CN203104592U - Device for cancelling echoes in miniature hands-free voice communication system - Google Patents

Abstract

The utility model discloses a device for cancelling echoes in a miniature hands-free voice communication system. The system comprises a voice receiver, a main voice transmitter and an auxiliary voice transmitter, wherein the distance between the main voice transmitter and the voice receiver is longer than the distance between the auxiliary voice transmitter and the voice receiver. The device comprises an array echo cancelling unit, a self-adaptive echo cancelling unit and a residual echo cancelling unit which are sequentially cascaded in the structure. The signal of the main voice transmitter and the signal of the auxiliary voice transmitter are inputted into the array echo cancelling unit, and one path of output signals is obtained through array filtering. The signal of the voice receiver, the output signal of the array echo cancelling unit and the signal of the auxiliary voice transmitter are inputted into the self-adaptive echo cancelling unit, and two paths of output signals are obtained through self-adaptive filtering. The two paths of output signals of the self-adaptive echo cancelling unit are inputted into the residual echo cancelling unit, and voice signals in which the echoes are cancelled are obtained through voice probability estimate and echo matching. Duplex performance can be improved, and the phase consistency of the voice transmitters is not strictly required.

Description

Echo eliminating device for small hands-free voice communication system

Technical Field

The utility model relates to an echo cancellation technical field, in particular to an echo cancellation device for among small-size hands-free voice communication system.

Background

The echo problem often exists in voice communication, and after being played by a receiver (also called a receiver, SPK, or an earphone), a signal at a receiving end is serially transmitted into a receiving signal at a transmitting end (also called a microphone or a sound pick-up) through a line and acoustic reflection, and is fed to a far end, so that a communication contact at the far end hears the echo. Echo can cause great interference to both parties of a call and influence the call quality. Echo can also cause howling and damage to the receiver when the echo is large. In order to ensure the communication quality and the equipment safety, echo suppression is required in voice communication.

The echo can be divided into two types from the generation mechanism, a linear echo component generated by the amplification of an electro-acoustic line and the acoustic transmission, and a nonlinear echo component generated by the nonlinear distortion of a receiver and the acoustic transmission. Linear echo component cancellation typically employs adaptive echo cancellation techniques, which are well established techniques that are widely used to cancel linear echo components without damaging the near-end speech. But the cancellation of the nonlinear echo component is likely to cause near-end speech impairment, resulting in poor duplex performance and even channel half-duplex.

The half-duplex phenomenon is common in small hands-free voice communication devices, such as mobile phones or speakerphones (speakerphones) with hands-free functions, because the receiver nonlinear distortion and nonlinear echo components of such devices are large. With the continuous improvement of the requirements for the smoothness and comfort of voice communication, the echo is suppressed, and meanwhile, the near-end voice is protected, and the duplex effect is ensured. Since duplex loss occurs mainly in nonlinear echo component cancellation, improvements in nonlinear echo component cancellation techniques are particularly desirable.

Echo cancellation on small hands-free voice communication equipment, one type of way to improve duplex performance is to combine echo filtering with array spatial filtering with the help of a microphone array, and to realize echo extraction and voice separation by using the signal difference of echo propagation to each microphone, for example, the method mentioned in the utility model patent application with chinese patent application No. 201110326010.0, can realize near full-duplex conversation by applying array signal processing and echo cancellation. However, this method requires accurate determination of the directions of arrival of the echo and near-end speech at the microphone array, and thus has a high requirement for microphone consistency, which requires not only microphone sensitivity consistency but also phase consistency, and thus requires strict acoustic design.

Disclosure of Invention

The utility model provides an echo cancellation device among small-size hands-free voice communication system alleviates the damage to near-end pronunciation when reducing the echo, promotes duplex performance, does not strictly require the phase consistency of microphone simultaneously.

The utility model discloses a device that is arranged in small-size hands-free voice communication system echo cancellation, small-size hands-free voice communication system includes receiver, main microphone and supplementary microphone, and the distance between main microphone and the receiver is greater than the distance between supplementary microphone and the receiver, and wherein, the device is including structurally cascading in proper order: the array echo cancellation unit inputs a main microphone signal and an auxiliary microphone signal, and removes partial linear echo components and partial nonlinear echo components in the main microphone signal through array filtering to obtain an output signal; the input signals of the self-adaptive echo cancellation unit are telephone receiver signals, output signals of the array echo cancellation unit and auxiliary microphone signals, and residual linear echo components in the main microphone signals and linear echo components in the auxiliary microphone signals are respectively removed from the output signals of the array echo cancellation unit and the auxiliary microphone signals through self-adaptive filtering to obtain two paths of output signals; the input signal of the residual echo eliminating unit is two paths of output signals of the self-adaptive echo eliminating unit, and residual nonlinear echo components in the signal of the main microphone are removed through voice probability estimation and echo matching to obtain one path of output signal which is used as a voice signal after echo elimination.

The embodiment of the utility model provides a beneficial effect is: according to the device for echo cancellation in the small hands-free voice communication system of the utility model, the acoustic characteristics of the small hands-free voice communication system and the position information of the transmitter and the receiver are fully utilized, and the amplitude matching echo is obtained by carrying out overall contour matching and amplitude matching on different transmitter signals; by utilizing the amplitude difference from the echo to different microphones, as the amplitude difference of signals of different microphones is larger and the probability of the near-end speech is smaller, the proportional speech probability information of the indication speech and the echo in each time-frequency region can be extracted, the speech region and the echo region are divided, the residual echo is effectively removed, the near-end speech is protected, the duplex performance is improved, and meanwhile, the phase consistency of the microphones is not strictly required.

