TW583851B - Fast converging affine projection based echo canceller for sparse multi-path channels - Google Patents

Abstract

A first adaptive filter having a first filter length estimates a channel weight vector of an echo channel using an affine projection (AP) update. The echo channel receives a send input sequence and a receive input sequence. The echo channel has a plurality of sparse multi-path channels. A delay estimator determines a plurality of delays in the echo channel using the estimated channel weight vector. A second adaptive filter estimates a plurality of short weight vectors using the AP update and the plurality of delays. The short weight vectors correspond to the sparse multi-path channels. Each of the short weight vectors has a short filter length.

Description

583851 A7 B7 V. Description of the Invention (Related Application This application claims the benefit of US Provisional Application No. 60 / 231,419, filed on September 8, 2000 (Agent Document No. 004419.P016Z) Background 1. Fields of the invention and signals Processing is concerned. In particular, the present invention is related to echo cancellation. *, 2 · Relevant technical description. In telephone technology, echo is usually unwanted. Echo is caused by many sources. These include the signal from the speaker back to the microphone. Multiple reflections' sonic coupling between the horn and microphone, as well as the surrounding bathing sound. Echobone cancellation is a technique to eliminate the desired echo effect. This echo canceller estimates the impulse response of this echo path and generates an echo The estimated echo is then subtracted from the near-end signal. Typically, an adaptive filter is used to estimate the echo because the echo path is usually unknown and changes randomly over time. The technique for echo cancellers is the normalized least mean square (nlms) method. This technique tries to minimize the expected value of this squared error. This normalization The small mean square method has many disadvantages.—The important one is the low convergence rate for color input. Moreover, when this echo channel has sparse multipath channels, this normalized minimum mean square method may not provide appropriate U results. Therefore, there is a need for a pair of effective techniques that perform an echo cancellation in a sparse multipath channel with a faster convergence rate than this normalized least-mean-square method. -4- This paper standard applies the Chinese National Standard (CNS) Α4 specification ( 21〇χ297mm) Binding 583851 A7 B7 Find 0 A brief description of the star A Through the following detailed description of the present invention, the features and advantages of the present invention will become more significant: Figure 1 is a diagram illustrating the present invention. An embodiment A diagram of a system that can be implemented on it. Fig. 2 is a diagram illustrating a system model for the affine projection-based echo canceller shown in Fig. 丨 according to an embodiment of the present invention. Fig. 3 is A diagram illustrating an echo canceler based on affine projection shown in FIG. 丨 according to an embodiment of the present invention. FIG. 4A is a diagram illustrating an embodiment according to the present invention, shown in FIG. 3 A diagram of an adaptive filter for an echo canceller based on affine projection. Fig. 4B is a diagram illustrating an example of an affine projection based on affine projection according to an embodiment of the present invention. Figure of an adaptive filter of an echo canceller. Fig. 5 is a diagram illustrating a delay estimator for an echo canceller based on affine projection shown in Fig. 2 according to an embodiment of the present invention. FIG. 6 is a flowchart illustrating a process for an echo canceller according to an embodiment of the present invention. FIG. 7 is a diagram illustrating an embodiment according to the present invention for use in FIG. 6. FIG. 8 is a flowchart illustrating a process of adaptive filtering updated using an affine projection. FIG. 8 is a diagram illustrating an embodiment according to the present invention for applying the Chinese National Standard (CNS) A4 specification to the paper size in FIG. 6 ( 210 X 297 mm) " " 1 ----

B7

A flowchart of the processing of the delay estimation. FIG. 9 is a flowchart illustrating a process for partially adaptive filtering using affine casting to update according to an embodiment of the present invention. FIG. 10A is a graph illustrating weights for non-sparse channels after the first adaptation process according to an embodiment of the present invention. FIG. 10B is a graph illustrating the mean square error (MS E) of the 10,000 learning samples before the non-sparse channel according to an embodiment of the present invention. FIG. 10C is a graph illustrating an impulse response estimation of a non-sparse channel after 10,000 samples of the channel according to an embodiment of the present invention. Fig. 11A is a graph illustrating weights for sparse channels after this first adaptive phase according to an embodiment of the present invention. Yes, illustrates a graph of the mean square error (M s Ε) for a sparse channel and then 10,000 learning samples according to an embodiment of the present invention.

Figure lie is a graph illustrating an estimate of the impulse response of a non-sparse channel for this channel after 10,000 samples according to an embodiment of the present invention. U In the following description, for the purpose of explanation, many are provided to provide completeness of the invention. However, some of these specific details may not be required to implement the invention. In the example of the dagger, the known electrical structure and circuit are shown in the block diagram of J-B0 M 7%, so as not to obscure the present invention. The present invention can be specified in hardware, software, or combination. achieve. When implemented in software, the elements of the present invention are code, or body, microcode, or any combination of these, or microcode, to implement the necessary code segments. 1-6-This paper size is in accordance with Chinese National Standard (CNS) A4 specification (210X 297 mm) — "--code section" can represent a program, a function, a subprogram, a program, a 2 formula , A routine, a module, a software package, a category or any instruction, a combination of data structures, or a program description. A code segment can send and / or receive information, data, arguments, parameters, or Memory content is coupled to another program section or a hardware circuit. Information, arguments, parameters, data, etc. can be included through memory sharing, message transmission, token transmission, network transmission, etc. < The device is transmitted, forwarded, or transmitted. This program or code segment can be stored in a processor-readable medium or by a computer data implemented on a carrier wave or a signal modulated by the carrier wave. "Transmission-readable media on a transmission medium" may include any medium that can store or transmit information. Examples of media readable by the processor include an electronic circuit, a semiconductor memory device, a read-only memory (ROM), a flash memory, and an erasable only memory (EROM), a floppy disk, and a read-only optical disk ( CD-ROM), optical discs, hard drives, fiber optic media, radio frequency (RF) links, and more. The computer data signal may include any signal that can be transmitted on a transmission medium such as an electronic network channel, optical fiber, radio, magnetic field, radio frequency link, and the like. These code segments ^ are downloaded via computer networks such as the Internet, corporate intranets. > Wangyi The present invention can be described as a process which is generally described as a flowchart, a flowchart, a structure diagram, or a block diagram. Although a flowchart can describe these operations as a series of processes, many of these operations can be performed in parallel or = time. In addition, the order of these operations can be rearranged. When its operation is completed, a process is terminated. A process can be mapped to a method, function, program, subroutine, subprogram, etc. When a process corresponds to a function, it terminates the function or main function corresponding to the return to call of this function

