`J. Aeoust. Soc. Jpn. (E) 10, 1 (1989)
`
`echo paths
`echo canceller with multiple
`Acoustic
`Acoustic echo canceller with multiple echo paths
`
`Nobuo Koizumi,* Shoji Makino,* and Hiroshi Oikawa**
`Nobuo Koizumi,* Shoji Makino,* and Hiroshi Oikawa**
`*NTT Human Interface Laboratories,
`* NTT Human Interface Laboratories,
`3-9-11, Midori-cho, Musashino, Tokyo, 180 Japan
`3-9-Il, Midori-eho, Musashino, Tokyo, ]80 Japan
`**NTT Customer Equipment Division,
`** NTT Customer Equipment Division,
`Yamato Seimei Building, 1-1-7, Uchisaiwai-cho, Chiyoda-ku, Tokyo, 100 Japan
`Yamato Seimei Building, ]-]-7, Uchisaiwai-cho, Chiyoda-ku, Tokyo, ]00 Japan
`(Received 31 May 1988)
`(Received 3] May ]988)
`
`A new configuration of acoustic echo canceller for multiple microphone teleconferencing
`A new configuration of acoustic echo canceller for multiple microphone teleconferencing
`systems is proposed. It is designed for use with microphones whose gains switch or
`It is designed for use with microphones whose gains switch or
`systems is proposed.
`vary during teleconferencing according to the talker. This system requires memory
`vary during teleconferencing according to the talker. This system requires memory
`for multiple echo paths, which enables the updating of filter coefficients when an echo
`for multiple echo paths, which enables the updating of filter coefficients when an echo
`path is changed due to the switching of the actuated microphone during talker alterna-
`path is changed due to the switching of the actuated microphone during talker alterna(cid:173)
`tion. In comparison to the single echo path model which uses only adaptation, this
`In comparison to the single echo path model which uses only adaptation, this
`tion.
`method maintains echo cancellation during abrupt changes of the echo path when the
`method maintains echo cancellation during abrupt changes of the echo path when the
`microphone alternates between talkers. Also in comparison to direct microphone
`microphone alternates between talkers. Also in comparison to direct microphone
`output mixing, this method reduces the stationary residual echo level by the reduction
`output mixing, this method reduces the stationary residual echo level by the reduction
`of acoustic coupling.
`of acoustic coupling.
`PACS number: 43. 55. Lb, 43. 88. Md, 43. 88. Si
`PACS number: 43.55. Lb, 43.88. Md, 43.88. Si
`
`INTRODUCTION
`1.
`INTRODUCTION
`1.
`In teleconferencing, hands-free telephony through
`In teleconferencing, hands-free telephony through
`multiple microphones
`is required
`to allow
`the
`multiple microphones
`is required to allow the
`participation of many people
`in each room.
`In
`participation of many people in each room.
`In
`such a system, acoustic feedback compensation and
`such a system, acoustic feedback compensation and
`room echo suppression are necessary for natural
`room echo suppression are necessary for natural
`two-way conversations.
`two-way conversations.
`To achieve
`this,
`the voice switching method,
`To achieve this,
`the voice switching method,
`which controls the send and receive signal levels, is
`which controls the send and receive signal levels, is
`commonly used. Although this method is reliable,
`commonly used. Although this method is reliable,
`it causes the chopping of speech when the send/
`it causes the chopping of speech when the send/
`receive status alternates. An echo canceller can
`receive status alternates. An echo canceller can
`eliminate this problem, and the recent implementa-
`eliminate this problem, and the recent implementa(cid:173)
`tion of digital signal processing developments makes
`tion of digital signal processing developments makes
`use of this method for treating long delay acoustic
`use of this method for treating long delay acoustic
`feedback practical.
`In the conventional configura-
`feedback practical.
`In the conventional configura(cid:173)
`tion, however, this method is applied
`to a single
`tion, however, this method is applied to a single
`echo path model. This means that it contains only
`echo path model. This means that it contains only
`a single echo path filter to compensate for the total
`a single echo path filter to compensate for the total
`acoustic feedback. When many microphones are
`acoustic feedback. When many microphones are
`used, direct mixing of multiple microphone outputs
`used, direct mixing of multiple microphone outputs
`decreases
`the signal-to-noise
`ratio, and degrades
`decreases the signal-to-noise ratio, and degrades
`
`in-
`this problem,
`eliminate
`To
`quality.
