CN113347536B - Acoustic feedback suppression algorithm based on linear prediction and sub-band adaptive filtering - Google Patents

Abstract

The invention discloses an acoustic feedback suppression method based on linear prediction and sub-band adaptive filtering. The linear prediction unit receives an external input sound source to suppress howling, and the adaptive filtering unit receives the signal processed by the linear prediction unit. The acoustic environment is modeled, the echo signal in the acoustic feedback is cancelled, and the echo is eliminated. The invention overcomes the biased estimation of the highly colored signal by the traditional sub-band adaptive filter, and can not only model the acoustic environment more accurately, but also effectively follow the changing acoustic environment. The invention can effectively suppress the acoustic feedback and reduce the influence on the original sound, and can be applied to medical hearing aids, sound reinforcement systems and the like.

Description

Acoustic Feedback Suppression Algorithm Based on Linear Prediction and Subband Adaptive Filtering

technical field

The invention relates to an acoustic feedback suppression algorithm in the field of digital signal processing, in particular to an acoustic feedback suppression method based on linear prediction and adaptive filtering that can be used for sound reinforcement systems, medical hearing aids and other equipment.

Background technique

A sound reinforcement system usually refers to a system that amplifies the speaker's voice in real time, and the speaker and the audience are usually in the same acoustic environment. It is also because the acoustic environment of the microphone and the speaker is not isolated that the problem of acoustic feedback has always existed since the sound reinforcement system was put into use.

During the working process of the sound reinforcement system, when the sound signal is captured by the microphone, amplified by the amplifier circuit, and played back by the speaker in the same environment, the speaker signal is inevitably fed back to the microphone to form a closed loop. Once the loop is formed, the echo tail will affect the audience's experience, and the howling effect will cause the sound reinforcement system to overload and damage the equipment.

Acoustic feedback contains a variety of frequency components, which can be divided into echo and howling according to the intensity and frequency spectrum width of the acoustic feedback. When the sound reinforcement system has a certain delay or the gain is small, and the acoustic feedback is not enough to destroy the Nyquist stability condition, an echo will be generated. The echo belongs to the repetition of the original sound signal and has a wide spectrum range. To suppress this type of acoustic feedback, a filter is generally used to model the sound field environment. By inputting the speaker signal to the filter to generate an estimate of the echo signal, the echo estimate is subtracted from the newly collected signal by the microphone to achieve the purpose of echo cancellation. If the gain of the system is large enough, the microphone is sensitive enough, or the distance between the microphone and the speaker is close, the input signal is likely to have frequency components that destroy the Nyquist amplitude and phase conditions at the same time and form positive feedback. Point will produce the so-called howling. According to the mechanism of howling, it can be known that howling is a narrowband signal. In practical applications, frequency shifters are often used to destroy the phase condition or notch filters are used to destroy the amplitude condition for suppression.

Since the adaptive filter can actively adapt to the changing acoustic environment, it is especially suitable for the modeling of the acoustic environment of hearing aids or the modeling of sound reinforcement systems where items and people change frequently. The sub-band adaptive filter can significantly reduce the computational complexity of the adaptive filter acoustic environment modeling, and at the same time, dividing the sub-band can effectively decorrelate the echo signal and speed up the convergence speed of the adaptive filter. The algorithm is widely used in the field of acoustic feedback suppression.

However, since the howling of the narrowband signal is highly colored, when the acoustic feedback contains howling that makes the sound reinforcement system fully loaded, the sub-band adaptive algorithm will produce a biased estimate of the acoustic environment, and the built model deviates from the real sound field, resulting in echo signals There is a deviation in the estimation of , which makes the sound after subtracting the echo estimation produce a certain degree of distortion, which seriously affects the speech recognition.

Contents of the invention

The purpose of the present invention is to solve the problem that the above-mentioned adaptive filter produces biased estimation of the acoustic environment, resulting in sound distortion, and provides an acoustic feedback suppression method that combines an adaptive linear predictor and a sub-band adaptive filter.

The technical scheme adopted for realizing the purpose of the present invention is:

The acoustic feedback suppression method based on linear prediction and sub-band adaptive filtering, the linear prediction unit receives an external input sound source for howling suppression, and the adaptive filtering unit receives the signal processed by the linear prediction unit to model the acoustic environment , to cancel the echo signal in the acoustic feedback and eliminate the echo.