Description of the invention (Network 145, one sending input decoder sending output encoder 180, one receiving input Figure 1 is a diagram illustrating the invention. This system is actually a system that can be implemented on it-遒 12〇, ~ Send output code; encoder U °, —echo pass—receive output encoder 15〇, —receive input decode Suo · 160, one echo channel 170… enter decoder 190, the first one is r ~ ... the side is W, two are missing And a receiving output encoder 195. Take out the input decoder 1 110 to receive the user and decode this encoded speech into a near-end encoded speech from the younger brother, this inverse speech data Sin. In this embodiment this time The 甬 decoder 110 is a -bull / A-law decoder. The second operation 120 includes an echo canceller using affine projection (AP). This echo canceller 125 based on affine projection (Ap) Since this time, the linear sample is Sin removed-the echo estimation signal is generated to generate the linear data sample S0t. The output encoder 13G provides the pre-packet speech compression. In the embodiment, this output encoder 130 Is a use including international two, Union (ITU) -T international standard G.7xx series of any compression standard for low bit rate sound (LBRV) to compress the G.7xx encoder from the sound channel data of 120. This compressed voice data is transmitted through a network. It is sent to the far end. The receiving input decoder 14 decompresses the voice data received from this first far end on the network 145. This decompression technology is compatible with the compression technology used on the output output encoder. The echo channel 120 receives Rin from the receive input decoder 14 and sends out Rout linear data samples. The receive output encoder 150 encodes the linear data samples Rout to be sent to this first near end; z-law and A-law Encoded voice. -8- * ^ [^ Degrees in use § National Standard (CNS) A4 specification (210 X 297 mm) binding € 5. Description of the invention (6) The ability to transmit or packetize data in IM30 capacity is self-contained This send 1 and receive input decoder, and any network on the receive output decoder 195. This network 145 can be the Internet, an enterprise intranet, an external network, a local area network ( LAN), or a wide area network (WAN ). Send the input decoder ⑽Receive = From the encoding head of 罔 Road 145 and decode this coded speech into the linear language Jinbei Sm. In one embodiment, the input decoder ⑽ is a rule of law decoder The echo channel 17 includes an echo canceller 175 using an affine projection. This echo canceler 175 based on affine projection samples the linear data from this sample and removes the echo estimation signal to generate a linear data sample. The output output encoder 18 provides speech compression before packetization. 纟-In the embodiment, this output output encoder is just used for low-bits including the International Telecommunication Union (ITU) _T international standardization g series. Any compression standard of rate sound (LBRV) is used to compress the voice data Sout G.7xx encoder from the echo channel 170. The compressed speech data is sent to the receiving device at the remote end. The receiving input decoder 190 decodes the voice data received from this second remote end. This decompression technique is compatible with the compression technique used on the output encoder 180. The echo channel 170 receives the scaling from the receiving input decoder 19 and sends this

Rout linear data sampling. The receiving output encoder 19 encodes this linear data sample Rout to encode the heart and A-laws to be sent to the network 145 and the second far end. In one embodiment, the input decoder 160 is sent out, the return channel 70 is sent, the output decoder 18 is received, the input input decoder 190 is received, and the receive output encoder 195 is integrated into a digital signal processor-9.-This paper The standard is applicable to China National Standard (CNS) A4 specification (21 × X 297 public directors) 583851 A7 B7 V. Description of invention (7) 165. Fig. 2 is a diagram illustrating a system model for the affine projection-based echo canceller 125 shown in Fig. 1 according to an embodiment of the present invention. The system model 200 includes a system 210, an adder 220, an adaptive filter 230, and a subtractor 240. This system model 200 builds a model of the echo cancellation process. This echo cancellation process can be modeled as a system identification problem identifying an unknown linear system. The input to this system is an input sequence u (k). System 2 1 0 characterizes the behavior of this system. System 2 10 can be shaped as a finite impulse response (FIR) filter with a weight vector w = [Wi, W2, ..., wN]. This system output is given by the formula y (k) = WTu (k) (1) where is the transpose of this weight vector W. This weight vector W is unknown. The adder 2 2 0 adds a random noise sequence v (k) to the system output y (k) to produce the desired output d (k). The adaptive filter 230 is shaped into a weighted vector w = [Wl, W2, ..., WN] finite impulse response filters. The subtractor 240 subtracts the desired output d (k) from the output of the adaptive filter to produce an error e (k). This error e (k) is used to adjust or update this weight vector W so that the error e (k) is minimized under an objective function (such as the least mean square). The weight vector is updated so that it approaches this weight vector W. This weight vector w therefore represents an estimate of this weight vector w. Fig. 3 is a diagram illustrating an echo canceller 125 based on affine projection shown in Fig. 1 according to an embodiment of the present invention. Echo Canceller -10- Paper Size + g g House Standard Half (CNS) A4 Specification (210X297 Male Eding