Drawings

Fig. 1 is a schematic diagram of the positions of a receiver and a transmitter applied to the echo cancellation device of the present invention for a small hands-free voice communication system;

fig. 2 is a schematic diagram of a desktop small hands-free voice communication system to which the echo cancellation device of the present invention is applied;

fig. 3 is a schematic diagram of a small hands-free voice communication system for a vehicle to which the echo cancellation device of the present invention is applied;

FIG. 4(a) is a graph showing the energy curve of echo components in two microphone signals;

FIG. 4(b) is a diagram illustrating energy curves of near-end speech components in the main and auxiliary microphone signals;

FIG. 4(c) is a diagram showing the energy curves of the echo component and the near-end speech component in the main microphone;

FIG. 5(a) is a schematic diagram of the signal energy curve of the main microphone;

FIG. 5(b) is a diagram illustrating the energy curve of the echo component of the main microphone;

FIG. 5(c) is a schematic diagram of the energy curve of the near-end speech component of the main microphone;

fig. 6 is a block diagram schematically illustrating an echo cancellation device for a small hands-free voice communication system according to a preferred embodiment of the present invention in a use state;

FIG. 7 is a diagram illustrating transfer functions of respective signal components from the auxiliary microphone to the main microphone;

FIG. 8 is a graph illustrating the energy curves of the main microphone signal and the array filtered output signal;

FIG. 9 is a schematic diagram of an energy curve of an output signal of the array filter and a first adaptive filtered signal obtained by adaptive echo filtering;

FIG. 10 is a graph illustrating an energy curve of a nonlinear echo component of a first adaptively filtered signal and a second adaptively filtered signal;

FIG. 11 is a graph illustrating an energy curve of a non-linear echo and an energy curve of a matched echo of a first adaptively filtered signal;

FIG. 12(a) is a schematic diagram of the energy curve of the main microphone signal;

FIG. 12(b) is a graph of the energy curve of the near-end speech component in the main microphone signal;

FIG. 12(c) is a graph showing the energy curve of the output signal after echo cancellation;

fig. 13 is a flowchart of a method for using the apparatus for echo cancellation in a small hands-free voice communication system according to a preferred embodiment of the present invention.

Fig. 14 is a detailed flowchart of a method for using the apparatus for echo cancellation in a small hands-free voice communication system according to a preferred embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention clearer, embodiments of the present invention will be described in further detail with reference to the accompanying drawings.

Fig. 1 is a schematic diagram of the positions of a receiver and a transmitter applied to the device for echo cancellation in a small hands-free voice communication system according to the present invention. Fig. 2 is a schematic diagram of a small hands-free voice communication system to which the echo cancellation device of the present invention is applied. Fig. 3 is a schematic diagram of the use of the hands-free small-sized voice communication system for vehicle, in which the echo cancellation device of the present invention is applied. The user's distance to the microphones (e.g., microphones) is approximately equal for the user, i.e., each microphone receives substantially the same near-end speech signal from the user. However, for the receivers (such as speakers), the distances from the receivers to the respective microphones are not equal, if the distances from the receivers to the main microphone are D1 and the distances from the auxiliary microphone are D2, when D1> D2, the two microphones receive echoes emitted from the receivers with power difference, and the near-end voices received by the two microphones and emitted from the users are almost the same, so that the voices and the echoes can be distinguished according to the difference of power relationship, and the purpose of separating the voices and the echoes is achieved. The present invention is directed to separating speech and echo by using this power difference. In this embodiment, D1> =2D2, for example, D1=13cm, D2=4cm, and at this time, due to the obvious difference between D1 and D2, the power difference between echoes emitted from receivers received by two microphones is obvious, and the effect is better.

FIG. 4(a) is a graph showing the energy curve of echo components in two microphone signals; FIG. 4(b) is a schematic diagram of the energy curves of the near-end speech component in the main microphone signal and the auxiliary microphone signal; fig. 4(c) is a diagram showing an energy curve of an echo component and a near-end speech component in the main microphone. FIG. 5(a) is a schematic diagram of the signal energy curve of the main microphone; FIG. 5(b) is a diagram illustrating the energy curve of the echo component of the main microphone; FIG. 5(c) is a diagram illustrating the energy curve of the near-end speech component of the main microphone. Summarizing, the microphone signal has the following characteristics:

first, as can be seen from fig. 4(a) and 4(b), the echo components in the two microphone signals have a significant energy difference, and the echo component in the auxiliary microphone signal is higher than the echo component in the main microphone signal by 6dB or more. This is because the echo energy is approximately inversely proportional to the distance between the microphone and the receiver, the auxiliary microphone is closer to the receiver, and the received echo is larger.

Second, it can be seen from fig. 4(b) that the near-end speech component energies in the microphone signals are close, because in general application, the mouth of the near-end speaker is close to the two microphones at an equal distance, and therefore the near-end speech energies received by the two microphones are also close.

Third, as can be seen from fig. 4(c), in the main microphone, the near-end speech is slightly lower than the echo energy, which is about 3 to 6dB lower. In addition, as can be seen from the spectrogram in fig. 5(a) -5 (c), when the near-end speech and the echo occur simultaneously, the near-end speech is masked by the echo in some time-frequency regions.

Fig. 6 is a block diagram of an apparatus for echo cancellation in a small hands-free voice communication system according to a preferred embodiment of the present invention in a use state. Fig. 7 is a diagram showing transfer functions of respective signal components from the auxiliary microphone to the main microphone. Fig. 8 is a graph of the energy curves of the main microphone signal and the array filtered output signal. Fig. 9 is a schematic diagram of an energy curve of an output signal of the array filter element and a first adaptive filtered signal obtained by adaptive echo filtering. Fig. 10 is a graph illustrating energy curves of nonlinear echo components of the first and second adaptively filtered signals. Fig. 11 is a diagram illustrating an energy curve of a non-linear echo and an energy curve of a matched echo of a first adaptively filtered signal. Fig. 12(a) is a schematic diagram of an energy curve of a main microphone signal. Fig. 12(b) is a diagram showing an energy curve of a near-end speech component in the main microphone signal. Fig. 12(c) is a diagram showing an energy curve of an output signal after echo cancellation.