Wherein, the linear prediction unit belongs to the inverse application of the adaptive spectral line enhancer, and is used for suppressing the narrow-band howling signal in the acoustic feedback.

Wherein, the linear prediction unit predicts the narrowband component in the current signal through the linear combination of the past signals, and removes the predicted narrowband signal from the current signal, and then supplements with an adaptive algorithm, based on the error signal that has eliminated the narrowband component The weight of the linear predictor is adjusted so that the linear predictor unit changes according to the change of the narrowband signal, and finally behaves as an adaptive narrowband notch filter.

Wherein, the adaptive algorithm includes least mean square algorithm LMS, normalized least mean square algorithm NLMS and algorithms derived from least mean square algorithm LMS and normalized least mean square algorithm NLMS.

Wherein, the adaptive filtering unit calculates an estimate of the echo signal by using a sub-band adaptive algorithm, obtains a difference between the input signal and the estimate of the echo, and restores the original speech.

Wherein, the subband adaptive algorithm includes least mean square algorithm LMS, normalized least mean square algorithm NLMS, and algorithms derived from least mean square algorithm LMS and normalized least mean square algorithm NLMS.

The invention overcomes the biased estimation of the highly colored signal by the traditional sub-band adaptive filter, and can not only model the acoustic environment more accurately, but also effectively follow the changing acoustic environment. The invention can effectively suppress the acoustic feedback and reduce the influence on the original sound, and can be applied to medical hearing aids, sound reinforcement systems and the like.

Description of drawings

FIG. 1 is a schematic structural diagram of a linear prediction unit provided by an embodiment of the present invention;

FIG. 2 is a schematic structural diagram of an adaptive filtering unit provided by an embodiment of the present invention;

FIG. 3 is a schematic structural diagram of a subband adaptive filtering unit provided by an embodiment of the present invention;

Fig. 4 is a structural diagram of an acoustic feedback suppression device based on linear prediction and sub-band adaptive filtering provided by an embodiment of the present invention.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, not to limit the present invention.

The acoustic feedback suppression algorithm based on linear prediction and sub-band adaptive filtering in the embodiment of the present invention is realized by a linear prediction unit and an adaptive filtering unit, the linear prediction unit receives an external input sound source to suppress howling, and the automatic The adaptive filtering unit receives the signal processed by the linear prediction unit, models the acoustic environment, and eliminates the echo.

In the following, the working principle of each unit will be described first, and then the overall working process will be described by combining the two.

The linear prediction unit in the embodiment of the present invention belongs to the inverse application of the adaptive spectral line enhancer, and is used for suppressing the narrow-band howling signal in the acoustic feedback. Since the adjacent samples of the narrowband signal are highly correlated, the narrowband component in the current signal can be predicted through a reasonable linear combination of the past signals, and the predicted narrowband signal can be removed from the current signal. The linear prediction unit acts as a notch filter; supplemented by an adaptive algorithm, the weight (tap coefficient) of the linear predictor is adjusted according to the error signal that has eliminated the narrowband component, and the linear prediction unit will change according to the change of the narrowband signal, and finally The linear prediction unit behaves as an adaptive narrowband notch filter.

The adaptive algorithm that can be adopted by the linear prediction unit in the embodiment of the present invention includes but not limited to least mean square algorithm (LMS), normalized least mean square algorithm (NLMS) and derived algorithms.

This unit belongs to the inverse application of the adaptive spectral line enhancer, which is used to remove the narrow-band howling signal in the acoustic feedback. In a preferred embodiment of the algorithm, the linear prediction unit adopts a finite impulse response (FIR) structure.

Fig. 1 is a schematic diagram of the structure of the linear prediction unit involved in the embodiment of the present invention, the original voice signal is marked as s(n), the howling signal introduced by the acoustic feedback is marked as g(n), and y(n) is doped with The howling voice signal, x(n) represents the speaker signal, and n represents that the signal is the sample at the nth moment.

For howling signal g(n), its samples are correlated. Therefore, according to the correlation between its past samples g(n-1), g(n-2), ... and the current sample g(n), the predicted value g'(n) of g(n) can be obtained by formula (1) Prediction is done in the shown way, where b _k is the weight of the past signal of g(n), and N is the length of the predictor.