Binding

583851 A8 B8 C8 D8

6. The patent application Fan_125 includes an adaptive filter 310 and a delay estimator 320. This adaptive filter 3 10 estimates a channel weight vector W of the echo channel 120 (Fig. 1) updated using the affine projection (a ρ). As shown in FIG. 1, the echo channel 120 receives a sending input sequence Sin and a receiving input sequence Rin. The adaptive filter 310 is used to borrow the delay estimator 320 to calculate these delays. The channel weight vector W has a first length mi. In one embodiment, the first length M1 is 1024. The delay estimator 3 2 0 uses the estimated channel weight vector W provided by the adaptive filter 3 1 0 to determine most of the delays in this echo channel 120. The delay estimator 3 2 0 provides these estimated delays to the adaptive filter 3 3 0. The adaptive filter 3 3 0 uses this affine projection update and the estimated delay from the delay estimator 3 2 0 to estimate multiple short weight vectors. These short weight vectors correspond to the impulse responses of these sparse multipath channels. The length of each of these short weight vectors is typically shorter than this first length M1. The adaptive filters 3 1 0 and 3 3 0 use an affine projection update rule to estimate the channel weight vector. This affine projection method is used to accelerate the convergence of the normalized least-mean-square (NLMS) technique especially for color input. This affine projection updates these weights based on multiple past input vectors, whereas this normalized least mean square updates these weights based on the most recent input vector. The rules for affine projection update are explained below. This input sequence u (k) can be shaped as a P-time automatic regression process, expressed as AR (P): ρ «(k) = Va; * u (ki) ^ z (k) (2) -11-this Paper size applies Chinese National Standard (CNS) A4 specification (210 x 297 mm) 583851 A7 B7 V. Description of invention (9) where z (k) is a white sequence with unit variation. Assuming that P is known in advance, the sampling of u (k) can be written as the quantity u (k), or: One (Mx 1) line direction uT (k) = [u (k), u (kl) ,. .., u (k-M + l)] (3) This AR (P) process can then be written as: P u (k) = [a! · * U (ki) + z (k) i = l = U (k) * a + z (k) where U (k) is the P aggregation of past vectors: (4) U (k) = [u (kl), u (k-2), ... ,, u (kP)] (5) and z (k) is a sampled (Mxl) row vector of a white random sequence. zT (k) = [z (k), z (kl),..., z ( k-M + l)] (6) The least square estimate of the parameter of a is given by the formula: a (k) = [U 丁 (k ^ l ^ k)] · 1 * uT (k) * u ( k) ⑺ where UT (k) * U (k) is assumed to be of order P and * indicates multiplication. The affine projection recursive update rule for // = 1 is defined by the following set of operations: φ (Ι〇 = u (k) -U (k) * a (k) = {IU (k) 3 ^ [ UT (k) * U (k)] · 1 * UT (k)} * u (k) (8) -12- This paper size applies to China National Standard (CNS) A4 (210x 297 mm) 583851 A7 B7

V. Description of the invention (1〇) e (k) = d (k) ~ uT (k) * W (k) (9) * (k ten 1) = w (k) + [φ (1ς) / (φτ (ΐζ) * (j) (k))] * e (k) (10) where you (K) is the channel weight vector estimated by the adaptive filter 3 1 〇 or by the adaptive filter 3 3 〇 The estimated individual short weight vectors. It has been proved that if u (k) is a ρ-time automatic regression process, then yK) is a least square estimate of its auto-learning coefficient and φ (k) is three z (k). In other words, φ (k) is a vector of estimates whose elements are a white random sequence. Fig. 4A is a diagram illustrating an adaptive filter 3 1 0 for an echo canceller based on affine projection shown in Fig. 3 according to an embodiment of the present invention. This affine projection-based adaptive filter 3 1 0 includes an automatic regression (AR) coefficient estimator 4 1 0, a random estimator 4 2 0, an error estimator 4 j 0 ', and a weight updater 4 4 0. The automatic regression coefficient estimator 4 10 estimates the automatic regression coefficient vector (k) at the first update rate using the past received input sequence and the current received sequence. This automatic regression coefficient estimator 4 1 0 generates this automatic regression coefficient vector using equation (7). The random estimator 4 2 0 estimates a random sequence φ (k) at a second update rate using the estimated automatic regression coefficient vector. The random estimator 4 2 0 uses equation (8) to determine this random sequence φ (k). The error estimator 430 estimates the error at the second update rate using the second sequence, the received sequence, and the estimated channel weight vector. This error estimator 4 3 0 uses equation (9) to calculate this error e (k). The weight updater 4 4 0 updates the error in using this estimate, and the estimated -13- This paper size applies the Chinese National Standard (CNS) A4 specification (210X297 mm) A7 A7

The channel weight vector on the second update rate of the k-machine sequence. This weight updater. 4 4 0 uses equation (1 0) to update this channel weight. The first and second update rates may be the same or different. In one embodiment, the first update rate is slower than the second update rate. This second update rate is the rate at which each iteration in the update cycle is repeated. This first update rate is the rate per k repeats in the update cycle, where k is a positive integer. In one embodiment, κ is equal to 100. In other words, the automatic regression coefficient estimator 4 1 0 produces a new result on every k repeats while the random estimator 4 2 0, the error estimator 4 3 0, and the weight updater 4 4 0 Repeatedly produce new results. Fig. 4B is a diagram illustrating an adaptive filter 3 for an echo canceller based on affine projection shown in Fig. 3 according to an embodiment of the present invention. The adaptive filter 33 includes an input separator 45, a partial automatic regression coefficient estimator 460! To 460s, S partial random estimators 470 #] 470S, an error estimator 48, and a weight The value updater 49q / input separator 450 divides the received input sequence u (k) into s partial input sequences. Each of these partial input sequences is delayed by a corresponding delay from the delay provided by the delay estimator 320 (FIG. 3) and is partially partially randomized by one of the following and an overlying partial input sequence weighted by a partial automatic regression coefficient vector And to represent:

Ui (k) = Ui (k) * ai + z, (k), i = 1..., S (11) The s partial automatic regression coefficient estimators 46〇i to 46〇s are coupled to the rotation The separator 45 0 receives this s part of the input sequence u 丨 ㈨, among which 丄 -14- 583851 A7 _____ B7 V. Description of the invention (12 1 ~ ": 1 ··. 'S. These parts are automatically regression coefficient estimated Each of the estimates of the generators 46〇1 to 46〇s automatically uses this part of the regression coefficient vector on this first part of the input sequence and the first update rate of this part of the input sequence:)) = (Ui'k) *! ^) ] · 1 * U, T (k) * u, (k), i = 1, ... S (12) The S partial random estimators 47 (^ to 470s are respectively coupled to these partial automatic regression coefficient estimators 46 (^ To 460s. These partial random estimators 47 (Each estimate of ^ to 470s is this partial random sequence at the second update rate of the partial automatic regression coefficient vector using this estimation: φ / k) —called! 〇 * Although), Bu 1, ... "(13) The error estimator 480 estimates the input sequence sin or d (k), these partial input sequences, and the second update rate of these estimated short weight vectors. Error: e ( k) = d (k)-Σ UiWiT, i = 1, ..., 5 (14) Weight updater 4 9 0 updates on the second update rate using the error of this estimate and the partial random sequence of this estimate Short weight vector Wi (k):

Wi (k + D = w 《k) + [Wk ^ 'k) * She Chuan ~ ⑻, I = 1, · ·, S (15) The first and second update rates may be the same or different. In one embodiment, the first update rate is slower than the second update rate. This second rate is the rate of each repetition in the update cycle. This first update rate 4 updates the rate of every K repetitions in the loop, where K is a positive integer. & -15- V. Description of the Invention (13) In one embodiment, K is equal to 100. In other words, the S partial automatic regression coefficient estimators 460] to 460s produce a new result on every k repeats and the S partial random estimators 46 ° # j46 ° s, the error estimator, and the push value The updater 490 produces new results on each iteration. Fig. 5 illustrates a graph of a delay estimator 3 2 0 for an echo canceller based on affine projection shown in Fig. 2 according to an embodiment of the present invention. The delay estimator 3 2 0 includes a peak locator 510 and a sinusoidal canceller 5 2 0. The delay estimator 3 2 0 藕 estimates the majority of the delay in the echo channel 120 (Figure 1) from the estimated channel weight vector w (k). The delay estimator 32 provides this estimated delay to the adaptive filter 3 3 0 (Fig. 3) for adaptation. The delay estimator 320 basically locates many peaks in the impulse response, and the components provided by the weight vector w (k) are thus estimated. The testicular foot positioner 510 determines that these peaks are the maximum value of the weights in a search area. L peaks are typically located, where L is a positive integer from 5 to 5. In one embodiment, L is equal to five. These peaks are located within a predetermined distance from other peaks. The peak locator 5 includes a square size computer 5 to calculate the square size of the elements of the estimated channel weight vector that determine the values of these peaks. First, this highest peak is determined as the highest value among the weights of the channel weight vector W (k). Then, the second highest t-peak is outside the area covering this highest peak. This area contains L i samples on either side of this highest peak. In one embodiment, B i -25. Then, the third highest peak is outside the area covering the second highest peak. This area contains -16 on either side of this second highest peak. This paper is sized for the National Standard (CNS) A4 size (210X 297 Public Love 583851)

L2 samples. In one embodiment, L2 = 25. This process continues until l symplectic values have been located. The peak canceller 5 2 0 eliminates one of the l peaks located by the peak locator 5] [〇. This error is detected when its value is less than a critical value. In one embodiment, this critical value is equal to the point * highest peak value, where cold is a number between 0.4 and 0.8. In one embodiment, /5=0.6. After all erroneous peaks have been eliminated, these remaining peaks provide a delay, or the location of these peaks. Fig. 6 illustrates a flowchart of a processing for echo cancellation 600 according to an embodiment of the present invention. At the beginning, process 6 0 0 to initialize this loop repeat index k = 1, filter length M = M1, number of repeats Q = Q1, channel weight vector w (k), automatic regression coefficient vector a, random sequence φ (k), error e (k), and past received input sequence U (k) (block 6 10). In one embodiment, M1 = 256, 512, or 104. Next, the process uses this affine projection update to perform an adaptive chirp to estimate the channel weight vector W (k) (block 620). Then, processing 6 0 0 determines whether this repeated index is equal to q 丨 (block 6 2 5). If not, the process 6 0 increases this repetition index k (block 6 2 7) and returns to block 6 2 0. Otherwise, process 600 uses the final estimated channel weight vector W (k) = W (Ql + l) (block 6 3 0) to estimate these delays. Next, process 6 0 0 to initialize this repetition index k = Q 1, filter length M = M2, number of repetitions Q = Q2, parameters W (k), a (k), φ (k), e (k ), U (k) 'and the partial input sequence Rn for each partial adaptive filter (block 640). M 2 may be the same as or smaller than M 1. In an embodiment -17- this paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 mm)