`speech
`in(cid:173)
`speech quality. To eliminate this problem,
`dividually
`controlled microphone
`gains are desir-
`dividually controlled microphone gains are desir(cid:173)
`able.
`This requires
`the use of multiple acoustic echo
`able. This requires the use of multiple acoustic echo
`paths,
`as stable operation
`cannot be expected by a
`paths, as stable operation cannot be expected by a
`single echo path configuration.
`single echo path configuration.
`which
`This
`paper
`proposes
`a
`configuration
`This paper proposes a configuration which
`achieves stable operation during the gain switching of
`achieves stable operation during the gain switching of
`multiple microphones,
`and the transient
`and station-
`multiple microphones, and the transient and station(cid:173)
`ary performance
`of this configuration
`is investigated.
`ary performance of this configuration is investigated.
`2. DESIGN
`OF THE MULTIPLE
`2. DESIGN OF THE MULTIPLE
`MICROPHONE
`SYSTEM
`MICROPHONE SYSTEM
`for
`Examples
`of
`the microphone
`arrangement
`Examples of the microphone arrangement for
`teleconferencing
`with many
`participants
`at each
`teleconferencing with many participants at each
`location
`is shown in Fig. 1. For best speech quality ,
`location is shown in Fig. 1. For best speech quality,
`the microphone
`should be set nearest
`the talker
`to
`the microphone should be set nearest the talker to
`the direct voice signal
`level higher
`than
`the
`keep
`keep the direct voice signal level higher than the
`reverberant
`portion
`of
`the
`signal.
`Therefore
`reverberant portion of
`the
`signal. Therefore
`microphones
`should be set up for each participant,
`microphones should be set up for each participant,
`or at least one
`for every two participants.
`If the
`or at least one for every two participants. If the
`outputs of these microphones
`are simply mixed
`into
`outputs of these microphones are simply mixed into
`a send signal, the signal-to-noise
`ratio will be reduced
`a send signal, the signal-to-noiseratio will be reduced
`by
`the ambient
`room
`noise,
`and distortion will
`by the ambient room noise, and distortion will
`
`39
`39
`
`RTL607_1026-0001
`
`Realtek 607 Ex. 1026
`
`
`
`~ talker
`
`no i s e ·
`
`(
`
`ambient G ...
`G.,,:':'pi microphone send- in) 0 send-out
`! /
`(1/
`
`/
`
`~receive-in
`loudspeaker(receive-out)
`]) Single microphone
`
`~talker
`
`ambient G
`noise n
`- ' 0
`
`microphone
`o~send-out
`
`7
`
`!
`o
`
`[)J----oreceive-in
`loudspeaker
`
`J. Acoust. Soc. Jpn. (E)10, 1 (1989)
`J. Acoust. Soc. Jpn. (E) 10, 1 (1989)
`
`simultaneously,
`speak
`two or more participants
`when
`when two or more participants speak simultaneously,
`it is acceptable,
`since
`in a real
`situation,
`this
`status
`it is acceptable, since in a real situation, this status
`will not
`last for
`long.
`will not last for long.
`micro-
`for an attenuated
`Therefore,
`let
`the gain
`Therefore, let the gain for an attenuated micro(cid:173)
`phone
`be normalized
`to unity.
`Represent
`the gain
`phone be normalized to unity. Represent the gain
`of
`the actuated
`microphones
`by ak', where
`k is the
`of the actuated microphones by a,,', where k is the
`number
`of actuated
`microphones,
`and
`then assume
`number of actuated microphones, and then assume
`that
`their gains have equal
`values
`according
`to each
`that their gains have equal values according to each
`k. When
`the
`total mixed
`signal
`level
`is kept
`con-
`k. When the total mixed signal level is kept con(cid:173)
`stant,
`the
`relation
`between
`the gain
`for when a single
`stant, the relation between the gain for when a single
`microphone
`is actuated
`and
`for when
`k microphones
`microphone is actuated and for when k microphones
`are actuated
`is
`are actuated is
`a!'2 + (N -1) = ka,,'2 + (N - k)
`
`2) Direct mixing of multiple microphones
`
`or
`or
`
`~talker microphone
`
`ambient G o,~send-out
`n
`nOIse
`- '0 0
`[)J------o recei ve- in
`1
`1
`A
`i
`loudspeaker
`'0 0
`
`/
`
`3) Gain switching
`systems.
