in the second acoustic signal in each speaker input signal by the number of speaker input signals. Of the matrix elements, the matrix elements other than the diagonal of the separation matrix are 0, and all the third matrix elements are all 0.

  More preferably, the separation matrix further includes a plurality of second matrix elements for each transfer characteristic from one or more sound sources present at the first location to a plurality of microphones provided at the first location, Of the second matrix elements, the matrix elements other than the diagonal of the separation matrix may be zero.

  The present invention is also directed to a multi-channel echo cancellation method, and the multi-channel echo cancellation method according to the present invention is present at the first location detected by a plurality of microphones provided at the first location. A first acoustic signal from one or more sound sources to be detected, and a second acoustic signal from one or more sound sources present at the second location detected by a plurality of microphones provided at the second location Is a multi-channel echo cancellation method used in an acoustic system that mutually transmits between the first and second locations by using a plurality of speakers provided in the first and second locations, respectively. In addition to the first acoustic signal, the plurality of microphones provided at one location further includes a second sound that is amplified by a plurality of speakers provided at the first location. The multi-channel echo canceling method is provided in the first location, including a speaker input signal including a second acoustic signal to be input to each of the plurality of speakers provided in the first location. Included in each detection signal by performing input processing for inputting detection signals of a plurality of microphones, and performing signal processing based on independent component analysis on each speaker input signal and each detection signal input in the input step. A separation step for separating the first acoustic signal and the second acoustic signal, and outputting only the first acoustic signal separated in the separation step to a plurality of speakers provided in the second location, A cancellation step of canceling the second acoustic signal included in the detection signal as an echo.

  More preferably, the separation step includes a first acoustic signal and a second acoustic signal included in each detection signal with respect to an input vector composed of each speaker input signal and each detection signal input in the input step. A separation matrix including a plurality of matrix elements relating to respective transfer characteristics from a plurality of speakers provided at the first location to a plurality of microphones provided at the first location. A first signal calculation step for calculating a plurality of output signals constituting an output vector as a result of the multiplication by multiplication, and a higher-order correlation between the output signals calculated in the first signal calculation step A matrix calculation step for calculating a correlation matrix having matrix elements in the matrix element, and a correlation matrix calculated in the matrix calculation step. A learning step for learning each matrix element, an update step for updating each matrix element included in the separation matrix used in the first signal calculation step to each matrix element learned in the learning step, and an input step A first acoustic signal included in each detection signal is obtained by multiplying an input vector composed of each input speaker input signal and each detection signal by a separation matrix in which each matrix element is updated in the update step. And a second signal calculation step for calculating a plurality of output signals constituting an output vector from which the second acoustic signal is separated, and the cancellation step includes the output signals calculated in the second signal calculation step. Among them, an output signal including only the first acoustic signal may be output to a plurality of speakers provided at the second location.

  The present invention is also directed to a program, the program according to the present invention from one or more sound sources present at the first location detected by a plurality of microphones provided at the first location. A first acoustic signal and a second acoustic signal from one or more sound sources present at the second location detected by a plurality of microphones provided at the second location are first and second. A program to be executed by a computer used in an acoustic system for mutual transmission between the first and second locations by using a plurality of speakers provided at each location, and a plurality of speakers provided at the first location In addition to the first acoustic signal, the microphone detects a second acoustic signal amplified by a plurality of speakers provided at the first location, and is provided to the computer at the first location. An input step for inputting a speaker input signal including a second acoustic signal to be input to each of the plurality of speakers, and detection signals of a plurality of microphones provided at the first location, and input in the input step A separation step for separating the first acoustic signal and the second acoustic signal included in each detection signal by performing signal processing based on independent component analysis on each speaker input signal and each detection signal, and a separation step A cancellation step of canceling the second acoustic signal included in each detection signal as an echo is executed by outputting only the first acoustic signal separated in step 2 to a plurality of speakers provided in the second location. It is a program.

  The present invention is also directed to an integrated circuit, and the integrated circuit according to the present invention includes one or more sound sources existing at the first location detected by a plurality of microphones provided at the first location. And a second acoustic signal from one or more sound sources present at the second location detected by a plurality of microphones provided at the second location, and An integrated circuit used in an acoustic system for mutual transmission between the first and second locations by using a plurality of speakers provided at each of the second locations, the plurality of speakers provided at the first location In addition to the first acoustic signal, the microphone detects a second acoustic signal amplified by a plurality of speakers provided at the first location, and the integrated circuit is provided at the first location. Input to multiple speakers. The speaker input signal including the second acoustic signal to be input and the detection signals of the plurality of microphones provided at the first location are input, and signal processing based on independent component analysis is performed to include the first signal included in each detection signal. The first acoustic signal and the second acoustic signal are separated, and only the separated first acoustic signal is output to a plurality of speakers provided at the second location, whereby the second included in each detection signal. An echo canceling unit for canceling the acoustic signal as an echo.

  Further, the multi-channel echo canceller according to the present invention, in addition to the acoustic signal of the near-end sound source to be detected, further detects a detection signal of one or more microphones including an acoustic signal amplified by a plurality of speakers. A multi-channel echo canceller that cancels the sound signals amplified by the plurality of speakers as echoes and outputs only the sound signal of the near-end sound source, and the sound signal of the near-end sound source and the sound signal amplified by each speaker The detection signal of one or more microphones including a speaker input signal to be input to each speaker having a sense of direction of sound, and an acoustic signal of a near-end sound source and an acoustic signal amplified by each speaker The signal to be output so that the signal to be output becomes a signal that retains the sound quality of the sound signal of the near-end sound source at the same time that is generated simultaneously. By performing an adaptive operation so that the sound signal amplified by each speaker is canceled, the sound signal of the near-end sound source included in the detection signal of one or more microphones and the sound signal amplified by each speaker are obtained. A sound source separation unit that separates and outputs only the acoustic signal of the separated near-end sound source is provided.

  The acoustic signal of the near-end sound source is a signal indicating a sound generated from one or more sound sources existing at a place where one or more microphones are provided, a statistic having characteristics of the sound, and the like. This is a signal obtained by canceling an acoustic signal amplified by a plurality of speakers with respect to detection signals of two or more microphones. In addition, a speaker input signal having a sense of direction of sound refers to, for example, a plurality of characteristics (level ratio, time delay, etc.) of an acoustic signal detected by a far-end microphone using a plurality of near-end speakers. A signal that can be reproduced. The signal retaining the sound quality means a signal retaining the frequency characteristics (amplitude frequency characteristics, amplitude phase frequency characteristics, etc.) of the acoustic signal of the near-end sound source input to the sound source separation unit.

  More preferably, the sound source separation unit estimates each transfer characteristic from each speaker to one or more microphones, and determines each transfer characteristic obtained by estimating an acoustic signal amplified by each speaker and detected by one or more microphones. The adaptive operation may be performed so that the calculated acoustic signal is subtracted from the detection signal of one or more microphones.

  According to the present invention, multi-channel echo that can always perform stable echo cancellation without causing deterioration in sound quality during multi-channel playback, and can perform stable echo cancellation regardless of double talk or single talk. A canceller can be provided.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(First embodiment)
With reference to FIG. 1, a configuration of a multi-channel echo canceller according to the first embodiment of the present invention will be described. FIG. 1 is a diagram illustrating a configuration example of a multichannel echo canceller according to a first embodiment used in an acoustic system. The acoustic system is a system that mutually transmits acoustic signals between locations on the near end side and the far end side. In the acoustic system shown in FIG. 1, speakers S1 and S2 exist as a plurality of different sound sources (near-end sound sources) on the near end side, and speakers S3 and S4 differ on the far end side. Assume that there are multiple sound sources (far-end sound sources). On the near end side, speakers 10 and 20 for amplifying far-end acoustic signals composed of the speech of the far-end speakers S3 and S4, and near-end sound composed of the speech of the near-end speakers S1 and S2 Microphones 11 and 21 for detecting signals are provided. On the far end side, speakers 30 and 40 for amplifying the near-end acoustic signal and microphones 31 and 41 for detecting the far-end acoustic signal are provided. The speakers (10, 20, 30, 40) and the microphones (11, 21, 31, 41) are the same as those described in FIG. 8, and are denoted by the same reference numerals. In the acoustic system shown in FIG. 1, as an example, it is assumed that the multichannel echo canceller according to the present embodiment is provided only on the near end side. In the acoustic system shown in FIG. 1, as an example, a state where a so-called double talk is performed in which the near-end speakers S1 and S2 and the far-end speakers S3 and S4 are talking simultaneously. Suppose that

  In FIG. 1, the multi-channel echo canceller according to the present embodiment is configured by an echo cancellation unit 1. The echo cancel unit 1 includes a sound source separation unit 100, conversion units 110 to 113, and inverse conversion units 120 and 121.

  The conversion unit 110 receives the speaker input signal sp2 (t) including the far-end acoustic signal to be input to the speaker 20, and converts the signal in the time domain (t) to the signal in the frequency domain (ω). Speaker input signal sp <b> 2 (ω) converted by conversion unit 110 is output to sound source separation unit 100. The conversion unit 111 receives a speaker input signal sp1 (t) including a far-end acoustic signal to be input to the speaker 10, and converts the signal (t) in the time domain into a signal (ω) in the frequency domain. The speaker input signal sp1 (ω) converted by the conversion unit 111 is output to the sound source separation unit 100. The conversion unit 112 receives the detection signal m2 (t), which is detected by the microphone 21 and includes the near-end acoustic signal and the far-end acoustic signal amplified by the speakers 10 and 20, and is input from the time-domain signal (t). Convert to frequency domain signal (ω). The detection signal m2 (ω) converted by the conversion unit 112 is output to the sound source separation unit 100. The conversion unit 113 receives the detection signal m1 (t) including the near-end acoustic signal detected by the microphone 11 and the far-end acoustic signal amplified by the speakers 10 and 20 as input, and from the time-domain signal (t). Convert to frequency domain signal (ω). The detection signal m1 (ω) converted by the conversion unit 113 is output to the sound source separation unit 100.

  The sound source separation unit 100 receives the detection signals (m1 (ω), m2 (ω)) and the speaker input signals (sp1 (ω), sp2 (ω)) as inputs. The sound source separation unit 100 performs sound source separation processing based on independent component analysis on the input signal. By this sound source separation process, the near-end acoustic signal and the far-end acoustic signal included in the detection signals (m1 (ω), m2 (ω)) are separated. The sound source separation process based on the independent component analysis will be described in detail later. The sound source separation unit 100 outputs only the separated near-end acoustic signals as output signals y1 (ω) and y2 (ω). Here, the far-end acoustic signal composed of the voices of the far-end speakers S3 and S4 corresponds to a signal unnecessary for the speakers S3 and S4, that is, an echo. Therefore, by outputting only the near-end acoustic signal from the sound source separation unit 100, the far-end acoustic signal included in the detection signals (m1 (ω), m2 (ω)) can be canceled as an echo.

