CN102300014A - Double-talk detection method applied to acoustic echo cancellation system in noise environment - Google Patents

Abstract

The invention discloses a double-talk detection method applied to an acoustic echo cancellation system in a noise environment. The method comprises the following steps of: obtaining an output error signal e(n) by utilizing an input reference signal x(n) and a signal d(n) acquired by a microphone by using an adaptive filtering module; further canceling residual echo delta(n) in e(n) to obtain ep(n) by using a residual echo suppression module; performing a voice enhancement operation on the signals d(n) and ep(n) to obtain signals de(n) and epe(n) to reduce the influence of background noises on double-talk detection; and detecting whether near-end voices exist or not by utilizing an energy ratio of the voice-enhanced signals epe(n) and de(n). The invention also discloses equipment for implementing the method. By the method and the equipment, the near-end voices can be accurately detected whether to exist or not in environments with continuous relatively stronger background noises.

Description

A Double Talk Detection Method for Acoustic Echo Cancellation System Applicable to Noisy Environment

1. Technical field

The present invention relates to a double-ended speech detection method applicable to an echo cancellation system in a noisy environment. The signal d(n) collected by a microphone and the signal e _p (n) obtained by an adaptive filtering module and a residual echo suppression module are separately calculated. Speech enhancement to reduce the influence of background noise on double-talk detection, and obtain signals d _e (n) and e _pe (n). Use the energy ratio of the signal e _pe (n) to de ₍ n) to detect whether there is near-end speech.

2. Background technology

With the continuous development of communication technology, people have higher and higher requirements for convenient communication methods, and video conferencing systems have been used more and more. In this type of communication terminal, the near-end voice is collected by the microphone and transmitted to the far end through the network (Internet or private network). Due to the coupling between the far-end microphone and the speaker, the near-end voice is transmitted back to the local area, forming an acoustic echo. Acoustic echo seriously affects the quality of speech transmission, so an echo cancellation system is indispensable in a video conferencing system.

的权值，使得

逐渐向h(n)逼近，得到包含残留回声δ(n)的输出误差信号e(n)，并通过引入后处理滤波器来提高回声抑制量得到回声抑制后的信号e_p(n)，具体工作流程如附图1所述。但在实际应用中，近端语音会导致自适应滤波误差过大甚至发散，因此必须引入双端说话检测器来检测本地信号。双端说话检测方法在噪声干扰下很难保证足够的准确度，为了应对这个问题，本发明提出了一种新的双端说话检测方法。本发明的双端说话检测方法需要对d(n)、e_p(n)分别做语音增强操作以降低本底噪声对双端说话检测的影响，经过该语音增强操作得到信号d_e(n)、e_pe(n)，用信号e_pe(n)与d_e(n)的能量比值检测是否存在近端语音，以检测结果决定自适应滤波设备的迭代步长，保证双端说话不被漏判，不影响自适应收敛过程，本发明的系统框图如附图2所示。The method of this problem has been discussed in a large number of literatures. A general echo cancellation system will include an adaptive filter, which is used to match the transfer function corresponding to the echo path as best as possible. x(n) is the input reference signal, d(n) is the signal collected by the microphone, that is, the expected response, and the purpose of adaptive filtering is to repeatedly adjust

weight, so that

Gradually approaching h(n), the output error signal e(n) containing the residual echo δ(n) is obtained, and the echo suppression amount is improved by introducing a post-processing filter to obtain the echo-suppressed signal e _p (n), specifically The workflow is as described in Figure 1. But in practical applications, the near-end speech will cause the adaptive filtering error to be too large or even divergent, so a double-talk detector must be introduced to detect the local signal. It is difficult for the double-talk detection method to ensure sufficient accuracy under noise interference. In order to deal with this problem, the present invention proposes a new double-talk detection method. The double-talk detection method of the present invention needs to do speech enhancement operation respectively to d (n), e _p (n) to reduce the impact of background noise on double-talk detection, obtain signal d _e (n) through this speech enhancement operation , e _pe (n), use the energy ratio of signal e _pe (n) and d _e (n) to detect whether there is near-end speech, and use the detection result to determine the iterative step size of the adaptive filtering device to ensure that double-ended speech is not missed Judgment, does not affect the adaptive convergence process, the system block diagram of the present invention is shown in Figure 2.