The utility model provides a device that is arranged in small-size hands-free voice communication system echo cancellation is by array echo cancellation unit 610, self-adaptation echo cancellation unit 620 and remain echo cancellation unit 630 and constitute, array echo cancellation unit 610, self-adaptation echo cancellation unit 620 and remain echo cancellation unit 630 are the cascade relation structurally, the input of array echo cancellation unit 610 is main microphone signal d1 and supplementary microphone signal d2, output signal d 1' all the way, through array filtering, get rid of some linear echo components and some nonlinear echo components in the main microphone signal; the input signals of the adaptive echo cancellation unit 620 are the receiver signal x, the output signal d1 'of the array echo cancellation unit 610 and the auxiliary microphone signal d2, and the two output signals e1 and e2, respectively remove the residual linear echo component in the main microphone signal from the output signal d 1' of the array echo cancellation unit 610 and the linear echo component in the auxiliary microphone signal from the auxiliary microphone signal d2 by adaptive filtering; the input signal of the residual echo cancellation unit 630 is two output signals e1 and e2 of the adaptive echo cancellation unit 620, and the residual nonlinear echo component in the main microphone signal is removed through speech probability estimation and echo matching, and the output signal is the output after echo cancellation, i.e. the speech signal after echo separation. The echo is cancelled and the near-end speech signal v is left more intact by the processing of the array echo cancellation unit 610, the adaptive echo cancellation unit 620 and the residual echo cancellation unit 630.

More specifically, the apparatus comprises: the system comprises an array filtering component, two self-adaptive filtering components, two time-frequency transformation components, a voice probability estimation component, a frequency spectrum filtering component and a frequency-time transformation component.

The array filter component includes an array filter 611 and a subtractor 612, the array filter 611 is used for performing array filtering on the auxiliary microphone signal d2 to obtain a second array filtered signal, and the subtractor 612 is used for subtracting the second array filtered signal from the main microphone signal d1 to remove a part of linear echoes and a part of non-linear echoes in the main microphone signal to obtain an output signal d 1'. The array filter has spatial directivity, and can detect the sound emitted from the receiver position, and can eliminate a part of the sound emitted from the receiver position after passing through the subtracter. Since both the linear echo component and the nonlinear echo component are transmitted from the receiver, both the linear echo component and the nonlinear echo component are partially removed after the array filter and the subtractor process.

d 1' the energy of the echo components is significantly attenuated relative to d1, while the speech energy does not change significantly. The principle is that the receiver is located close to the two microphones and has a significant difference in distance, and the near-end speech is located far away from the two microphones and close to each other. The propagation characteristics of near-end speech and echo to the microphone are quite different, which is reflected in the transfer function of the two microphone signals. By the difference in transfer functions, speech and echo can be distinguished. The array filter 611 is designed according to the transfer function between echo components, and the echo is removed by the way of array filtering cancellation, while the near-end speech component is not affected.

Let d_echo1Being the echo component in the main microphone, d_sph1Is the near-end speech component in the main microphone signal. d_echo2As echo component in the auxiliary microphone, d_sph2Is the near-end speech component in the auxiliary microphone signal. The primary and secondary microphone signals may be expressed as:

d_i＝d_echoi+d_sphi,i＝1,2（1）

let h be the transfer function between echo components and hN be the transfer function between near-end speech components, then there is the relation:

d_echo1＝d_echo2*h,d_sph1＝d_sph2*hN（2）

if h is not equal to hN, and approximately obtain the array filter similar to h

Then d 1' is obtained by means of filter elimination:

${{\mathrm{<figure-callout\; id="1"\; label="d"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">d</figure-callout>}}_1}^{\&prime;}=d_1-{\mathrm{<figure-callout\; id="2"\; label="d"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">d</figure-callout>}}_2*\hat{h}---\left(3\right)$

combining the above equations (1), (2) and (3), the relation between the echo component and the near-end speech component in d 1' can be obtained:

${{\mathrm{<figure-callout\; id="1"\; label="d"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">d</figure-callout>}}_1}^{\&prime;}=d_{\mathrm{sph}1}-{\mathrm{<figure-callout\; id="2"\; label="d\; sph"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">d</figure-callout>}}_{\mathrm{sph}}*\hat{h}$

$=d_{\mathrm{echo}1}-{\mathrm{<figure-callout\; id="2"\; label="d\; echo"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">d</figure-callout>}}_{\mathrm{echo}}*\hat{h}+d_{\mathrm{sph}1}-{\mathrm{<figure-callout\; id="2"\; label="d\; sph"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">d</figure-callout>}}_{\mathrm{sph}}*\hat{h}$ （4）

$={\mathrm{<figure-callout\; id="2"\; label="d\; echo"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">d</figure-callout>}}_{\mathrm{echo}}*(h-\hat{h})+{\mathrm{<figure-callout\; id="2"\; label="d\; sph"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">d</figure-callout>}}_{\mathrm{sph}}*(\hat{h}-\mathrm{hN})$

${d_{\mathrm{echo}1}}^{\&prime;}+{{\mathrm{<figure-callout\; id="1"\; label="d\; sph"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">d</figure-callout>}}_{\mathrm{sph}}}^{\&prime;}$

comparing equations (1) and (4) shows that if equations (5) and (6) below or equations (5) and (7) below are satisfied, the energy of the echo component in d 1' can be significantly attenuated without significant change in the speech energy, and the objective of reducing the echo while protecting the speech can be achieved, where equations (6) and (7) are approximately equivalent forms:

$E{(h-\hat{h})}^2<E{\left(h\right)}^2---\left(5\right)$

$E{(\hat{h}-\mathrm{hN})}^2>=E{\left(\mathrm{hN}\right)}^2---\left(6\right)$

$E(\hat{h}\&CenterDot;\mathrm{hN})<\frac{E\left[{\hat{h}}^2\right]}{2}---\left(7\right)$

if it is not

If the estimate is more accurate, equation (5) may be satisfied, if this is the case

Completely different from hN, then

The condition of equation (7) can be satisfied. It is ensured that the array filtering cancels the echo without attenuating the speech.