`Fig. 1 Arrangement
`of microphone
`Fig. 1 Arrangement of microphone systems.
`
`occur when
`the reverberant portions are added.
`occur when the reverberant portions are added.
`As the returned echo path signal level and the am-
`As the returned echo path signal level and the am(cid:173)
`bient noise signal level increase, acoustic coupling
`bient noise signal level increase, acoustic coupling
`increases and
`the system will become unstable.
`increases and the system will become unstable.
`Degradation of the S/N or increase of acoustic coupl-
`Degradation of the SIN or increase of acoustic coupl(cid:173)
`ing is estimated as 10 log10 N(dB), where N is the
`ing is estimated as 10 log!o N(dB), where N is the
`number of the connected microphones. Also, it is
`number of the connected microphones. Also, it is
`known that the speech transmission
`index, the ob-
`known that the speech transmission index, the ob(cid:173)
`jective measure of speech signal intelligibility, will
`jective measure of speech signal intelligibility, will
`be reduced by approximately 0.1 when six micro-
`be reduced by approximately 0.1 when six micro(cid:173)
`phones are connected in a room with 1 s reverbera-
`phones are connected in a room with 1 s reverbera(cid:173)
`tion time.1) However, with regard to microphone
`tion time.!) However, with regard to microphone
`usage, usually there is only one talker and occasion-
`usage, usually there is only one talker and occasion(cid:173)
`ally two or more when another talker interrupts.
`ally two or more when another talker interrupts.
`Considering
`these parameters,
`the best speech
`Considering these parameters, the best speech
`quality can be attained by actuating only the micro-
`quality can be attained by actuating only the micro(cid:173)
`phone nearest
`the talker and attenuating
`the rest.
`phone nearest the talker and attenuating the rest.
`When
`the talker changes, microphone gains must
`When the talker changes, microphone gains must
`also change
`to maintain a constant
`total signal
`also change to maintain a constant total signal
`level. Also, we should avoid the chopping of the
`level. Also, we should avoid the chopping of the
`voice signals. This means that an attenuated micro-
`voice signals. This means that an attenuated micro(cid:173)
`phone gain should maintain some value.2) Though
`phone gain should maintain some value. 2) Though
`this switching of microphone gains will cause a
`this switching of microphone gains will cause a
`decrease of individual
`talker's speech signal level
`decrease of individual talker's speech signal level
`
`40
`40
`
`(1)
`/a!'2+k-1
`,
`(1)
`a" =V
`k
`there are no correlations
`where we assume
`that
`where we assume that there are no correlations
`between microphone
`signals. We consider
`this the
`between microphone signals. We consider this the
`ideal gain switching condition.
`ideal gain switching condition.
`3. CONFIGURATION
`OF THE ECHO
`3. CONFIGURATION OF THE ECHO
`CANCELLER
`CANCELLER
`the portion
`An echo canceller
`is used to eliminate
`An echo canceller is used to eliminate the portion
`of the send signal which comes from a loudspeaker
`of the send signal which comes from a loudspeaker
`as a returned
`echo.
`It is constructed
`of an echo
`It is constructed of an echo
`as a returned echo.
`path filter which simulates
`the total of all acoustic
`path filter which simulates the total of all acoustic
`echo paths from the loudspeaker
`to each microphone.
`echo paths from the loudspeaker to each microphone.
`However, when a conventional
`echo canceller
`that
`However, when a conventional echo canceller that
`is unable
`to memorize several sets of filter coefficients
`is unable to memorize several sets of filter coefficients
`is applied
`to a switched gain microphone
`system,
`the
`is applied to a switched gain microphone system, the
`amount of cancellation
`abruptly decreases when
`the
`amount of cancellation abruptly decreases when the
`microphone
`gain switches,
`and recovery
`is delayed
`microphone gain switches, and recovery is delayed
`until the convergence
`of the adaptation.
`Therefore,
`until the convergence of the adaptation. Therefore,
`a system
`is proposed with sets of acoustic echo path
`a system is proposed with sets of acoustic echo path
`transfer
`functions
`contained
`in memory.
`The
`transfer
`functions contained
`in memory. The
`filter coefficients of the total acoustic echo path are
`filter coefficients of the total acoustic echo path are
`calculated
`by a linear combination
`of the values
`calculated by a linear combination of the values
`stored
`in
`the memory.