Since the correlation of y(n) is mainly introduced by the howling signal g(n), the predicted value y′(n) of y(n) can be approximately considered as the predicted value g′(n) of g(n), namely g'(n)≈y'(n). Therefore, g'(n) can also be predicted by y(n), as shown in formula (2), where a _k is the weight of the past signal of y(n).

As shown in Figure 1, an adaptive algorithm is introduced to dynamically adjust the weights of past samples y(n-1), y(n-2), ... according to the error signal z(n) of y(n) and g′(n). a _k , and taking the error signal z(n) as the output signal, the linear prediction unit is equivalent to a notch filter that can be adaptive according to the howling frequency. G in Fig. 1 represents the FIR filter with a _k as the tap coefficient.

Specifically, if the adaptive algorithm adopts the NLMS algorithm, set the number of taps as N, and define the following vector:

Then, the sample weights can be iterated according to the following steps:

The first step, (the weight is assigned the initial value):

The second step, (howling estimation):

The third step, (calculate the error signal): z(n)=y(n)-g'(n)

The fourth step, (weight iteration):

代表向量，μ为NLMS算法迭代步长。Among them, subscript T represents the transpose of matrix or vector,

Represents a vector, and μ is the iteration step size of the NLMS algorithm.

The second to fourth steps are performed in a loop, and the linear prediction unit equivalent to a notch filter can converge to the howling frequency point. Taking the error signal z(n) as the output signal, y(n) with the howling component removed can be obtained.

Since the linear prediction unit is only sensitive to highly correlated narrowband signals, if y(n) does not contain howling components, the linear prediction unit is embodied as an all-pass filter. Looking at Figure 1, there is only one subtraction operation between the output signal z(n) of the linear prediction unit and the input signal y(n), so the linear prediction unit is approximately without delay, which is compared with the traditional notch filter Another advantage, which makes the linear prediction unit especially suitable for fields such as medical hearing aids that require relatively high real-time performance.

The adaptive filtering unit in the embodiment of the present invention is used to model the acoustic environment and cancel the echo signal in the acoustic feedback. Although there is a certain correlation between the echo signal and the original speech signal, but because there is a certain time interval between the echo signal and the original speech signal, and it has a wide spectrum range, the linear prediction unit represented as a notch filter cannot It is effectively controlled. The usual way to suppress echo is to use adaptive filters to model the acoustic environment. The adaptive filtering unit in the embodiment of the present invention calculates the estimate of the echo signal by using the sub-band adaptive algorithm, and obtains the estimated value of the input signal and the echo estimate. difference to restore the original voice.

In the embodiment of the present invention, the adaptive filter unit also introduces an adaptive filter to model the acoustic environment, and adopts a sub-band adaptive filter algorithm with better performance to offset the noise with a wide spectral range in the acoustic feedback. echo signal. On the one hand, the sub-band adaptive filtering algorithm divides the input signal into sub-bands, which can effectively remove the correlation between the original speech and the echo signal, and speed up the convergence of the adaptive filter; on the other hand, each sub-band signal can be sampled to a smaller Compared with the traditional adaptive filter, the computational complexity is greatly reduced.

Fig. 2 is a schematic structural diagram of an adaptive filtering unit involved in the present invention. There are various structures of the sub-band adaptive filter. In the following, a typical adaptive filter structure shown in FIG. 3 is taken as an example, and illustrated in conjunction with FIG. 2 . In Figure 2, A represents the analysis filter bank that divides the full-band signal into M sub-band signals, and S is a synthesis filter bank that is approximately the inverse matrix of A, which is used to reconstruct the sub-band signal into a full-band signal, and the signal is sequentially passed through A After , S, the Perfect Reconstruction condition is satisfied, that is, only a pure time delay is introduced.

The original speech signal is denoted as s(n), the echo signal introduced by the sound field F is denoted as f(n), d(n) is the speech signal mixed with the echo signal, x(n) represents the loudspeaker signal, f′(n ) is the echo estimation signal. d _i (n) and x _i (n) in Figure 2 are respectively the i-th subband component after d(n) and x(n) are divided into subbands by A; f′ _i (n) is x _i (n ) is the echo estimation signal of the i-th sub-band obtained after F′ _i filtering; e _i (n) represents the error signal of the i-th sub-band, which can be synthesized by S into a full-band error signal e(n). F' is the sound field model established by the adaptive algorithm.

