An echo cancellation method and system based on a non-canonical fir structure

Abstract

The application provides an echo cancellation method and system based on a non-standard FIR structure, and relates to the technical field of telephone communication.The method comprises the following steps: sampling a far-end signal to obtain a discrete value of the far-end signal; filtering an input vector of an adaptive filter at a current time through the adaptive filter to obtain an output vector; wherein the adaptive filter is a non-standard FIR filter, and a tap weight coefficient of an i-th tap of the non-standard FIR filter is calculated at an n-i time; obtaining a near-end sampling signal at the current time, subtracting the near-end sampling signal from the obtained output vector to obtain a residual signal at the current time, and returning the residual signal to the far-end; updating a tap weight coefficient vector of the adaptive filter, and repeating the above steps until the call is ended.The application can reduce the error of echo cancellation in a pulse noise environment.

Description

Technical Field

[0001] This invention relates to the field of telephone communication technology, and specifically to an echo cancellation method and system based on a non-standard FIR structure. Background Technology

[0002] Acoustic echo cancellation (AEC) refers to eliminating noise generated by the feedback path created between the microphone and speaker. In voice communication, if the sound of the distant caller emitted by the speaker is picked up by the microphone, an acoustic loop is formed between the speaker and the microphone, causing the distant caller to hear their own echo, thus affecting call quality. Therefore, it is necessary to cancel the sound from the speaker in the near-end signal picked up by the microphone.

[0003] Adaptive filters can adaptively filter near-end signals to achieve echo cancellation. Adaptive filtering can be achieved through system identification. The core principle of adaptive filtering system identification technology is to construct a speech model based on the input signal and its related characteristics. By analyzing the residual between the input signal and the unknown system, the filter coefficients are continuously adjusted so that the estimated value gradually approaches the parameters of the unknown system. FIR (Finite Impulse Response) filters, also known as non-recursive filters, can maintain strictly linear phase frequency characteristics while ensuring arbitrary amplitude-frequency characteristics. Furthermore, their unit sample response is finite, making them stable systems. FIR filters can be used for adaptive filtering of near-end signals.

[0004] The Maximum Correlation Entropy (MCC) algorithm has attracted increasing attention in sparse system identification due to its robustness to non-Gaussian noise and its ability to handle impulse noise. The MCC algorithm is characterized by its ability to maintain a relatively fast convergence speed in impulse noise environments and effectively suppress fluctuations in algorithm performance.

[0005] However, the drawback of the FIR-based MCC algorithm is its inability to operate at higher frequencies, limiting its use in high-frequency applications. Furthermore, the performance of the MCC algorithm degrades under extremely sparse or noisy conditions. Impulse noise has always been a significant challenge in system identification. Although it has a low probability of occurrence and short duration, its large amplitude can cause drastic fluctuations in the convergence performance of traditional algorithms. The high amplitude of impulse noise may superimpose with the output of the unknown system, causing the coefficients of the adaptive filter to be updated towards a larger error. Therefore, developing and optimizing adaptive filter algorithms with excellent noise resistance for sparse system identification under impulse noise environments has become a challenge in the field of system identification. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide an echo cancellation method and system based on a non-standard FIR structure, so as to reduce the error of echo cancellation in impulse noise environments.

[0007] To achieve the above objectives, according to some embodiments, a first aspect of the present invention provides an echo cancellation method based on a non-canonical FIR structure, comprising:

[0008] Step A: Sample the remote signal to obtain the discrete value of the remote signal. Use the discrete values ​​of the remote signal at the current time n and the previous L-1 times as the input vector of the adaptive filter at the current time, where L is the tap length of the adaptive filter.

[0009] Step B: Filter the input vector of the adaptive filter at the current time to obtain the output vector; wherein, the adaptive filter is a non-normalized FIR filter, and the tap weight coefficients of the i-th tap of the non-normalized FIR filter are calculated at time ni.

[0010] Step C: Obtain the near-end sampling signal at the current moment, subtract the near-end sampling signal from the output vector obtained in step B, obtain the residual signal at the current moment, and send it back to the far end;

[0011] Step D: Update the tap weight coefficient vector of the adaptive filter. The tap weight coefficient vector of the adaptive filter at the next time step is obtained from the tap weight coefficient vector of the adaptive filter at the current time step and the cost function. The cost function is obtained according to the maximum correlation entropy criterion.

[0012] Step E: Let n = n + 1, and repeat steps A, B, C, and D until the call ends.