  The inverse conversion unit 120 receives the output signal y1 (ω) from the sound source separation unit 100 and converts the signal in the frequency domain (ω) into a signal in the time domain (t). The acoustic signal y <b> 1 (t) converted by the inverse conversion unit 120 is output to the speaker 30 and is amplified by the speaker 30. The inverse conversion unit 121 receives the output signal y2 (ω) from the sound source separation unit 100 and converts the signal in the frequency domain (ω) into a signal in the time domain (t). The output signal y <b> 2 (t) converted by the inverse conversion unit 121 is output to the speaker 40 and is amplified by the speaker 40.

式（４）において、話者Ｓ１の音声をｓ１（ω）、話者Ｓ２の音声をｓ２（ω）、話者Ｓ３の音声をｓ３（ω）、話者Ｓ４の音声をｓ４（ω）とする。また、話者Ｓ１からマイクロホン１１までの伝達特性をａ１１（ω）、話者Ｓ１からマイクロホン２１までの伝達特性をａ１２（ω）、話者Ｓ２からマイクロホン１１までの伝達特性をａ２１（ω）、話者Ｓ２からマイクロホン２１までの伝達特性をａ２２（ω）、話者Ｓ３からマイクロホン３１までの伝達特性をａ３１（ω）、話者Ｓ３からマイクロホン４１までの伝達特性をａ３２（ω）、話者Ｓ４からマイクロホン３１までの伝達特性をａ４１（ω）、話者Ｓ４からマイクロホン４１までの伝達特性をａ４２（ω）とする。 Hereinafter, the sound source separation process based on the independent component analysis performed by the sound source separation unit 100 will be described in detail. First, detection signals (m1 (ω), m2 (ω)) and speaker input signals (sp1 (ω), sp2 (ω)) input to the sound source separation unit 100 will be described in detail. The detection signals (m1 (ω), m2 (ω)) and the speaker input signals (sp1 (ω), sp2 (ω)) are expressed as in Expression (4).

In the equation (4), the voice of the speaker S1 is s1 (ω), the voice of the speaker S2 is s2 (ω), the voice of the speaker S3 is s3 (ω), and the voice of the speaker S4 is s4 (ω). To do. Further, the transfer characteristic from the speaker S1 to the microphone 11 is a11 (ω), the transfer characteristic from the speaker S1 to the microphone 21 is a12 (ω), the transfer characteristic from the speaker S2 to the microphone 11 is a21 (ω), The transfer characteristic from the speaker S2 to the microphone 21 is a22 (ω), the transfer characteristic from the speaker S3 to the microphone 31 is a31 (ω), the transfer characteristic from the speaker S3 to the microphone 41 is a32 (ω), and the speaker The transfer characteristic from S4 to the microphone 31 is a41 (ω), and the transfer characteristic from the speaker S4 to the microphone 41 is a42 (ω).

  Next, a detailed configuration of the sound source separation unit 100 will be described with reference to FIG. FIG. 2 is a diagram illustrating a detailed configuration of the sound source separation unit 100. In FIG. 2, the sound source separation unit 100 includes a separation unit 101 and a learning unit 102.

In the separation unit 101, a separation matrix W (4, 4) configured by matrix elements wij (the number of rows i and the number of columns j is an integer of 1 to 4) is set. In the initial state, for example, it is assumed that the unit matrix is set as the separation matrix W (4, 4). The separation unit 101 receives detection signals (m1 (ω), m2 (ω)) and speaker input signals (sp1 (ω), sp2 (ω)). The separation unit 101 calculates the output signals y1 to y4 according to the equation (5) based on the set separation matrix W (4, 4), and outputs the calculated output signals y1 to y4, respectively. Specifically, the separation unit 101 includes a detection signal (m1 (ω), m2 (ω)) and a speaker input signal (sp1 (ω), sp2 (ω)) as shown in Expression (5). Is multiplied by the set separation matrix W (4, 4) to calculate an output vector composed of the output signals y1 (ω) to y4 (ω).

  The learning unit 102 receives the output signals y1 (ω) to y4 (ω) and learns the separation matrix W (4, 4) according to the independent component analysis. Specifically, the learning unit 102 learns the separation matrix W (4, 4) so that the output signals y1 (ω) to y4 (ω) are independent signals. Here, independent means that there is no correlation, that is, the correlation is 0 (zero). The learning unit 102 updates the separation matrix W (4, 4) set in the separation unit 101 to the learned separation matrix W (4, 4).

式（６）において、出力信号ｙ１（ω）〜ｙ４（ω）の要素は、周波数領域の複素信号になっており、分離行列Ｗ（４、４）ｉ、Ｗ（４、４）ｉ―１を構成する各行列要素は、複素数の係数になっている。Ｉは４×４の単位行列を示し、ε｛・｝は時間平均を示し、＊は複素共役信号を示す。φ（・）は非線形関数を示す。非線形関数としては、信号の確率密度関数の対数部分を微分したものに対応したものを用いるのがよく、一般的にはｔａｎｈ（・）を用いる。αは学習速度を制御するためのステップサイズパラメータを示す。ｉは学習回数を示し、右辺のＷ（４、４）ｉを左辺のＷ（４、４）ｉ―１に代入することで学習が行われる。εの括弧内に示される行列は、高次の相関行列である。 Hereinafter, the learning method of the learning unit 102 will be described more specifically. A learning formula generally used for frequency domain independent component analysis using the gradient method is as shown in Formula (6). Note that the learning formula used for the independent component analysis is not limited to the formula (6), and may be another learning formula.

In Expression (6), the elements of the output signals y1 (ω) to y4 (ω) are complex signals in the frequency domain, and the separation matrices W (4, 4) i, W (4, 4) i−1. Each matrix element that constitutes is a complex coefficient. I represents a 4 × 4 unit matrix, ε {·} represents a time average, and * represents a complex conjugate signal. φ (·) indicates a nonlinear function. As the nonlinear function, it is preferable to use a function corresponding to a derivative of the logarithm part of the probability density function of the signal, and generally tanh (·) is used. α represents a step size parameter for controlling the learning speed. i indicates the number of learning times, and learning is performed by substituting W (4, 4) i on the right side into W (4, 4) i-1 on the left side. The matrix shown in parentheses for ε is a higher-order correlation matrix.

  Here, the speakers S1 to S4 are all different speakers and are independent sound sources. Therefore, it can be said that s1 (ω) to s4 (ω) in the equation (4) are independent from each other and have uncorrelated sounds. Further, two detection signals (m1 (ω), m2 (ω)) are input, and the number of detection signals matches the number of speakers (S1 and S2) on the near end side. Also, two speaker input signals (sp1 (ω), sp2 (ω)) are input, and the number of speaker input signals matches the number of far-end speakers (S3 and S4). Therefore, when the learning unit 102 learns the separation matrix W (4, 4) under these conditions, and the learning converges, the separation matrix W (4, 4) becomes the detection signals (m1 (ω), m2 (ω )) And the loudspeaker input signals (sp1 (ω), sp2 (ω)), a matrix capable of separating each of s1 (ω) to s4 (ω). That is, the output signal y1 calculated by the separation unit 101 based on the separation matrix W (4, 4) where learning has converged is included in the detection signal (m1 (ω), m2 (ω)). ), And the output signal y2 includes only s2 (ω) included in the detection signals (m1 (ω), m2 (ω)). Similarly, the output signal y3 includes only s3 (ω) included in the speaker input signals (sp1 (ω), sp2 (ω)), and the output signal y4 includes the speaker input signal (sp1 (ω). ), Sp2 (ω)), only s4 (ω) included in sp2 (ω)) is included.

  Actually, as sounds from independent sound sources other than the speakers S1 and S2 on the near end side, there are environmental noises on the near end side. The same applies to the far end side. However, these environmental noises are generally signals closer to a Gaussian distribution than the speaker's voice. For this reason, in the learning based on the equation (6), that is, the learning based on the independent component analysis, the speech of a speaker having a large non-Gaussian property is preferentially processed. That is, since the learning unit 102 preferentially processes s1 (ω) to s4 (ω), the detection signal (m1 (ω), m2 (ω)) and the speaker input signal (sp1 (ω), sp2 (Ω)), a separation matrix capable of separating each of s1 (ω) to s4 (ω) is learned.

  In this manner, the learning unit 102 learns the separation matrix W (4, 4) according to the independent component analysis, so that the separation unit 101 obtains the near-end acoustic signal from the detection signals (m1 (ω), m2 (ω)). The output signals y1 and y2 can be separated, and the far-end acoustic signal can be separated from the speaker input signals (sp1 (ω), sp2 (ω)) as output signals y3 and y4. Note that the output signals y3 and y4, which are far-end acoustic signals, are not output from the sound source separation unit 100 and are used only for learning by the learning unit 102.

Hereinafter, what kind of signal the output signals y1 to y4 will be will be described using Expression (7). Expression (7) is a more detailed representation of the signal input to the separation unit 101 by substituting Expression (4) into Expression (5). In Expression (7), the description of (ω) shown in Expression (4) is omitted.

  In the state where learning of the separation matrix W (4, 4) has converged, in the output signal y1 output from the separation unit 101, the voices of the speakers S3 and S4 included in the detection signal m1 (s3 · a31, s3 · a32, s4 · a41, s4 · a42) are canceled as echoes, the voice of the speaker S2 (s2 · a21) included in the detection signal m1 is canceled, and the voice of the speaker S1 included in the detection signal m2 (s1 · a12) Will be added. Finally, the output signal y1 is a signal including only s1, and a signal including only the voice of the speaker S1. Similarly, the output signal y2 is a signal including only s2, and includes only the voice of the speaker S2. The output signal y3 is a signal including only s3, and includes only the voice of the speaker S3. The output signal y4 is a signal including only s4, and includes only the voice of the speaker S4.

  Here, for example, in the output signal y1, in order to cancel the voices of the speakers S3 and S4 included in the detection signal m1 as echoes, w13 is obtained by multiplying the transfer characteristic h11 from the speaker 10 to the microphone 11 by minus ( -H11). Further, w14 needs to be a value obtained by multiplying the transfer characteristic h21 from the speaker 20 to the microphone 11 by minus (−h21). Therefore, in the state where learning of the separation matrix W (4, 4) has converged, even if the speaker input signals (sp1, sp2) include correlated sounds (s3, s4), the microphones from the speaker 10 and the speaker 20 are used. It can be said that the transfer characteristics h11 and h21 up to 11 are correctly estimated.