3. Contents of the invention

The invention proposes a method for detecting double-talk in an acoustic echo cancellation system suitable for background noise environments. There is background noise in the working environment of echo cancellation system in the present invention, therefore need to do speech enhancement operation to d(n), e _p (n) respectively to reduce the influence of background noise on double-talk detection, through this speech enhancement The operation obtains the signals d _e (n) and e _pe (n); the energy ratio of the signals e _pe (n) to d _e (n) is used to detect whether there is near-end speech.

The purpose of the present invention is achieved through the following technical solutions:

的权值，尽可能好的匹配回声路径所对应的传递函数h(n)，得到期望响应的估计值y(n)和包含残留回声δ(n)的输出误差信号e(n)；A) The adaptive filtering device uses the input reference signal x(n) and the signal d(n) collected by the microphone to adjust the transfer function

The weight value of the echo path corresponds to the best possible matching transfer function h(n), and the estimated value y(n) of the expected response and the output error signal e(n) including the residual echo δ(n) are obtained;

B) The residual echo suppression equipment further eliminates the residual echo δ(n) contained in e(n), and obtains the signal e _p (n);

C) In view of the background noise in the working environment of the system, it is necessary to perform speech enhancement operations on d(n) and e _p (n) to reduce the influence of the background noise on double-talk detection, and obtain the speech-enhanced signal d _e ( n) and e _pe (n);

D) Utilize the energy ratio of e _pe (n) and d _e (n) of the signal after speech enhancement to detect whether there is near-end speech;

E) Determine the iterative step size of the adaptive filtering based on the detection results, so as to ensure that double-ended speech is not missed and does not affect the adaptive convergence process.

The present invention relates to a double-end speech detection method device in an echo cancellation system working in a continuous strong background noise environment, which is realized by the following technical solutions:

A) The receiving unit receives the signal d(n) collected by the microphone and the output error signal e _p (n) obtained by the adaptive filtering device and the residual echo suppression device;

B) speech enhancement unit, signal d (n) and e _p (n) are done speech enhancement respectively, obtain the signal d _e (n) and e _pe (n) after suppressing the background noise;

C) calculation unit, calculates the root of the energy ratio of e _pe (n) and d _e (n) to obtain parameter h;

D) The detection unit detects whether the current state is double-talking by using the range of the h value.

E) The adjustment unit determines the iterative step size of the adaptive filtering device based on the result of the detection unit, so as to ensure that double-ended speech is not missed and does not affect the adaptive convergence process.

4. Description of drawings

Figure 1 is a block diagram of adaptive filtering and post-processing work in an echo cancellation system.

Figure 2 is a working block diagram of adaptive filtering, post-processing, and double-talk detection in the echo cancellation system.

Fig. 3 is a schematic flowchart of an implementation of the double-talk detection method provided by the present invention.

FIG. 4 is a schematic diagram of a signal processing flow implemented by the double-talk detection method provided by the present invention.

FIG. 5 is a schematic flow diagram of an implementation of the double-talk detection device provided by the present invention.

FIG. 6 is a schematic diagram of the workflow of the echo cancellation system in the present invention.

Fig. 7 is a set of waveform diagrams obtained after the present invention works according to the process described in Fig. 6 .

5. Implementation

The present invention will be described below by way of example with reference to the accompanying drawings.

