GB2348091A - Subband echo canceller - Google Patents

Abstract

A subband acoustic echo canceller (120) includes an adaptation speed controller (240). The adaptation speed controller measures the energies of the echo residual signals r<SB>m</SB>(n) for each subband. Those subbands having high echo residual energies are allocated normal adaptation rates; whereas those subbands having low echo residual energies are allocated slower adaptation rates. The adaption rates are controlled by signals s<SB>m</SB>(n) sent from the adaptation speed controller (240) to adaptive filters (201, 202, 203) for the respective subbands.

Description

SUBBAND ECHO CANCELLER AND METHOD THEREFOR FIELD OF THE INVENTION The present invention pertains generally to two-way communication systems and more particularly to subband echo cancellers therefor.

BACKGROUND OF THE INVENTION The need for echo cancellation arises in many full-duplex communication systems. One particularly challenging environment where the need for reliable echo cancellation exists is full-duplex hands-free operation of cellular radiotelephone devices and teleconferencing devices. During hands-free operation of such devices, signals from the speaker are fed back into the microphone through various acoustic paths and are subject to delay before reaching the original speaker. These feedback signals are perceived by far-end users as echo signals. The echo signals, commonly referred to as acoustic echo, are very annoying to the participants involved in two-way communication and difficult to eliminate.

One of the most effective solutions for eliminating echo signals is use of an echo canceller having an adaptive filter. A least means square (LMS) adaptive filter is the most common type of filter used. An LMS filter is a finite impulse response (FIR) filter which models an echo path through adaptively adjusted coefficients. The coefficients of the filter are adaptively trained using the far-end signal, which drives the loudspeaker of a hands-free communication device, and the near-end signal, which is output from the microphone ofthe hands-free communication device. In a handsfree device, the adaptive filter adaptively synthesizes a replica of the acoustic echo from the far end signal, which replica is subtracted from the near end signal output by the microphone at the near end. The result is a substantially echo-free signal which is further transmitted to the far end.

Subband echo cancellation circuits have been used in environments subject to a long echo path. In these circuits, the far-end signals are segmented into a set of signals, each of which represents the portion of the far-end signal in one subband. The near-end signal is also segmented in the same way as the far-end signal. There are a number of ways to segment the far-end signals into subbands. A basic approach uses a filter bank to separate the full band of the transmission signal into adjacent segments.

The output of each filter is the signal from that subband. Thus, for each subband, a. pair of signals are obtained which correspond to the far-end and near-end signals.

A subband echo canceller includes an echo canceller for each subband. A respective adaptive filter is provided for each echo canceller. Echo cancellation is performed either in the time-domain or the frequency-domain. The echo cancelled signals produced by the respective subband echo cancellers are combined into a fullband near end signal using a synthesis filter. Adaptive subband filtering offers an efficient way to achieve both performance enhancement and complexity reduction.

US Patent 5,001,701, entitled SUBBAND ECHO CANCELLER INCLUDING REAL-TIME ALLOCATION AMONG THE SUBBANDS, issued to Steven L. Gay, proposed a scheme to improve the performance of such a subband echo canceller. In that patent, it is assumed that the total computational resources available for echo cancellation are not fixed. Those echo cancellers associated with subbands having larger adaptive filter misalignment are given more iterations per second to adapt than those echo cancellers in a subband having a smaller adaptive filter misalignment to realize faster convergence of overall subband echo canceller.

The misalignment of the adaptive filters is a weighted norm of the difference between the current and past filter coefficients.

Although such a subband echo cancellation technique reduces the computational complexity of the acoustic echo canceller, its computational complexity is still too high for some applications. Further reductions of the computational complexity of subband acoustic echo cancellers is therefor demanded to meet the requirements of many applications. It is also desirable to reduce the computational complexity as much as possible while maintaining the overall performance as perceived by a user.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a circuit schematic illustrating a full-duplex communication device employing a subband echo canceller.

FIG. 2 is a circuit schematic illustrating a subband echo canceller for fullduplex communication.

FIG. 3 is a schematic illustrating an adaptation speed control device for a subband echo canceller.

FIG. 4 is a circuit schematic illustrating a residual energy calculator in the adaptation speed control device.

FIG. 5 is a circuit schematic illustrating an adaptive filter for one subband.