Equipment double 583851 A7 __________ B7 5. In the description of the invention (π), M1 1024 ′ Μ 2 = 60 to 100. Then, process 600 uses this affine / renewal to perform adaptive filtering to estimate the short weight vector Wi (k) (, ghost 50). The calculation and update in block 650 is basically the same as block 6 The calculations and updates in 2: are the same. In this way, the same filter with monthly note κ is now available on both stages. Alternatively, implement two different adaptive filters. It then processes 6 0 0 to determine whether this repeat indicator is equal to Q 2 (block 6 5 5). If not, process 6 0 increases this repeating index k (block 6 5 7) and returns to block 6 5 0. Otherwise, the processing is terminated 600. FIG. 7 is a flowchart illustrating a process 6 2 0 for adaptive filtering not using affine projection update in FIG. 6 according to an embodiment of the present invention. Initially, process 6 2 0 calculates this short-term average power, stavp, and long-term average power, Rinphat (block 71). Then, the process 62 determines whether stavp is less than Rinphat (block 710) by more than a predetermined number (block 715). In one embodiment, this number is 20 decibels. If so, the process 6 2 0 freezes this estimated channel weight vector (block 7 2 0) and estimates the error e (k) as shown in equation (9) (block 7 2 5), and then terminates. Otherwise, processing 6 2 0 stores the estimated automatic regression coefficient vector a (n) (block 7 3 0). Then processing 6 2 0 determines whether it is time to update the automatic regression coefficient vector a (block 7 3 5). . In one embodiment, this first update rate corresponds to every R repetitions, where R = 1 0 0. If it is not the time to update at this time, the estimated automatic regression coefficient vector is maintained as described in box 7-30. ^ The paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 mm) ~ " " '

Binding

583851 A7 B7 V. Description of the invention (the previous values stored are the same (block 7 4 5). Otherwise, according to equation (7) to update this automatic regression coefficient vector (block 7 4 0). Then, process 6 2 0 using the equation (8) to estimate this random sequence φ (k) (block 7 5 0). Then, process 6 2 0 uses equation (9) to estimate this error e (k) (block 7 5 5). Then, process 620 uses Equation (10) updates this channel weight vector W (k) (block 7 6 0) and then terminates. Figure 8 is a diagram illustrating an embodiment according to the present invention for the delay shown in Figure 6. Estimated flowchart of processing 6 3 0. Initially, processing 6 3 0 initializes the peak index i = i and determines the highest peak, peak (k) = max {W (k)}, where W (k ) Is the channel weight vector provided by this adaptive filter (block 8 1 0). The size of the estimated channel weight vector is used to determine this highest peak. Then, the processing 6 3 0 is determined to be located around the peak ( k) the next peak outside the window, next peak = max {W (k) -R}, where R is the area around peak (i), this most recently located peak (block 8 2 0 In one embodiment, r covers 25 samples on either side of a total of 5 i samples of peak (i). Then, process 63 determines whether next-peak is greater than the highest value of stone * (block 830). In one embodiment, cold = 0.6. If next-peak is not greater than 0 * highest peak, then next-peak is assumed to be a false peak and eliminated (block 8 4 0). Otherwise 'Process 6 3 0 decision This peak (i) is next_peak (block 850). Then, the process 630 decides whether the number of pinpoints that have been located so far is equal to $, where s is the total number of peaks that have been pre-positioned (block 86). In one embodiment, S is equal to 3 to 5. If not, process 6 3 0 increases the peak index id + 1 (block 8 7 0), and then returns to block 820. Otherwise, process 63 0-19-this Paper size applies Chinese National Standard (CNS) A4 specification (210X297 public love ·)

Mounting line 583851

The delay corresponds to the position of these peak (i), where 丨 = 丨, s, (block 8 8 0). The processing is then terminated 6 3 0. Fig. 9 illustrates a flowchart of a process 650 for partially adaptive filtering updated using affine projection according to an embodiment of the present invention. Initially, process 6 50 divides this received input sequence into multiple partial input sequences (box 9 0 2). Each of these partial input sequences is delayed by one of the delays provided by this delay estimator. Each partial input sequence is represented by the sum of a partial random sequence and a past partial input sequence weighted by a portion of the automatic regression coefficient vector. Next, the process 65 starts the filter index for the loops of the S partially adaptive filter elements (block 905). It should be noted that although this flowchart shows a sequential operation of this part of the adaptive filter, these operations can be performed sequentially or in parallel. Then, process 65 0 calculates the short-term average power, StaVPi, and long-term average power, Rinphati (block 9), which are associated with this part of the input sequence. Then, processing 6 50 determines whether stavp is less than Rinnap by more than a predetermined number (block 9 1 5). In the example, this number is 20 decibels. If so, the process 65 0 freezes this estimated channel weight vector (block 920) and estimates the error e (k) as shown in equation (14) (block 92 5) and then terminates. Otherwise, process 6 50 stores this estimated auto-regression coefficient vector (box 93 0) and then process 6 50 to determine whether it is time to update this part of the auto-regression coefficient vector a (box 9 35). In one implementation In the example, this first update rate corresponds to every R repetitions, where ⑽. If it is not updated at this time -20- This paper size applies the Chinese National Standard (CNS) A4 specification (210 X 297 public attack ^ ---- -------

Binding € 583851 V. Description of the invention (At 18, the estimated partial automatic regression system, the number vector is kept in the previous value stored in the block :: 0 (block 945). Otherwise, update according to equation (1-) This part of the automatic regression coefficient vector § (block 94). Then, the process 650 uses equation (13) to estimate the random sequence φ (10 (box 9)). Then, the process 65 0 determines whether the filter index is Equal to 5 (10,000 blocks 95 2). If not, process 65 to increase the filter index to reach the next partial filter (block 9 5 3) and return to block 9 丨 〇. Otherwise, process 6 50 0 user-side private ( 9) to estimate this error e (k) (block 9 5 5). Then, process 6 5 0 uses equation (10) to update this channel weight vector (block 9 6 0) and then terminate. Use a partially adaptive filter The echo cancellation of the affine projection update of the router converges faster than the normalized minimum mean square technique used for sparse multipath channels. This echo cancellation based on affine projection can be specifically implemented in different ways. In one embodiment, The length of this adaptive filter 3 1 〇 The degree M 1 (that is, the size of the channel weight vector) is 10 2 4 and the length M 2 of the element of the adaptive filter 3 3 0 ranges from 60 to 100, depending on the multipath. It depends on the characteristics. It should be noted that each sub-filter of this adaptive filter 3 3 0 can have the same or different length. The echo canceller 1 2 5 can be specified by hardware or software or a combination of both. In addition, the adaptive filters 3 1 0 and 3 3 0 can be implemented as a single adaptive filter or separate filters. These results of the AP (2) technique for echo cancellation are shown in the figure 10A, 10B, 10C, 11A, 11B, and 11C. The standard AP (2) method is tested using the data file on G. 168 test 2b (Rin and Sin). -21-This paper standard applies to China National Standard (CNS) A4 Specification (210X297 Public Love) Binding