Accompanying drawing 3 is a schematic flow chart of an implementation of the double-talk detection method provided by the present invention, and the implementation of this example includes the following steps:

S1: The adaptive filtering device uses the input reference signal x(n) and the signal d(n) collected by the microphone to adjust the transfer function The weight value of the echo path corresponds to the best possible matching transfer function h(n), and the estimated value y(n) of the expected response and the output error signal e(n) including the residual echo δ(n) are obtained;

S2: The residual echo suppression device further eliminates the residual echo δ(n) included in e(n), to obtain a signal e _p (n);

S3: In view of the background noise existing in the system working environment, it is necessary to perform speech enhancement operations on d(n) and e _p (n) to reduce the influence of the background noise on double-talk detection, and obtain the speech-enhanced signal d _e ( n) and e _pe (n);

S4: Utilize the energy ratio of e _pe (n) and d _e (n) of the signal after speech enhancement to detect whether there is near-end speech;

S5: Determine the iterative step size of the adaptive filtering device based on the detection results, so as to ensure that double-ended speech is not missed and does not affect the adaptive convergence process.

In order to prevent the block effect caused by the excessive change of the speech enhancement filtering result, it is necessary to perform overlapping accumulation and smoothing processing on the post-processing result. Figure 4 is a schematic diagram of a signal processing flow implemented by the double-talk detection method provided by the present invention.

为全频带自适应滤波器系数，自适应滤波设备通过反复调整的权值，使得逐渐向h(n)逼近，得到输出误差信号e(n)＝d(n)-y(n)，残留回声抑制设备将包含在e(n)中的残留回声δ(n)进一步消除，得到信号e_p(n)，以上所述变量均是时域信号；x(n) is the input reference signal, d(n) is the signal collected by the microphone, that is, the expected response, h(n) is the impulse response of the echo path,

For the full-band adaptive filter coefficients, the adaptive filtering device adjusts weight, so that Gradually approaching h(n), the output error signal e(n)=d(n)-y(n) is obtained, the residual echo suppression equipment will further eliminate the residual echo δ(n) contained in e(n), and get Signal e _p (n), the variables mentioned above are all time domain signals;

50% overlapping of d(n) and e _p (n) constitutes frame signals d(m) and e _p (m), one frame is M points, and speech enhancement is performed on d(m) and e _p (m) Operation, obtain the signal frame signal d _e (m) and e _pe (m) after suppressing the background noise;

Add hamming window smoothing to the frame signals d _e (m) and e _pe (m) respectively, align and add them, and reconstruct the processed time domain sequences d _e (n) and e _pe (n).

The variables p e and _{p d} representing the energy sum of each M/2 points of the speech-enhanced signal e _pe (n) and d _e (n) are calculated by formulas (1) and (2):

${\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; e</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; e</span>}}_{\mathrm{<span\; class="notranslate">\; pe</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

${\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; e</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

_Select the appropriate frame energy smoothing factor according to whether it was in the state of double-talking at the previous _moment to obtain p _es and p _ds , and the system is in the non-double-talking state at the current moment, use formula (3), ( 4) get:

p _es =λ ₁ p _es +(1-λ ₁ )p _e (3)

p _ds =λ ₁ p _es +(1-λ ₁ )p _d (4)

Wherein the smoothing factor is that λ ₁ is small, and the dependence of the detection parameters p _es and p _ds on the previous moment can be reduced, so that the detection parameters can quickly catch up with the double-ended speaking state;

At the current moment when the system is in the double-talking state, p _es and p _ds can be calculated using formulas (5) and (6):

p _es =λ ₂ p _es +(1-λ ₂ )p _e (5)

p _ds =λ ₂ p _es +(1-λ ₂ )p _d (6)

Where the smoothing factor is λ ₂ , and the condition λ ₂ > λ ₁ is satisfied, the dependence of the detection parameters p _es and p _ds on the previous moment is enhanced, the sensitivity of the detection parameters to parameter changes is weakened, and double-ended speech is not missed. Does not affect the adaptive convergence process;

According to the formula (7), the parameter h is calculated to detect whether there is a near-end voice:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; =</span>\sqrt{\frac{{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; es</span>}}}{{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; ds</span>}}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$