FIG. 6 is a schematic illustrating an adaptive filter control device.

DETAILED DESCRIPTION OF THE DRAWINGS By utilizing adaptive filtering in individual subbands, subband echo cancellation techniques offer an opportunity to use computational resources of a device more efficiently. An efficient adaptation method reduces the computational complexity as much as possible while maintaining the overall user perceived performance. The method identifies those subbands with high echo residual energy.

For the subbands with high echo residual energy, adaptation occurs at a normal speed.

Other subbands adapt at a lower speed. This adaptation provides echo cancellation across all the subbands, but with more efficiency than prior art echo cancellers. Those skilled in the art will recognize that although in the illustrated embodiment, timedomain adaptive filtering is used due to its structural simplicity and conceptual clearness, other forms of adaptive filtering could be used.

Because human perception of echo disturbance comes mainly from the subband signals with high energies, the reduction of the computational complexity is based upon a criteria that the user perceives. It has been found that a few of the subbands will produce echo residual signals that have dominant energies. The adaptive filter coefficients in the other subbands with lower echo residual energies adapt at lower speeds, whereas the adaptive filter coefficients in those subbands with higher echo residual energies adapt at normal speeds. Because few subbands have high echo residual energies, most of the subbands have low echo residual energies and thus adapt at slower speeds.

The echo canceller is very efficient as the coefficient adaptation of adaptive filters is at normal speed in only in a few subbands, which significantly reduces the computational complexity of the overall subband echo canceller. The present inventions takes advantage of the efficiency obtained through use of different adaptation speeds for adaptive filters in different subbands to obtain uniform echo suppression over the subbands. The most efficient way to perform echo cancellation is to suppress the echo in all subbands uniformly.

A subband acoustic echo canceller 120 (FIG. 1) is employed in a two-way communication device 100 for improving hands-free operation. The full-duplex communication device 100 is illustrated as a hands-free device, and may be a handsfree radiotelephone, a hands-free teleconferencing device, a hands-free satellite telephone, a hands-free cordless telephone, a personal computer (PC) multimedia communication device, or any other suitable communication device. The communication device 100 includes a transceiver 114 to receive remote signals (far end signals) from remote sites and transmit local signals (near-end signals) to remote sites.

The transceiver can be any suitable transceiver for cable, optical, wire-less, wire line or satellite communication, the operation of which are well known to those skilled in the art and are not described in greater detail herein for brevity. In the illustrated embodiment, transceiver 114 is coupled to an antenna 116 for wireless communication in a cellular system. Transceiver 114 transmits an echo-cancelled near-end signal output by a D/A converter 110 to a far-end communication device via antenna 116 and inputs a received signal from a remote device detected by antenna 116 to an A/D converter 112.

The receive path includes an analog-to-digital (A/D) converter 112 and a digital-to-analog (D/A) converter 108 connected to a loudspeaker 102. The transceiver 114 output is converted to a digital signal in A/D converter 112 for subband echo canceller 120. A digital-to-analog (D/A) converter 108 converts the digitized far-end signal into an analog format which drives a local loudspeaker 102.

The transmit path includes a microphone 104, which converts local acoustic signals into electrical signals for transmission to remote sites, and an A/D converter 106 converts the microphone electrical signal into a digital signal supplied to subband acoustic echo canceller 120. The echo-cancelled signal output from the subband acoustic echo canceller 120 is converted into analog format in D/A converter 110 for transmission via transceiver 114.

Those skilled in the art will recognize that A/D converter 112 and a D/A converter 110 are used in an analog system. Alternatively A/D converter 112 may be replaced by a speech decoder, and a D/A converter 110 may be replaced by a speech encoder, in digital implementations. For example, converters 112 and 110 can be a speech decoder and encoder of a digital interface in a communication device for the global system for mobile communications (GSM) or an integrated services digital network (ISDN). It is further noted that the combination of an antenna 116 and a transceiver 114 may be replaced by a network interface device in some applications, such as applications for wire line or optical communication systems.

During operation, a far-end voice signal, x (n), in digital format at output 101 of an A/D converter 112, is input to a D/A converter 108, which generates an analog signal to drive a loudspeaker 102. A portion of the far-end signal output by speaker 102 is detected by a microphone 104 with the near-end audible signals and converted to digital format in an A/D converter 106. The resulting near-end signal y (n) at output 103 is an input to a subband acoustic echo canceller 120. An echo-cancelled speech signal r (n), in digital format, is output at output 105 of a subband acoustic echo canceller 120 and input to a D/A converter 110.

