CN1780165A - Apparatus and method for echo cancellation - Google Patents

Abstract

An echo eliminating device is composed of near-end signal state judging module, far-end signal calculating and searching module and NLP module. Its method features that the state of near-end input signal is judged according to the maximal values of the energy parameters for near-end and far-end input signals and the combination with the primary judged result of near-end input signal to control the on and off of NLP for eliminating echo.

Description

Apparatus and method for echo cancellation

technical field

The invention relates to the technical field of voice transmission, in particular to an echo cancellation device and method.

Background technique

When using wired and wireless communication systems, echoes are often heard, which degrades the communication voice quality. The echo is actually a sound processed by a variable filter after a certain delay. To improve voice quality, echo cancellation must be performed during voice transmission. Echo is divided into electrical echo and acoustic echo according to its generation principle. In the Public Switched Telephone Network (PSTN, Public Switched Telephone Network), the impedance mismatch caused by the "two-four wire" mixing of the converter causes a small part of the received signal to "leak" into the transmission path, so that the far end The signal returns part of the signal at the same time, so the speaker hears the echo of his own speech, resulting in an electrical echo, also known as a hybrid echo. Acoustic echoes are formed when sound waves originating from the receiver are reflected by solids located in the sound path into the microphone of the mobile phone.

In the wired communication system, the receiving end will generate both electrical echo and acoustic echo, but because the distance between the receiver and the microphone is slightly farther, the acoustic echo is relatively weak and can be ignored compared with the electrical echo.

In the wireless communication system, since there is no "two-four-wire" mixing in the PSTN network, the wireless equipment will not generate electrical echo; however, because many wireless phone manufacturers do not comply with the official standard of the wireless phone: Proper isolation between the receiver and microphone is required, and as a result, the short distance between the mobile phone receiver and microphone makes acoustic echo an urgent concern for wireless service providers.

In addition, the voice compression technology of wireless networks and the compression algorithm of nonlinear voice introduce a long processing delay time. The round-trip delay is the time it takes for the echo to reflect. If the round-trip delay exceeds 10 milliseconds, the human ear will notice the presence of the echo , causing voice quality to degrade. The voice compression technology makes the round-trip delay greater than 200 milliseconds, so in wireless applications, the acoustic echo problem is more obvious.

Fig. 1 is a schematic diagram of an echo generated in a mobile communication system, in which both an acoustic echo and an electrical echo are illustrated. As shown in Figure 1, the echo heard by the fixed phone side is the acoustic echo generated by the mobile phone side, and the echo heard by the mobile phone side is mainly the electrical echo.

An echo canceller is a device that detects and cancels the echo of the signal from the far end after it echoes at the local end device. Due to the simple linear relationship between the electrical echo and the source speech signal, an electrical echo canceller usually consists of two main functional components: an adaptive filter and a subtractor. The specific process is as follows: the adaptive filter monitors the receiving path and dynamically builds a mathematical model for the echo generating line. This line model is convolved with the voice stream on the receiving path, so that an estimated value of the echo is obtained and input to the Subtractor, this subtracts the linear portion of the echo from the transmit path line.

Due to the nonlinear relationship between the acoustic echo and the source speech signal, the electrical echo cancellation method cannot be simply applied to the acoustic echo cancellation. The existing acoustic echo cancellation generally uses algorithms to control nonlinear processors (NLP, Non -Linear Processor), Fig. 2 is an existing acoustic echo cancellation device, Rin in the figure is the far-end input signal of the AEC module, that is, the far-end input signal, and Rout is the far-end output acoustic echo controller (AEC, Acoustic Echo The signal of the Controller) module is the far-end output signal, Sin is the signal of the near-end input AEC module, that is, the near-end input signal, and Sout is the signal of the near-end output AEC module, that is, the near-end output signal. As shown in Figure 2, the device includes The following modules:

Noise reduction module 201 : used to perform noise reduction processing on the near-end input signal, and output the noise-reduced near-end input signal to the intelligent voice level detection module 202 and the near-end signal short-term power calculation module 204 .

Intelligent voice level detection module 202: for detecting the relative level difference between the far-end input signal and the near-end input signal when in the mute state, and output the level difference to the near-end signal state judging module 205; The signal levels of the terminal input signal and the near-end input signal in the non-mute state, and output the levels of the far-end and near-end input signals to the near-end signal state judging module 205 .

The near-end noise threshold input module 203 : used to save the near-end signal noise threshold, and output the near-end signal noise threshold to the near-end signal state judging module 205 .

The near-end signal short-term power calculation module 204 : used to calculate the smooth short-term power of the near-end input signal, and output the smooth short-term power of the near-end input signal to the near-end signal state judgment module 205 .

The near-end signal state judgment module 205: used to receive and save the near-end signal noise threshold output by the near-end noise threshold input module 203, and compare the output of the near-end input signal short-term power calculation module 204 with the near-end input signal noise threshold , to determine whether the near-end input signal is voice, if the former is greater than the latter, then the near-end input signal is voice, and the close command is output to the NLP module 210; otherwise, the near-end input signal is non-speech, and the intelligent voice level detection The mute level difference that module 202 outputs revises the far-end and near-end mute level difference, then calculates the signal level of the far-end input signal that intelligent voice level detection module 202 outputs at pure delay moment and current near-end input signal Level difference, the pure delay time is the time obtained by subtracting the pure time delay from the time of the current near-end input signal. If the difference is within the predetermined range, it is determined that the near-end input signal is an echo and will start The instruction is output to the NLP module 210, and the state of the near-end input signal is output to the adaptive echo path delay estimation module 209; otherwise, the near-end input signal is determined to be noise, and the closing instruction is output to the adaptive echo path delay estimation module 209.