`Filter
`coefficients
`are
`stored
`in
`the memory. Filter coefficients are
`changed when
`talkers
`alternate
`and microphone
`changed when talkers alternate and microphone
`gain changes.
`Because
`this gain switching
`is dis-
`gain changes. Because this gain switching is dis(cid:173)
`tinguished
`from
`environmental
`variations
`in
`the
`tinguished from environmental variations in the
`acoustic echo path, cancellation
`degradation
`due to
`acoustic echo path, cancellation degradation due to
`change
`of speaker
`can be avoided.
`The
`system
`change of speaker can be avoided. The system
`configuration
`is shown
`in Fig. 2. Though
`it
`is
`configuration is shown in Fig. 2. Though it is
`assumed
`that
`the number of echo path sets is equal
`assumed that the number of echo path sets is equal
`to the number
`of microphones,
`it is not necessary
`to the number of microphones, it is not necessary
`that the impulse
`response stored
`in a memory
`is that
`that the impulse response stored in a memory is that
`of the echo path to a single microphone.
`of the echo path to a single microphone.
`the
`of
`In
`this
`configuration,
`when
`the gain
`In this configuration, when the gain of the
`
`RTL607_1026-0002
`
`
`
`PATHS
`ECHO
`WITH MULTIPLE
`ECHO CANCELLER
`et al.: ACOUSTIC
`N. KOIZUMI
`N. KOIZUMI et at.: ACOUSTIC ECHO CANCELLER WITH MULTIPLE ECHO PATHS
`
`microphone (send-in)
`
`send-out
`
`0r-1 ~!\~ching M;o<+ .. o-------------r--r
`o
`
`memory
`
`memory
`
`echo path
`estimator
`
`receive-in
`echo paths.
`for multiple
`of echo canceller
`Fig. 2 Configuration
`Fig. 2 Configuration of echo canceller for multiple echo paths.
`
`a,-l
`(5)
`a,,-l=--
`(5)
`k
`arbitrary
`an
`for when
`response
`impulse
`total
`the
`the total impulse response for when an arbitrary
`number
`of microphones
`is actuated
`can be obtained
`number of microphones is actuated can be obtained
`by
`the following
`equation.
`by the following equation.
`
`(6)
`1 "
`H = - 2J H* j
`( 6 )
`k j,on
`If only single microphone is actuated when there is
`If only single microphone is actuated when there is
`no near-end talker, the impulse response H*j can be
`no near-end talker, the impulse response H*j can be
`identified during operation. Therefore,
`if they are
`identified during operation. Therefore, if they are
`stored in memory, H can be estimated on demand.
`stored in memory, H can be estimated on demand.
`The gain ak given by Eq. (5) does not coincide with
`The gain a" given by Eq. (5) does not coincide with
`the ideal constant
`level gain ak' given by Eq. (1).
`the ideal constant level gain a,,' given by Eq. 0).
`However, the total mixed signal level can be inter-
`However, the total mixed signal level can be inter(cid:173)
`preted as a constant if the difference is small. The
`preted as a constant if the difference is small. The
`values of Eq. (1) and Eq. (5) are compared in Fig. 3.
`values of Eq. (1) and Eq. (5) are compared in Fig. 3.
`When single actuated gain a1 is set to 2 (6 dB) or 3
`When single actuated gain a, is set to 2 (6 dB) or 3
`(9.5 dB), the difference between ak and ak' is within
`(9.5 dB), the difference between a" and a,,' is within
`1 dB. This implies that using Eq. (5) as the gain
`1 dB. This implies that using Eq. (5) as the gain
`setting is acceptable if a1 is appropriately given, and
`setting is acceptable if a, is appropriately given, and
`acoustic coupling can be maintained approximately
`acoustic coupling can be maintained approximately
`
`is set to zero, the total echo
`attenuated microphones
`attenuated microphones is set to zero, the total echo
`path can be formulated
`as a linear combination
`of
`path can be formulated as a linear combination of
`each microphone's
`echo path
`for arbitrary
`gain
`each microphone's echo path for arbitrary gain
`settings. However, when attenuated microphones
`settings. However, when attenuated microphones
`have gain values such as those
`found
`in practical
`have gain values such as those found in practical
`situations,
`some constraints
`are required
`to obtain
`situations, some constraints are required to obtain
`the total echo path by a linear combination.