In the apparatus shown in FIG. 1, since the ratio of D1 to D2 is large, for example, in the present embodiment, D1>2D2 and D1-D2>6cm, the condition that h is completely different from hN is satisfied. Both h and hN are close to a single peak function with a peak width around 0.25 ms and a half-width around 0.125 ms. Due to the definition of D1>2D2, the energy difference between h and hN is more than 6dB, and the peak value absolute value difference is more than 2 times. The definition of D1-D2>6cm allows the peak positions of h and hN to be different by more than 0.17 ms, exceeding the half-width of the peak, and the peaks to be completely staggered in time. The transfer function of the individual signal components from the auxiliary microphone to the main microphone can be seen in fig. 7, where the transfer function h between the echo components is a solid line and is a unimodal curve with a peak at a time delay of (D1-D2)/c, where c is the speed of sound traveling in air and its maximum amplitude is approximately D2/D1. With D1=13cm, D2=4cm, the peak height is about 0.3 and the peak position is at 0.26 milliseconds. The transfer function hN between near-end speech components is a dotted curve, also shaped as a single peak, with a peak at 0 milliseconds and a peak height of 1. It can be seen that the two transfer functions are distinct.

In practical systems, the transfer function of the array filter

Can be calculated off-line and fixed in advance. A more accurate calculation may use a minimum mean square error criterion, such as equation (8), whereIs the transfer function of the array filter, d₁Is the main microphone signal, d₂Is the auxiliary microphone signal E.]To evaluate the desired operation. Is the convolution operation:

$\frac{\&PartialD;E\left[{(d_1-{\mathrm{<figure-callout\; id="2"\; label="d"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">d</figure-callout>}}_2*\hat{h})}^2\right]}{\&PartialD;\hat{h}}=0---\left(8\right)$

the output D1 'of the array filter module 610 is formula (3), and taking D1=13cm and D2=4cm as examples, the effect of the array echo cancellation can be seen in fig. 8, where the solid line is the energy curve of the main microphone signal D1 and the dotted line is the energy curve of the output D1' of the array filter module, and it can be seen that the echo energy variation causes the energy average reduction of about 9 dB.

The two adaptive filtering components, each including a filter 621, a filtering controller 622 and a subtractor 623, are configured to perform adaptive filtering on the auxiliary microphone signal d2 to obtain a second adaptive filtered signal e2 to remove linear echo components in the auxiliary microphone signal, and perform adaptive filtering on the signal d 1' obtained by subtracting the second array filtered signal from the main microphone signal to obtain a first adaptive filtered signal e1 to remove residual linear echo components in the main microphone signal.

The adaptive echo cancellation unit has three inputs, a receiver signal x and an auxiliary microphone signal d2, and an output signal d 1' of the array filter module. The outputs are the first adaptive filtered signal e1 and the second adaptive filtered signal e2 after adaptive filtering. The working principle of the part is similar to that of the general adaptive echo filtering, and the time domain or frequency domain filtering form can be adopted. By comparing the similarity of the receiver end signal x and the d2 and d1 ', the echo signal is matched and eliminated from d2 and d 1'. The effect can be seen in fig. 9, where the solid line is the energy curve of the output d 1' of the array filter assembly and the dashed line is the energy curve of the result after adaptive echo filtering, i.e. the first adaptively filtered signal e 1. The energy variation can be seen, with the energy of e1 being reduced by about 5dB on average over the energy of d 1' in the echo region.

The two time-frequency transformation components, each of which includes a data buffer 631 and a time-frequency transformer 632, are configured to perform time-frequency transformation on the first adaptive filtered signal E1 to obtain a first adaptive frequency domain signal E1, and perform time-frequency transformation on the second adaptive filtered signal E2 to obtain a second adaptive frequency domain signal E2, respectively.

The data buffer 631 is used to form signal vectors for the time-frequency transformer 632, and the length of the data buffer 631 is set to L, which is related to the computing resources and can be generally set to 256 or 512. If the input signals are e1(n) and e2(n) at the current time n, the vectors formed in the two data buffers 631 are [ e1(n-L +1), e1(n-L +2) …. e1(n) ] and [ e2(n-L +1), e2(n-L +2) …. e2(n) ], respectively.

The time-frequency transformer 632 transforms the signal from the time domain to the frequency domain, which may be implemented by fourier transform, or by modified discrete digital cosine transform. Taking the example of fourier transform, the frequency domain signal is:

$E_1\left(k\right)=\underset{m=1}{\overset{L}{\mathrm{\&Sigma;}}}{e}_1(n-L+m)W^{k(m-1)},W=\exp (-\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">j</figure-callout>}\frac{2\mathrm{\&pi;}}{M})$

$E_2\left(k\right)=\underset{m=1}{\overset{L}{\mathrm{\&Sigma;}}}{e}_2(n-L+m)W^{k(m-1)},W=\exp (-\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">j</figure-callout>}\frac{2\mathrm{\&pi;}}{M})$

the speech probability estimating component 633 is configured to perform frequency domain speech probability estimation according to the amplitudes of the first adaptive frequency domain signal E1 and the second adaptive frequency domain signal E2 to obtain a frequency domain speech probability signal pF, where the frequency domain speech probability signal pF represents a proportion of a near-end speech signal in the first adaptive frequency domain signal E1. This is because:

the speech probability estimation component 633 compares the amplitude relationship of the two signals to obtain frequency domain speech probability information pF, which is a time-frequency function indicating the ratio of speech and echo in each time-frequency region. pF is 1, which indicates that the near-end speech signals exist in the region, pF is 0, which indicates that the echo signals exist in the region, the value between 0 and 1 indicates that the near-end speech signals exist in the region as a mixture, the probability of the existence of the near-end speech signals is high when the value is close to 1, and the existence of the near-end speech signals is low when the value is close to 0.