The far-end noise threshold input module 206 : used to save the far-end signal noise threshold, and output the far-end signal noise threshold to the far-end signal state judging module 208 .

Remote signal short-term power calculation module 207 : used to receive the remote input signal, calculate the smooth short-term power of the remote input signal, and output it to the remote signal state judgment module 208 .

The remote signal state judgment module 208: used to receive and save the remote signal noise threshold output by the remote noise threshold input module 206, and compare the remote input signal noise threshold with the smooth short time of the remote input signal at pure time delay time power to obtain the state of the far-end input signal at the time of pure delay, and output the state of the far-end input signal to the adaptive echo path delay estimation module 209 .

Adaptive echo path delay estimation module 209: when receiving the output of the far-end input signal as voice, record the time when the far-end voice is generated, when the output of the near-end signal state judgment module 205 is received for the first time thereafter as echo , record the time when the near-end echo is generated, calculate and save the difference between the above two moments, that is, the echo path delay, when the output of the near-end signal status judgment module 205 is received as noise, according to the saved last echo path delay , combined with the latest voice occurrence time of the far-end input signal, estimate the echo occurrence point of the current near-end input signal, that is, the "echo point". If the "echo point" happens to be near the near-end noise moment, the near-end noise is determined is the residual echo, and output the start instruction to the NLP module 210, otherwise, determine that the near-end input signal is noise.

NLP module 210: when receiving the startup instruction of the near-end signal status module 205 or the adaptive echo path delay estimation module 209, the module starts; when receiving the shutdown instruction of the near-end signal status judging module 205, the module shuts down .

In the existing acoustic echo canceller, because it is difficult to eliminate white noise, the noise reduction algorithm is complicated and time-consuming; because the adaptive echo path delay estimation module needs to dynamically monitor and process the far-end and near-end input signals, and in In a wireless system, the movement of the mobile phone will cause the change of the echo reflection path, which increases the complexity of the algorithm and the processing time. In existing wireless communication systems, AEC processing is often bundled with codec modules, and too long AEC processing time will degrade system performance.

Contents of the invention

In view of this, the main purpose of the present invention is to provide an echo cancellation device and method with simple algorithm and relatively short processing time, so as to quickly judge the state of the near-end input signal, thereby determining whether to activate or deactivate the NLP.

In order to achieve the above object, technical solution of the present invention is achieved in that way:

An echo canceling device, comprising a nonlinear processor NLP module for performing nonlinear processing on the echo, further comprising:

The far-end signal calculation search module is used to search for the maximum value of the energy parameter of the far-end input signal within the predetermined time delay range according to the current time of the near-end input signal, and output the maximum value to the near-end signal state judgment module ;

The near-end signal state judging module is used to output the moment of the current near-end input signal to the far-end signal calculation and search module to initially judge whether the near-end input signal is voice or noise, and save the preliminary judgment result. The maximum value of the energy parameter of the end input signal and the energy parameter of the far-end input signal within the predetermined delay range further judges whether the near-end input signal is an echo, and if so, outputs the start command to the NLP module, otherwise, if the preliminary judgment result is voice, then output the shutdown command to the NLP module, if the preliminary judgment result is noise, then keep the current state of the NLP module.

The remote signal calculation search module includes:

The remote signal energy parameter calculation module is used to calculate the energy parameter of the remote input signal, and output the energy parameter of the remote input signal to the remote signal energy parameter search module;

The far-end signal energy parameter search module is used to receive and save the far-end signal energy parameters, and search for the maximum value of the energy parameters of the far-end input signal within the predetermined time delay range according to the moment of the current near-end input signal, and store the maximum The value is output to the near-end signal state judging module.

The near-end signal state judgment module includes:

The near-end noise threshold input module is used to save the near-end signal noise threshold and output it to the near-end comparison module;

The near-end signal energy parameter calculation module is used to receive the near-end input signal, and output the time of the current near-end input signal to the far-end signal calculation and search module, and calculate the energy parameter of the near-end input signal at the same time, and output the near-end input signal The signal energy parameter is output to the near-end comparison module;

The near-end comparison module is used to preliminarily judge whether the near-end input signal is voice or noise according to the energy parameter of the near-end input signal and the noise threshold of the near-end signal, and is used to determine the maximum value of the energy parameter of the far-end input signal within a predetermined time delay range and the energy parameters of the near-end input signal to further judge whether the near-end input signal is an echo, and if it is an echo, the start command will be output to the NLP module; otherwise, if it is initially judged as voice, the shutdown command will be output to the NLP module; , to maintain the current state of the NLP module.

The near-end signal state judgment module further includes a near-end signal dynamic change trend judgment module, which is used to receive and save the output of the near-end comparison module, and judge the state of the near-end input signal in combination with the saved state of the last near-end input signal. Dynamic change trend, when the change of the near-end input signal tends to have an echo, the start command is output to the NLP module, and when the change of the near-end input signal tends to be voice, the close command is output to the NLP module.

The echo canceling device further includes a comfort noise generation module, configured to generate comfort noise, and output the comfort noise to the NLP module,

At the same time, when the NLP module receives the activation instruction of the near-end signal state judgment module, it outputs the received comfort noise instead of the echo to the far-end.

The comfort noise generation module includes:

The clamped amplitude input module is used to save the clamped amplitude of the comfort noise, and output the clamped amplitude to the comfort noise amplitude update module;

The comfort noise amplitude update module updates the current comfort noise amplitude according to the output of the near-end signal state judgment module and the output of the clamped amplitude input module;

The comfort noise determination module generates comfort noise according to the noise amplitude output by the comfort noise update module, and outputs it to the NLP module;

At the same time, when the near-end signal status judging module determines that the near-end input signal is noise, the near-end signal status judging module further outputs the noise amplitude as a new comfort noise amplitude to the comfort noise amplitude updating module for updating the comfort noise amplitude. noise amplitude.