  Of the matrix elements constituting the separation matrix W (4, 4), (w11, w12, w21, w22) relate to the transfer characteristics from the speakers S1 and S2 on the near end side to the microphones 11 and 21. Is. (W11, w12, w21, w22) are used to separate the speech of the speaker S1 and the speech of the speaker S2 included in the detection signals (m1, m2) as output signals y1 and y2. Further, (w13, w14, w23, w24) relates to transfer characteristics from the speakers 10 and 20 on the near end side to the microphones 11 and 21. (W13, w14, w23, w24) are used to cancel the voices of the speakers S3 and S4, which are echo components, from the detection signals (m1, m2). Further, (w33, w34, w43, w44) relates to transfer characteristics from the far-end speakers S3 and S4 to the microphones 31 and 41. (W33, w34, w43, w44) are used to separate the speech of the speaker S3 and the speech of the speaker S4 included in the speaker input signals (sp1, sp2) as output signals y3 and y4. (W31, w32, w41, w42) uses the detection signals (m1, m2) to output the voice of the speaker S3 and the voice of the speaker S4 included in the speaker input signals (sp1, sp2) as the output signal y3. And y4 to separate.

  As described above, in the present embodiment, the sound source separation unit 100 cancels the far-end acoustic signal included in the detection signal (m1, m2), so that the near-end acoustic signal included in the detection signal (m1, m2). And the far-end acoustic signal are separated. Then, the sound source separation unit 100 outputs only the separated near-end acoustic signals as output signals y1 and y2. Thereby, the far-end acoustic signal included in the detection signal (m1, m2) can be canceled as an echo regardless of whether or not the correlated sound is included in the speaker input signals (sp1, sp2). As a result, in the present embodiment, it is possible to always perform stable echo cancellation without causing deterioration in sound quality while solving indefiniteness of solutions during multi-channel reproduction.

  Further, in the present embodiment, since a conventional adaptive filter is not used, it is possible to estimate a correct transfer characteristic regardless of double talk or single talk.

  In the above description, the output signals y1 and y2 from the sound source separation unit 100 include the voices of the speakers S1 and S2 as near-end acoustic signals. However, the present invention is not limited to this. The output signals y1 and y2 may include a statistic indicating the characteristics of the voices of the speakers S1 and S2 as the near-end acoustic signal. That is, the near-end acoustic signal may be an acoustic signal composed of statistics indicating characteristics of the voices of the speakers S1 and S2 instead of the voices of the speakers S1 and S2.

  In the above description, the processing at the time of double talk has been described. However, even at the time of single talk (when only the speakers S3 and S4 are talking), echo is canceled by performing the same processing as at the time of double talk. Needless to say. However, during single talk, since the detection signals (m1, m2) do not include the voices of the speakers S1 and S2, the sound source separation unit 100 outputs the output signals y1 and y2 as silence signals. . Actually, the near-end environmental noise that is a sound from a sound source independent of the speakers S3 and S4 is output as a silence signal.

  In the above description, there are two speakers S1 and S2 on the near end side, but the present invention is not limited to this. There may be one speaker on the near end side or three or more speakers.

  First, a case where only one speaker S1 exists will be described as a case where there is only one speaker on the near end side. The sound source separation unit 100 separates the near-end acoustic signal by the number of input microphone detection signals. In the present embodiment, the number of microphone detection signals input to the sound source separation unit 100 is two, m1 and m2. Therefore, in this case, the sound source separation unit 100 outputs the output signal y1 as a signal including only the voice of the speaker S1, and outputs the output signal y2 as a silence signal including only the environmental noise on the near end side. In this case, since the voice of the speaker S1 and the environmental noise are independent from each other, the output signal y1 including only the voice of the speaker S1 and the output signal y2 including only the environmental noise are independent. Further, the output signals y1 and y2 and the output signals y3 and y4 are also independent. Therefore, even in this case, the sound source separation unit 100 can separate the output signals y1 to y4 that are independent from each other from the input signal, and the near-end acoustic signal included in the detection signal (m1, m2). The far-end acoustic signal can be separated.

  Next, as a case where there are three speakers on the near end side, for example, a case where speaker S5 further exists will be described. In this case, for example, if the speaker S5 is present at a position close to the speaker S1, the sound source separation unit 100 outputs the output signal y1 as a signal including only the voices of the speakers S1 and S5, and outputs the output signal y2 as the speaker. It is output as a signal including only the sound of S2. When the speaker S5 exists at a position close to the speaker S1, the transfer characteristics from the speaker S1 to the microphone 11 and the transfer characteristics from the speaker S5 to the microphone 11 are approximated, and the speaker S1 to the microphone 21 is approximated. The transfer characteristic and the transfer characteristic from the speaker S5 to the microphone 21 are approximated. Therefore, the voice of the speaker S5 is included in the output signal y1 including the voice of the speaker S1 whose transfer characteristics are approximate. In this case, since the voices of the speakers S1, S2, and S5 are independent from each other, the output signal y1 that includes only the voices of the speakers S1 and S5 and the output signal y2 that includes only the voices of the speaker S2 Become independent. Further, the output signals y1 and y2 and the output signals y3 and y4 are also independent. Therefore, the sound source separation unit 100 can separate the output signals y1 to y4 that are independent from each other from the input signal, and separates the near-end acoustic signal and the far-end acoustic signal included in the detection signals (m1, m2). can do.

  In the above description, there are two speakers S3 and S4 on the far end side, but the present invention is not limited to this. There may be one speaker on the far end side or three or more speakers.

  First, a case where only one speaker S3 exists will be described as a case where there is one speaker on the far end side. The sound source separation unit 100 separates the far-end acoustic signal by the number of input speaker input signals. In the present embodiment, the number of speaker input signals input to the sound source separation unit 100 is two, sp1 and sp2. Therefore, in this case, the sound source separation unit 100 outputs the output signal y3 as a signal including only the voice of the speaker S3, and outputs the output signal y4 as a silence signal including only environmental noise on the near end side. In this case, since the voice of the speaker S3 and the environmental noise are independent from each other, the output signal y3 including only the voice of the speaker S3 and the output signal y4 including only the environmental noise are independent. Further, the output signals y1 and y2 and the output signals y3 and y4 are also independent. Therefore, even in this case, the sound source separation unit 100 can separate the output signals y1 to y4 that are independent from each other from the input signal, and the near-end acoustic signal included in the detection signal (m1, m2). The far-end acoustic signal can be separated.

  Next, a case where there are three speakers S6 on the far end side, for example, will be described. In this case, for example, if the speaker S6 is present at a position close to the speaker S3, the sound source separation unit 100 outputs the output signal y3 as a signal including only the voices of the speakers S3 and S6, and outputs the output signal y4. It is output as a signal including only the sound of S4. When the speaker S6 is present at a position close to the speaker S3, the transfer characteristic from the speaker S3 to the microphone 31 and the transfer characteristic from the speaker S6 to the microphone 31 are approximated, and the transfer from the speaker S3 to the microphone 41 is approximated. The transfer characteristic and the transfer characteristic from the speaker S6 to the microphone 41 are approximated. For this reason, the voice of the speaker S6 is included in the output signal y3 including the voice of the speaker S3 whose transfer characteristics are approximate. In this case, since the voices of the speakers S3, S4, and S6 are independent from each other, the output signal y3 that includes only the voices of the speakers S3 and S6 and the acoustic signal y4 that includes only the voices of the speaker S4. Become independent. Further, the output signals y1 and y2 and the output signals y3 and y4 are also independent. Therefore, the sound source separation unit 100 can separate the output signals y1 to y4 that are independent from each other from the input signal, and separates the near-end acoustic signal and the far-end acoustic signal included in the detection signals (m1, m2). can do.

  In the acoustic system shown in FIG. 1, as an example, the multi-channel echo canceller according to the present embodiment is provided only on the near end side, but it goes without saying that the multi-channel echo canceller may also be installed on the far end side.

(Second Embodiment)
With reference to FIG. 3, the configuration of a multi-channel echo canceller according to the second embodiment of the present invention will be described. FIG. 3 is a diagram illustrating a configuration example of the multi-channel echo canceller according to the second embodiment used in the acoustic system. In the acoustic system shown in FIG. 3, the speaker S1 exists as a sound source (near-end sound source) on the near end side, and a plurality of sound sources (far end) from which the speakers S3 and S4 are different from each other exist on the far end side. Suppose that it exists as a sound source. On the near-end side, speakers 10 and 20 for amplifying the far-end acoustic signal composed of the speech of the far-end speakers S3 and S4 and the near-end acoustic signal composed of the speech of the near-end speaker S1 are provided. Microphones 11 and 21 for detection are provided. On the far end side, speakers 30 and 40 for amplifying the near-end acoustic signal and microphones 31 and 41 for detecting the far-end acoustic signal are provided. The speakers (10, 20, 30, 40) and the microphones (11, 21, 31, 41) are the same as those described in FIG. 8, and are denoted by the same reference numerals. In the acoustic system shown in FIG. 3, as an example, it is assumed that the multichannel echo canceller according to the present embodiment is provided only on the near end side. In the acoustic system shown in FIG. 3, as an example, the near-end speaker S1 and the far-end speakers S3 and S4 are talking at the same time, so-called double talk is being performed. And

  In FIG. 3, the multi-channel echo canceller according to the present embodiment is configured by an echo cancellation unit 2. The echo cancellation unit 2 includes a first sound source separation unit 210, a second sound source separation unit 220, conversion units 230 to 235, and inverse conversion units 240 and 241.

  In the echo cancellation unit 1 according to the first embodiment described above, one sound source separation unit 100 is provided for the microphones 11 and 21. On the other hand, in the echo cancellation unit 2 according to the present embodiment, the first sound source separation unit 210 and the second sound source separation unit 220 are provided so as to correspond to the microphones 11 and 21, respectively. That is, in the present embodiment, one sound source separation unit is provided for one microphone on the near end side. Note that the conversion units 230 to 235 perform the same operation as the conversion units 110 to 113 of the echo cancellation unit 1, but in FIG. The inverse transform units 240 and 241 perform the same operation as the inverse transform units 120 and 121 of the echo cancel unit 1, but the reference numerals are changed for convenience. Hereinafter, a description will be given focusing on differences from the first embodiment.