The subband acoustic echo canceller 120 can be implemented in a digital signal processor (DSP), a microprocessor, a programmable logic device, or the like.

The subband acoustic echo canceller 120 includes an analysis filter 220 (FIG. 2) to decompose a full-band far-end signal x (n) on bus 101 into M adjacent subband farend signals xi (n) through xM (n) on outputs 221-223, an analysis filter 242 to decompose a full-band near-end signal y (n) on bus 103 into M adjacent subband nearend signals y, (n) through yM (n) on outputs 210-212, and a synthesis filter 230 to combine the M subband echo-cancelled signals ri (n) through rM (n) on outputs 260262 into a full-band echo-cancelled signal r (n) on bus 105. M adaptive filters 201-203 are employed to cancel echo signals in M subband near-end signals y, (n) through yM (n) at outputs 210-212, and adaptation speed controller 240 outputs M adaptation speed control signals sl (n) through sM (n) on outputs 241-243 to M adaptive filters 201 through 203. It should be noted that the time index n for subband signals may or may not be the same as the time index n for full band signals : for example, it is envisioned that the time index (sampling intervals) for subband signals can be smaller intervals relative to the time index of full band signals such that adaptation occurs more often in subband echo cancellers.

Ideally, the analysis filters 220,242 output M subband signals wherein each subband signal has no overlap with adjacent subbands in the frequency-domain.

However, since absolute isolation between adjoining subbands is extremely difficult to achieve, some small overlap is allowed. This will have little impact on the performance of an acoustic echo canceller as long as the overlap is small.

Additionally, operation in synthesis filter 230 that is perfectly inverse to operation of analysis filters 220,242 will be difficult or too expensive for practical implementation. Consequently, although ideally the output signal of a synthesis filter will be a delayed version of the signal input to analysis filter, in practice, a small amount of distortion in amplitude and phase is allowed in order to accommodate a simpler design and reduced computational complexity.

There are a number of ways to implement the analysis filters 220,242 and the synthesis filter 230. These implementations are well known to those skilled in the art, and for brevity they are not further described herein.

Adaptive filter 201 (FIG. 2) is used to cancel echo signal in near-end signal yi (n) on output 210 in subband 1, adaptive filter 202 is used to cancel echo signal in the near-end signal y2 (n) on output 211 in subband 2, and adaptive filter 203 is used to cancel echo signal in the near end signal yM (n) on output 212 in the Mth subband.

The Mth subband is the 32nd subband in one embodiment. Adaptive filters 201 through 203 can be either LMS or RLS algorithms, for example. The analysis filter 220,242, and synthesis filter 230, and adaptive filters 201 through 203 can be implemented either in time-domain or in the frequency-domain.

The operation is described with respect to a time-domain system using an LMS algorithm for each adaptive filter 201 through 203 and the number of subbands M is equal to 32 (three adaptive filters 201-203 are shown in FIG. 2, but those skilled in the art will recognize that an additional twenty-nine adaptive filters are provided for the other twenty-nine subbands in an embodiment having 32 subbands). LMS adaptive filters are widely used due to their simplicity and robustness. The M adaptive filters may be different or the same depending upon the implementation. By using the same structure for all of the M adaptive filters, the system can take advantage of program reuse in a DSP processor or implementation of one circuit for M adaptive filters implementations using other hardware. It is envisioned that each of the adaptive filters 201-203 includes a finite impulse response (FIR) filter with adjustable coefficients to generate an echo estimate, a subtractor to generate an echo-cancelled signal, and a control device to update its coefficients.

The illustrated embodiment assumes that M (the number of subbands) is thirty-two, and N (the number of groups of subbands) is 8, although any other number of subbands and/or groups could be used. The adaptation speed control device 240 includes thirty-two energy calculators 310 through 317 (FIG. 3), eight of which are shown, to find thirty-two echo residual energies lei (n): i = 1,2,..., 32} on outputs 320 through 327 for the thirty-two subband echo residual signals fri (n): i = 1,2,..., 32} on ports 260,261,302-306,262. Eight adders 330-331, two of which are shown, are each used to sum M/N subband echo residual energies. A maximum group selector 340 selects the maximum value among the eight group energies {Ci (n): i = 1,2,..., 8} at ports 332 through 333. An output speed control generator 350 generates thirty-two adaptation speed control signals s (n), s2 (n), sM (n) on outputs 241-243 for the thirtytwo adaptive filters 201-203 (FIG. 2).