[0013] A second aspect of the present invention provides an echo cancellation system based on a non-standard FIR structure, comprising:

[0014] The remote signal sampling module is configured to sample the remote signal to obtain the discrete value of the remote signal, and use the discrete values ​​of the remote signal at the current time n and the previous L-1 times as the input vector of the adaptive filter at the current time, where L is the tap length of the adaptive filter.

[0015] An adaptive filtering module is configured to filter the input vector of the adaptive filter at the current time to obtain the output vector; wherein the adaptive filter is a non-normalized FIR filter, and the tap weight coefficients of the i-th tap of the non-normalized FIR filter are calculated at time ni.

[0016] The residual signal calculation module is configured to acquire the near-end sampled signal at the current moment, subtract the near-end sampled signal from the obtained output vector to obtain the residual signal at the current moment, and send it back to the far end.

[0017] The adaptive filter tap weight coefficient update module is configured to update the tap weight coefficient vector of the adaptive filter. The tap weight coefficient vector of the adaptive filter at the next time step is obtained based on the tap weight coefficient vector of the adaptive filter at the current time step and the cost function. The cost function is obtained based on the maximum correlation entropy criterion.

[0018] The iterative processing module is configured to set n=n+1 and repeat the operations of the remote signal sampling module, the adaptive filtering module, the residual signal calculation module, and the adaptive filter tap weight coefficient update module until the call ends.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] This invention provides an echo cancellation method and system based on a non-canonical FIR structure. By employing a maximum correlation entropy algorithm based on a non-canonical FIR (NCFIR) structure, it overcomes the problems of high time processing cost and inability to operate at high frequencies caused by the structural complexity of existing FIR filters for echo cancellation. This invention achieves lower time processing cost and stable operation at higher frequencies. Compared with existing algorithms, this invention not only significantly reduces additional errors but also does not significantly increase computational complexity.

[0021] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a schematic diagram of a non-standard FIR filter in Embodiment 1 of the present invention;

[0024] Figure 2 This is a schematic diagram of the MSD curves of the method provided in Embodiment 1 of the present invention and existing algorithms for identifying sparse systems under Gaussian noise conditions;

[0025] Figure 3This is a schematic diagram comparing the accuracy of the method provided in Embodiment 1 of the present invention and the MCC algorithm in tracking the standard solution; wherein (a) is a schematic diagram of the accuracy of the MCC algorithm in tracking the standard solution, and (b) is a schematic diagram of the accuracy of the method provided in Embodiment 1 of the present invention in tracking the standard solution. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Example 1

[0028] Embodiment 1 of the present invention provides an echo cancellation method based on a non-canonical FIR structure, comprising:

[0029] Step A: Sample the remote signal to obtain the discrete value of the remote signal. Use the discrete values ​​of the remote signal at the current time n and the previous L-1 times as the input vector of the adaptive filter at the current time, where L is the tap length of the adaptive filter.

[0030] Step B: Filter the input vector of the adaptive filter at the current time to obtain the output vector; wherein, the adaptive filter is a non-normalized FIR filter, and the tap weight coefficients of the i-th tap of the non-normalized FIR filter are calculated at time ni.

[0031] Step C: Obtain the near-end sampling signal at the current moment, subtract the near-end sampling signal from the output vector obtained in step B, obtain the residual signal at the current moment, and send it back to the far end;

[0032] Step D: Update the tap weight coefficient vector of the adaptive filter. The tap weight coefficient vector of the adaptive filter at the next time step is obtained from the tap weight coefficient vector of the adaptive filter at the current time step and the cost function. The cost function is obtained according to the maximum correlation entropy criterion.

[0033] Step E: Let n = n + 1, and repeat steps A, B, C, and D until the call ends.

[0034] Existing maximum correlation entropy (MCC) algorithms based on standard FIR structures, while possessing strong robustness, suffer from high processing costs due to their structural complexity, limiting their applicability in high-frequency applications. Therefore, this embodiment provides an echo cancellation method based on non-canonical FIR (NCFIR) maximum correlation entropy (MCC), abbreviated as NCMCC. This method achieves faster convergence speed and lower steady-state error in impulse noise environments, reduces processing latency, supports higher frequency operations, improves efficiency, and expands the possibilities of signal processing. Compared to existing algorithms, the method provided in this embodiment not only significantly reduces additional errors but also does not significantly increase computational complexity.

[0035] Specifically, in step A, the discrete value of the remote signal transmitted from the remote end is obtained by sampling the remote signal at the current time n. , take the current time n and the previous time n Discrete values ​​of the far-end signal at each time point The input vector that makes up the adaptive filter at the current time n Wherein, the superscript T indicates transpose, and L is the tap length of the adaptive filter. In this embodiment, L is 32.