  The first sound source separation unit 210 includes a speaker input signal sp2 (ω) converted into the frequency domain (ω) by the conversion unit 230 and a speaker input signal sp1 (ω) converted into the frequency domain (ω) by the conversion unit 231. ) And the detection signal m1 (ω) converted into the frequency domain (ω) by the conversion unit 232 as inputs. The first sound source separation unit 210 performs sound source separation processing based on independent component analysis on the input signal. By this sound source separation process, the near-end acoustic signal and the far-end acoustic signal included in the detection signal m1 (ω) are separated. The sound source separation process based on the independent component analysis is substantially the same as the process of the first embodiment, but will be described in detail later. The first sound source separation unit 210 outputs only the separated near-end acoustic signal as the output signal y1a (ω). Here, the far-end acoustic signal corresponds to an echo. Therefore, by outputting only the near-end acoustic signal from the first sound source separation unit 210, the far-end acoustic signal included in the detection signal m1 (ω) can be canceled as an echo. The output signal y1a (ω) output from the first sound source separation unit 210 is converted into a signal in the time domain (t) by the inverse conversion unit 240. The output signal y1a (t) converted into the time domain (t) is output to the speaker 30 and is amplified by the speaker 30.

  The second sound source separation unit 220 includes the speaker input signal sp2 (ω) converted into the frequency domain (ω) by the conversion unit 233 and the speaker input signal sp1 (ω) converted into the frequency domain (ω) by the conversion unit 234. ) And the detection signal m2 (ω) converted into the frequency domain (ω) by the conversion unit 235. The second sound source separation unit 220 performs sound source separation processing based on independent component analysis on the input signal. By this sound source separation process, the near-end acoustic signal and the far-end acoustic signal included in the detection signal m2 (ω) are separated. The sound source separation process based on the independent component analysis is the same process as the first sound source separation unit 210. The second sound source separation unit 220 outputs only the separated near-end acoustic signal as the output signal y1b (ω). Here, the far-end acoustic signal corresponds to an echo. Therefore, by outputting only the near-end acoustic signal from the second sound source separation unit 220, the far-end acoustic signal included in the detection signal m2 (ω) can be canceled as an echo. The output signal y1b (ω) output from the second sound source separation unit 220 is converted into a signal in the time domain (t) by the inverse conversion unit 241. The output signal y1b (t) converted to the time domain (t) is output to the speaker 40 and is amplified by the speaker 40.

式（８）において、話者Ｓ１の音声をｓ１（ω）、話者Ｓ３の音声をｓ３（ω）、話者Ｓ４の音声をｓ４（ω）とする。また、話者Ｓ１からマイクロホン１１までの伝達特性をａ１１（ω）、話者Ｓ１からマイクロホン２１までの伝達特性をａ１２（ω）、話者Ｓ３からマイクロホン３１までの伝達特性をａ３１（ω）、話者Ｓ３からマイクロホン４１までの伝達特性をａ３２（ω）、話者Ｓ４からマイクロホン３１までの伝達特性をａ４１（ω）、話者Ｓ４からマイクロホン４１までの伝達特性をａ４２（ω）とする。 Hereinafter, a sound source separation process based on independent component analysis performed by the first and second sound source separation units 210 and 220 will be described in detail. Here, the first sound source separation unit 210 will be described as an example. First, the detection signal m1 (ω) and speaker input signals (sp1 (ω), sp2 (ω)) input to the first sound source separation unit 210 will be described in detail. The detection signal m1 (ω) and the speaker input signals (sp1 (ω), sp2 (ω)) are expressed as Expression (8).

In the equation (8), the voice of the speaker S1 is s1 (ω), the voice of the speaker S3 is s3 (ω), and the voice of the speaker S4 is s4 (ω). Further, the transfer characteristic from the speaker S1 to the microphone 11 is a11 (ω), the transfer characteristic from the speaker S1 to the microphone 21 is a12 (ω), the transfer characteristic from the speaker S3 to the microphone 31 is a31 (ω), The transfer characteristic from the speaker S3 to the microphone 41 is a32 (ω), the transfer characteristic from the speaker S4 to the microphone 31 is a41 (ω), and the transfer characteristic from the speaker S4 to the microphone 41 is a42 (ω).

  Next, a detailed configuration of the first sound source separation unit 210 will be described with reference to FIG. FIG. 4 is a diagram illustrating a detailed configuration of the first sound source separation unit 210. In FIG. 4, the first sound source separation unit 210 includes a separation unit 211 and a learning unit 212.

In the separation unit 211, a separation matrix W (3, 3) configured by matrix elements wij (the number of rows i and the number of columns j is an integer of 1 to 3) is set. In the initial state, for example, it is assumed that the unit matrix is set as the separation matrix W (3, 3). The separation unit 211 receives a detection signal m1 (ω) and speaker input signals (sp1 (ω), sp2 (ω)). The separation unit 211 calculates the output signals y1a to y3a according to Equation (9) based on the set separation matrix W (3, 3), and outputs the calculated output signals y1a to y3a, respectively. Specifically, the separation unit 211 is set with an input vector composed of the detection signal m1 (ω) and the speaker input signals (sp1 (ω), sp2 (ω)) as shown in Expression (9). The output vector composed of the output signals y1a (ω) to y3a (ω) is calculated by multiplying the separation matrix W (3, 3).

  The learning unit 212 receives the output signals y1a (ω) to y3a (ω) and learns the separation matrix W (3, 3) according to the independent component analysis. Specifically, the learning unit 212 learns the separation matrix W (3, 3) so that the output signals y1a (ω) to y3a (ω) are independent signals. The learning unit 212 updates the separation matrix W (3, 3) set in the separation unit 211 to the learned separation matrix W (3, 3).

式（１０）において、出力信号ｙ１ａ（ω）〜ｙ３ａ（ω）の要素は、周波数領域の複素信号になっており、分離行列Ｗ（３、３）ｉ、Ｗ（３、３）ｉ―１を構成する各行列要素は、複素数の係数になっている。Ｉは３×３の単位行列を示し、ε｛・｝は時間平均を示し、＊は複素共役信号を示す。φ（・）は非線形関数を示す。非線形関数としては、信号の確率密度関数の対数部分を微分したものに対応したものを用いるのがよく、一般的にはｔａｎｈ（・）を用いる。αは学習速度を制御するためのステップサイズパラメータを示す。ｉは学習回数を示し、右辺のＷ（３、３）ｉを左辺のＷ（３、３）ｉ―１に代入することで学習が行われる。εの括弧内に示される行列は、高次の相関行列である。 Hereinafter, the learning method of the learning unit 212 will be described more specifically. A learning formula generally used for frequency-domain independent component analysis using the gradient method is as shown in Formula (10). Note that the learning formula used for the independent component analysis is not limited to the formula (10), as in the first embodiment, and may be another learning formula.

In Expression (10), the elements of the output signals y1a (ω) to y3a (ω) are complex signals in the frequency domain, and the separation matrices W (3, 3) i, W (3, 3) i−1. Each matrix element that constitutes is a complex coefficient. I represents a 3 × 3 unit matrix, ε {·} represents a time average, and * represents a complex conjugate signal. φ (·) indicates a nonlinear function. As the nonlinear function, it is preferable to use a function corresponding to a derivative of the logarithm part of the probability density function of the signal, and generally tanh (·) is used. α represents a step size parameter for controlling the learning speed. i represents the number of learnings, and learning is performed by substituting W (3,3) i on the right side into W (3,3) i-1 on the left side. The matrix shown in parentheses for ε is a higher-order correlation matrix.

  Here, the speakers S1, S3, and S4 are all different speakers and are independent sound sources. Therefore, it can be said that s1 (ω), s3 (ω), and s4 (ω) in the equation (8) are independent from each other and the speech has no correlation with each other. One detection signal m1 (ω) is input, and the number of detection signals matches the number of near-end speakers S1. Also, two speaker input signals (sp1 (ω), sp2 (ω)) are input, and the number of speaker input signals matches the number of far-end speakers (S3 and S4). Therefore, when the learning unit 212 learns the separation matrix W (3, 3) under these conditions, and the learning converges, the separation matrix W (3, 3) becomes the detection signal m1 (ω) and the speaker input signal ( This is a matrix that can separate s1 (ω), s3 (ω), and s4 (ω) from sp1 (ω) and sp2 (ω)). That is, the output signal y1a calculated by the separation unit 101 based on the separation matrix W (3, 3) in which learning converges includes only s1 (ω) included in the detection signal m1 (ω). Become. Similarly, the output signal y2a includes only s3 (ω) included in the speaker input signals (sp1 (ω), sp2 (ω)), and the output signal y3a includes the speaker input signal (sp1 (ω ), Sp2 (ω)), only s4 (ω) included in sp2 (ω)) is included.

  Thus, the learning unit 212 learns the separation matrix W (3, 3) according to the independent component analysis, so that the separation unit 211 separates the near-end acoustic signal from the detection signal m1 (ω) as the output signal y1a. In addition, the far-end acoustic signal can be separated from the speaker input signals (sp1 (ω), sp2 (ω)) as output signals y2a and y3a. Note that the output signals y2a and y3a, which are far-end acoustic signals, are not output from the first sound source separation unit 210 and are used only for learning by the learning unit 212.

  Note that the second sound source separation unit 220 also performs sound source separation processing similar to that of the first sound source separation unit 210. As a result, the second sound source separation unit 220 outputs an output signal y1b including only s1 (ω) included in the detection signal m2 (ω).

Hereinafter, what kind of signal the output signals y1a to y3a become will be described using Equation (11). Expression (11) is obtained by substituting Expression (8) into Expression (9) and further represents a signal input to the separation unit 211 in more detail. In Expression (11), the description of (ω) shown in Expression (8) is omitted.

  In the state where learning of the separation matrix W (3, 3) has converged, w11 = γ (arbitrary real number), w12 = −h11γ, w13 = −h21γ, and finally the output signal y1a is y1a = s1 · a11. Become. That is, in the output signal y1a, the voices (s3 · a31, s3 · a32, s4 · a41, s4 · a42) of the speakers S3 and S4 included in the detection signal m1 are canceled as echoes. In the state where learning of the separation matrix W (3, 3) has converged, w21 and w31 are w21 = w31 = 0, and the matrix elements (w22, w23, w32, w33) are s3 (ω) and s4 (ω). And transfer characteristics that can be separated from each other. Thereby, the output signal y2a finally becomes a signal including only s3, and the output signal y3a becomes a signal including only s4. In this state where the separation matrix W (3, 3) is converged, even if the speaker input signals (sp1, sp2) include correlated sounds (s3, s4), the speaker 10 and the speaker 20 to the microphone 11 are included. It can be said that the transfer characteristics h11 and h21 are correctly estimated.