The echo canceling device further includes a far-end signal state judging module, which is used to judge whether the far-end input signal is voice or noise, and outputs the far-end input signal state to the near-end signal state judging module,

Simultaneously, before further judging whether the near-end input signal is an echo, the near-end signal status judging module further judges whether the received far-end input signal status is voice or noise, and if it is voice, directly judges whether the near-end input signal is an echo, If it is noise, reset the preliminary judgment result of the current near-end input signal as speech, and then further judge whether the near-end input signal is an echo.

The remote signal state judgment module includes:

The far-end noise threshold input module is used to save the far-end signal noise threshold and output it to the far-end comparison module;

The far-end comparison module judges whether the current far-end input signal is voice or noise by comparing the current far-end signal energy parameter and the far-end signal noise threshold, and outputs the judgment result to the near-end signal state judgment module,

At the same time, the remote signal calculation and search module further outputs the energy parameter of the current remote input signal to the remote comparison module.

A method of echo cancellation, the method comprising:

A. Preliminarily judge whether the near-end input signal is voice or noise, and save the preliminary judgment result;

B. Based on the current moment of the near-end input signal, search for the maximum value of the energy parameter of the far-end input signal within the predetermined time delay range;

C. Calculate the energy parameter of the near-end input signal, and further judge whether the near-end input signal is an echo according to the energy parameter of the near-end input signal and the maximum value of the energy parameter of the far-end input signal obtained in step B, and if so, start NLP processing , otherwise, if the preliminary judgment result is speech, turn off the NLP processing, and if the preliminary judgment result is noise, then maintain the current NLP processing state.

The predetermined time delay range in step B is from the moment obtained by subtracting the pure time delay and then subtracting half of the dynamic time delay from the moment of the near-end input signal to the moment of the near-end input signal minus the pure time delay plus Half of the upper dynamic delay is obtained at the moment.

The pure delay is any value between 100 and 150 milliseconds, and the dynamic delay is 80 milliseconds.

Further include between the steps B and C, judging whether the current far-end input signal is voice or noise, if the current far-end input signal is voice, step C is directly executed; otherwise, the preliminary judgment result of resetting the near-end input signal is After speaking, go to step C.

The echo cancellation method further includes pre-setting a far-end signal noise threshold, and the judging whether the current far-end input signal is speech or noise is obtained by comparing the current far-end input signal energy parameter with the far-end signal noise threshold, if the former is smaller than the latter Otherwise, it is determined that the far-end input signal is noise; otherwise, it is determined that the far-end input signal is speech.

The step C further includes before starting or closing the NLP processing: saving the state of the current near-end input signal, and judging the dynamic change trend of the near-end input signal in combination with the state of the previous near-end input signal. , the NLP process is started, and the NLP transition time variable is decremented from the smooth transition length value to 0. When the change tends to be speech, the NLP process is turned off, and the NLP transition time variable is incremented from 0 to the smooth transition length value.

After the step A preliminarily determines that the near-end input signal is noise, it further includes using the noise amplitude as a new comfort noise amplitude for NLP processing.

The echo cancellation method further includes setting a clamping amplitude in advance, and after starting the NLP processing in step C, it further includes: judging whether the comfort noise amplitude is greater than the clamping amplitude, and if so, using the clamping amplitude as the new comfort noise amplitude value; otherwise, keep the current comfort noise amplitude.

The echo cancellation method further includes presetting a near-end signal noise threshold, and the judgment of whether the near-end input signal is voice or noise is obtained by comparing the energy parameter of the near-end input signal with the near-end signal noise threshold, if the former is smaller than the latter , determine that the near-end input signal is noise; otherwise, determine that the near-end input signal is speech.

The echo cancellation method further includes presetting a near-end signal noise threshold, and the judging whether the near-end input signal is speech or noise is obtained by comparing the maximum value of the energy parameter of the near-end segmented signal with the near-end signal noise threshold, if If the former is smaller than the latter, it is determined that the near-end input signal is noise; otherwise, it is determined that the near-end input signal is speech.

The near-end input signal in step C is a near-end segmented signal.

The near-end input signal described in step C is the first near-end segmented signal after the near-end input signal in step A is segmented,

Simultaneously, after said step C, it further includes, in turn, judging whether the second near-end segment signal and the following near-end segment signal are echoes, if yes, then close the NLP processing, otherwise, if the current far-end input signal is If the current far-end input signal is voice and the previous near-end segment signal is voice or if the current far-end input signal is noise, NLP processing is turned off. If the current far-end input signal is speech and the previous near-end segmented signal is noise, the current NLP processing state is maintained.

In step C, further judging whether the near-end input signal is an echo according to the energy parameter of the near-end input signal and the maximum value of the energy parameter of the far-end input signal obtained in step B is as follows: comparing the energy parameter of the near-end input signal with that obtained in step B The attenuated value of the maximum value of the energy parameter of the far-end input signal. If the former is smaller than the latter, it is determined that the near-end input signal is an echo.