In the illustrated embodiment, the thirty-two subband far-end signals {xi (n): i = 1,2,..., 32} generated in analysis filter 220 (FIG. 2) are output at outputs 221-223. These outputs of analysis filter 220 are input to 32 adaptive filters 201-203.

The 32 subband near-end signals {ys (n): i = 1,2,..., 32} at outputs 210 through 212 of analysis filter 242 are also input to the 32 adaptive filters 201-203. The 32 subband adaptive filters 201-203 generate thirty-two subband echo-cancelled signals fri (n): i = 1,2,..., 32} at outputs 260-262. These subband echo cancelled signals are input to a synthesis filter 230, which outputs a full-band echo-cancelled signal r (n) at an output 105.

The subband acoustic echo canceller 120 employs an adaptation speed controller 240. The rate at which the adaptive filters 201-203 adapt is set by this adaptation speed controller 240. Thirty-two adaptation speed control signals fsi (n): i = 1,2,..., 32} at outputs 241-243 of adaptation speed controller 240 are input to thirty-two adaptive filters 201-203. The signals x, (n), yi (n), ri (n), and s, (n) are used for one subband i.

The adaptation speed control 240 assigns adaptation speeds to each of the M adaptive filters 201 through 203. The adaptation speed control 240 includes energy calculators 310 through 317. The energy calculators 310-317 are connected to receive the echo residual signals at ports 262,302-306,261,260. Each of the energy calculators includes a squaring or absolute value generating circuit 410 (FIG. 4) to output a positive value. The output of circuit 410 is filtered in a low pass filter 412 to generate an average value. The thirty-two energy calculators 310 through 317 calculate echo residual energies {ei (n): i = 1,2,..., 32} at ports 320 through 327 according to the following equation: ei (n) = (1-g) e i (n) + g rs (n) r i (n) i = 1, 2,..., 32 where g is a scalar between 0 and 1 and can be chosen, for example, 0.01. The 32 subband echo residual energies {ei (n): i = 1,2,..., 32} at outputs 320 through 327 are divided into the following eight groups of four adjacent subbands: group 1: {ei (n), e2 (n), e3 (n), e4 (n)} group 2: {es (n), e6 (n), e7 (n), es (n)} group 3: {es (n) e (n) , (n), en (n)} group 4: {e13(n), e14(n), e15(n), e16(n)} group 5: {tel7 (n), e, 8 (n), e, 9 (n), e2o (n)} group 6: {e2l (n), e22 (n), e23 (n), e24 (n)} group 7: {e25(n) (n), e26 (n), e27 (n), e2s (n)} group 8: {e29 (n) e3o (n), e31 (n), e32 (n)).

For each group, 4 adjoining subband echo residual energies are added together to generate one group energy. Thus, eight adders 330 through 331 generate eight group energies (Ci (n): i = 1,2,..., 8} at outputs 332 through 333, such that: Ci (n) = e4i-3 (n) + e4i-2 2 (n) + e4j (n) + e4i (n) i = 1, 2,..., 8.

A maximum group energy signal Cmax (n) at an output 341 is generated by a maximum selector 340 responsive to the inputs {C, (n): i = 1,2,..., 8} from outputs 332 through 333, such that: Cmax (n) = max {Cl (n), C2 (n),..., Cs (n)}.

The output speed control generator 350 is responsive to the output 341 of the maximum selector 340 to generate thirty-two adaptation speed control signals {ss (n): i = 1,2,..., 32} at outputs 241 through 243. These adaptation speed control signals are input to the M adaptive filters 201 through 203 (FIG. 2). In operation, if Ci (n) from group i has the maximum group energy, then, adaptation speed control signals at outputs 241 to 243 are generated as follows: s4s 3 (n) = s4s 2 (n) = s4s 1 (n) = S4i (n) = 1 for adaptive filters associated with the group i having the maximum value; and sj (n) = K and j w 4i, 4i-l, 4i-2,4i-3 and K > 1 for the adaptive filters associated with the other groups where K is the number of samples for which one coefficient adaptation will occur. For example, if K = 2, the adaptation speed for the adaptive filters other than selected group are half-speed (updates once every two samples), or if K = 4, the adaptation speed for the adaptive filters other than the selected group are quarter-speed (updates once every four samples).