  Of the matrix elements constituting the converged separation matrix W (3, 3), (w11) relates to the transfer characteristics from the near-end speaker S1 to the microphone 11. (W11) is used to define the signal level of s1 · a11. (W12, w13) relates to each transfer characteristic from the speakers 10 and 20 on the near end side to the microphone 11. (W12, w13) is used to cancel the voices of the speakers S3 and S4, which are echo components, from the detection signal m1. (W22, w23, w32, w33) relate to respective transfer characteristics from the far-end speakers S3 and S4 to the microphones 31 and 41. (W22, w23, w32, w33) are used to separate the speech of the speaker S3 and the speech of the speaker S4 included in the speaker input signals (sp1, sp2) as output signals y2a and y3a. (W21, w31) are used to separate the speech of the speaker S3 and the speech of the speaker S4 included in the speaker input signals (sp1, sp2) as the output signals y2a and y3a using the detection signal m1. It is done.

  As described above, in the present embodiment, the first sound source separation unit 210 separates the near-end acoustic signal and the far-end acoustic signal included in the detection signal m1, and outputs only the separated near-end acoustic signal as the output signal y1a. Output as. The second sound source separation unit 220 separates the near-end acoustic signal and the far-end acoustic signal included in the detection signal m2, and outputs only the separated near-end acoustic signal as the output signal y1b. Thereby, the far-end acoustic signal included in the detection signal (m1, m2) can be canceled as an echo regardless of whether or not the correlated sound is included in the speaker input signals (sp1, sp2). As a result, in the present embodiment, it is possible to always perform stable echo cancellation without causing deterioration in sound quality while solving indefiniteness of solutions during multi-channel reproduction.

  Further, in the present embodiment, since a conventional adaptive filter is not used, it is possible to estimate a correct transfer characteristic regardless of double talk or single talk.

  In the present embodiment, a first sound source separation unit 210 and a second sound source separation unit 220 are provided so as to correspond to the microphones 11 and 21, respectively. Therefore, the first sound source separation unit 210 outputs an output signal y1a including only the voice S1 of the speaker S1 included in the detection signal m1, and the second sound source separation unit 220 outputs the detection signal m2. The output signal y1b including only the voice s1 of the speaker S1 included in the output is output. Since the output signal y1a includes only the voice s1 of the speaker S1 included in the detection signal m1, the output signal y1a is a signal having a sense of direction from the speaker S1 toward the microphone 11. Similarly, since the output signal y1b includes only the voice s1 of the speaker S1 included in the detection signal m2, the output signal y1b is a signal having a sense of direction from the speaker S1 to the microphone 21. Therefore, when these output signals y1a and y1b are louded by the far-end speakers 30 and 40, the speakers S3 and S4 can feel a sense of direction with respect to the voice of the loudspeaker S1.

  Although the processing at the time of double talk has been described above, the echo is canceled by performing the same processing at the time of single talk (when only the speakers S3 and S4 are talking). Needless to say. However, during single talk, the detection signal m1 does not include the voice of the speaker S1, so the first sound source separation unit 210 outputs the output signal y1a as a silence signal. Actually, the near-end environmental noise that is a sound from a sound source independent of the speakers S3 and S4 is output as a silence signal.

  In the above description, it is assumed that there is one speaker S1 on the near end side, but the present invention is not limited to this. There may be two or more speakers on the near end.

  As an example in which there are two speakers on the near end side, for example, a case where a speaker S2 further exists will be described. The first sound source separation unit 210 and the second sound source separation unit 220 separate the near-end acoustic signals by the number of input microphone detection signals. In the present embodiment, the number of microphone detection signals input to the first sound source separation unit 210 is one of m1, and the number of microphone detection signals input to the second sound source separation unit 220 is One of m2. Therefore, in this case, the first sound source separation unit 210 outputs the output signal y1a including only the voices of the speakers S1 and S2 included in the detection signal m1, and the second sound source separation unit 220 The output signal y1b including only the voices of the speakers S1 and S2 included in the detection signal m2 is output. In this case, since the voices of the speakers S1 to S4 are independent from each other, the output signal y1a including only the voices of the speakers S1 and S2, the output signal y2a including only the voice of the speaker S3, and the speaker S4 And the output signal y3a including only the voice of. Therefore, the first sound source separation unit 210 can separate the output signals y1a to y3a that are independent from each other from the input signal, and separates the near-end acoustic signal and the far-end acoustic signal included in the detection signal m1. be able to. The same applies to the second sound source separation unit 220.

  In the above description, there are two speakers S3 and S4 on the far end side, but the present invention is not limited to this. There may be one speaker on the far end side or three or more speakers. Since this case is the same as that of the first embodiment described above, a description thereof will be omitted.

  In the above description, one conversion unit 230 to 235 is provided for each signal input to the first sound source separation unit 210 and the second sound source separation unit 220. However, some conversion units may be shared as shown in FIG. FIG. 5 is a diagram illustrating a case where some conversion units are shared. In FIG. 5, the conversion unit 233 converts the speaker input signal sp2 into the frequency domain (ω) and outputs the converted signal to the first sound source separation unit 210 and the second sound source separation unit 220. The conversion unit 234 converts the speaker input signal sp1 into the frequency domain (ω) and outputs it to the first sound source separation unit 210 and the second sound source separation unit 220, respectively. As described above, in FIG. 5, the conversion units 233 and 234 are shared with the first sound source separation unit 210 and the second sound source separation unit 220. In this way, by using the conversion units 233 and 234 in common, the processing amount of the entire multichannel echo canceller can be reduced.

  In the acoustic system shown in FIG. 3, as an example, the multi-channel echo canceller according to the present embodiment is provided only on the near end side, but it goes without saying that the multi-channel echo canceller may also be installed on the far end side.

(Third embodiment)
The first sound source separation unit 210 and the second sound source separation unit 220 described above are configured to update all matrix elements constituting the separation matrix. On the other hand, among the matrix elements constituting the separation matrix, some matrix elements may be constrained (some matrix elements are set to 0). Hereinafter, with reference to FIG. 6, a case in which some matrix elements of the first sound source separation unit 210 and the second sound source separation unit 220 are separated will be described as a third embodiment. FIG. 6 is a diagram illustrating a configuration of the first sound source separation unit 210a in which a part of the separation matrix set in the first sound source separation unit 210 is constrained.

式（１２）に示された分離行列Ｗａ（３、３）において、行列要素である（ｗ２１、ｗ３１）、（ｗ２３、ｗ３２）は０となり、拘束されている。拘束型分離部２１１ａから出力された出力信号ｙ１ａ（ω）〜ｙ３ａ（ω）のうち、近端音響信号である出力信号ｙ１ａ（ω）のみ、第１の音源分離部２１０ａから出力される。 In FIG. 6, the first sound source separation unit 210a includes a constraint type separation unit 211a and a constraint type learning unit 212a. In the constraining separation unit 211a, a separation matrix Wa (3, 3) configured by matrix elements wij (the number of rows i and the number of columns j is an integer of 1 to 3) is set. In the initial state, for example, it is assumed that the unit matrix is set as the separation matrix Wa (3, 3). A detection signal m1 (ω) and speaker input signals (sp1 (ω), sp2 (ω)) are input to the constraining separation unit 211a. The constraining separation unit 211a calculates the output signals y1a to y3a according to Expression (12) based on the set separation matrix Wa (3, 3), and outputs the calculated output signals y1a to y3a, respectively. Specifically, as shown in Expression (12), the constraining separation unit 211a includes an input vector composed of a detection signal m1 (ω) and speaker input signals (sp1 (ω), sp2 (ω)); By multiplying the set separation matrix Wa (3, 3), an output vector composed of the output signals y1a (ω) to y3a (ω) is calculated.

In the separation matrix Wa (3, 3) shown in the equation (12), the matrix elements (w21, w31), (w23, w32) are 0 and are constrained. Of the output signals y1a (ω) to y3a (ω) output from the constraining separation unit 211a, only the output signal y1a (ω) that is a near-end acoustic signal is output from the first sound source separation unit 210a.

式（１３）に従えば、拘束型学習部２１２ａは、分離行列Ｗａ（３、３）を構成する各行列要素のうち、非ゼロとなる行列要素のみを更新することとなる。このような分離行列Ｗａ（３、３）の学習によっても、入力された信号に含まれる近端音響信号と遠端音響信号とを分離することができる。 The constraint learning unit 212a receives the output signals y1a (ω) to y3a (ω) as input, and performs independent component analysis to learn the separation matrix Wa (3, 3) set in the constraint separation unit 211a. Specifically, the constraint learning unit 212a learns the separation matrix Wa (3, 3) according to the equation (13). The constraint learning unit 212a updates the separation matrix Wa (3, 3) set in the constraint separation unit 211a to the learned separation matrix Wa (3, 3).

According to Expression (13), the constraint learning unit 212a updates only non-zero matrix elements among the matrix elements constituting the separation matrix Wa (3, 3). The learning of the separation matrix Wa (3, 3) can also separate the near-end acoustic signal and the far-end acoustic signal included in the input signal.

  Hereinafter, the purpose of constraining a part of the matrix elements of the separation matrix and the reason why the near-end acoustic signal and the far-end acoustic signal can be separated even if the part of the matrix element is constrained will be described. First, in the above-described second embodiment, no particular mention has been made of the case where the far-end sound source moves or the number of far-end sound sources is large when the separation matrix has once converged. However, when the far-end sound source moves or when the number of far-end sound sources is large, the actual transfer characteristics (a31 (ω), a32 (ω), etc.) vary. Thereby, in the configuration in which all the coefficients of the separation matrix are updated as in the second embodiment, the matrix elements (w12, w13) used for canceling the echo may temporarily change through learning. . When the matrix elements (w12, w13) are temporarily changed, the separation between the near-end acoustic signal and the far-end acoustic signal is incomplete, and the echo cancellation effect is temporarily deteriorated.

  The matrix elements (w21, w31) use the detection signal m1 to separate the speech of the speaker S3 and the speech of the speaker S4 included in the speaker input signals (sp1, sp2) as output signals y2a and y3a. Is a matrix element used to Here, the output signals y2a and y3a, which are far-end acoustic signals, are not output from the first sound source separation unit 210 but are used only for learning by the learning unit 212. For this reason, the detection signal m1 does not include a signal that contributes to separation of the output signals y2a and y3a, which are far-end acoustic signals. Therefore, in the state where learning of the separation matrix has converged, w21 = w31 = 0 should be obtained. However, in the configuration in which all matrix elements of the separation matrix are updated as in the second embodiment, when the far-end sound source moves or the actual transfer characteristics (a31 (ω), a32 (ω), etc.) vary. The matrix elements (w21, w31) also temporarily change through learning. When the matrix elements (w21, w31) are temporarily changed, the matrix element (w11) is also temporarily changed through the next learning. As a result, the separation between the near-end acoustic signal and the far-end acoustic signal becomes incomplete, and the echo cancellation effect temporarily deteriorates.

  In order to prevent such a temporary deterioration of the echo cancellation effect, in this embodiment, some matrix elements of the separation matrix are constrained.