583851 A7 _________ B7 ___ V. Description of the invention (19) FIG. 10A is a graph illustrating the weights for non-sparse channels after this first adaptation phase according to an embodiment of the present invention. FIG. 10B is a graph illustrating the mean square error (MSE) of 10,000 learning samples before this non-sparse channel according to an embodiment of the present invention. FIG. 10C is a graph illustrating an estimate of the impulse response of the non-sparse channel for this channel after 10,000 samples according to an embodiment of the present invention. Fig. 11A is a graph illustrating weights for sparse channels after this first adaptive phase according to an embodiment of the present invention. Fig. 11B is a graph illustrating the mean square error (MSE) of the first 10,000 learning samples for this sparse channel according to an embodiment of the present invention. Figure ΠC is a graph illustrating an impulse response estimate for a non-sparse channel for this channel after 10,000 samples according to an embodiment of the present invention. Although the invention has been described with reference to the illustrated embodiments, this description is not intended to be construed in a limiting sense. Different modifications of these illustrated embodiments, and other embodiments of the invention, can be considered within the spirit and scope of the invention. -22- This paper size applies to China National Standard (CNS) A4 (210 X 297 mm)

Claims (1)

B8 C8 May D8 Amendment ^ Supplement No. 091118539 Patent Application Chinese Patent Application Replacement (U-92) VI. Application Patent Scope 1. A device for echo cancellation, comprising: a device having a first length to A first adaptive filter for estimating the channel weight vector of an echo channel using an affine projection (AP) update. The echo channel receives a send-out input sequence and a receive input sequence. The return channel has multiple sparse and multiple Path channel; a delay estimator coupled to the adaptive filter to use the estimated channel weight vector to determine multiple delays on the echo channel; and a delay estimator coupled to the delay estimator to use the simulation A second adaptive filter for the projection projection update and the multiple delays to estimate multiple short weight vectors corresponding to the sparse multipath channels, each short weight vector having a length shorter than the first The second length of the length. 2. The device according to item 1 of the patent application scope, wherein the received input sequence is represented by a sum of a random sequence and a past received input sequence weighted by an automatic regression (AR) coefficient vector. 3. The device according to item 2 of the patent application scope, wherein the first adaptive filter includes: an automatic estimation of the automatic regression coefficient vector on the use of the past received input sequence and a first update rate of the received input sequence A regression coefficient estimator; a random estimator that is rounded to the automatic regression coefficient estimator to estimate the random sequence at a second update rate using the estimated automatic regression coefficient vector; an estimate using the sent-out input sequence, The error between the received input sequence and the second update rate of the channel weight vector estimated by the child is 297 mm) 583851 A8 B8 C8 08 Apply for a patent range estimator; and a weight update that integrates the long-distance estimation and the 4 卩 remote machine estimator to update the channel weight vector at the second update rate using the error of the remote estimation and the estimated random sequence Device. 4. The device according to item 1 of the scope of patent application, wherein the delay estimator includes a peak locator using the estimated channel weight vector to locate the peak position of the impulse response of the echo channel, the positions corresponding to The delays; and a derivative canceller to eliminate erroneous peaks in the octave values. 5. The device according to the first item of the patent application scope, wherein the second adaptive filter comprises: an input that separates the received input sequence into a plurality of partial input sequences is A delay that is delayed by one of the delays provided by the delay estimator and is represented by the sum of a part of the random sequence and the past part of the input sequence weighted by the partial knife automatic regression (AR) coefficient vector; The partial automatic regression coefficient estimator of the input separator, each of the partial automatic regression coefficient estimators estimates the partial automatic regression coefficient vector at the first update rate using the past partial input sequence and the partial input sequence; Multiple partial random estimators that are combined into the partial automatic regression coefficient estimator, each of these partial random estimators estimates the partial randomness on the second update rate of the estimated partial knife automatic regression coefficient vector using − 2- This paper rule reduces wealth S ® miscellaneous standard ((: redundant iron full-583851 Λ8 B8 C8 D8 month a) patent application scope supplementary sequence one estimate An error estimator using the output input sequence, the partial input sequence, and the error at the second update rate of the far-estimated moment weights to f; and an error estimator coupled to the error estimator and the random estimator to update the A weight updater using the estimated error and the short weight vector on a second update rate of the estimated partial random sequence. Among them, the first update rate is the first update rate. The device according to item 5 is slower than the second update rate. 7. The device according to item 5 of the patent application scope is slower than the second update rate. Among them, when a short-term average power 8. The device rate according to item 3 of the patent application scope is less than a long-term average. When the power is higher than a critical power number, the weight updater freezes the channel weight vector. 9. According to the device in the scope of the patent application, the package is further included. A plurality of calculations are related to the partial input sequence. An associated partial short-term average power computer; and a plurality of partial short-term average power computers used to calculate the partial short-term average power associated with the partial input sequence Long-term partial power computer. 10. The device according to item 9 of the scope of the patent application, wherein the weight updater freezes when the short-term average power of the associated part is less than the long-term average power of the associated part above a bounded power The short weight vector. '^ 11. A method for echo cancellation, including: The first-adaptive filter updated by using an affine projection (AP) 3- This paper scale applies the Chinese National Standard (CNS) Α4 size (210 × 297 mm) 583851 80 0 ^ Y: ', B8 £ J = Estimating the channel weight vector of a Jin channel, the echo channel receiving a knock-out input sequence and a receiving input sequence, the echo channel has multiple sparse multi-path channels; by using one of the estimated channel weight vectors A delay estimator to determine a plurality of delays in the echo channel; and to estimate a plurality of short weight vectors by using an edge affine projection update and a second adaptation of the plurality of delays ^ Corresponding to the sparse multi-path channel, each of the short weight vectors has a second length shorter than the first length. 