Compared with the prior art, the present invention first preliminarily judges whether the near-end input signal is voice or noise, and then further judges the near-end Whether the input signal is an echo realizes the control of NLP startup and shutdown. Because the noise reduction processing will suppress a part of the echo, the present invention does not perform this processing, which saves time while simplifying the algorithm; because the present invention judges the state of the near-end input signal through two comparisons, the accuracy of the algorithm is guaranteed. And it saves the detection and correction of the silence level, which saves time; at the same time, the present invention limits the search range to the predetermined time delay, compared with the prior art to estimate the echo time delay every time to determine the "echo point", The complexity of the algorithm is reduced and the processing time is reduced. From the above analysis, it can be seen that the algorithm of the present invention is simple and saves processing time, so that the echo controller (EC, Echo Controller) module can be flexibly configured in the system.

Description of drawings

Fig. 1 is a schematic diagram of an echo in a digital wireless network;

Fig. 2 is a structural block diagram of an acoustic echo cancellation device in the prior art;

Fig. 3 is a block diagram of the overall structure of the echo canceling device of the present invention;

FIG. 4 is a structural block diagram of an embodiment of an echo cancellation device according to the present invention;

Fig. 5 is a detailed structural block diagram of the echo canceling device shown in Fig. 4;

Fig. 6 is the overall flow chart of realizing echo cancellation in the present invention;

FIG. 7 is a flow chart of a specific embodiment of the present invention for implementing acoustic echo cancellation.

Detailed ways

In order to make the purpose, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

Fig. 3 is the overall structural block diagram of the echo canceling device of the present invention, and this structural block diagram has included the necessary module that the device of the present invention realizes echo canceling, and as shown in Fig. 3, the device of the present invention comprises:

Remote signal calculation and search module 31: used to search for the maximum value of the energy parameter of the remote input signal within the predetermined time delay range according to the time of the near-end input signal, and output the maximum value to the near-end signal state judgment module 32.

Near-end signal state judging module 32: at first, this module is exported to far-end signal calculation search module 31 with the moment of current near-end input signal; The preliminary judgment result; finally, it is used to further judge whether the near-end input signal is an echo according to the energy parameter of the near-end input signal and the maximum value of the energy parameter of the far-end input signal within the predetermined delay range, and if so, the instruction will be started Output to NLP module 33, otherwise, judge that near-end input signal state is identical with preliminary judgment result, if preliminary judgment result is voice, then will close instruction output to NLP module 33, if preliminary judgment result is noise, keep current NLP module 33 status.

NLP module 33: when receiving the startup instruction of the near-end signal state judging module 32, the module starts; when receiving the shutdown instruction of the near-end signal state judging module 32, the module shuts down.

Fig. 4 is a structural block diagram of an embodiment of the echo cancellation device of the present invention, as shown in Fig. 4, in practical applications, the present invention may also include the following two modules:

Far-end signal state judging module 34: be used for receiving the current far-end signal energy parameter of far-end signal calculation search module 31 output, and judge whether current far-end input signal is voice or noise according to this energy parameter, and present far-end input signal The state is output to the near-end signal state judging module 32, and when the near-end signal state judging module 32 receives the far-end input signal as voice, it starts to further judge whether the near-end input signal is an echo; the near-end signal state judging module 32 receives the far-end When the input signal is noise, the preliminary judgment result of the near-end input signal is reset as speech, and then further judged whether the near-end input signal is an echo.

Comfort noise generation module 35: used to generate comfort noise, and when the near-end signal state judgment module 32 judges that the near-end input signal is noise, the noise amplitude is output to the comfort noise generation module 35 as a new comfort noise amplitude After updating the comfort noise amplitude, the comfort noise generation module 35 outputs the new comfort noise to the NLP module 35 for NLP processing.

Fig. 5 is a detailed structural block diagram of the echo cancellation device shown in Fig. 4, as shown in Fig. 5, the far-end signal calculation search module 31 includes the following submodules:

The remote signal energy parameter calculation module 301 : used to receive the remote input signal, calculate the smooth short-term power of the remote input signal, and output it to the remote signal energy parameter search module 302 and the remote comparison module 304 .

The far-end signal energy parameter search module 302: used to receive the output of the far-end signal energy parameter calculation module 301, and save it in the circular queue, and search for the far-end input signal based on the time of the current near-end input signal The maximum value of the segment-smoothed short-term power within the predetermined delay range is output to the near-end comparison module 307 .

The far-end signal state judging module 34 includes the following submodules:

The far-end noise threshold input module 303 : used to save the far-end signal noise threshold, and output the far-end signal noise threshold to the far-end comparison module 304 .

Far-end comparison module 304 : used to judge whether the current far-end input signal is voice or noise according to the current far-end input signal energy parameter and the far-end signal noise threshold, and output the status of the far-end input signal to the near-end comparison module 307 .

The near-end signal state judgment module 32 includes the following submodules:

The near-end noise threshold input module 305 is used to store the near-end signal noise threshold and output the near-end signal noise threshold to the near-end comparison module 307 .

The near-end signal energy parameter calculation module 306: used to receive the near-end input signal, and output the time of the current near-end input signal to the far-end signal energy parameter search module 302, and then calculate the smooth short-term power of the current near-end input signal Or calculate the segmental short-term smooth power of the near-end input signal after rough segmentation or detailed segmentation, and calculate the maximum value of the rough segmental smooth short-term power of the near-end input signal smoothed short-term power or the detailed segmental smooth short-term power The power is output to the near-end comparison module 307 .