Hereinafter, the phenomenon that the echo canceling effect is temporarily deteriorated depending on the state of the far-end sound source will be described using mathematical formulas. Expression (14) is an expansion of the first term on the right side of the update expression of Expression (10).

The first term on the right side of Equation (14) indicates the update amount ΔW in learning of the separation matrix W. The matrix element ε {φ (y _i ) y _j *} _i ≠ _j becomes ε {φ (y _i ) y _j *} _i ≠ _j ≈ 0 if the output signals y _{i and} y _j become independent from each other. In the state where learning of the separation matrix W has converged, the update amount ΔW oscillates in the vicinity of 0 (zero). That is, all matrix elements of the update amount ΔW are zero.

Here, let us consider a case where the far-end transfer characteristics (a31 (ω), a32 (ω), etc.) have changed from the state where learning of the separation matrix W has converged and echo cancellation has been performed satisfactorily. In this case, the estimated values of the matrix elements (w22, w23, w32, w33) of the converged separation matrix W do not match the actual transfer characteristics. Thereby, the separation of the output signals y2a and y3a relating to the far-end acoustic signal becomes incomplete. In other words, the independence between the output signals y2a and y3a is reduced, and they are correlated with each other. In the expression (14), ε {φ (y _2a ) y _3a *} and ε {φ (y _3a ) y _2a *} have values. In particular, when the number of far-end sound sources is large, the transfer characteristics on the far-end side always change, so ε {φ (y _2a ) y _3a *} and ε {φ (y _3a ) y _2a *} always have values. Will have. It should be noted that ε {φ (y _2a ) y _3a *}, ε {φ (y _3a ) y _2a *} are included in the matrix elements in the second and third rows of the first term on the right side of Equation (14). It is included. For this reason, the fact that ε {φ (y _2a ) y _3a *} and ε {φ (y _3a ) y _2a *} fluctuate means that the second and third rows of the first term on the right side of equation (14) It means that each matrix element of the eye fluctuates.

  When the matrix elements in the second and third rows of the first term on the right side of Equation (14) change, the matrix elements in the second and third rows of the separation matrix W learned based on the change ( w21-w23, w31-w33) also vary. Among these, when the next learning is performed based on the change of the matrix elements (w23, w32), the matrix elements (w12, w13) of the separation matrix W will change. In addition, when the next learning is performed based on the change of the matrix elements (w21, w31), the matrix element (w11) of the separation matrix W changes. Due to such a variation of each matrix element in the first row of the separation matrix W, the echo cancellation effect is temporarily deteriorated.

  Therefore, in this embodiment, the matrix elements (w32, w23) and the matrix elements (w21, w31) of the separation matrix are constrained to 0, respectively. As a result, even if the transfer characteristic on the far end side fluctuates, it is possible to prevent fluctuation of each matrix element in the first row of the separation matrix W through learning, and to prevent temporary deterioration of the echo cancellation effect. Can do.

式（１５）の右辺の第１項を一旦展開して再度整理すると、式（１３）と同じ式になる。つまり、式（１３）は、分離行列の一部の行列要素を拘束することによって得られる学習式であるといえる。 Next, let us consider a learning formula when the matrix elements (w32, w23) and the matrix elements (w21, w31) are constrained to 0 as in the separation matrix Wa. If the learning formula of Formula (10) is simply applied to the separation matrix Wa of Formula (12), Formula (15) is obtained.

If the first term on the right side of Expression (15) is once expanded and rearranged, the same expression as Expression (13) is obtained. That is, equation (13) can be said to be a learning equation obtained by constraining some matrix elements of the separation matrix.

In Expression (13), when the learning of the separation matrix Wa converges, the update amount ΔW becomes 0 matrix. The update amount ΔW becomes a zero matrix means that {φ (y _1a ) y _2a *} = {φ (y _2a ) y _1a *} = {φ (y _1a ) y _3a *} = {φ (y _3a ) Y _1a *} = 0 and 1−ε {φ (y _1a ) y _1a *} = 1−ε {φ (y _2a ) y _2a *} = 1−ε {φ (y _3a ) y _3a * } = 0. Therefore, when the learning of the separation matrix Wa converges, the output signal y1a that is the near-end acoustic signal and the output signal y1a that is the near-end acoustic signal and the output signal y1a that is the near-end acoustic signal and the far-end acoustic signal It can be seen that the output signal y3a which is a signal becomes independent. That is, it can be seen that when the learning based on Expression (13) converges, the near-end acoustic signal and the far-end acoustic signal are separated.

Note that the values of {φ (y _2a ) y _3a *} and {φ (y _3a ) y _2a *} are irrelevant for the update amount ΔW to be a zero matrix. For this reason, at the time of convergence of learning, the output signal y2a and the output signal y3a, which are far-end acoustic signals, may or may not be independent. That is, for the far-end acoustic signal, the output signal y2a and the output signal y3a that are independent from each other are not necessarily output.

  As described above, by constraining some matrix elements of the separation matrix, even if the transfer characteristic on the far end side varies, the variation of each matrix element in the first row of the separation matrix W through learning is prevented. And temporary deterioration of the echo canceling effect can be prevented. In addition, when a part of the matrix elements of the separation matrix is constrained, the amount of calculation can be reduced as compared with the case where it is not constrained.

In addition, although Formula (13) was used as a learning formula in this embodiment, Formula (16) may be used. Even using Equation (16), the near-end acoustic signal and the far-end acoustic signal can be separated.

  In the present embodiment, the separation matrix set in the first sound source separation unit 210 and the second sound source separation unit 220 described above has been described. However, the present invention is not limited to this. Even if a part of the coefficients of the separation matrix W (4, 4) set in the sound source separation unit 100 described above is constrained, the same effect as in the present embodiment can be obtained. Hereinafter, common constraint conditions for separation matrices having different numbers of rows and columns are shown in Equations (17) to (19). Note that the number of rows and the number of columns of the separation matrix are each (M + K). M indicates the number of near-end microphones that input detection signals to the sound source separation unit, and K indicates the number of near-end speakers that input speaker input signals to the sound source separation unit. Further, the input vector multiplied by the separation matrix is such that the speaker input signal is from the (M + 1) th row to the (M + K) th row of the separation matrix so that the detection signal of the near-end microphone corresponds to the first row to the Mth row of the separation matrix. Suppose that it is configured to respond to the eyes. In the equations (17) to (19), i (i = 1 to M + K) represents the number of rows, and j (j = 1 to M + K) represents the number of columns.

式（１７）に示す拘束条件は、遠端音響信号について互いに独立な複数の出力信号にさらに分離させる必要がない場合に有効である。この場合であっても、近端音響信号と遠端音響信号とを分離することはできる。 Expression (17) is a matrix element (w22, w23, w32, w33 in the 3 × 3 matrix) regarding each transfer characteristic from the far-end speaker (S3 and S4) to the far-end microphone (31 and 41). It is a formula which shows the constraint conditions about.

The constraint condition shown in Expression (17) is effective when it is not necessary to further separate the far-end acoustic signal into a plurality of output signals independent of each other. Even in this case, the near-end acoustic signal and the far-end acoustic signal can be separated.

式（１８）に示す拘束条件は、システム処理上、音源分離処理とは関係のない行列要素に関する条件であるため、式（１８）に示す拘束条件を適用しても、近端音響信号と遠端音響信号とを分離することはできる。 Equation (18) uses the near-end microphone detection signals (m1, m2, etc.) to calculate the far-end speaker's speech (S3 and S4) contained in the speaker input signals (sp1, sp2). It is an expression showing the constraint conditions for the matrix elements (w21, w31 in the 3 × 3 matrix) used to separate each output signal (y3, y4, etc.).

Since the constraint condition shown in Expression (18) is a condition related to matrix elements that are not related to the sound source separation process in system processing, even if the constraint condition shown in Expression (18) is applied, The edge acoustic signal can be separated.

式（１９）に示す拘束条件は、近端音響信号について互いに独立な複数の出力信号にさらに分離させる必要がない場合に有効である。この場合であっても、近端音響信号と遠端音響信号とを分離することはできる。 Equation (19) expresses the constraint on the matrix elements (w11 in the 3 × 3 matrix) for each transfer characteristic from the near-end speakers (such as S1 and S2) to the near-end microphones (11 and 21). It is a formula which shows.

The constraint condition shown in Expression (19) is effective when it is not necessary to further separate the near-end acoustic signal into a plurality of output signals independent from each other. Even in this case, the near-end acoustic signal and the far-end acoustic signal can be separated.

(Fourth embodiment)
In the fourth embodiment, a case will be described in which the sound source separation processing in the sound source separation unit 100 described above is realized on a computer system using a computer program. The computer system includes a microprocessor, ROM, RAM, and the like. A computer program is stored in the RAM. The sound source separation process of the sound source separation unit 100 is realized by the microprocessor operating according to the computer program. Note that the computer program is configured by combining a plurality of instruction codes indicating commands to the computer system in order to realize the sound source separation processing of the sound source separation unit 100. Further, the sound source separation processing of the first sound source separation unit 210 and the second sound source separation unit 220 may be realized on a computer system using a computer program.

A program processing flow for realizing the sound source separation processing of the sound source separation unit 100 will be described with reference to FIG. FIG. 7 is a diagram showing a program processing flow for realizing the sound source separation processing of the sound source separation unit 100. In FIG. 7, for example, a unit matrix is set as an initial matrix in the separation unit 101 of the sound source separation unit 100 (step S1). After step S1, the separation unit 101 inputs the detection signals (m1 (ω), m2 (ω)) and the speaker input signals (sp1 (ω), sp2 (ω)) as input vectors (step S2). . Following step S2, the separation unit 101 calculates output signals y1 to y4, which are output vectors, according to Expression (5) based on the currently set separation matrix W (step S3). After step S3, the learning unit 102 learns the separation matrix W according to the equation (6) based on the output signals y1 to y4 calculated in step S3 (step S4). Specifically, the learning unit 102 calculates a high-order correlation matrix by calculating high-order correlations (for example, {φ (y ₃ ) y ₂ *}, etc.) between the output signals y1 to y4. To do. Then, the learning unit 102 learns the separation matrix to be updated using the calculated higher-order correlation matrix. After step S4, the learning unit 102 updates the separation matrix W currently set in the separation unit 101 to the separation matrix W learned in step S4 (step S5). After step S5, the learning unit 102 determines whether or not the update has been performed N (an integer equal to or greater than 1) times (step S6). If the update has not been performed N times (No in step S6), the process returns to step S2. When the update has been performed N times (Yes in step S6), the separation unit 101 calculates the output signals y1 to y4, which are output vectors, according to Expression (5) based on the updated separation matrix W (step S7). ). By this step S7, the near-end acoustic signal and the far-end acoustic signal included in the detection signals (m1 (ω), m2 (ω)) are separated. After step S7, the separation unit 101 outputs only the output signals y1 and y2 that are near-end acoustic signals (step S8). By this step S8, the far-end acoustic signal included in the detection signals (m1 (ω), m2 (ω)) is canceled as an echo.