12. The method according to the scope of the patent application, wherein the received input sequence is represented by a sum of a random sequence and a past received input sequence weighted by an automatic regression (AR) coefficient vector. 13. The method according to item 11 of the scope of patent application, wherein the first adaptive filter is used to estimate the channel weight vector packet of an echo channel: an automatic regression estimator is used to estimate the received input in the past. The automatic regression coefficient vector at the first update rate of the sequence and the received input sequence; a random estimator coupled to the automatic regression coefficient estimator to estimate a second update using the estimated automatic regression coefficient vector The random sequence at a rate; an error estimator to estimate an error in the second update rate using the send-out input sequence, the received input sequence, and the estimated channel weight vector; and a combination To the violation > Wu difference estimate and the weight of the random estimate is more 4- The paper size applies Chinese National Standard (CNS) A4 specification (21〇x 297 mm) 583851 Year M Day Supplement I Λ 8 R8 C8 The patent application scope new device to update the channel weight vector on the second update rate using the estimated error and the estimated random sequence. H. The method according to item u of the patent application range, wherein determining the plurality of delays includes: using a peak locator of the estimated channel weight vector to locate the peak positions of the impulse response of the echo channel, these positions Corresponds to the delays; and a peak canceller is used to eliminate erroneous values in the peaks. 15. The method according to item 11 of the scope of patent application, wherein estimating the channel weight vector comprises: separating the received input sequence into a plurality of partial input sequences by an input separator, each of the partial input sequences Represented by the sum of a past partial input sequence delayed by one of the delays provided by the delay estimator and weighted by a portion of a random sequence and a portion of an automatic regression (AR) coefficient vector; the estimation is using the past The partial input sequence and the partial automatic regression coefficient vector on the first update rate of the partial input sequence; estimating the partial random sequence on the second update rate using the estimated partial automatic regression coefficient vector; with an error estimate To estimate the error on the input sequence using the send part, and the suitable round-in sequence, the Update the short weight vectors at the second update rate using the error of the estimate and the estimate. Partially random sequences of this paper are in accordance with China National Standard (CNS) A4 (210X 297 mm) 583851 The first update rate is slower than the second update rate according to the method in item 13 of the scope of patent application. 17. Slower than the second update rate according to item 15 of the scope of patent application. Method, wherein the first update rate 18. According to the information in item 13 of the request, updating the channel weight vector includes freezing the short-term average power when the short-term average power is less than a long-term average power higher than a critical power number. Channel weight vector. 19. The method according to item 5 of the scope of patent application, further comprising: • using a plurality of partial short-term power computers to calculate a partial short-term average power associated with the partial input sequences; and using a plurality of long-term partial powers The computer calculates the partial long-term average power associated with the partial input sequence. 20. The method according to item 9 of the scope of patent application, wherein updating the channel weighting order vector includes freezing the short weight when the short-term average power of the associated part is less than the long-term average power of the associated part above a critical power number Value vector. 21. —A system for echo cancellation, comprising: a first decoder coupled to the far end of a sound wave channel to decode a far-end signal, the first decoder generating a received input sequence; A second decoder for the near end of an sonic channel to decode a near-end signal, the second decoder generating a send-out input sequence; and a turn in the echo channel to the first and second decoders to perform echo cancellation The echo canceller of the echo channel receives the receiving and sending input sequences. The echo canceller includes: -6- This paper size applies to the Chinese National Standard (CNS) Α4 specification (210X297 mm) Λ8 B8 C8 D8 No supplement 9 ^ 7th correction The scope of the patent application ^-a first adaptive filter having a first length to estimate a channel weight vector of an echo channel using an affine projection (A p) update, the echo channel receives an input sequence and a Receive input sequence 'the echo channel has multiple sparse multipath channels; a delay estimator coupled to the adaptive filter to use the estimated channel weight vector to determine multiple delays on the echo channel ; And a second adaptive filter that is rounded to the delay estimator to use the affine projection update and the multiple delays to estimate multiple short weight vectors' the short weight vectors correspond to the sparse multiples Path path, each short weight vector has a second length shorter than the first length. 22. The system according to item 21 of the scope of patent application, wherein the received input sequence is represented by the sum of a random sequence and a past received input sequence weighted by an automatic regression (AR) coefficient vector. 23 According to the application The system of claim 21, wherein the first adaptive filter includes: an automatic regression coefficient estimation that estimates the automatic regression coefficient vector at the first update rate using the past received input sequence and the received input sequence -A random estimator coupled to the automatic regression coefficient estimator to estimate the random sequence at a second update rate using the estimated automatic regression coefficient vector; an estimate using the send-out input sequence, the received input sequence , And the error of the error on the second update rate of the channel weight vector estimated by this paper The paper size is suitable mmt ^ (CNS) A4 ^ (2i〇7i ^ y 5 8 3 8 5 AB c D Patch 7 8 1 · June 92. Apply for a patent range estimator; and a coupling to the error estimator and the random estimator to update the error and estimated follow-up The weight updater of the channel weight vector at the second update rate of the sequence. 24. The system according to item 21 of the patent application range, wherein the delay estimator includes:-using the estimated channel weight vector to locate the A peak locator of the peak position of the impulse response of the echo channel, the positions corresponding to the delays; and a peak r value canceller to eliminate the erroneous peaks among the peaks. 