The output speed control generator 350 thus generates a respective control signal (si (n): i = 1,2,..., 32} at outputs 241 through 243 for each adaptive filter 201 through 203 which indicates whether coefficient adaptation is performed in a particular sampling instant. A normal adaptation speed, in which adaptation occurs every available sampling instant, is given to the group, or groups, of subbands with high residual echo group energzes, whereas low adaptation speeds, wherein adaptation is skipped for some sampling intervals, are assigned to the groups of subbands with low group residual echo energies. By delegating allocated adaptation energy to those subbands having the most echo residual energy, subband echo canceller 120 can use less processing resources without significantly degrading performance of the echo canceller.

The adaptation speed control 240 can assign control signals {ss (n): i = 1,2,..., 32} at outputs 241 through 243 in a number of alternative ways.

Other methods of assigning control signals could use individual subband energies, and permit selection of those bands that have more than a threshold level of residual echo energy. Alternatively, the bands having the greatest residual energy can be selected without regard to the value of the energy in adjacent bands. The advantage to selecting groups of subbands is it simplifies processing and takes advantage of the likelihood that adjacent subbands will have the most residual echo energy.

Referring to FIG. 5, adaptive filter 201 is shown. The adaptive filter 201 includes a FIR filter 500 and an adaptive filter control device 502. The adaptive filter control device 502 generates adaptable coefficients W (n) at an output 504. The FIR filter 500 uses the coefficients W (n) at an output 504 to generate a subband echo estimate zi (n) at an output 508. At the current sampling instant, n, a subband far-end speech sample xi (n) is received at a port 221 from analysis filter 220, and a subband near-end speech sample y) (n) is received from a port 210 as an output of an analysis filter 242. Subband signals X) (n) and y, (n) at inputs 221 and 210 are synchronized as A/D converter 106 (FIG. 1) and D/A converter 108 use the same clock.

The echo estimate z (n) at output 508 is synthesized by FIR filter 500 based on the following equation:

where superscript T means the transpose of a vector or a matrix, L is the order of a FIR filter 500, XI (n) = [xl (n) xi (n-l)... xl (n-L+l)] T is L most recent subband far-end speech samples for the subband.

Adaptive filter 201 (FIG. 5) further includes a subtractor 506 which generates an echo-cancelled signal ri (n) in a double-talk condition (or an echo residual signal in single-talk condition) at output 260 by subtracting the echo estimate zl (n) at output 508 from the near-end signal y, (n) at input 210, such that: r, (n) = yz (n)-z (n) where a single-talk condition is defined that no local speech exists and a near-end signal includes echo signals only; whereas a double-talk condition is defined as nearend speech exists in addition to echo signals. The coefficient adaptation must stop in a double-talk condition. A double-talk detector can be employed to distinguish these two situations as is known in the art. A single-talk condition is assumed herein, such that r, (n) is a subband echo residual signal for subband 1.

Adaptive filter 201 further includes a control device 502 to output coefficients W, (n) at output 504 for FIR filter 500. The coefficients W, (n) at an output 504 are generated based upon a subband far-end speech signal Xi (n) at a port 221, a near-end speech signal y, (n) at a port 210, and an adaptation speed control signal s, (n) at a port 241.

With reference to FIG. 6, the operation of the adaptive filter 201 will now be described (the description of the adaptive filter 201 applies to each of the other 31 adaptive filters). A vector XI (n) which is the L most recent subband far-end signal samples from a subband far-end signal x (n) input at a 221 is stored in a buffer (not shown) in adaptive filter 201 as indicated by block 602: XI (n) = [x (n), xi (n-l),... xi (n-L+l)] T.

An energy estimate El for the subband far-end signal energy is generated according to the following equation in adaptation filter controller 502 (shown in FIG.

5) as indicated in block 604: Es = (1-a) El + a x (n) xi (n) where a is a scalar between 0 and 1, for example a = 0.01.