Near-end comparison module 307: this module has two functions,

Function 1, used to receive and save the near-end signal noise threshold output by the near-end noise threshold input module 305, and preliminarily judge whether the near-end input signal is voice or noise, and save the preliminary judgment result of the near-end input signal, if the near-end input signal is judged to be noise, then input the noise amplitude as a new comfort noise amplitude to the comfort noise amplitude updating module 310;

Function two, when receiving the far-end input signal output by the far-end comparison module 304, the state is voice, this module is used to further judge whether the near-end input signal is an echo, when receiving the far-end input signal output by the far-end comparison module 304 When it is noise, the module resets the preliminary judgment result of the current near-end input signal as speech, and then further judges whether the near-end input signal is an echo; if the near-end input signal is not an echo and the near-end input signal is not segmented, according to The preliminary judgment result determines the state of the near-end input signal. If the near-end input signal is not an echo and the near-end input signal is segmented, the near-end segment is determined according to the preliminary judgment result or comprehensively considering the preliminary judgment result and the judgment result of the previous near-end segmented signal. Finally, the judgment result of the near-end input signal state is output to the near-end signal dynamic change trend judging module 308 .

Near-end signal dynamic change trend judgment module 308: this module is used to receive and save the output of the near-end comparison module 307, and judge the near-end input according to the state of the current near-end input signal and the saved state of the previous near-end input signal The dynamic change trend of the signal, when the change trend of the near-end input signal is that echo occurs, the start command is output to the NLP module 312; when the change trend of the near-end input signal is that there is voice, the close command is output to NLP module 312.

In practical applications, the near-end signal state judging module 32 may not include the near-end signal dynamic change trend judging module 308. At this time, if the near-end comparison module 307 determines that the segmented or non-segmented near-end input signal is an echo, then Output the start command to the NLP module 312; if it is judged as voice, output the shutdown command to the NLP module 312; if it is judged as noise, keep the current state of the NLP module 312.

The comfort noise generating module 35 includes the following submodules:

Clamped amplitude input module 309 : used to save the clamped amplitude of the comfort noise and output it to the comfort noise amplitude updating module 310 .

Comfort noise amplitude update module 310 : used to update the current comfort noise amplitude according to the output of the near-end comparison module 307 and the output of the clamped amplitude input module 309 , and output the updated comfort noise amplitude to the comfort noise determination module 311 .

Comfort noise determination module 311 : used to generate initial comfort noise, update the current comfort noise according to the comfort noise amplitude output by the comfort noise amplitude update module 310 , and output the comfort noise to the NLP module 312 .

The NLP module 33 is the NLP module 312: when receiving the activation instruction from the near-end signal dynamic change trend module 308 or the near-end comparison module 307, the module is started, and the comfort noise output by the comfort noise determination module 311 replaces the echo and then outputs to The remote end: when receiving the shutdown command from the near-end signal dynamic change trend module 308 or the near-end comparison module 307, the module is turned off; otherwise, the module maintains the current state.

Fig. 6 is the overall flow chart of realizing echo cancellation in the present invention, and the flow chart includes the necessary steps for realizing echo cancellation in the method of the present invention. As shown in Fig. 6, the specific steps for realizing acoustic echo cancellation are as follows:

Step 601: The near-end signal state judging module 32 receives a near-end input signal, and calculates an energy parameter of the near-end input signal.

Step 602: the near-end signal state judging module 32 outputs the moment of the current near-end input signal to the far-end signal calculation and search module 31, then judges whether the near-end input signal is voice, if yes, execute step 603; otherwise, execute step 603; 604.

Step 603 : The near-end signal state judging module 32 preliminarily judges that the near-end input signal is voice, saves the state of the near-end input signal, and executes step 605 .

Step 604: The near-end signal state judging module 32 preliminarily judges that the near-end input signal is noise, and saves the state of the near-end input signal.

Step 605: After the far-end signal calculation and search module 31 receives the moment of the current near-end input signal, according to this moment, search for the maximum value of the energy parameter of the far-end input signal within the predetermined time delay range, and output the maximum value to the near-end signal state judging module 32.

Step 606: The near-end signal state judging module 32 compares the energy parameter of the near-end input signal with the attenuated value of the maximum value of the energy parameter of the far-end input signal. If the former is smaller than the latter, go to step 607; otherwise, go to step 608.

Step 607: The near-end signal state judging module 32 judges that the current state of the near-end input signal is an echo, and the NLP module 33 starts NLP processing.

Step 608: The near-end signal status judging module 32 judges whether the preliminary judging result is voice, if yes, go to step 609; otherwise, go to step 610.

Step 609: The NLP module 33 closes the NLP processing.

Step 610: The NLP module 33 maintains the current processing state.

Specific embodiment one:

When implementing echo cancellation in this embodiment, when judging whether the near-end input signal is an echo, it is judged for the entire input frame signal, that is, the frame signal is not segmented, and the unsegmented frame signal is called For near-end input frame signal, Fig. 7 is the flow chart of a specific embodiment that the present invention realizes acoustic echo cancellation, is the specific description of Fig. 6 step, as shown in Fig. 7, the concrete steps of realizing acoustic echo cancellation are as follows:

Step 701: After the near-end signal energy parameter calculation module 306 receives the near-end input frame signal, it outputs the current time of the near-end input frame signal to the far-end signal energy parameter search module 302, and calculates the smooth short-term value of the near-end input frame signal. time power SPower, and output SPower to the near-end comparison module 307.

Step 702: The near-end comparison module 307 compares SPower with the saved noise threshold SNoise of the near-end input signal output by the near-end noise threshold input module 305, and if SPower>SNoise, execute step 703; otherwise, execute step 704.

Step 703: The near-end comparison module 307 preliminarily determines that the near-end input frame signal is speech, saves the state of the near-end input frame signal, and executes step 705 .

Step 704: The near-end comparison module 307 preliminarily determines that the near-end input frame signal is noise, saves the state of the near-end input frame signal, and outputs the noise amplitude of the near-end input frame signal as a new comfort noise amplitude to the comfort Noise magnitude update module 310 .