The working principle of the voice probability estimation is that two input signals E1 and E2 of the voice probability estimation component 633 both contain nonlinear echo components and near-end voice signals, the energy of the nonlinear echo components in E1 is low, and the energy of the nonlinear echo components in E2 is about 20dB higher than that of the nonlinear echo components in E1. Therefore, in the time-frequency region where the echo is located, the amplitude of E2 is much higher than that of E1, and in the time-frequency region where the near-end speech is located, the energy of E1 and E2 are closer. By comparing the amplitudes of E1 and E2 at the respective frequency points, the distribution of the nonlinear echo components and near-end speech in frequency can be known.

The specific implementation method in this embodiment is to calculate the amplitude ratio of E1 and E2, and obtain the speech probability according to the mapping relationship:

$\mathrm{pF}\left(f\right)=\left\{\begin{array}{cc}1& |{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2/{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1|<T_S\\ (T_E-|{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2/{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1\left|\right)/(T_E-T_S)& |{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2/{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1|\&Element;[T_S,T_E]\\ 0& |{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2/{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1|>T_E\end{array}\right.$

wherein T is_SIs the average amplitude difference, T, of the speech components in the two microphone signals_EIs the average amplitude difference of the echo components in the two microphone signals. After the voice probability judgment, the voice probability of each frequency point can be calculated. Where f is the frequency.

T_SAnd T_EIs related to the structure, taking the structure of FIG. 1 as an example, T_S=1.4, where 1.4 corresponds to 3dB, and represents the difference in the amplitude of the near-end speech component received by the two microphones, which is the sensitivity tolerance of the same type of microphone in industrial production. If D1=13cm and D2=4cm, T_E=4, 4 is an approximation of the amplitude ratio average of the echo components in the two microphones, which is about (D1/D2) × 1.4. The echoes propagating to two microphones, the echo signalsThe difference in amplitude was D1/D2, with a sensitivity tolerance of 1.4 being added, and 4 being rounded down. 2.6 is T_SAnd T_EThe difference between the two is set so as to make the function curve continuous. The meaning of this equation is that when the amplitude difference of the signals is within 1.4, i.e. not exceeding the sensitivity tolerance, it is likely to be a speech component, with a speech probability of 1. When the signal amplitude difference exceeds 4, much like the echo component amplitude difference value, it is likely to be an echo component, and the speech probability is zero. The middle portion is fitted with a one-time slope curve, with lower probability values closer to 4.

The calculation method is as follows:

$\mathrm{pF}\left(f\right)=\left\{\begin{array}{cc}1& |{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2/{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1|<1.4\\ (4-|{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2/{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1\left|\right)/2.6& |{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2/{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1|\&Element;\left[\mathrm{1.4,4}\right]\\ 0& |{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2/{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1|>4\end{array}\right.$

the spectral filtering component comprises an echo matcher 634, a subtracter 635 and a multiplier 636, wherein the echo matcher 634 is used for performing echo matching according to the energy of the nonlinear echo components of the first adaptive frequency domain signal E1 and the second adaptive frequency domain signal E2 to obtain matched echo; the subtractor 635 is configured to subtract the first adaptive frequency domain signal E1 from the matched echo; the multiplier 636 is used to multiply the result of the subtraction with the frequency domain speech probability signal pF.

The echo matcher 634 spectrally estimates the nonlinear echo component and the linear echo component from the correlation of the nonlinear echo linear echo components in the two microphone signals and suppresses them by filtering. The nonlinear echo component and the linear echo component are generated by the receiver and propagated to the main microphone and the auxiliary microphone, and the nonlinear echo component and the linear echo component in the main microphone signal d1 and the auxiliary microphone signal d2 have high similarity, which is mainly represented by the consistency of the spectral peak positions of the nonlinear echo component and the linear echo component. Since the spectral peak concentrates almost all the energy of the nonlinear echo component and the linear echo component, if the spectral peak positions coincide, it can be considered that the frequency distribution laws of the nonlinear echo component and the linear echo component are consistent. If the spectral peaks of the nonlinear echo component and the linear echo component are suppressed, most of the nonlinear echo component and the linear echo component can be removed. Both the array echo cancellation unit 610 and the adaptive echo cancellation unit 620 perform only linear filtering, only the amplitudes and spectral envelope shapes of the nonlinear echo components and the linear echo components are changed, but the spectral peak positions are not changed, i.e., the similarity relationship between the nonlinear echo components and the linear echo components is still maintained. The spectral peak positions of the nonlinear echo components are thus highly approximated in the first and second adaptive frequency-domain signals E1 and E2. It can be seen on the spectrum shown in fig. 10 that the peak positions of the nonlinear echo component and the linear echo component in E1 and E2 are the same or close, but the overall undulation shape and the signal energy are different.

Therefore, the nonlinear echo linear echo components of E1 and E2 can be subjected to overall contour matching to obtain matched echoes, and then multiplied by a certain factor Ag to be subjected to amplitude matching to obtain amplitude matched echoes. The factor Ag decreases as the speech probability pF increases. In the time-frequency region with lower voice probability and higher echo probability, the amplitude of the matched echo is higher than the nonlinear echo linear echo component, and in the region with higher voice probability and lower echo probability, the amplitude obtained by multiplying the matched echo by Ag is equal to or slightly lower than the nonlinear echo linear echo component. The residual echo may be removed by subtracting the amplitude matched echo from the first adaptive frequency domain signal E1.

Generally, in order to remove the residual echo completely, the stronger the residual echo, the larger Ag, and the stronger the spectrum filtering, but the more the damage to the near-end speech. In addition, interference of the residual echo to communication is different between pure echo and double-talk, when pure echo exists, people are sensitive to the residual echo, little residual echo causes discomfort, but in a double-talk area where echo and near-end voice occur simultaneously, people are not sensitive to the residual echo, and the requirement on the near-end voice quality is high. The utility model discloses in owing to used the pronunciation probability estimation subassembly, estimate and the echo matching combines together through the pronunciation probability for it is less that the value can be got to match the echo, and not too strong spectrum filtering just can get rid of the linear echo composition of nonlinear echo, and spectrum filtering intensity changes along with near-end pronunciation probability, and when near-end pronunciation probability is higher, the spectrum filtering dynamics reduces, with better protection near-end pronunciation. Therefore, the filtering strength is dynamically adjusted along with the voice probability, so that the comfort level and the voice quality can be simultaneously improved, and the near-end voice can be well reserved.