When the system is in the non-double talk state at the previous moment and h<h _max , it is judged that the current system is in the non-double talk state;

When the system is in the non-double talk state at the previous moment and h>h _max , it is judged that the current system is in the double talk state;

When the system was in the double-talking state at the previous moment and h<h _min , it is judged that the current system is in the non-double-talking state;

When the system is in the state of double-talk at the previous moment and h>h _min , it is judged that the system is in the state of double-talk;

Similarly, choosing different judgment thresholds h _max and h _min according to the state of the system at the previous moment is also to ensure that double-talking is not missed, so that double-talking does not affect the adaptive convergence process.

Accompanying drawing 5 is a schematic flow chart of an implementation of the double-talk detection device provided by the present invention, and the implementation of this example includes the following steps:

The receiving unit receives the signal d(n) collected by the microphone and the output error signal e _p (n) obtained by the adaptive filtering device and the residual echo suppression device;

Speech enhancement unit, signal d (n) and e _p (n) are done speech enhancement respectively, obtain the signal d _e (n) and e _pe (n) after suppressing the background noise;

Calculation unit, calculating the root of the energy ratio of e _pe (n) and d _e (n) to obtain parameter h;

The detection unit uses the range of the h value to detect whether the current state is double-talk, and determines the iterative step size of the adaptive filtering device based on the detection result, so as to ensure that the double-talk is not missed and does not affect the adaptive convergence process.

The working process of the echo cancellation system in the present invention is as shown in accompanying drawing 6, and this system considers the complexity of the echo system operation, and the computation amount and the real-time requirement of the system realized on the DSP, the self-adaptive algorithm adopts multi-delay time-frequency domain algorithm (MDF), while down-sampling the original sampling signal to 16kHz sampling rate for processing. Such a processing method is easily extended to a multi-channel echo cancellation system:

The AD of the DSP collects the reference signal x(n) and the signal d(n) collected by the microphone at a sampling rate of 48khz;

Downsample the signal obtained at the 48kHz sampling rate to obtain the signals x(n) and d(n) at the 16kHz sampling rate;

The adaptive algorithm in the adaptive filtering device uses the signals x(n) and d(n) to obtain an output error signal e(n) including residual echo δ(n);

The residual echo suppression equipment further eliminates the residual echo δ(n) included in e(n), and obtains the signal e _p (n);

The double-ended speech detection device performs speech enhancement operations on the signals d(n) and e _p (n) to obtain the signals d _e (n) and e _pe (n), and uses the energy of the sum of e _pe (n) and d _e (n) The ratio detects whether there is near-end speech, and the detection result determines the iterative step size of the adaptive filtering device to ensure that double-end speech is not missed and does not affect the adaptive convergence process;

The signal e _pe (n) calculated under the 16kHz sampling rate is up-sampled to obtain the signal under the 48kHz sampling rate;

Send the up-sampled signal to the DA output of DSP.

Accompanying drawing 7 is a group of waveform diagrams obtained after the present invention works according to the process described in accompanying drawing 6. Wherein (1) is the signal that includes echo and near-end voice collected by the microphone; (2) is the actual near-end voice signal; (3) is the near-end voice signal obtained after passing through the adaptive equipment and residual echo suppression equipment; (4 ) is the near-end speech signal waveform obtained after the speech augmentation device and the double-talk detection result. Comparing the waveform diagrams of accompanying drawing 7(3) and accompanying drawing 7(4), it can be seen that the present invention can better eliminate the background noise and restore the original voice more accurately. Visible by the rectangular line of accompanying drawing 7 (4), the present invention can reach the accurate detection of the double-end speech state of the echo cancellation system working in the environment of continuous strong background noise, and determine the iterative step of the adaptive filtering device with the detection result. Long, to ensure that double-ended speech will not be missed, and will not affect the adaptive convergence process.