`the total echo path by a linear combination.
`Consider
`the
`impulse
`response
`of
`the acoustic
`Consider the impulse response of the acoustic
`echo path from
`the loudspeaker
`to the j-th micro-
`echo path from the loudspeaker to the j-th micro(cid:173)
`phone, Hj.
`When
`the
`system
`is operating,
`Hj
`phone, Hj. When the system is operating, Hj
`cannot be observed or estimated
`because attenuated
`cannot be observed or estimated because attenuated
`microphones
`still keep some gain.
`The
`total
`im-
`microphones still keep some gain. The total im(cid:173)
`pulse
`response
`from
`the
`loudspeaker
`to the mixed
`pulse response from the loudspeaker to the mixed
`output
`of the microphone
`system when k micro-
`output of the microphone system when k micro(cid:173)
`phones are actuated
`can be described as
`phones are actuated can be described as
`
`"
`N
`H=(ale-l) 2J Hj+ 2JHi
`
`j:on
`
`i=l
`
`(2)
`(2)
`
`to
`gain is normalized
`where attenuated microphone
`where attenuated microphone gain is normalized to
`unity
`and
`ak
`indicates
`the actuated microphone
`unity and ale
`indicates the actuated microphone
`gain when k microphones
`are actuated.
`In
`this
`gain when k microphones are actuated.
`In this
`equation,
`the first summation
`on the right side rep-
`equation, the first summation on the right side rep(cid:173)
`resents
`the summation
`of impulse
`responses
`for k
`resents the summation of impulse responses for k
`actuated microphones.
`When only a single
`j-th
`actuated microphones. When only a single j-th
`microphone
`is actuated,
`total
`impulse
`response,
`microphone is actuated, total impulse response,
`which is indicated as H*j, becomes
`which is indicated as H*j, becomes
`
`N
`H*j=(al-l)Hj+ 2JHi
`i=l
`micro-
`for all actuated
`summation
`the
`take
`If we
`If we take the summation for all actuated micro(cid:173)
`phones, we have
`phones, we have
`
`(3)
`(3 )
`
`k
`"N
`2J H*j=(a,-l) 2J Hj+k2JHi
`
`(4)
`(4)
`
`j:on
`J:on
`i=l
`al and ak
`gains
`between
`relation
`the
`if
`Therefore,
`Therefore, if the relation between gains a, and a"
`is given as
`is given as
`
`!J. ak
`Oak
`... ak
`• ak
`
`} a, : 9.5 dB
`
`} a, : 6 dB
`
`~
`
`oJ
`:5
`-.:; 10
`'"
`'" 8
`"'
`"
`";;j
`bJJ 6
`"
`.c 4
`:::
`·s 2
`<!) ;;; 0
`" £j
`
`<!)
`
`0
`p.
`
`0
`
`'C
`
`0
`
`!J.
`
`!J.
`0
`
`• 0
`•
`
`lI.
`
`• .....
`
`!J.
`0
`
`!J.
`0
`
`!J.
`0
`
`• • •
`
`2
`
`3
`
`4
`
`5
`
`6
`
`Number of actuated microphones
`Fig. 3 Values of actuated microphone
`gain.
`Fig. 3 Values of actuated microphone gain.
`
`41
`41
`
`RTL607_1026-0003
`
`
`
`constant. Therefore we adopt ak defined by Eq.
`constant. Therefore we adopt a" defined by Eq.
`(5) as the gain setting of this proposed configuration.
`(5) as the gain setting of this proposed configuration.
`The block diagram of the operation
`is shown in
`The block diagram of the operation is shown in
`Fig. 4. At
`the startup, microphone gains are
`Fig. 4. At the startup, microphone gains are
`switched
`serially
`to actuate
`each microphone.
`to actuate each microphone.
`switched serially
`Using a test signal, initial values of H*j (j=1,...,
`Using a test signal, initial values of H*j U= 1, ... ,
`N) are identified and stored in memory. When
`N) are identified and stored in memory. When
`the conferencing starts, one stored response is called
`the conferencing starts, one stored response is called
`up as the set of filter coefficients for a single talker,
`up as the set of filter coefficients for a single talker,
`and for multiple talkers, filter coefficients are cal-
`and for multiple talkers, filter coefficients are cal(cid:173)
`culated from stored responses. When speech stops,
`culated from stored responses. When speech stops,
`the impulse response is adaptively
`identified using
`the impulse response is adaptively identified using
`the speech signals of the far-end talker at the opposite
`the speech signals of the far-end talker at the opposite
`site.