[0036] In step B, the adaptive filter used in this embodiment is a non-canonical FIR (NCFIR) filter. Based on a standard FIR filter, the non-canonical FIR filter first reverses the direction of the addition lines to align the input and output; then it moves the position of the delay unit so that the delay unit is on the addition line, i.e., a retiming technique; finally, it changes the output position and the order of the filter coefficients to optimize the critical processing path and improve processing efficiency. Taking a 5-tap non-canonical FIR filter as an example, its structure is as follows... Figure 1 As shown in the figure The input signal to the filter. , , , , Here are the tap weight coefficients for each tap. This is the output signal of the filter. Compared to standard FIR filters, non-standard FIR filters have less output delay because their structure allows the signal to pass through the filter earlier. In echo cancellation, also known as system identification, reducing delay improves real-time performance and is helpful for online identification and control systems.

[0037] The NCFIR filter has better numerical stability because it can reduce the accumulated error in the multiplication operation of the filter coefficients. Numerical stability directly affects the accuracy of the identification results, so NCFIR has better system identification accuracy than traditional FIR.

[0038] Taking a 5-tap filter as an example, for a standard FIR filter, the signal needs to go through one multiplier and four adders from input to output, resulting in a higher time delay. However, for an NCFIR filter, the signal only needs to go through one multiplier and one adder from input to output, resulting in a shorter time delay. Therefore, using an NCFIR filter can save computation time and achieve stable operation at high frequencies.

[0039] After the non-canonical FIR (NCFIR) filter structure transformation, the filter error feedback path includes: This time delay leads to unrealizable non-causal time variations. To address this, this embodiment further constructs a causal adaptive framework, the output of which is:

[0040] ;

[0041] in, This represents the output vector at the current time n; Represents the discrete value of the remote model at time ni; This represents the input vector of the adaptive filter at the current time n; , represents the tap weight coefficient vector of the adaptive filter after the time offset at the current time n; express The tap weight coefficients at time i are calculated. At this time, the tap weight coefficients are offset compared to the tap weight coefficients of the FIR filter. The tap weight coefficients of the traditional FIR filter are from a unified time, while the tap weight coefficients after the structural transformation are from different times. This time offset will reduce the additional error, thereby improving the accuracy of system identification.

[0042] Echo signal estimation is performed based on the constructed causal adaptive framework. The input vector of the adaptive filter at the current time n is... The output vector at current time n is obtained through the NCFIR filter. That is, the estimated value of the echo signal:

[0043] ;

[0044] in, This is the time offset coefficient. ; This is the time offset matrix. ,in Represents a diagonal matrix; This is the tap weight coefficient vector of the adaptive filter at the current time n; This represents the tap weight coefficient calculated at time n on the i-th tap; This represents the discrete value of the remote model at time ni.

[0045] Introducing time offset coefficients and time offset matrices into echo signal estimation can yield an output form similar to that of FIR structures, thus meeting the computational requirements for echo cancellation.

[0046] In step C, the echo signal is cancelled. The far-end input signal, i.e., the input vector... Through unknown sparse systems and interference noise The near-end sampling signal is generated after superposition. The echo-bearing near-end sampling signal acquired at the current time n. The output vector of the adaptive filter at the current time n Subtracting them yields the residual signal at the current time n. This is the signal after echo cancellation; then the residual signal at the current time n... Send back to the remote end.

[0047] In step D, the tap weight coefficient vector of the adaptive filter is updated. The tap weight coefficient vector of the adaptive filter at the next time step n+1 is... The formula is calculated based on the tap weight coefficient vector of the adaptive filter at the current time n, combined with the error function.

[0048] ;

[0049] in, Let n be the tap weight coefficient vector of the adaptive filter at the current time n. This indicates taking the derivative of the error function. It is the step size parameter.

[0050] Based on the maximum correlation entropy algorithm, the cost function is... for:

[0051] ;

[0052] in, Indicates the kernel width.

[0053] right Differentiation yields:

[0054] ;

[0055] At this point, the step size parameter That is, corresponding part; That is, corresponding Partial. Therefore, the tap weight coefficient vector of the adaptive filter at the next time step n+1. The formula can be expressed as:

[0056] .

[0057] This embodiment integrates the maximum correlation entropy algorithm into the NCFIR filter, further improving the performance of echo cancellation based on the NCFIR filter.

[0058] To verify the effectiveness of the method provided in this embodiment, experiments were conducted, and its performance was compared with that of the SA algorithm, LMS algorithm, MCC algorithm, and LMP algorithm.

