CN1299254C - Approaching end voice detection realizing method for echo inhibitor - Google Patents

Abstract

The present invention relates to a method for realizing near-end speech detection in an echo suppressor, adopting Geigel algorithm for near-end speech detection, taking M sampling points in the far-end input sequence y(i) as a subframe, and M is less than the The filter length N in the echo suppressor, when calculating the input of the first sampling point of each subframe, first obtain |y(i)| to |y(i -N+M)|The maximum value of the latest data of a total of N-M+1 sampling points, and save it as Frame_MAX, and then perform operations on the M sampling points of the subframe one by one, and take the maximum value Frame_MAX and the current The maximum value of the data of the remaining M-1 sampling points in the sliding window is taken as the maximum value MAX of each sampling point of the subframe, if the near-end input sequence s(i) of the sampling points of the subframe satisfies the condition| s(i)|≥u ^* MAX, then the near-end has voice. The method of the invention can greatly reduce the amount of calculation required for program calculation and reduce the cost under the condition that the calculation result is completely consistent with the existing method.

Description

A Realization Method of Near-End Voice Detection in Echo Suppressor

technical field

The invention relates to a method for realizing near-end voice detection in an echo suppressor in a communication device.

Background technique

Echo suppression is a technology widely used in the field of communication at present. The purpose of this technology is to eliminate the echo caused by the impedance mismatch of the circuit or the acoustic reflection, so as to provide higher quality voice communication services for both parties in the call.

The most widely used echo suppressor is composed of the following three parts, which are duplex detector, echo estimation canceller and nonlinear processor. The function of the duplex detector is to instruct the echo estimation canceller whether to update the filter coefficients and whether to enable the nonlinear processor. The core structure of the echo estimation canceller is an adaptive filter. The function of the filter is to estimate the echo data according to the far-end input voice and subtract the data from the near-end input to weaken or eliminate the echo. The estimated error is the difference between the near-end input and the estimated data, and the filter coefficients are updated according to a certain convergence algorithm, so that the filter coefficients are closer to the parameters of the echo path so that the filter converges. The function of the nonlinear processor is to eliminate the residual echo through some nonlinear processing methods, so as to further improve the voice quality.

The most important job of a duplex detector is to detect whether there is voice input at the near end. If a near-end voice input is detected, the coefficient update of the adaptive filter is stopped immediately, and the non-linear processor is stopped. Because the difference between the estimated echo of the filter and the near-end input is no longer the estimation error but the sum of the estimation error and the near-end speech, if the filter coefficients are updated at this time, the filter will be in an abnormal working state. Certain convergent algorithms can even cause the filter to diverge. If the nonlinear processing is still performed at this time, the near-end voice will be clipped and the voice quality will be affected. It can be seen that the near-end voice detection plays a very important role in the echo suppressor. In order to implement echo suppression processing of more channels or integrate more functions on the same hardware platform, it is necessary to provide an efficient and stable near-end voice detection method.

At present, the most classic algorithm in near-end voice detection is the Geigel algorithm. The premise of this algorithm is that the echo has at least 6dB attenuation relative to the voice, that is, the echo amplitude will be at least half of the far-end voice amplitude. Assuming that the filter length is N, the current sampling data at the near end is s(i), and the input sequence at the far end is y(i) to y(i-N-1), then if the inequality |s(i)|≥0.5max(|y (i)|, |y(i-1)|, .

It can be seen from the above formula that if the method 1 described in Figure 1 is used to directly judge according to the above judgment conditions, then corresponding to each near-end input sampling point, it is necessary to calculate the maximum value of the N-point far-end input sequence, and the calculation amount is similar to that of filtering It is proportional to the order N of the device. Taking the implementation of 16ms echo suppression in a telephone system sampled at 8KHz as an example, it takes at least 2N instruction cycles to calculate the maximum value of N-point data on a general-purpose processor or DSP, and the amount of calculation required per second is 16*8 *2*8000=2048000, 2MIPS, the amount of calculation is close to or even greater than that of the core filter, so it is very uneconomical to use this method directly in a general-purpose processor or DSP.