`In this case, the microphone of the last talker
`In this case, the microphone of the last talker
`site.
`remains actuated. The adaptation will work well,
`remains actuated. The adaptation will work well,
`and more adaptation
`time will be assigned to fre-
`and more adaptation time will be assigned to fre(cid:173)
`quently
`actuated microphones. Because
`this
`quently
`actuated microphones. Because
`this
`scheme only requires
`the addition of memory to
`scheme only requires the addition of memory to
`the hardware configuration and a simple calculation
`the hardware configuration and a simple calculation
`scheme, it can be easily implemented.
`scheme, it can be easily implemented.
`
`send-in
`
`,1) Initial setup Cafter training)
`send-in
`
`receive-in
`2) First microphone is actuated (single participant is
`talking, or he stops talking, which initiates adaptation).
`send- in
`
`H,
`
`J. Acoust. Soc. Jpn. (E)10, 1 (1989)
`J. Acoust. Soc. Jpn. (E) 10, 1 (1989)
`
`STUDY
`4. SIMULATION
`4. SIMULATION STUDY
`The convergence of echo cancellation is simulated
`The convergence of echo cancellation is simulated
`using real acoustic echo path data. For the measure-
`using real acoustic echo path data. For the measure(cid:173)
`ment of impulse response data, 3 microphones are
`ment of impulse response data, 3 microphones are
`arrayed in a line 50 cm apart as shown in Fig. 5,
`arrayed in a line 50 cm apart as shown in Fig. 5,
`in a room with a reverberation time of 0.46 s. The
`in a room with a reverberation time of 0.46 s. The
`stored impulse response word length is 4,000 samples,
`stored impulse response word length is 4,000 samples,
`which can deal with 0.5 s data at an 8 kHz sampling
`which can deal with 0.5 s data at an 8 kHz sampling
`rate.
`rate.
`The change of echo cancellation due to micro-
`The change of echo cancellation due to micro(cid:173)
`phone switching using the conventional single echo
`phone switching using the conventional single echo
`path model
`is shown
`in Fig. 6. The horizontal
`path model is shown in Fig. 6. The horizontal
`axis is the iteration, and the microphone switching
`axis is the iteration, and the microphone switching
`occurs at a constant rate. The first six switches are
`occurs at a constant rate. The first six switches are
`by single microphone and last three are by two
`by single microphone and last three are by two
`actuated microphones. The echo cancellation
`is
`actuated microphones. The echo cancellation is
`defined as 10 log [(power of returned signals)/(power
`defined as 10 log [(power of returned signals)!(power
`of
`the
`residual
`signals)]. For
`the adaptation
`of the residual signals)]. For
`the adaptation
`algorithm,
`the adaptive
`learning method3) is used
`algorithm, the adaptive learning method 3) is used
`and white noise is applied as the test signal. The
`and white noise is applied as the test signal. The
`figure shows
`that cancellation decreases abruptly
`figure shows that cancellation decreases abruptly
`when
`the distance
`from
`the loudspeaker
`to the
`when the distance from the loudspeaker to the
`microphone varies only slightly. Though the con-
`microphone varies only slightly. Though the con(cid:173)
`vergence performance depends on the adaptation
`vergence performance depends on the adaptation
`algorithm,
`the convergence will deteriorate when a
`algorithm, the convergence will deteriorate when a
`speech signal is used for adaptation during real use.
`speech signal is used for adaptation during real use.
`This kind of degradation can be avoided if the
`This kind of degradation can be avoided if the
`proposed configuration
`is applied with the same
`proposed configuration is applied with the same
`adaptation algorithm, and only the time period for
`adaptation algorithm, and only the time period for
`the convergence of
`the variation of
`the room
`the convergence of the variation of the room
`acoustics
`is required. This feature
`is shown
`in
`acoustics is required. This feature is shown in
`Fig. 7.
`In this figure, the horizontal axis represents
`In this figure, the horizontal axis represents
`Fig. 7.
`
`reverberation time (500 Hz) 0.46s
`
`I
`i
`
`_.-15,. I loudspeaker
`i height ll5em
`!