The adaptive filter 201 is responsive to an adaptation control signal si (n) at an input 241 to set the adaptation speed. The adaptive filter controller 502 selects either to adapt during a particular sampling instant or to use the previous coefficients, as indicated in decision block 606. If in a sampling instant, [n mod si (n)] = 1, updated coefficients Wl (n) are generated in adaptive filter controller 502 as indicated in block 608 based on input ri (n) at input 260, XI (n) from block 602, and Et from block 604 as follows: W, (n) = Wl (n-l) + b ri (n) X, (n) E,-' where b is a step size and Wi (n) = [w'o (n) w (n)... wlL l (n)] T. Otherwise coefficients WI (n) remain unchanged when [n mod s (n)] = 0 as indicated in block 610. In either case the coefficients W1 (n) are output at 504 for use by the FIR filter 500.

Thus for an adaptive filter having normal adaptation speed si (n) = 1 will adapt for every sampling instant. For those adaptive filters having a slower adaptation speed, for example 1/4 speed, s, (n) will equal 4 and only adapt once in every 4 sample instants.

Accordingly, the adaptation control signals {si (n): i = 1,2,..., 32} at outputs 241 through 243 for each of the adaptive filter 201-203 are generated each sampling instant to enable or disable adaptation for the respective adaptive filter 201 through 203. Those subbands having high echo residual signal energies will adapt each sampling instant whereas those subbands having low echo residual energies will adapt only during a portion of the available sampling instants. By adapting during fewer sampling instants, the complexity of the echo canceller as reflected by the number of machine instructions per seconds needed to perform echo cancellation can be significantly reduced without a significant degradation in the performance of the subband echo canceller. The computational complexity is reduced while maintaining original performance which is highly desirable for practical applications having limited processing capacity. Additionally, the use of the echo residual energy as the measure for determining which subbands receive the most adaptation resources affects a reduction in the complexity of the echo canceller while maintaining its performance as perceived by users.

Claims (10)

1. A method of operating adaptive filters of a subband echo canceller, the method comprising the steps of : determining the echo residual energy in each of a plurality of the subbands; assigning respective adaptation speeds to adaptive filters associated with respective subbands as a function of the echo residual energies in the respective subbands; and updating coefficients of adaptive filters at the respective adaptation rates selected for the subbands.

2. The method as defined in claim 1, wherein the step of assigning adaptation speeds for M adaptive filters includes assigning adaptation speeds to M adaptive filters based on M energies of M subband echo residual signals.

3. The method as defined in claim 2, wherein the step of calculating M energies of M subband echo residual signals includes low-pass filtering the square of echo residual signals for each subband.

4. The method as defined in claim 2, wherein the step of assigning adaptation speeds for M adaptive filters includes the step of dividing M subband echo residual signals into several groups and the echo residual energies in each group are added together to output a set of group energies.

5. The method as defined in claim 2, wherein the step of assigning adaptation speeds for M adaptive filters includes selecting a maximum value among a set of group energies.

6. The method as defined in claim 2, wherein the step of assigning adaptation speeds for M adaptive filters further including the step of outputting normal adaptation speed control signals for the subbands with higher group energies.

7. The method as defined in claim 2 wherein the step of assigning adaptation speeds for M adaptive filters further including the step of outputting slow adaptation speed control signals for the subbands with lower group energies.

8. The method as defined in claim 6, wherein the step of outputting normal adaptation speed control signals including the step of outputting signals which enable adaptation, wherein outputting normal adaptation speed control signals enable coefficient adaptation of the corresponding adaptive filter every sample interval.

9. The method as defined in claim 7, wherein the step of outputting slow adaptation speed control signals includes the step of outputting signals which disables adaptation during some adaptation intervals wherein outputting slow adaptation speed control signals enables the corresponding adaptive filter to perform adaptation once every D sample instants, wherein D is an integer.

10. A subband echo canceller comprising: a plurality of adaptive filters associated with a plurality of subbands; and an adaptation speed controller coupled to the plurality of adaptive filters, wherein the adaptation speed controller measures the echo residual energy in each subband and assigns normal adaptation speeds to those subbands having higher echo residual energies and assigns slow adaptation speeds to those subbands having lower echo residual energies.