Step 705 : After receiving the remote input signal, the remote signal energy parameter calculation module 301 calculates the smoothed short-term power RPower of the remote input signal, and outputs RPower to the remote signal energy parameter search module 302 and the remote comparison module 304 .

Step 706: The remote comparison module 304 compares RPower with the saved noise threshold RNoise of the remote input signal output by the remote noise threshold input module 303. If RPower<RNoise, execute steps 707-708; otherwise, execute step 709.

Step 707: The far-end comparison module 304 determines that the far-end input signal is noise, and outputs the state of the far-end signal to the near-end comparison module 307 .

Step 708 : After receiving the output of the far-end comparison module 304 as noise, the near-end comparison module 307 resets the preliminary state of the near-end input frame signal as speech, and saves the preliminary state, and goes to step 710 .

Step 709: The far-end comparison module 304 determines that the far-end input signal is speech, and outputs the status to the near-end comparison module 307 .

Step 710: The far-end signal energy parameter search module 302 receives and saves the far-end signal energy parameters, and searches for the smooth short-term power RPower of the far-end input signal within a predetermined time delay range according to the current moment of the near-end input frame signal. the maximum value RMPower, and output RMPower to the near-end comparison module 307 . The predetermined delay range is from the moment obtained by subtracting the pure delay and then subtracting half of the dynamic delay from the moment of the current near-end input signal to the moment of the near-end input signal minus the pure delay plus the dynamic delay. Half of the delay is obtained at a moment. Generally, the value of the pure delay is any value between 100 and 150 milliseconds, and the value of the dynamic delay is 80 milliseconds.

Step 711: The near-end comparison module 307 compares SPower and RMPower/B, where B is an attenuation value determined by experience, and if SPower<RMPower/B, then execute step 712; otherwise, execute step 713.

Step 712: The near-end comparison module 307 determines that the current state of the near-end input frame signal is echo, and outputs the state of the near-end input frame signal to the near-end signal dynamic change trend judgment module 308, and then goes to step 714.

Step 713: The near-end comparison module 307 determines that the state of the current near-end input frame signal is the same as the preliminary judgment result of the near-end input frame signal, and outputs the state of the near-end input frame signal to the near-end signal dynamic change trend judgment module 308 .

Step 714: The near-end signal dynamic change trend judging module 308 saves the state of the current near-end input frame signal, and combines the saved state of the previous near-end input frame signal to determine the dynamic change trend of the near-end input frame signal. There are four dynamic trends of the input frame signal at the terminal: when there is an echo, it transitions from voice or noise to echo; if there is a continuous echo, it transitions from echo or noise to voice; voice.

Step 715: The near-end signal dynamic change trend judging module 308 judges whether the dynamic change trend of the near-end input frame signal is echo, and if yes, execute steps 716-719; otherwise, execute steps 720-722.

Step 716 : The comfort noise amplitude updating module 310 compares the current comfort noise amplitude with the clamped amplitude outputted by the clamped amplitude input module 309 , if the former is greater than the latter, execute step 717 ; otherwise, execute step 718 .

Step 717: The comfort noise amplitude update module 310 replaces the current comfort noise amplitude with the clamped amplitude and outputs it to the comfort noise determination module 311, and the comfort noise determination module 311 outputs the new comfort noise to the NLP module 312, and then goes to step 719.

Step 718 : the comfort noise amplitude update module 310 keeps the current comfort noise amplitude unchanged, and the comfort noise determination module 311 outputs the current comfort noise to the NLP module 312 .

Step 719: NLP module 312 starts NLP processing, that is, switches between comfort noise c(n) and residual echo e(n), and the switching output is Sout(n), in order to obtain c(n) and e(n ) to smoothly switch between, use the following interpolation function:

$\mathrm{<span\; class="notranslate">\; Sout</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; ramp</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; ramp</span>}<span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; m</span>}}$

Among them, M is the length of the smooth transition; ramp is the transition time variable, and its variation range is from 0 to M. Obviously, the switching smoothness is determined by both ramp and M. When an echo appears, the ramp decreases from M to 0.

Step 720: The near-end signal dynamic change trend judging module 308 judges whether the near-end input frame signal dynamic change trend is voice, if yes, go to step 721; otherwise, go to step 722.

Step 721: the NLP module 312 closes the NLP processing, at this time, the interpolation function

$\mathrm{<span\; class="notranslate">\; Sout</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; ramp</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; ramp</span>}<span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; e</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; no</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; m</span>}}$

The ramp in is incremented from 0 to M.

Step 722: The NLP module 312 keeps the NLP processing state unchanged.

It should be noted that after the near-end comparison module 307 receives the output of the far-end comparison module 304 as voice, it starts to judge the near-end input signal according to the energy parameter of the near-end signal and the maximum value of the far-end signal within the dynamic delay range. Whether to echo.

Specific embodiment two:

When implementing echo cancellation in this embodiment, before judging whether the near-end input signal is an echo, the near-end input frame signal is segmented, and each segmented signal is called a near-end segmented signal. The difference between this embodiment and specific embodiment one is:

Between step 710 and step 711, it further includes that the near-end signal energy parameter calculation module 306 further divides the near-end input frame signal into segments, and calculates the segmental smoothing short-term smoothing power SLPower of the first near-end segmented signal, And output SLPower to the near-end comparison module 307 . At the same time, the SPower in step 711 is changed to "SLPower", that is, the near-end comparison module 307 judges whether the first near-end segment signal is an echo.

The near-end input frame signals in steps 712-722 are all replaced with "the first near-end segmented signal".

The rest of the steps are the same as in the first embodiment.