The process of spectrum filtering is as follows: the E2 signal is amplitude matched with the echo signal in E1. Amplitude matching can be done in the following way: dividing the full frequency into M sub-bands with sub-band boundary B₁～B_M+1M of this embodiment may be 32 or 16. E2 and E1 calculate energy in each sub-band, divide the energy and perform an evolution operation to obtain a matching function H_m. Multiplication of E2 by the matching function H_mAnd multiplying by a factor Ag to obtain a matched echo Y_m。

The matching effect can be seen in fig. 11, which shows that in the frequency range where the echo is located, the echo Y is matched_mThe echo contribution to E1 is close (the difference between around 300Hz and around 3800Hz is due to the noise floor and not to the matching error).

The matching function is calculated as the following formula:

$H_m\left(f\right)=\sqrt{\frac{\underset{k=B_i}{\overset{k=B_{i+1}}{\mathrm{\&Sigma;}}}{{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1}^2\left(k\right)}{\underset{k=B_i}{\overset{k=B_{i+1}}{\mathrm{\&Sigma;}}}{{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2}^2\left(k\right)}},B_{i+1}>f\&GreaterEqual;B_i$

wherein i ∈ [1, M ]]Is the subband number, f is the frequency, k is the frequency sampling point within the subband, E₁(k) For the amplitude of the first adaptive frequency domain signal at the frequency sampling points, E₂(k) Amplitude of the second adaptive frequency domain signal at the frequency sampling point

Second adaptive frequency domain signal E₂Multiplication by a matching function H_mAnd multiplying by a factor Ag for amplitude matching to obtain a matched echo Y_m：

Y_m(f)＝Ag(f)·E₂(f)H_m(f)

After the echo matching and the voice probability estimation are completed, the method is carried outE1 subtracting the matched echo Y_mAnd then multiplying the result by a voice probability function pF to obtain the result:

E_out(f)＝[E₁(f)-Y_m(f)]·pF(f)。

and a frequency-time transform component 637 configured to perform frequency-time transform on the result of the multiplication. The frequency-time transform component 637 transforms the digital signal from the frequency domain to the time domain, and may be implemented by an inverse fourier transform, or an inverse discrete digital cosine transform.

After frequency-time conversion, the frequency domain signal Eout is converted into a time domain signal Eout, which is the total output of the system.

The final effect can be seen in the main microphone signal d1 in fig. 12(a), the near-end speech component in the main microphone signal d1 in fig. 12(b), and the energy curve of eout in fig. 12(c), which shows that in eout, the echo component is cancelled, and the near-end speech remains intact, and the signal energy is not significantly attenuated compared to the original near-end speech component. The full duplex requirement is achieved.

Fig. 13 is a flowchart of a method for using the apparatus for echo cancellation in a small hands-free voice communication system according to a preferred embodiment of the present invention. The small hands-free voice communication system comprises a receiver, a main transmitter and an auxiliary transmitter, wherein the distance between the main transmitter and the receiver is larger than the distance between the auxiliary transmitter and the receiver. The method comprises the following steps:

s1301: the main transmitter signal and the auxiliary transmitter signal are input into an array echo cancellation unit for array filtering, and part of linear echo components and part of nonlinear echo components in the main transmitter signal are removed to obtain an output signal.

S1302: the receiver signal, the output signal of the array echo cancellation unit and the auxiliary microphone signal are input into the self-adaptive echo cancellation unit for self-adaptive filtering, and the residual linear echo component in the main microphone signal and the linear echo component in the auxiliary microphone signal are respectively removed from the output signal of the array echo cancellation unit, so that two paths of output signals are obtained.

S1303: and inputting the two paths of output signals of the self-adaptive echo cancellation unit into a residual echo cancellation unit, and removing residual nonlinear echo components in the main microphone signal through voice probability estimation and echo matching to obtain one path of output signal as a voice signal after echo cancellation.

Fig. 14 is a detailed flowchart of a method for using the apparatus for echo cancellation in a small hands-free voice communication system according to a preferred embodiment of the present invention. The small hands-free voice communication system comprises a receiver, a main transmitter and an auxiliary transmitter, wherein the distance between the main transmitter and the receiver is larger than the distance between the auxiliary transmitter and the receiver. The method comprises the following steps:

s1401: carrying out array filtering on the auxiliary microphone signal to obtain a second array filtering signal, and subtracting the second array filtering signal from the main microphone signal to remove part of linear echo components and part of nonlinear echo components in the main microphone signal;

s1402: carrying out self-adaptive filtering on the auxiliary microphone signal to remove a linear echo component in the auxiliary microphone signal to obtain a second self-adaptive filtering signal; carrying out self-adaptive filtering on the signal obtained by subtracting the second array filtering signal from the main microphone signal so as to remove residual linear echo components in the main microphone signal to obtain a first self-adaptive filtering signal;

s1403: performing time-frequency transformation on the first adaptive filtering signal and the second adaptive filtering signal to obtain a first adaptive frequency domain signal and a second adaptive frequency domain signal;

s1404: performing frequency domain voice probability estimation according to the amplitude of the first adaptive frequency domain signal and the amplitude of the second adaptive frequency domain signal to obtain a frequency domain voice probability signal, wherein the frequency domain voice probability signal represents the proportion of a near-end voice signal in the first adaptive frequency domain signal;

s1405: according to the amplitude of nonlinear echo components of the first adaptive frequency domain signal and the second adaptive frequency domain signal, performing echo matching on the second adaptive frequency domain signal to obtain matched echo, subtracting the matched echo from the first adaptive frequency domain signal, and multiplying the result obtained by subtraction by the frequency domain voice probability signal;

s1406: and performing frequency-time conversion on the result obtained by multiplying to output the echo cancellation result.