[0059] In the simulation experiment, the tap length L of the adaptive filter is 32, the input signal at the far end is a first-order autoregressive (AR(1)) signal, the input variance of Gaussian white noise is 1, and the background noise is a Gaussian Bernoulli distribution (where the signal-to-noise ratio of Gaussian noise is 30dB and the collision probability Pr is 0.01). The specific values ​​of the step size parameters of each algorithm in the experiment are shown in Table 1.

[0060] Table 1. Step size parameter values ​​for each algorithm

[0061]

[0062] Figure 2 This is a schematic diagram showing the MSD (Mean Squared Deviation) curves of the LMS algorithm, LMP algorithm, SA algorithm, MCC algorithm, and the method of this invention (NCMCC) for identifying sparse systems under Gaussian noise conditions. Figure 2 As can be seen, the LMS algorithm exhibits the slowest convergence rate and the lowest accuracy. In contrast, the SA and LMP algorithms perform moderately well in terms of convergence performance and accuracy, and are relatively better than the LMS algorithm. Further observation shows that the method provided in this embodiment and the MCC algorithm have significantly improved convergence speed and accuracy, especially in the early stage of algorithm iteration, where their performance is particularly outstanding.

[0063] Figure 3 This is a schematic diagram comparing the accuracy of the MCC algorithm and the method provided in this embodiment in tracking the standard solution. Figure 3 In this context, RKM (Runge-Kutta Method) is a method for numerically solving ordinary differential equations; here, it is considered the standard solution. From... Figure 3 As can be seen, with the progress of the iteration process, the method provided in this embodiment, namely... Figure 3 In (b), compared to the MCC algorithm, it exhibits higher recognition accuracy and stability, and a higher fit to the RKM curve, while the result of the MCC algorithm, i.e. Figure 3 In (a), significant fluctuations were observed, revealing its relatively low system stability when dealing with the same problem.

[0064] Example 2

[0065] This embodiment provides an echo cancellation system based on a non-standard FIR structure, including:

[0066] The remote signal sampling module is configured to sample the remote signal to obtain the discrete value of the remote signal, and use the discrete values ​​of the remote signal at the current time n and the previous L-1 times as the input vector of the adaptive filter at the current time, where L is the tap length of the adaptive filter.

[0067] An adaptive filtering module is configured to filter the input vector of the adaptive filter at the current time to obtain the output vector; wherein the adaptive filter is a non-normalized FIR filter, and the tap weight coefficients of the i-th tap of the non-normalized FIR filter are calculated at time ni.

[0068] The residual signal calculation module is configured to acquire the near-end sampled signal at the current moment, subtract the near-end sampled signal from the obtained output vector to obtain the residual signal at the current moment, and send it back to the far end.

[0069] The adaptive filter tap weight coefficient update module is configured to update the tap weight coefficient vector of the adaptive filter. The tap weight coefficient vector of the adaptive filter at the next time step is obtained based on the tap weight coefficient vector of the adaptive filter at the current time step and the cost function. The cost function is obtained based on the maximum correlation entropy criterion.

[0070] The iterative processing module is configured to set n=n+1 and repeat the operations of the remote signal sampling module, the adaptive filtering module, the residual signal calculation module, and the adaptive filter tap weight coefficient update module until the call ends.

[0071] It should be noted that each module in this embodiment corresponds one-to-one with the steps of the method in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An echo cancellation method based on a non-standard FIR structure, characterized in that, include: Step A: Sample the remote signal to obtain the discrete value of the remote signal. Use the discrete values ​​of the remote signal at the current time n and the previous L-1 times as the input vector of the adaptive filter at the current time, where L is the tap length of the adaptive filter. Step B: Filter the input vector of the adaptive filter at the current time to obtain the output vector; wherein, the adaptive filter is a non-normalized FIR filter, and the tap weight coefficients of the i-th tap of the non-normalized FIR filter are calculated at time ni. A denormalized FIR filter, based on a standard FIR filter, first reverses the direction of the addition lines to align the input and output; then it moves the position of the delay unit so that it lies on the addition line; finally, it changes the output position and the order of the filter coefficients to optimize the critical processing path and improve processing efficiency. After the denormalized FIR filter structure transformation, the filter error feedback path includes... For each time delay, a causal adaptive framework was constructed, and its output is: ； in, This represents the output vector at the current time n. This represents the discrete value of the remote model at time ni. Let n represent the input vector of the adaptive filter at the current time n. This represents the tap weight coefficient vector of the adaptive filter after the time offset at the current time n. express The tap weight coefficient at time i is calculated. Echo signal estimation is performed based on the constructed causal adaptive framework: the input vector of the adaptive filter at the current time n is... The output vector at time n is obtained through a non-normal FIR filter. That is, the estimated value of the echo signal: ； in, This represents the output vector at the current time n. This is the time offset coefficient. This is the time offset matrix. ,in Represents a diagonal matrix. This is the tap weight coefficient vector of the adaptive filter at the current time n; This represents the tap weight coefficient calculated at time n on the i-th tap; express The discrete value of the remote model at any given time; The formula for calculating the time offset factor is: ； in, express The tap weight coefficient at time i is calculated. This represents the tap weight coefficient calculated at time n on the i-th tap. Step C: Obtain the near-end sampling signal at the current moment, subtract the near-end sampling signal from the output vector obtained in step B, obtain the residual signal at the current moment, and send it back to the far end; Step D: Update the tap weight coefficient vector of the adaptive filter. The tap weight coefficient vector of the adaptive filter at the next time step is obtained based on the tap weight coefficient vector of the adaptive filter at the current time step and the error function, as shown in the formula: ； in, Let n be the tap weight coefficient vector of the adaptive filter at the current time n. This indicates taking the derivative of the error function. It is the step size parameter; Based on the maximum correlation entropy algorithm, the cost function is... for: ； in, Indicates the kernel width; right Differentiation yields: ； At this point, the step size parameter That is, corresponding part; That is, corresponding In part, the formula for calculating the tap weight coefficient vector of the adaptive filter at the next time step is: ； in, For residual signals, Indicates the kernel width. This is the time offset matrix. This represents the input vector of the adaptive filter at the current time n; Step E: Let n = n + 1, and repeat steps A, B, C, and D until the call ends.