  By performing the program processing shown in FIG. 7, the sound source separation process in the sound source separation unit 100 described above can be realized on a computer system.

  Note that the first sound source separation unit 210a and the like in which the matrix elements are partly constrained can also be realized on a computer system using a computer program. The program processing in this case differs from the program processing shown in FIG. 7 in that the predetermined matrix elements shown in the equations (17) to (19) are processed while being restricted to 0. That is, in steps S4 and S5, the learning unit 102 processes only matrix elements other than predetermined matrix elements, and performs processing while restricting predetermined matrix elements to 0.

(Other variations)
Although the multichannel echo canceller according to the present invention has been described in the first to third embodiments described above, the echo canceller according to the present invention is not limited to the contents described in the first to third embodiments. . The multi-channel echo canceller according to the present invention may have the following form.

  (1) Some or all of the constituent elements of the multichannel echo canceller according to the first to third embodiments described above may be configured by one system LSI (Large Scale Integration). Good. The system LSI is a super multifunctional LSI that can be manufactured by integrating a plurality of components on a single chip. In the system LSI, for example, a computer system including a microprocessor, a ROM, a RAM, and the like can be realized. A computer program is stored in the RAM. As the microprocessor operates in accordance with the computer program, the system LSI realizes a function as a computer system.

  (2) Part or all of the components constituting the multichannel echo canceller according to the first to third embodiments described above are configured by an IC card that can be attached to and detached from the multichannel echo canceller, or a single module. Also good. Note that the IC card or module can also realize a computer system including a microprocessor, ROM, RAM, and the like. A computer program is stored in the RAM. The IC card or the module realizes a function as a computer system by the microprocessor operating according to the computer program. Further, the IC card or the module may include the super multifunctional LSI of the above (1). Further, the IC card or the module may have tamper resistance.

  (3) The present invention may be a multi-channel echo cancellation method based on the first to third embodiments described above. The present invention may be a computer program for realizing the multi-channel echo cancellation method on a computer, or may be a digital signal composed of the computer program. The present invention also provides a computer-readable recording medium (for example, a flexible disk, a hard disk, a CD-ROM, an MO, a DVD, a DVD-ROM, a DVD-RAM, a BD (Blu-ray Disc). ), A semiconductor memory, etc.). In the present invention, the computer program or the digital signal is transmitted via an electric communication line (wireless communication line, wired communication line, network line represented by the Internet, data broadcasting line, etc.). Also good. The present invention is realized on a computer system including a microprocessor and a memory, and may be realized by the microprocessor operating in accordance with a computer program stored in the memory. Further, the present invention may be realized on another independent computer system by recording the computer program or the digital signal on a recording medium and transferring it (or by transferring it via a network or the like). Good.

  (4) You may combine suitably the 1st-3rd embodiment mentioned above and the modification of (1)-(3) mentioned above.

  The multi-channel echo canceller according to the present invention can always perform stable echo cancellation without deterioration of sound quality during multi-channel playback, and can perform stable echo cancellation regardless of double talk or single talk. In addition to a conference system and a hands-free phone, it is applied to a voice recognition system at the time of guide announcement playback and music playback.

The figure which shows the structural example of the multichannel echo canceller which concerns on 1st Embodiment used for an acoustic system. The figure which shows the detailed structure of the sound source separation part 100. The figure which shows the structural example of the multichannel echo canceller which concerns on 2nd Embodiment used for an acoustic system. The figure which shows the detailed structure of the 1st sound source separation part 210. Figure showing a case where some conversion units are shared The figure which shows the structure of the 1st sound source separation part 210a which restrained a part of separation matrix set to the 1st sound source separation part 210. FIG. The figure which shows the program processing flow which implement | achieves the sound source separation process of the sound source separation part 100 The figure which shows the structure of the multichannel echo canceller 9 using the conventional adaptive filter.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1, 2, Echo cancellation part 10, 20, 30, 40 Speaker 11, 21, 31, 41 Microphone 100 Sound source separation part 101, 211, 211a Separation part 102, 212, 212a Learning part 110-113, 230-235 Conversion part 120 , 121, 240, 241 Inverse conversion unit 210, 210a First sound source separation unit 220 Second sound source separation unit

Claims (13)

An acoustic signal is transmitted between the first location and the second location, and the first acoustic signal emitted from one or more sound sources existing at the first location is provided at the first location. It is detected by a plurality of microphones for a second acoustic signal generated from one or more sound sources present in the second location is detected by a plurality of microphones provided in the second location, the first location acoustic signal transmitted to the second location from a loudspeaker of a plurality of speakers provided in the second location, an acoustic signal transmitted to said first location from the second location the first used in acoustic systems loudspeaker from a plurality of speakers provided in the first place, a multi-channel echo canceller that provided in the first location,
Detecting an acoustic signal in which a speaker input signal transmitted from the second location to the first location is amplified by a plurality of speakers provided in the first location, and the first acoustic signal ; An acquisition unit for acquiring detection signals of a plurality of microphones provided in the first location and the speaker input signal;
A separation matrix for separating a first acoustic signal included in each of the detection signals and a second acoustic signal included in the amplified acoustic signal included in each of the detection signals, A first submatrix composed of a plurality of matrix elements relating to respective transfer characteristics from a plurality of speakers provided at one place to a plurality of microphones provided at the first place;
A second sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the second location to a plurality of microphones provided at the second location;
A third sub-matrix composed of a plurality of matrix elements for separating the second acoustic signal included in the amplified acoustic signal included in the detection signal from each of the detection signals;
A fourth sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the first location to a plurality of microphones provided at the first location;
For a separation matrix containing
The first acoustic signal and the second acoustic signal included in each detection signal are separated by multiplying the separation matrix by an input vector constituted by each detection signal and each speaker input signal. An echo canceling unit for outputting the separated first acoustic signal;
With
The separation matrix is
The multi-channel echo canceller, wherein the second sub-matrix is learned according to independent component analysis as a diagonal matrix in which components other than the diagonal are zero . An acoustic signal is transmitted between the first location and the second location, and the first acoustic signal emitted from one or more sound sources existing at the first location is provided at the first location. The second acoustic signal detected by the plurality of microphones and emitted from one or more sound sources existing at the second location is detected by the plurality of microphones provided at the second location, and the first location The acoustic signal transmitted from the second location to the second location is amplified from a plurality of speakers provided at the second location, and the acoustic signal transmitted from the second location to the first location is the first location. A multi-channel echo canceller provided in the first location, which is used in an acoustic system that is loudspeaked from a plurality of speakers provided in the location;
Detecting an acoustic signal in which a speaker input signal transmitted from the second location to the first location is amplified by a plurality of speakers provided in the first location, and the first acoustic signal; A plurality of microphones provided in the first location;
An acquisition unit for acquiring detection signals detected by a plurality of microphones provided in the first place, and the speaker input signal;
A separation matrix for separating a first acoustic signal included in each of the detection signals and a second acoustic signal included in the amplified acoustic signal included in each of the detection signals,
A first submatrix composed of a plurality of matrix elements relating to respective transfer characteristics from a plurality of speakers provided at the first location to a plurality of microphones provided at the first location;
A second sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the second location to a plurality of microphones provided at the second location;
A third sub-matrix composed of a plurality of matrix elements for separating the second acoustic signal included in the amplified acoustic signal included in the detection signal from each of the detection signals;
A fourth sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the first location to a plurality of microphones provided at the first location;
For a separation matrix containing
The first acoustic signal and the second acoustic signal included in each detection signal are separated by multiplying the separation matrix by an input vector constituted by each detection signal and each speaker input signal. An echo canceling unit;
An output unit for outputting the separated first acoustic signal;
With
The multi-channel echo canceller is characterized in that the separation matrix is learned according to an independent component analysis, with the second sub-matrix being a diagonal matrix whose components other than the diagonal are zero. An acoustic signal is transmitted between the first location and the second location, and the first acoustic signal emitted from one or more sound sources existing at the first location is provided at the first location. The second acoustic signal detected by the plurality of microphones and emitted from one or more sound sources existing at the second location is detected by the plurality of microphones provided at the second location, and the first location The acoustic signal transmitted from the second location to the second location is amplified from a plurality of speakers provided at the second location, and the acoustic signal transmitted from the second location to the first location is the first location. A multi-channel echo canceller provided in the first location, which is used in an acoustic system that is loudspeaked from a plurality of speakers provided in the location;
A plurality of speakers provided at the first location for amplifying speaker input signals transmitted from the second location to the first location;
A plurality of microphones provided at the first location for detecting an acoustic signal amplified by a plurality of speakers provided at the first location and the first acoustic signal;
An acquisition unit for acquiring detection signals detected by a plurality of microphones provided in the first place, and the speaker input signal;
A separation matrix for separating a first acoustic signal included in each of the detection signals and a second acoustic signal included in the amplified acoustic signal included in each of the detection signals,
A first submatrix composed of a plurality of matrix elements relating to respective transfer characteristics from a plurality of speakers provided at the first location to a plurality of microphones provided at the first location;
A second sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the second location to a plurality of microphones provided at the second location;
A third sub-matrix composed of a plurality of matrix elements for separating the second acoustic signal included in the amplified acoustic signal included in the detection signal from each of the detection signals;
A fourth sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the first location to a plurality of microphones provided at the first location;
For a separation matrix containing
The first acoustic signal and the second acoustic signal included in each detection signal are separated by multiplying the separation matrix by an input vector constituted by each detection signal and each speaker input signal. An echo canceling unit;
An output unit for outputting the separated first acoustic signal;
With
The multi-channel echo canceller is characterized in that the separation matrix is learned according to an independent component analysis, with the second sub-matrix being a diagonal matrix whose components other than the diagonal are zero. The separation matrix is
The multi-channel echo according to any one of claims 1 to 3, wherein the third sub-matrix is learned according to independent component analysis as a zero matrix in which each component is all zero. Canceller. The separation matrix is
The said 4th submatrix is learned according to an independent component analysis as a diagonal matrix whose components other than a diagonal are 0 , The any one of Claim 1- Claim 4 characterized by the above-mentioned. Multi channel echo canceller. An acoustic signal is transmitted between the first location and the second location, and the first acoustic signal emitted from one or more sound sources existing at the first location is provided at the first location. The second acoustic signal detected by the two microphones detected and emitted from one or more sound sources existing at the second location is detected by the two microphones provided at the second location, and the first location The acoustic signal transmitted from the second location to the second location is amplified by two speakers provided at the second location, and the acoustic signal transmitted from the second location to the first location is the first location. A multi-channel echo canceller provided in the first location, which is used in an acoustic system that is loudspeaked from a plurality of speakers provided in the location;
An acoustic signal in which two speaker input signals sp1 and sp2 transmitted from the second location to the first location are amplified by two speakers provided in the first location, and the first acoustic signal Detecting signals m1, m2 of two microphones provided in the first place,
The speaker input signal and
An acquisition unit for acquiring
First acoustic signals y1 and y2 included in the detection signals of the two microphones, and second acoustic signals y3 and y4 included in the amplified acoustic signal included in the detection signals of the two microphones. A separation matrix W for separation,

A separation matrix W represented by:
An input vector I composed of detection signals of the two microphones and the two speaker input signals,

And the first acoustic signals y1 and y2 and the second acoustic signals y3 and y4 included in the detection signals of the two microphones.