`Mic. I Mic.2 Mic.3
`,--0' 0 0
`l~oc~l:oc~j
`
`receive-in
`3) First and second are actuated Cno adaptation).
`Fig. 4 Operational
`block diagram.
`Fig. 4 Operational block diagram.
`
`response
`impulse
`for
`Fig. 5 Location
`Fig. 5 Location for impulse response
`measurement.
`measurement.
`
`42
`42
`
`RTL607_1026-0004
`
`
`
`ECHO PATHS
`WITH MULTIPLE
`ECHO CANCELLER
`et al.: ACOUSTIC
`N. KOIZUMI
`N. KOIZUMI et at.: ACOUSTIC ECHO CANCELLER WITH MULTIPLE ECHO PATHS
`
`CD Mic. 1 actuated
`® Mic. 2 actuated
`® Mic. 3 actuated
`50
`
`G) Mic. 1 actuated
`® Mic. 2 actuated
`® Mic. 3 actuated
`
`CD Mic. 1, 2 actuated
`® Mic. 2,3 actuated
`® Mic. 3, 1 actuated
`
`40
`
`r-,
`p:]
`""0
`'J
`
`c:
`0 ....,
`.!:!
`a;
`u c:
`'" u
`0 .c
`u
`(i1
`
`0L-----~------~------~3------~4~----~5------~-
`
`(XIO') ITERATION
`Fig. 6 Variation
`of echo cancellation.
`Fig.6 Variation of echo cancellation.
`
`'-../ 40
`c:
`·3 .!:! 30
`a;
`u c: 20
`'" u
`~ 10
`u
`(i1
`
`4000 taps
`8 kHz sampling
`
`I ~V'f--. by adaptation
`V i
`
`T change of talker
`
`Time (s)
`over
`of echo cancellation
`Fig. 7 Variation
`Fig.7 Variation of echo cancellation over
`time.
`time.
`
`or
`talker
`a
`after
`real environment,
`In a
`time.
`time.
`In a real environment, after a talker or
`actuated microphone
`changes,
`the adaptation
`stops
`actuated microphone changes, the adaptation stops
`and
`restarts
`after
`the
`talker
`stops
`talking.
`How-
`and restarts after the talker stops talking. How(cid:173)
`ever, to show only the difference of the convergence,
`ever, to show only the difference of the convergence,
`in the figure,
`this
`interval
`is abridged,
`and adapta-
`in the figure, this interval is abridged, and adapta(cid:173)
`tion
`is shown
`to start
`immediately
`after the alterna-
`tion is shown to start immediately after the alterna(cid:173)
`tion.
`Even
`though
`the
`conventional
`adaptation
`tion. Even though the conventional adaptation
`method
`greatly
`reduces
`cancellation
`when
`talker
`method greatly reduces cancellation when talker
`alternates
`frequently,
`the proposed method main-
`alternates frequently, the proposed method main(cid:173)
`tains
`a
`stable
`cancellation
`performance
`during
`tains a stable cancellation performance during
`talker alternation.
`talker alternation.
`PERFORMANCE
`5.
`STATIONARY
`5. STATIONARY PERFORMANCE
`If it is assumed
`that
`the delay processing
`time of
`If it is assumed that the delay processing time of
`the echo path filter, which
`is determined by the order
`the echo path filter, which is determined by the order
`
`r-. s
`
`0
`I
`
`(Q
`'D
`'J
`Q)
`:0-
`~
`-;;;
`" .~
`'"
`;
`'D " <J>
`'"
`'D " '"
`
`.S
`I
`'D " <J>
`
`(fJ
`
`-10
`
`-20
`
`-30
`
`-40
`
`-50
`
`-60
`
`-70
`
`delay processing time: 0.14s
`room reverberation time: O.ls
`
`am bient noise level
`
`-40
`
`-30
`
`-20 -15
`
`Receive-in signal level (dBm )
`Fig. 8 Send-in
`and send-out
`signal
`level of
`Fig. 8 Send-in and send-out signal level of
`echo canceller
`in room
`environment.
`
`echo canceller in room environment.
`
`exceeds
`sufficiently
`filter,
`response
`impulse
`of finite
`of finite impulse response filter, sufficiently exceeds
`reverberation
`decay
`time,
`the
`cancellation
`per-
`reverberation decay
`time,
`the cancellation per(cid:173)
`formance
`in a
`real
`room's
`acoustics
`depends
`on
`formance in a real room's acoustics depends on
`ambient
`noise
`level.