Specifically, in step S1401, the performing array filtering on the auxiliary microphone signal to obtain a second array filtered signal specifically includes: determining the transfer function of an array filter

And using a transfer function

The array filter of (1) array-filtering the auxiliary microphone signal, wherein the transfer function of the array filter is determined according to the following formula

$\frac{\&PartialD;E\left[{(d_1-{\mathrm{<figure-callout\; id="2"\; label="d"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">d</figure-callout>}}_2*\hat{h})}^2\right]}{\&PartialD;\hat{h}}=0$

Wherein,

is the transfer function of the array filter, d₁Is the main microphone signal, d₂Is the auxiliary microphone signal, E.]To find the expected operand, it is the convolution operand.

In step S1404, performing frequency-domain speech probability estimation according to the amplitudes of the first adaptive frequency-domain signal and the second adaptive frequency-domain signal to obtain a frequency-domain speech probability signal includes:

calculating a frequency domain speech probability signal according to the following formula:

$\mathrm{pF}\left(f\right)=\left\{\begin{array}{cc}1& |{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2/{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1|<T_S\\ (T_E-|{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2/{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1\left|\right)/(T_E-T_S)& |{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2/{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1|\&Element;[T_S,T_E]\\ 0& |{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2/{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1|>T_E\end{array}\right.$

wherein f is frequency, E₁For the amplitude of the first adaptive frequency domain signal, E₂For the amplitude of the second adaptive frequency domain signal, pF is the frequency domain speech probability signal, T_SIs the average amplitude ratio, T, of the near-end speech signal in the auxiliary microphone signal and the main microphone signal_EIs the average amplitude ratio of the nonlinear echo component signals in the auxiliary microphone signal and the main microphone signal, where T_E＞T_S＞1。

In step S1405, performing echo matching on the second adaptive frequency domain signal according to the energy of the nonlinear echo components of the first adaptive frequency domain signal and the second adaptive frequency domain signal, specifically including two steps:

(1) dividing the full frequency into M sub-bands with sub-band boundary B₁～B_M+1Within each sub-band, to the first adaptive frequency-domain signal E₁And a second adaptive frequency domain signal E₂Solving the energy, dividing the energy and performing an evolution operation to obtain a matching function H_m：

$H_m\left(f\right)=\sqrt{\frac{\underset{k=B_i}{\overset{k=B_{i+1}}{\mathrm{\&Sigma;}}}{{\mathrm{<figure-callout\; id="1"\; label="E"\; filenames="HDA00002619339700021.png,HDA00002619339700022.png"\; state="\{\{state\}\}">E</figure-callout>}}_1}^2\left(k\right)}{\underset{k=B_i}{\overset{k=B_{i+1}}{\mathrm{\&Sigma;}}}{{\mathrm{<figure-callout\; id="2"\; label="E"\; filenames="HDA00002619339700011.png,HDA00002619339700021.png"\; state="\{\{state\}\}">E</figure-callout>}}_2}^2\left(k\right)}},B_{i+1}>f\&GreaterEqual;B_i;$

Wherein i ∈ [1, M ]]Is the subband number, f is the frequency, k is the frequency sampling point within the subband, E₁(k) For the amplitude of the first adaptive frequency domain signal at the frequency sampling points, E₂(k) The amplitude of the second adaptive frequency domain signal at the frequency sampling point is obtained;

(2) second adaptive frequency domain signal E₂Multiplication by a matching function H_mAnd multiplying by a factor Ag for amplitude matching to obtain a matched echo Y_m：

Y_m(f)＝Ag(f)·E₂(f)H_m(f)；

In step S1405, the first adaptive frequency domain signal and the matched echo Y are combined_mSubtracting and multiplying the result of the subtraction with the frequency domain speech probability signal as:

E_out(f)＝[E₁(f)-Y_m(f)]·pF(f)；

where pF is the frequency domain speech probability signal, and Ag is the matching factor associated with pF, and decreases as pF increases, for example, Ag (f) max (2-pF (f) 2.5, 0).

When pF is close to 0, Ag is greater than 1, so that the amplitude of the matched echo and Ag is higher than that of the first adaptive frequency domain signalE₁The non-linear echo component of (b) may cancel the echo, and when pF is close to 1, Ag is a value less than 1, so that the amplitude of the matching echo is lower than the first adaptive frequency domain signal E₁The non-linear echo component of (2) may enable speech to be preserved. In other words, the matching echo is very close to the residual nonlinear echo component in the first adaptive frequency domain signal, after subtraction, the nonlinear echo component in the first adaptive frequency domain signal is almost completely subtracted, but there may still be a weak residual, and the subtraction result is multiplied by the probability of speech, so that the nonlinear echo component can be completely removed.

The embodiment of the utility model has following advantage:

by adopting the technical scheme of the utility model, the voice probability estimation is used, and echo matching is combined, so that the damage to near-end voice can be reduced while echo is reduced, and the duplex performance is improved;

for a small hands-free voice communication system with a plurality of microphones, the technical scheme of the utility model can be realized only by using the microphone which is farthest from and closest to the microphone, and the implementation is easy;

thirdly, because the technical scheme of the utility model is used, the phase of the voice does not need to be distinguished, so that the phase consistency of the microphone is not strictly required, the acoustic design is less limited, and the product design is easy;

and (IV) the frequency domain voice probability signal is used for echo matching operation, so that the spectrum filtering strength is changed along with the near-end voice probability, and when the near-end voice probability is higher, the spectrum filtering strength is reduced, and the near-end voice can be better protected.