And an echo canceling unit that outputs the separated first acoustic signals y1 and y2 by a matrix operation represented by
With
The separation matrix is learned according to independent component analysis with W34 = W43 = 0.
Multi-channel echo canceller characterized by that. The multi-channel echo canceller according to claim 6, wherein the separation matrix is learned according to independent component analysis as W31 = W32 = W41 = W42 = 0. The multi-channel echo canceller according to claim 6 or 7, wherein the separation matrix is learned according to independent component analysis with W12 = W21 = 0. An acoustic signal is transmitted between the first location and the second location, and the first acoustic signal emitted from one or more sound sources existing at the first location is provided at the first location. The second acoustic signal detected by the plurality of microphones and emitted from one or more sound sources existing at the second location is detected by the plurality of microphones provided at the second location, and the first location The acoustic signal transmitted from the second location to the second location is amplified from a plurality of speakers provided at the second location, and the acoustic signal transmitted from the second location to the first location is the first location. A multi-channel echo canceller provided in the first location, which is used in an acoustic system that is loudspeaked from a plurality of speakers provided in the location;
Detecting an acoustic signal in which a speaker input signal transmitted from the second location to the first location is amplified by a plurality of speakers provided in the first location, and the first acoustic signal; An acquisition unit for acquiring detection signals of a plurality of microphones provided in the first location and the speaker input signal;
A separation matrix for separating a first acoustic signal included in each of the detection signals and a second acoustic signal included in the amplified acoustic signal included in each of the detection signals, A first submatrix composed of a plurality of matrix elements relating to respective transfer characteristics from a plurality of speakers provided at one place to a plurality of microphones provided at the first place;
A second sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the second location to a plurality of microphones provided at the second location;
A third sub-matrix composed of a plurality of matrix elements for separating the second acoustic signal included in the amplified acoustic signal included in the detection signal from each of the detection signals;
A fourth sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the first location to a plurality of microphones provided at the first location;
For a separation matrix containing
The first acoustic signal and the second acoustic signal included in each detection signal are separated by multiplying the separation matrix by an input vector constituted by each detection signal and each speaker input signal. An echo canceling unit for outputting the separated first acoustic signal;
With
The separation matrix is
A multi-channel echo canceller, wherein the third sub-matrix is learned in accordance with independent component analysis as a zero matrix in which each component is all zero. An acoustic signal is transmitted between the first location and the second location, and the first acoustic signal emitted from one or more sound sources existing at the first location is provided at the first location. The second acoustic signal detected by the plurality of microphones and emitted from one or more sound sources existing at the second location is detected by the plurality of microphones provided at the second location, and the first location The acoustic signal transmitted from the second location to the second location is amplified from a plurality of speakers provided at the second location, and the acoustic signal transmitted from the second location to the first location is the first location. A multi-channel echo canceller provided in the first location, which is used in an acoustic system that is loudspeaked from a plurality of speakers provided in the location;
Detecting an acoustic signal in which a speaker input signal transmitted from the second location to the first location is amplified by a plurality of speakers provided in the first location, and the first acoustic signal; An acquisition unit for acquiring detection signals of a plurality of microphones provided in the first location and the speaker input signal;
A separation matrix for separating a first acoustic signal included in each of the detection signals and a second acoustic signal included in the amplified acoustic signal included in each of the detection signals, A first submatrix composed of a plurality of matrix elements relating to respective transfer characteristics from a plurality of speakers provided at one place to a plurality of microphones provided at the first place;
A second sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the second location to a plurality of microphones provided at the second location;
A third sub-matrix composed of a plurality of matrix elements for separating the second acoustic signal included in the amplified acoustic signal included in the detection signal from each of the detection signals;
A fourth sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the first location to a plurality of microphones provided at the first location;
For a separation matrix containing
The first acoustic signal and the second acoustic signal included in each detection signal are separated by multiplying the separation matrix by an input vector constituted by each detection signal and each speaker input signal. An echo canceling unit for outputting the separated first acoustic signal;
With
The separation matrix is
The multi-channel echo canceller, wherein the fourth sub-matrix is learned according to independent component analysis as a diagonal matrix in which components other than the diagonal are zero. An acoustic signal is transmitted between the first location and the second location, and the first acoustic signal emitted from one or more sound sources existing at the first location is provided at the first location. The second acoustic signal detected by the plurality of microphones and emitted from one or more sound sources existing at the second location is detected by the plurality of microphones provided at the second location, and the first location The acoustic signal transmitted from the second location to the second location is amplified from a plurality of speakers provided at the second location, and the acoustic signal transmitted from the second location to the first location is the first location. A multi-channel echo cancellation method for performing multi -channel echo cancellation for a first location, used in an acoustic system that is loudspeaked from a plurality of speakers provided at a location ,
Detecting an acoustic signal in which a speaker input signal transmitted from the second location to the first location is amplified by a plurality of speakers provided in the first location, and the first acoustic signal; An acquisition step of acquiring detection signals of a plurality of microphones provided in the first location and the speaker input signal;
A separation matrix for separating a first acoustic signal included in each of the detection signals and a second acoustic signal included in the amplified acoustic signal included in each of the detection signals, A first submatrix composed of a plurality of matrix elements relating to respective transfer characteristics from a plurality of speakers provided at one place to a plurality of microphones provided at the first place;
A second sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the second location to a plurality of microphones provided at the second location;
A third sub-matrix composed of a plurality of matrix elements for separating the second acoustic signal included in the amplified acoustic signal included in the detection signal from each of the detection signals;
A fourth sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the first location to a plurality of microphones provided at the first location;
For a separation matrix containing
The first acoustic signal and the second acoustic signal included in each detection signal are separated by multiplying the separation matrix by an input vector constituted by each detection signal and each speaker input signal. Steps,
Canceling the second acoustic signal included in each of the detection signals as an echo by outputting the separated first acoustic signal; and
A learning step of learning the second sub-matrix of the separation matrix as a diagonal matrix whose components other than the diagonal are 0 according to independent component analysis . An acoustic signal is transmitted between the first location and the second location, and the first acoustic signal emitted from one or more sound sources existing at the first location is provided at the first location. The second acoustic signal detected by the plurality of microphones and emitted from one or more sound sources existing at the second location is detected by the plurality of microphones provided at the second location, and the first location The acoustic signal transmitted from the second location to the second location is amplified from a plurality of speakers provided at the second location, and the acoustic signal transmitted from the second location to the first location is the first location. A program for causing a computer to perform multi-channel echo cancellation for a first location, used in an acoustic system that is loudspeaked from a plurality of speakers provided at a location ,
In the computer,
Detecting an acoustic signal in which a speaker input signal transmitted from the second location to the first location is amplified by a plurality of speakers provided in the first location, and the first acoustic signal; An acquisition step of acquiring detection signals of a plurality of microphones provided in the first location and the speaker input signal;
A separation matrix for separating a first acoustic signal included in each of the detection signals and a second acoustic signal included in the amplified acoustic signal included in each of the detection signals, A first submatrix composed of a plurality of matrix elements relating to respective transfer characteristics from a plurality of speakers provided at one place to a plurality of microphones provided at the first place;
A second sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the second location to a plurality of microphones provided at the second location;
A third sub-matrix composed of a plurality of matrix elements for separating the second acoustic signal included in the amplified acoustic signal included in the detection signal from each of the detection signals;
A fourth sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the first location to a plurality of microphones provided at the first location;
For a separation matrix containing
The first acoustic signal and the second acoustic signal included in each detection signal are separated by multiplying the separation matrix by an input vector constituted by each detection signal and each speaker input signal. Steps,
Canceling the second acoustic signal included in each of the detection signals as an echo by outputting the separated first acoustic signal; and
A program for executing a learning step of learning the second sub-matrix of the separation matrix as a diagonal matrix in which components other than the diagonal are 0 according to independent component analysis . An acoustic signal is transmitted between the first location and the second location, and the first acoustic signal emitted from one or more sound sources existing at the first location is provided at the first location. The second acoustic signal detected by the plurality of microphones and emitted from one or more sound sources existing at the second location is detected by the plurality of microphones provided at the second location, and the first location The acoustic signal transmitted from the second location to the second location is amplified from a plurality of speakers provided at the second location, and the acoustic signal transmitted from the second location to the first location is the first location. An integrated circuit that performs multi-channel echo cancellation for a first location, used in an acoustic system that is loudspeaked from a plurality of speakers provided at the location ,
Detecting an acoustic signal in which a speaker input signal transmitted from the second location to the first location is amplified by a plurality of speakers provided in the first location, and the first acoustic signal; An acquisition unit for acquiring detection signals of a plurality of microphones provided in the first location and the speaker input signal;
A separation matrix for separating a first acoustic signal included in each of the detection signals and a second acoustic signal included in the amplified acoustic signal included in each of the detection signals, A first submatrix composed of a plurality of matrix elements relating to respective transfer characteristics from a plurality of speakers provided at one place to a plurality of microphones provided at the first place;
A second sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the second location to a plurality of microphones provided at the second location;
A third sub-matrix composed of a plurality of matrix elements for separating the second acoustic signal included in the amplified acoustic signal included in the detection signal from each of the detection signals;
A fourth sub-matrix composed of a plurality of matrix elements relating to each transfer characteristic from one or more sound sources present at the first location to a plurality of microphones provided at the first location;
For a separation matrix containing
The first acoustic signal and the second acoustic signal included in each detection signal are separated by multiplying the separation matrix by an input vector constituted by each detection signal and each speaker input signal. An echo canceling unit for outputting the separated first acoustic signal;
With
The separation matrix is
An integrated circuit in which the second sub-matrix is learned according to independent component analysis as a diagonal matrix in which components other than the diagonal are zero .

Images