`Here,
`some
`stationary
`charac-
`ambient noise level. Here, some stationary charac(cid:173)
`teristics
`from
`the
`experimental
`results
`of a single
`teristics from the experimental results of a single
`microphone
`system
`are
`considered.
`These
`results
`microphone system are considered. These results
`were
`obtained
`from
`equipment
`which
`used
`digital
`were obtained from equipment which used digital
`signal
`processor
`chips.4)
`The
`set has
`a passband
`signal processor chips.4) The set has a passband
`of 4 kHz
`and
`a delay
`processing
`time
`of 144 ms.
`of 4 kHz and a delay processing time of 144 ms.
`It
`is designed
`for a conventional
`single
`echo
`path
`It is designed for a conventional single echo path
`configuration.
`The
`room
`used
`this experiment
`for
`configuration. The room used for this experiment
`
`43
`43
`
`RTL607_1026-0005
`
`
`
`delay processing time: 0.14 s
`room reverberation. time: 0.1 s
`
`40
`
`30
`
`20
`
`10
`
`- 20 -15
`- 30
`Receive-in signal level (dBm)
`Fig. 9 Cancellation
`performance
`of echo
`Fig. 9 Cancellation performance of echo
`canceller
`in room environment.
`canceller in room environment.
`
`level
`time of 0.1 s. Send-in
`reverberation
`has a
`has a reverberation time of 0.1 s. Send-in level
`and send-out
`level after cancellation
`of the returned
`and send-out level after cancellation of the returned
`signal versus
`receive
`level
`is shown
`in Fig. 8 for
`signal versus receive level is shown in Fig. 8 for
`various ambient noise
`levels.
`various ambient noise levels.
`to the
`The send-in
`level is of course proportional
`The send-in level is of course proportional to the
`receive
`level. However,
`the
`send-out
`level as a
`receive level. However, the send-out level as a
`residual
`level
`is approximately
`constant
`and
`in-
`residual level is approximately constant and in(cid:173)
`dependent
`of the receive signal
`level, and
`is deter-
`dependent of the receive signal level, and is deter(cid:173)
`mined by the ambient
`noise
`level. The echo can-
`mined by the ambient noise level. The echo can(cid:173)
`cellation,
`the subtraction
`of the send-out
`level from
`cellation, the subtraction of the send-out level from
`the send-in
`level,
`is shown
`in Fig. 9. The echo
`the send-in level, is shown in Fig. 9. The echo
`cancellation
`depends on the receive-in
`level and
`is
`cancellation depends on the receive-in level and is
`determined
`by the ratio of the returned
`signal
`level
`determined by the ratio of the returned signal level
`to the ambient noise level.
`From
`these results,
`the
`to the ambient noise level. From these results, the
`performance
`of the proposed method when multiple
`performance of the proposed method when multiple
`microphones
`are used can be projected.
`An es-
`microphones are used can be projected. An es(cid:173)
`timated comparison
`between
`the proposed
`switching
`timated comparison between the proposed switching
`and direct microphone
`output mixing for a multiple
`and direct microphone output mixing for a multiple
`microphone
`system
`is shown
`in Fig. 10.
`microphone system is shown in Fig. 10.
`is
`The absolute value used as a scale in the figure
`The absolute value used as a scale in the figure is
`the nominal
`value
`for
`the hardware
`used
`in
`the
`the nominal value for the hardware used in the
`experiment.
`Also,
`as an
`example,
`the
`ambient
`experiment. Also, as an example,
`the ambient
`noise
`level is set to 40 dB, the number of connected
`noise level is set to 40 dB, the number of connected
`microphones
`is 6, and
`the difference between
`the
`microphones is 6, and the difference between the
`single actuated
`gain and
`the attenuated
`gain
`is set
`single actuated gain and the attenuated gain is set
`at 9.5 dB (a1=3).
`In this case, the proposed
`con-
`at 9.5 dB (a, = 3).
`In this case, the proposed con(cid:173)
`figuration
`reduces
`the returned
`signal
`level by ap-
`figuration reduces the returned signal level by ap(cid:173)
`proximately
`6 dB. The ambient
`noise
`level is also
`proximately 6 dB. The ambient noise level is also
`reduced by