It should be pointed out that the above is the process of processing the first near-end segmented signal. After that, continue to repeat steps 706 to 722 for the second segment and subsequent segmental signals until the near-end input Until all the segmented signals of the frame signal are processed.

In practical applications, for all specific embodiments, when initially judging whether the near-end input frame signal is speech or noise, in order to make the judgment more accurate, steps 701 and 702 can also be replaced as follows:

Step 701 : After the near-end input frame signal is roughly segmented, the smoothed short-term power of each segment is calculated sequentially, and then the maximum value SMPower of all smoothed short-term powers is searched.

In step 702, SPower is changed to "SMPower", that is, it is preliminarily judged whether the near-end input frame signal is speech or noise by comparing the maximum value SMPower of the smooth short-term power of the near-end input frame signal with the noise threshold of the near-end signal.

In addition, for a specific embodiment in which the near-end input frame signal is segmented before further judging whether the near-end input signal is an echo, considering the continuity of the voice frame signal, steps 707 and 709 further include after the existing steps , the near-end comparison module 307 saves the current far-end input signal state, and at the same time, step 713 is replaced as follows:

Step 713: the near-end comparison module 307 first judges whether the received current far-end input signal state is speech, if the current far-end input signal state is speech, then the current near-end segment signal state is the same as that of the previous near-end segment signal The state is the same, and if the first segment of the near-end segment signal is non-echo, it is determined that the state of the first segment of the near-end segment signal is the same as the preliminary judgment result of the near-end input signal, and the status of the current near-end segment signal Output to the near-end signal dynamic change trend judging module 308; If the current far-end input signal state is noise, then determine the state of the current near-end segment signal as voice, and output the state of the current near-end segment signal to the near-end signal Dynamic change trend judgment module 308 .

In addition, in all specific embodiments, the energy parameters of the near-end input signal and the far-end input signal can also be represented by energy or level.

At the same time, the present invention is also applicable to the elimination of electrical echo.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims (21)

1, a kind of echo cancelling device comprises the nonlinear processor NLP module that is used for echo is carried out Nonlinear Processing, it is characterized in that, further comprises:

Remote signaling calculates search module, is used for according to the current near end input signal moment of living in, searches for the maximum of the energy parameter of remote end input signal in the preset time delay scope, and this maximum is outputed to the near end signal condition judgment module;

The near end signal condition judgment module, be used for that the current near end input signal moment of living in is outputed to remote signaling and calculate search module, preliminary judgement near end input signal is voice or noise, preserve this preliminary judged result, simultaneously, the maximum that is used for the energy parameter in the preset time delay scope according near end input signal energy parameter and remote end input signal judges further whether near end input signal is echo, if, enabled instruction is outputed to the NLP module, otherwise,, then out code is outputed to the NLP module if preliminary judged result is voice, if preliminary judged result is a noise, then keep the state of current NLP module.

2, device as claimed in claim 1 is characterized in that, described remote signaling calculates search module and comprises:

Remote signaling energy parameter computing module is used to calculate the energy parameter of remote end input signal, and the energy parameter of remote end input signal is outputed to remote signaling energy parameter search module;

Remote signaling energy parameter search module, be used for receiving and preserving the remote signaling energy parameter, and according to the maximum of current near end input signal moment of living in search remote end input signal, and this maximum outputed to the near end signal condition judgment module in preset time delay scope self-energy parameter.

3, device as claimed in claim 1 is characterized in that, described near end signal condition judgment module comprises:

Near-end noise threshold value input module is used to preserve the near end signal noise threshold and it is outputed to the near-end comparison module;

Near end signal energy parameter computing module, be used to receive near end input signal, and the current near end input signal moment of living in is outputed to remote signaling calculate search module, calculate the energy parameter of near end input signal simultaneously, and the near end input signal energy parameter is outputed to the near-end comparison module;

The near-end comparison module, be used for judging tentatively that according near end input signal energy parameter and near end signal noise threshold near end input signal is voice or noise, be used for simultaneously judging further according to the energy parameter maximum and the near end input signal energy parameter of remote end input signal in the preset time delay scope whether near end input signal is echo, if be echo then enabled instruction is outputed to the NLP module, otherwise, if tentatively be judged as voice then out code outputed to the NLP module, if tentatively be judged as noise, keep the state of current NLP module.

4, device as claimed in claim 3, it is characterized in that, described near end signal condition judgment module further comprises near end signal dynamic change trend judge module, be used to receive and preserve the output of near-end comparison module, and judge the dynamic change trend of near end input signal in conjunction with the state of the preceding near end input signal of having preserved, when the tendency of changes of near end input signal occurs for echo is arranged, just enabled instruction is outputed to the NLP module, when the tendency of changes of near end input signal occurs for voice are arranged, just out code is outputed to the NLP module.

5, device as claimed in claim 1 is characterized in that, described device further comprises the comfort noise generation module, is used to produce comfort noise, and this comfort noise is outputed to the NLP module,

Simultaneously, when described NLP module receives the enabled instruction of near end signal condition judgment module, the comfort noise that receives is substituted echo output to far-end.

6, device as claimed in claim 5 is characterized in that, described comfort noise generation module comprises:

Clamp down on the amplitude input module, be used to preserve the amplitude of clamping down on of comfort noise, and will clamp down on amplitude and output to comfort noise amplitude update module;

Comfort noise amplitude update module is upgraded the amplitude of current comfort noise according to the output of near end signal condition judgment module and the output of clamping down on the amplitude input module;

The comfort noise determination module according to the noise amplitude generation comfort noise of comfort noise update module output, and outputs to the NLP module;

Simultaneously, described near end signal condition judgment module judges that the near end signal condition judgment module further outputed to comfort noise amplitude update module with this noise amplitude as new comfort noise amplitude and is used to upgrade the comfort noise amplitude when near end input signal was noise.