2. An echo cancellation system based on a non-standard FIR structure, characterized in that, include: The remote signal sampling module is configured to sample the remote signal to obtain the discrete value of the remote signal, and use the discrete values ​​of the remote signal at the current time n and the previous L-1 times as the input vector of the adaptive filter at the current time, where L is the tap length of the adaptive filter. An adaptive filtering module is configured to filter the input vector of the adaptive filter at the current time to obtain the output vector; wherein the adaptive filter is a non-normalized FIR filter, and the tap weight coefficients of the i-th tap of the non-normalized FIR filter are calculated at time ni. A denormalized FIR filter, based on a standard FIR filter, first reverses the direction of the addition lines to align the input and output; then it moves the position of the delay unit so that it lies on the addition line; finally, it changes the output position and the order of the filter coefficients to optimize the critical processing path and improve processing efficiency. After the denormalized FIR filter structure transformation, the filter error feedback path includes... For each time delay, a causal adaptive framework was constructed, and its output is: ； in, This represents the output vector at the current time n. This represents the discrete value of the remote model at time ni. Let n represent the input vector of the adaptive filter at the current time n. This represents the tap weight coefficient vector of the adaptive filter after the time offset at the current time n. express The tap weight coefficient at time i is calculated. Echo signal estimation is performed based on the constructed causal adaptive framework: the input vector of the adaptive filter at the current time n is... The output vector at time n is obtained through a non-normal FIR filter. That is, the estimated value of the echo signal: ； in, This represents the output vector at the current time n. This is the time offset coefficient. This is the time offset matrix. ,in Represents a diagonal matrix. This is the tap weight coefficient vector of the adaptive filter at the current time n; This represents the tap weight coefficient calculated at time n on the i-th tap; express The discrete value of the remote model at any given time; The formula for calculating the time offset factor is: ； in, express The tap weight coefficient at time i is calculated. This represents the tap weight coefficient calculated at time n on the i-th tap. The residual signal calculation module is configured to acquire the near-end sampled signal at the current moment, subtract the near-end sampled signal from the obtained output vector to obtain the residual signal at the current moment, and send it back to the far end. The adaptive filter tap weight coefficient update module is configured to update the tap weight coefficient vector of the adaptive filter. The tap weight coefficient vector of the adaptive filter at the next time step is obtained based on the tap weight coefficient vector of the adaptive filter at the current time step and the error function, as shown in the formula: ； in, Let n be the tap weight coefficient vector of the adaptive filter at the current time n. This indicates taking the derivative of the error function. It is the step size parameter; Based on the maximum correlation entropy algorithm, the cost function is... for: ； in, Indicates the kernel width; right Differentiation yields: ； At this point, the step size parameter That is, corresponding part; That is, corresponding In part, the formula for calculating the tap weight coefficient vector of the adaptive filter at the next time step is: ； in, For residual signals, Indicates the kernel width. This is the time offset matrix. This represents the input vector of the adaptive filter at the current time n; The iterative processing module is configured to set n=n+1 and repeat the operations of the remote signal sampling module, the adaptive filtering module, the residual signal calculation module, and the adaptive filter tap weight coefficient update module until the call ends.

Images