An improved implementation method is the method 2 described in Figure 2. This implementation method is to calculate the maximum value of the remote input sequence at the time of initialization, and in the subsequent processing, only need to calculate |y(i)| and the previous The maximum value between the frame maximum values is the maximum value of the data in this frame, but after the calculation, it is necessary to check whether |y(i-N+1)| is the maximum value of this frame, and if it is, it means that the next time it will slide The value moved out of the window is the maximum value, and the maximum value of N-1 points needs to be recalculated for use in the next frame. It can be seen from the principle of this method that in the case of normal speech, this case can greatly reduce the amount of calculation. However, in an extreme case, for example, when the remote input is silent for a period of time, that is, the absolute values of all remote input sequences are equal, the calculation amount at this time and the above calculation amount remain unchanged. That is to say, this method can reduce the average calculation volume to an uncertain level, but the peak calculation volume has not changed. In real-time operations such as echo suppression, it is necessary to ensure that the operation of the previous data point or data frame must be completed before the arrival of the next data point or data frame. Therefore, developers must allocate computing resources according to the peak processing capacity when designing. While this method provides the same operational peak value as Method 1, there is no real improvement over Method 1. At the same time, the average calculation value of this method varies with the remote input signal, and the peak value of the actual running calculation cannot be obtained through the test method, so it is not convenient for developers to perform performance statistics.

Contents of the invention

The technical problem to be solved by the present invention is to provide a low-computing solution for near-end voice detection, so that more channels of echo suppression solutions can be realized on the same computing platform, or more processors with more balance can be calculated for other modules. resources in order to reduce costs. At the same time, it can accurately calculate the calculation amount of the module during the development stage, which is convenient for developers to perform performance statistics and calculation resource allocation, and is conducive to product stability and compatibility.

In order to achieve the above object, the present invention provides a method for implementing near-end speech detection in an echo suppressor, using the Geigel algorithm for near-end speech detection, which is characterized in that M samples are taken in the far-end input sequence y(i) point as a subframe, M is less than the filter length N in the echo suppressor, when calculating the input of the first sampling point of each subframe, first obtain the sliding window | The maximum value of the latest data from y(i)| to |y(i-N+M)| with a total of N-M+1 sampling points, and save it as Frame_MAX, and then carry out the M sampling points of the subframe one by one Operation, take the maximum value of the maximum value Frame_MAX and the remaining M-1 sampling points in the current sliding window as the maximum value MAX of each sampling point of the subframe, if the sampling points of the subframe The near-end input sequence s(i) satisfies the condition |s(i)|≥u*MAX, then the near-end has voice.

The implementation method of near-end speech detection in the above-mentioned echo suppressor is characterized in that, the method for carrying out near-end speech detection to the subframe in described far-end input sequence y (i) comprises the following steps:

Step 1, initialize a subframe counter with a count value of 0;

Step 2, take the absolute value of the near-end input data |s(i)|, take the absolute value of the far-end input data |y(i)|, and save the data in the far-end data buffer, the buffer A digital sequence of N sampling points from |y(i)| to |y(i-N+1)| is saved;

Step 3, if the subframe counter is 0, it means that the current remote data sampling point is the first sampling point of the subframe, calculate the latest from |y(i)| to |y(i -N+M)|A total of the maximum value of N-M+1 sampling point data, and save this value as Frame_MAX; if the count value of the subframe counter is not 0, then perform step 4;

Step 4, calculate the maximum value MAX of Frame_MAX and the remaining M-1 sampling points in the remote data buffer;

Step 5: Compare the product u*MAX of the attenuation factor u and MAX with the absolute value |s(i)| of the near-end input data. If the condition |s(i)|≥u*MAX is satisfied, it indicates that there is voice at the near-end enter;

Step 6, updating the subframe counter, adding 1 to the subframe counter value;

Step 7, if the subframe counter value is equal to M, then reset the subframe counter value to 0;

Step eight, skip to step two, and calculate the next sampling point data of the subframe.

The feature of the implementation method of the near-end voice detection in the above-mentioned echo suppressor is that when the method is applied to an application with the concept of speech frame length itself, M should be the common divisor of the speech frame length at the same time.

The implementation method of near-end voice detection in the above-mentioned echo suppressor is characterized in that, for a fixed system, according to different design requirements, there are one or two optimal M values, and when the optimal M value is adopted The number of processor operation instruction cycles consumed is the smallest.

The above method for implementing near-end voice detection in the echo suppressor is characterized in that the optimal M value is obtained through actual testing.