The above description is only for the specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention should be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (5)

1. An apparatus for echo cancellation in a small hands-free voice communication system, wherein the small hands-free voice communication system comprises a receiver, a main microphone and an auxiliary microphone, and a distance between the main microphone and the receiver is greater than a distance between the auxiliary microphone and the receiver, wherein the apparatus comprises, in a cascade structure: an array echo cancellation unit, an adaptive echo cancellation unit, and a residual echo cancellation unit,

the input of the array echo cancellation unit is a main microphone signal and an auxiliary microphone signal, and partial linear echo components and partial nonlinear echo components in the main microphone signal are removed through array filtering to obtain an output signal;

the input signals of the self-adaptive echo cancellation unit are telephone receiver signals, output signals of the array echo cancellation unit and auxiliary microphone signals, and residual linear echo components in the main microphone signals and linear echo components in the auxiliary microphone signals are respectively removed from the output signals of the array echo cancellation unit and the auxiliary microphone signals through self-adaptive filtering to obtain two paths of output signals;

the input signal of the residual echo eliminating unit is two paths of output signals of the self-adaptive echo eliminating unit, and residual nonlinear echo components in the signal of the main microphone are removed through voice probability estimation and echo matching to obtain one path of output signal which is used as a voice signal after echo elimination.

2. The apparatus of claim 1,

the array echo cancellation unit comprises an array filtering component, the array filtering component comprises an array filter and a subtracter, the array filter is used for carrying out array filtering on the auxiliary microphone signal to obtain a second array filtering signal, and the subtracter is used for subtracting the second array filtering signal from the main microphone signal to remove part of linear echo components and part of nonlinear echo components in the main microphone signal;

the adaptive echo cancellation unit comprises two adaptive filtering components, wherein the two adaptive filtering components are used for respectively carrying out adaptive filtering on the auxiliary microphone signal to obtain a second adaptive filtering signal so as to remove the linear echo component in the auxiliary microphone signal, and carrying out adaptive filtering on the signal obtained by subtracting the second array filtering signal from the main microphone signal to obtain a first adaptive filtering signal so as to remove the residual linear echo component in the main microphone signal;

the residual echo eliminating unit comprises two time-frequency transformation components, a voice probability estimation component, a frequency spectrum filtering component and a frequency-time transformation component, wherein the two time-frequency transformation components are used for respectively carrying out time-frequency transformation on a first self-adaptive filtering signal to obtain a first self-adaptive frequency domain signal and carrying out time-frequency transformation on a second self-adaptive filtering signal to obtain a second self-adaptive frequency domain signal;

the voice probability estimation component is used for performing frequency domain voice probability estimation according to the amplitude of the first adaptive frequency domain signal and the amplitude of the second adaptive frequency domain signal to obtain a frequency domain voice probability signal, wherein the frequency domain voice probability signal represents the proportion of a near-end voice signal in the first adaptive frequency domain signal;

the frequency spectrum filtering component comprises an echo matcher, a subtracter and a multiplier, wherein the echo matcher is used for performing echo matching on a second self-adaptive frequency domain signal according to the amplitudes of the first self-adaptive frequency domain signal and the second self-adaptive frequency domain signal; the subtracter is used for subtracting the result of the first adaptive frequency domain signal after echo matching so as to remove a nonlinear echo component signal; the multiplier is used for multiplying the result obtained by subtraction with the frequency domain voice probability signal to suppress the echo signal and protect the near-end voice signal;

and the frequency-time conversion component is used for carrying out frequency-time conversion on the result obtained by multiplying so as to output the echo cancellation result.

3. The apparatus of claim 2,

d1>2D2 and D1-D2>6 cm;

wherein D1 is the distance between the main microphone and the receiver, and D2 is the distance between the auxiliary microphone and the receiver;

the transfer function of the array filter is determined by the following equation:

$\frac{\&PartialD;E\left[{(d_1-d_2*\hat{h})}^2\right]}{\&PartialD;\hat{h}}=0$

wherein,

is the transfer function of the array filter, d₁Is the main microphone signal, d₂Is the auxiliary microphone signal, E.]To find the expected operand, it is the convolution operand.

4. The apparatus of claim 2, wherein the speech probability estimating component is configured to calculate the frequency domain speech probability signal according to the following formula:

$\mathrm{pF}\left(f\right)=\left\{\begin{array}{cc}1& |E_2/E_1|<T_S\\ (T_E-|E_2/E_1\left|\right)/(T_E-T_S)& |E_2/E_1|\&Element;[T_S,T_E]\\ 0& |E_2/E_1|>T_E\end{array}\right.$

wherein f is frequency, E₁For the amplitude of the first adaptive frequency domain signal, E₂For the amplitude of the second adaptive frequency domain signal, pF is the frequency domain speech probability signal, T_SIs the average amplitude ratio, T, of the near-end speech signal in the auxiliary microphone signal and the main microphone signal_EAs auxiliary microphone signal and main microphone signalAverage amplitude ratio of intermediate non-linear echo component signals, where T_E＞T_S＞1。

5. The apparatus according to any one of claims 2 to 4,

the echo matcher is used for dividing full frequency into M sub-bands, and the sub-band boundary is B₁～B_M+1Within each subband, the following calculations are performed:

calculating a matching function H_m：

$H_m\left(f\right)=\sqrt{\frac{\underset{k=B_i}{\overset{k=B_{i+1}}{\mathrm{\&Sigma;}}}{E_1}^2\left(k\right)}{\underset{k=B_i}{\overset{k=B_{i+1}}{\mathrm{\&Sigma;}}}{E_2}^2\left(k\right)}},B_{i+1}>f\&GreaterEqual;B_i$

Wherein i ∈ [1, M ]]Is the subband number, f is the frequency, k is the frequency sampling point within the subband, E₁(k) For the amplitude of the first adaptive frequency domain signal at the frequency sampling points, E₂(k) The amplitude of the second adaptive frequency domain signal at the frequency sampling point is obtained;

second adaptive frequency domain signal E₂Multiplication by a matching function H_mAnd multiplying by a factor Ag for amplitude matching to obtain a matched echo Y_m：

Y_m(f)＝Ag(f)·E₂(f)H_m(f)；

The subtracter combines the first adaptive frequency domain signal with the matched echo Y_mSubtracting and multiplying the result of subtracting by the frequency domain speech probability signal by a multiplier as:

E_out(f)＝[E₁(f)-Y_m(f)]·pF(f)；

where pF is the frequency domain speech probability signal and Ag is a factor related to pF, decreasing as pF increases.

Images