25. According to the scope of the patent application A system according to 1, wherein the second adaptive waver includes: an input splitter that separates the received input sequence into a plurality of partial input sequences, and each of the remote partial input sequences is processed by the delay estimator One of the provided delays and is represented by the sum of a partial random sequence and a past partial input sequence weighted by a partial automatic regression coefficient vector; multiple partial automatic regression coefficient estimators coupled to the input separator , Each of the partial partial automatic regression coefficient estimators uses the first partial input sequence and the first partial input sequence to estimate The partial automatic regression coefficient vector at the new rate; multiple partial random estimators coupled to the partial partial automatic regression coefficient estimator, each of these partial random estimators estimates the = Part of the update rate ± random binding§ 8- Λ8 B8 C8 Scope of patent application ►2. I amended supplementary sequence in January, 2011; an error estimator that estimates the error on the second update rate of the input sequence, the partial input sequence, and the estimated short weight vector ; And a weight updater coupled to the error estimator and the random estimator to update the short weight vector at a second update rate using the estimated error and the estimated partial random sequence. 26. The system according to item 23 of the scope of patent application, wherein the first update rate is slower than the second update rate. 27. The system according to item 25 of the patent application scope, wherein the first update rate is slower than the second update rate. 28. The system according to item 23 of the scope of patent application, wherein when a short-term average power is less than a long-term average power is higher than a critical power number, the weight updater freezes the channel weight vector. 29. The system according to item 25 of the scope of patent application, further comprising: a plurality of partial short-term power computers for calculating a partial short-term average power associated with the partial input sequences; and ^ for calculating the short-term power associated with the partial input sequences; A partial input sequence has a long-term partial power computer with associated partial long-term average power. 30. The system according to item 29 of the scope of patent application, wherein the weight updater freezes the short weight vector when the short-term average power of the related part is less than the long-term average power of the related part is higher than a Junjie power number . 31 · — A kind of computer program product, including a computer usable medium with code embedded in it, the code package-9- 583851 A8 B8 C8 D8 92. 11-year Λ yue 7 Patent application scope amendments ^^ ν »% I complements nothing; includes: computer-readable code for estimating a channel weight vector of an echo channel by using a first adaptive filter updated by an affine projection (AP), the echo channel receiving a send An input sequence and a received input sequence, the echo channel having multiple sparse multipath channels; computer-readable by using a delay estimator of the estimated channel weight vector to satisfy multiple delays in the echo channel Code; and computer-readable code for estimating a plurality of short weight vectors by a second adaptive filter using the affine projection update and the plurality of delays, the short weight vectors corresponding to the For sparse multi-path channels, each of the short weight vectors has a second length shorter than the first length. 32. The computer program product according to item 31 of the scope of patent application, wherein the received input sequence is represented by a sum of a random sequence and a past received input sequence weighted by a part of an automatic regression (AR) coefficient vector. 33. The computer program product according to item 31 of the scope of patent application, wherein the computer-readable code used to estimate the channel weight vector includes: used to estimate the received input sequence and the first Computer-readable code for the automatic regression coefficient vector at the update rate; computer-readable code for estimating the random sequence at the second update rate using the estimated automatic regression coefficient vector; used to estimate the in-use rate Computer-readable code for the error at the second update rate of the send input sequence, the receive input sequence, and the estimated channel weight vector; and Binding -10- 583851 7 Amendment; Japanese Patent Application No. J Patent Application Range Computer-readable code for updating the channel weight vector at the second update rate using the estimated error and the estimated random sequence. 34. The computer program product according to item 3i of the scope of patent application, wherein the computer-readable code for determining the plurality of delays includes: using the estimated channel weight vector to locate the impulse response of the echo channel ♦ Computer-readable code for the value locations, the locations corresponding to the delays; and computer-readable code for eliminating false peaks among the peaks. 35. The computer program product according to item 31 of the scope of patent application, wherein the computer-readable code for estimating a plurality of short weight vectors includes: a computer-readable program that separates the received input sequence into a plurality of partial input sequences Code, each of these partial input sequences is delayed by one of the delays and is represented by the sum of a partial random sequence and a past partial input sequence weighted by a partial automatic regression (AR) coefficient vector; it is estimated to be in use The computer-readable code of the partial automatic regression coefficient vector on the past partial input sequence and the first update rate of the partial input sequence estimates a distant portion on the second update rate using the estimated partial automatic regression coefficient vector Computer-readable code for a random sequence, a computer-readable program for estimating the error by an error estimator at the second update rate of the output input sequence, the partial input sequence, and the estimated short weight vector Yards; and -11-this paper is suitable for financial standards s SZJ (CNS) A4 size (210X ^^ 57 583851 A computer-readable program for updating short weights and quantities at a first update rate using a child estimation error and a weighted updater of the random estimator coupled to the child estimation and the random estimator code. 36. The computer program product according to item 33 of the scope of patent application, wherein the first update rate is slower than the second update rate. 37. The computer program product according to item 35 of the scope of patent application, wherein the first update rate is slower than the second update rate. 38. The computer program product according to item 33 of the patent application scope, wherein the computer-readable code for updating the channel weight vector includes freezing a short-term average power when the long-term average power is less than a critical power number Channel weight vector.口 39. The computer program product according to item 35 of the scope of patent application, further comprising: computer-readable code for calculating a part of the short-term average power associated with the part of the input sequence; and for calculating the part of the short-term average power; Part of the input sequence has computer-readable code associated with part of the long-term average power. 40. The computer program product according to item 39 of the scope of the patent application, wherein the computer-readable code for updating the short weight vectors includes when the short-term average power of a part of the coupler is less than the long-term average power of the part For a critical power number, a computer-readable program using the short weight vector to compile the short weight vector 12- This paper size applies Chinese national standards ((;;) ^ $ A4 specification (21〇x 297 mm) Binding i