The above method for implementing near-end voice detection in the echo suppressor is characterized in that, in the method, the calculation amount of all subframes is the same, but the calculation amount of each sampling point in each subframe is inconsistent.

The above-mentioned implementation method of near-end voice detection in the echo suppressor is characterized in that the echo of the user circuit is attenuated by at least 6 dB relative to the voice, and the echo amplitude attenuation factor u is set to 0.5.

Compared with the prior art, the method proposed by the present invention is characterized in that it can greatly reduce the amount of calculation required for program calculations when the calculation results are completely consistent with the existing methods, so more can be realized on the same processing platform. The echo suppression processing of the channel, or provide enough processing power to realize more functions, so the purpose of cost reduction can be achieved. At the same time, using this method, the maximum value of computation is consistent with the average value. When the parameters are constant, the computation is a certain value, which is convenient for developers to perform performance statistics and allocation of computing resources, and is conducive to the stability and compatibility of application products.

The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments, but not as a limitation of the present invention.

Description of drawings

Fig. 1 is a kind of flow chart of the implementation method of near-end voice detection in existing echo suppression;

Fig. 2 is another kind of flow chart of the implementation method of near-end voice detection in the existing echo suppression;

Fig. 3 is a flowchart of the method of the present invention.

Detailed ways

A method for realizing near-end voice detection of an echo suppressor in the field of communication provided by the present invention, the principle of the method is described as follows:

(1) Geigel algorithm is used for near-end speech detection, the filter length is N, the near-end input sequence is s(i), the far-end input sequence is y(i), and the echo amplitude attenuation factor is u. Then when the following inequality holds true, there is voice at the near end:

|s(i)|≥u*max(|y(i)|, |y(i-1)|, ..., |y(i-N+1)|)

(2) Take M sampling points as a subframe, and M is smaller than the filter length N. If the present invention is used in VOIP and other applications that have the concept of voice frame length, M should be the common divisor of the frame length. For example, in G.729 applications, the frame length is 10ms, corresponding to 80 sampling points of 8000Hz sampling, Then M should be divisible by 80. In the present invention, it can be guaranteed that the amount of computation of data in each subframe is completely consistent.

(3) When calculating the first point input of each subframe, first calculate the maximum value of |y(i)| to |y(i-N+M)| totaling N-M+1 points, and save is Frame_MAX, then take the maximum value of this value and the remaining M-1 points in the sliding window as the maximum value of the sampling point.

(4) When calculating the remaining M-1 points of the subframe one by one, the N-M+1 points used to calculate Frame_MAX are still kept in the sliding window at this time (same as when calculating the first point, no need to repeat operation), so it is only necessary to calculate the maximum value of Frame_MAX and the remaining M-1 points in the sliding window at this time, which is equivalent to taking the maximum value of M points.

Please refer to Fig. 3, will realize the inventive method, concrete operating steps are as follows:

Step 1: Initialize the subframe counter COUNT value to 0.

Step 2: Take the absolute value of the near-end input data |s(i)|. Take the absolute value of the remote input data |y(i)|, and save the data into the buffer BUFFER. The buffer stores a sequence of N points from |y(i)| to |y(i-N+1)|.

Step 3: Determine whether the subframe counter is 0? If the subframe counter COUNT is 0, perform step 4; if the subframe counter COUNT is not 0, directly perform step 5.

Step 4: If the subframe counter is 0, it means that the current remote data sampling point is the first sampling point of the subframe, and calculate the latest N-M+1 points in the remote input sliding window, ie |y(i )| to |y(i-N+M)| totals the maximum value of N-M+1 point data, and saves this value as Frame_MAX.

Step 5: Calculate the maximum value of Frame_MAX and the remaining M-1 points in the remote data buffer BUFFER, which is MAX.

Step 6: Compare the product u*MAX of the attenuation factor u and MAX with the absolute value |s(i)| of the near-end input data.

Step 7: If the inequality |s(i)|≥u*MAX holds true, it indicates that there is voice input at the near end (referring to the attenuation of the user circuit echo relative to the voice by at least 6dB, u can be taken as 0.5). If not, go directly to step 8.

Step 8: Update the subframe counter, and add 1 to the value of the subframe counter.