7, device as claimed in claim 1 is characterized in that, described device further comprises the remote signaling condition judgment module, be used to judge that remote end input signal is voice or noise, and the remote end input signal state is outputed to the near end signal condition judgment module,

Simultaneously, the near end signal condition judgment module is before further judging whether near end input signal is echo, judge that further the remote end input signal state that receives is voice or noise, if be voice, just directly judge whether near end input signal is echo, if be noise, the preliminary judged result of just resetting current near end input signal is voice, judges further then whether near end input signal is echo.

8, device as claimed in claim 7 is characterized in that, described remote signaling condition judgment module comprises:

Far-end noise threshold value input module is used to preserve the remote signaling noise threshold and it is outputed to the far-end comparison module;

The far-end comparison module by more current remote signaling energy parameter and remote signaling noise threshold, is judged that current remote end input signal is voice or noise, and judged result is outputed to the near end signal condition judgment module,

Simultaneously, described remote signaling calculates search module and further current remote end input signal energy parameter is outputed to the far-end comparison module.

9, a kind of echo cancel method is characterized in that, this method comprises:

A, tentatively judge that near end input signal is voice or noise, preserves this preliminary judged result;

B, with the current near end input signal moment of living in be basis, the maximum of the search energy parameter of remote end input signal in the preset time delay scope;

The energy parameter of C, calculating near end input signal, the maximum of the remote end input signal energy parameter that obtains according to the energy parameter and the step B of near end input signal, further judge whether near end input signal is echo, if, start NLP and handle, otherwise, if preliminary judged result is voice, then close NLP and handle,, then keep current NLP treatment state if preliminary judged result is a noise.

10, method as claimed in claim 9, it is characterized in that, the described preset time delay scope of step B is for deducting half moment that obtains that pure time delay deducts dynamic delay again from the near end input signal moment of living in, only deducts half moment that obtains that pure time delay adds dynamic delay to the near end input signal moment of living in.

11, method as claimed in claim 10 is characterized in that, described pure time delay is the arbitrary value between 100 to 150 milliseconds, and described dynamic delay is 80 milliseconds.

12, method as claimed in claim 9 is characterized in that, further comprises between described step B and the C, judges that current remote end input signal is voice or noise, if current remote end input signal is voice, and just direct execution in step C; Otherwise, after the preliminary judged result of resetting near end input signal is voice, execution in step C.

13, method as claimed in claim 12, it is characterized in that, this method further comprises and sets in advance a remote signaling noise threshold, the current remote end input signal of described judgement is that voice still are that noise obtains by more current remote end input signal energy parameter and remote signaling noise threshold, if the former is less than the latter, the judgement remote end input signal is a noise; Otherwise the judgement remote end input signal is voice.

14, method as claimed in claim 9, it is characterized in that, described step C is starting or is closing NLP and handle to take a step forward and comprise: the state of preserving current near end input signal, and judge the dynamic change trend of near end input signal in conjunction with the state of a preceding near end input signal, when tendency of changes is that echo is when occurring, just starting NLP handles, and NLP variable transit time is decremented to 0 by seamlessly transitting length value, when tendency of changes is that voice are when occurring, just close NLP and handle, NLP variable transit time is incremented to by 0 seamlessly transits length value.

15, method as claimed in claim 9 is characterized in that, described steps A further comprises after the preliminary judgement near end input signal is noise, with this noise amplitude as being used for the new comfort noise amplitude that NLP handles.

16, method as claimed in claim 9, it is characterized in that, this method comprises that further setting in advance one clamps down on amplitude, and after starting the NLP processing, step C further comprises: judge that whether the comfort noise amplitude is greater than clamping down on amplitude, if then will clamp down on amplitude as new comfort noise amplitude; Otherwise, keep current comfort noise amplitude.

17, method as claimed in claim 9, it is characterized in that, this method further comprises and sets in advance a near end signal noise threshold, described judgement near end input signal be voice still be noise be by relatively the energy parameter and the near end signal noise threshold of near end input signal obtain, if the former is less than the latter, the judgement near end input signal is a noise; Otherwise the judgement near end input signal is voice.

18, method as claimed in claim 9, it is characterized in that, this method further comprises and sets in advance a near end signal noise threshold, described judgement near end input signal be voice still be noise be by relatively the maximum and the near end signal noise threshold of the energy parameter of near-end block signal obtain, if the former is less than the latter, the judgement near end input signal is a noise; Otherwise the judgement near end input signal is voice.

19, method as claimed in claim 9 is characterized in that, the described near end input signal of step C is the near-end block signal.

20, method as claimed in claim 9 is characterized in that, the described near end input signal of step C is the first section near-end block signal of near end input signal after segmentation in the steps A,

Simultaneously, further comprise after the described step C, judge successively whether second section near-end block signal and subsequent near-end block signal are echo, if, then closing NLP handles, otherwise, if current remote end input signal be voice and the last period the near-end block signal be echo, then starting NLP handles, if current remote end input signal be voice and the last period the near-end block signal be if that voice or current remote end input signal are noise, then close NLP and handle, if current remote end input signal be voice and the last period the near-end block signal be noise, then keep current NLP treatment state.

21, method as claimed in claim 9, it is characterized in that, the maximum of the described remote end input signal energy parameter that obtains according to the energy parameter and the step B of near end input signal of step C judges further whether near end input signal is echo being: the value of maximum after decaying of the energy parameter of near end input signal and the remote end input signal energy parameter that step B obtains relatively, if the former is less than the latter, the judgement near end input signal is an echo.

Images