At the same time, by adopting the above-mentioned double-talk detection method, the output of the acoustic echo cancellation system can be flexibly selected, which can be the signal e _p (n) before speech enhancement, or the signal e _pe (n) after speech enhancement. If e _p (n) is selected, the noise floor at the output can also be suppressed, but speech enhancement needs to take into account the voice quality; if e _pe (n) is selected, the speech enhancement operation is completely to eliminate the noise floor For the purpose of impact, the background noise can be eliminated more thoroughly without taking into account the voice quality, which will help to further improve the accuracy of double-talk detection.

Claims (4)

1. an acoustic echo bucking-out system both-end that is applicable to noise circumstance detection method of speaking is characterized in that:

A) the adaptive-filtering module is utilized signal d (n) the adjusting transfer function that input reference signal x (n) and microphone collect

Weights, the pairing transfer function h of coupling echo path (n) as well as possible obtains the estimated value y (n) of Expected Response and comprises the output error signal e (n) of residual echo δ (n);

B) the residual echo suppression equipment will be included in the further elimination of residual echo δ (n) among the e (n), obtain signal e _p(n);

C) background noise that exists at the system works environment need be to d (n) and e _p(n) carry out voice and strengthen operation, obtain the signal d after voice strengthen to reduce the influence that background noise is spoken and detected both-end _e(n) and e _Pe(n);

D) the signal e after utilizing voice to strengthen _Pe(n) and d _e(n) energy ratio detects whether there is near-end speech;

E) with the iteration step length of testing result decision adaptive-filtering, guarantee that both-end is not failed to judge in a minute, do not influence the self adaptation convergence process.

2. the acoustic echo bucking-out system both-end that the is applicable to noise circumstance as claimed in claim 1 detection method of speaking is characterized in that: voice are strengthened back signal e _Pe(n) the signal d that collects with microphone _e(n) with 50% overlapping configuration frame signal e _Pe(m) and d _e(m), calculate the variable p that characterizes every frame energy summation _eAnd p _d, and whether be in the both-end state of speaking according to previous moment and select suitable frame energy smoothing factor to obtain p _EsAnd p _Ds, by p _EsAnd p _DsUtilize formula

Calculate parameter h and detect whether there is near-end speech.

3. the acoustic echo bucking-out system both-end that the is applicable to noise circumstance as claimed in claim 1 checkout equipment that both-end is spoken in a kind of echo cancelltion system that is operated in continuously under the strong background noise environment of detection method of speaking is characterized in that:

A) receiving element, the output error signal e that signal d (n) that the reception microphone collects and process adaptive-filtering equipment and residual echo suppression equipment obtain _p(n);

B) voice enhancement unit is with signal d (n) and e _p(n) do voice respectively and strengthen the signal d behind the background noise that is inhibited _e(n) and e _Pe(n);

C) computing unit calculates e _Pe(n) and d _eThe evolution of energy ratio (n) obtains parameter h;

D) detecting unit utilizes whether the residing range detection current state of h value is that both-end is spoken.

E) adjustment unit determines the iteration step length of adaptive-filtering equipment with the detecting unit result, guarantees that both-end is not failed to judge in a minute, does not influence the self adaptation convergence process.

4. the acoustic echo bucking-out system both-end that the is applicable to noise circumstance as claimed in claim 1 detection method of speaking, it is characterized in that: adopt the above-mentioned both-end detection method of speaking, the output of acoustic echo bucking-out system can be selected flexibly, both can be the signal e before voice strengthen _p(n), also can be signal e after voice strengthen _Pe(n).If select e _p(n), then the output background noise also can be inhibited, but the voice enhancing need take into account speech quality; If select e _Pe(n), then voice enhancing operation is a purpose with the influence that the elimination background noise is spoken to both-end fully, can eliminate background noise more completely under the prerequisite of not taking into account speech quality, helps further to promote the both-end accuracy of detection in a minute.

Images