Specifically, as shown in Figure 3, an _adaptive algorithm is introduced in each sub-band, and in the i-th sub-band, the error signal e i (n) of d _i (n) and f′ _i (n) is used to dynamically adjust the sound field model F' _i can make the full-band error signal e(n) gradually approach the original speech signal s(n).

If the NLMS algorithm is used in the sub-band, the number of sub-band adaptive filter taps is set to L, and the j-th tap coefficient of the i-th sub-band adaptive filter F′ _i (n) is denoted as F′ _{i, j} (n) , define the following vectors:

F′ _i (n)=[F′ _i,0 (n)F′ _i,1 (n)...F′ _i,L-1 (n)]

Then, the tap coefficient F′ _i (n) of the i-th subband can be iterated according to the following steps:

The first step, (assign the initial value of the tap coefficient):

The second step, (echo estimation):

The third step, (calculate error signal): e _i (n)=d _i (n)-f' _i (n)

The fourth step, (weight iteration):

和粗体字母代表向量，α为NLMS算法迭代步长。Among them, subscript T represents the transpose of matrix or vector,

and bold letters represent vectors, and α is the iteration step size of the NLMS algorithm.

The second to fourth steps are performed circularly in each subband, and F' can gradually converge to the acoustic environment F. Taking the error signal e(n) as the output signal can effectively cancel the echo component in d(n).

Fig. 4 is an overall structural diagram of a device for implementing an acoustic feedback suppression method based on linear prediction and sub-band adaptive filtering provided by an embodiment of the present invention. The present invention combines the above two units to simultaneously remove the narrow-band howling signal and the wide-spectrum echo signal in the acoustic feedback. The linear prediction unit receives an external input sound source and performs howling suppression, and the sub-band adaptive filtering unit receives the signal processed by the linear prediction unit, models the acoustic environment, and suppresses the echo signal.

The above is only a preferred embodiment of the present invention, it should be pointed out that, for those of ordinary skill in the art, without departing from the principle of the present invention, some improvements and modifications can also be made, these improvements and Retouching should also be regarded as the protection scope of the present invention.

Claims (1)

1. The acoustic feedback suppression method based on linear prediction and sub-band adaptive filtering, is characterized in that, the external input sound source is received by the linear prediction unit for howling suppression, and the signal processed by the linear prediction unit is received by the adaptive filtering unit, Model the acoustic environment, cancel the echo signal in the acoustic feedback, and eliminate the echo; finally realize the simultaneous removal of the narrow-band howling signal and the wide-spectrum echo signal in the acoustic feedback; The linear prediction unit belongs to the inverse application of the adaptive spectral line enhancer, and is used to suppress the narrow-band howling signal in the acoustic feedback; the linear prediction unit is equivalent to a notch filter that can be adaptive according to the howling frequency point; The linear prediction unit equivalent to a notch filter can converge to the howling frequency point; if the input signal y(n) does not contain a howling component, the linear prediction unit is embodied as an all-pass filter; the output signal z( There is only one subtraction operation between n) and the input signal y(n); The linear prediction unit predicts the narrowband component in the current signal through the linear combination of the past signals, removes the predicted narrowband signal from the current signal, and then supplements it with an adaptive algorithm to adjust the linearity according to the error signal from which the narrowband component has been removed. The weight of the predictor makes the linear prediction unit change according to the change of the narrowband signal, and finally behaves as an adaptive narrowband notch filter; The adaptive algorithm includes least mean square algorithm LMS, normalized least mean square algorithm NLMS and least mean square algorithm LMS, the algorithm derived from normalized least mean square algorithm NLMS; The adaptive filtering unit introduces an adaptive filter to model the acoustic environment, calculates an estimate of the echo signal by using a sub-band adaptive algorithm, obtains the difference between the input signal and the echo estimate, and restores the original voice ; On the one hand, the sub-band adaptive algorithm divides the input signal into sub-bands to remove the correlation between the original voice and the echo signal, and accelerate the convergence speed of the adaptive filter; on the other hand, each sub-band signal is sampled to a predetermined low Sampling rate processing; The sub-band adaptive algorithm includes least mean square algorithm LMS, normalized least mean square algorithm NLMS, and algorithms derived from least mean square algorithm LMS and normalized least mean square algorithm NLMS.

Images