Step 9: Determine whether the subframe counter value is equal to M? If the subframe counter value is not equal to M, then jump to step 2 to calculate the next sampling point data; if the subframe counter value is equal to M, then perform step 10.

Step 10: Reset the value of the subframe counter to 0, and jump to step 2 to calculate the data of the next sampling point.

According to the above implementation steps, the amount of computation consumed for the entire subframe requires a total of N+M*(M-1) two-point maximum calculations instead of N*M two-point maximum calculations, thus achieving The purpose of reducing the amount of calculation is to ensure that the amount of calculation of each sub-frame is equal and the amount of calculation can be accurately obtained through calculation in the design stage.

Taking the implementation of a 16ms echo suppressor as an example, taking two instruction cycles as an example to realize the maximum value of two points inside the processor, for a system with 8000Hz sampling, the filter length N is (8000/1000)*16=128. If get M to be 16, then realize the near-end voice detection of a subframe M point according to the method in Fig. 1, the instruction cycle number consumed at calculating the maximum value is N*M*2=128*16*2=4096, and if according to The number of instruction cycles consumed by the method provided by the present invention for operation is (N+M*(M-1))*2=(128+16*(16-1))*2=736. Compared with the original method, more than 80% of processor computing power can be saved.

For a fixed system, according to different design requirements, there is one or two optimal M values, and the number of processor operation instruction cycles consumed when using this M value is the smallest. That is, (N+M*(M-1))/M is the smallest among all value ranges of M that meet the system requirements at this time. And the optimal M value can also be obtained through actual testing.

In summary, the features of the present invention are as follows:

(1) The present invention adopts the subframe concept, and processes the content of the previous subframe before the data of the next subframe arrives. The length M of the subframe is smaller than the filter order N. If it is applied to the application with the concept of speech frame, the length M of the subframe should be divisible by the number of sampling points FRAME of the speech frame.

(2) The present invention guarantees that the calculation amount of all subframes is consistent, but the calculation amount of each sampling point in each subframe is inconsistent.

(3) In the calculation of each frame in the present invention, the latest maximum value Frame_MAX of N-M+1 points is always calculated first, and only the maximum value of M points needs to be calculated in the subsequent sampling points of this subframe.

Of course, the present invention can also have other various embodiments, and those skilled in the art can make various corresponding changes and deformations according to the present invention without departing from the spirit and essence of the present invention, but these corresponding Changes and deformations should belong to the scope of protection of the appended claims of the present invention.

Claims (7)

1, an implementation method of near-end speech detection in an echo suppressor, adopts Geigel algorithm to carry out near-end speech detection, it is characterized in that, get M sampling points as a subframe in far-end input sequence y (i), M is less than the filter length N in the echo suppressor, when calculating the input of the first sampling point of each subframe, first obtain |y(i)| to | y(i-N+M)|The maximum value of the latest data of a total of N-M+1 sampling points, and save it as Frame_MAX, and then perform operations on the M sampling points of the subframe one by one, and take the maximum value Frame_MAX and the maximum value of the data of the remaining M-1 sampling points in the current sliding window are used as the maximum value MAX of each sampling point of the subframe, if the near-end input sequence s(i ) satisfies the condition |s(i)|≥u*MAX, then the near-end has voice, where u is the echo amplitude attenuation factor. 2, the implementation method of near-end voice detection in the echo suppressor according to claim 1, is characterized in that, the method for carrying out near-end voice detection to the subframe in described far-end input sequence y (i) comprises the following steps : Step 1, initialize a subframe counter with a count value of 0; Step 2: Take the absolute value of the near-end input data |s(i)|, take the absolute value of the far-end input data |y(i)|, and save the data |y(i)| to the remote data buffer In the area, the buffer stores a digital sequence of N sampling points from |y(i)| to |y(i-N+1)|; Step 3, if the subframe counter is 0, it means that the current remote data sampling point is the first sampling point of the subframe, calculate the latest from |y(i)| to |y(i -N+M)|A total of the maximum value of N-M+1 sampling point data, and save this value as Frame_MAX; if the count value of the subframe counter is not 0, then perform step 4; Step 4, calculate the maximum value MAX of Frame_MAX and the remaining M-1 sampling points in the remote data buffer; Step 5: Compare the product u*MAX of the attenuation factor u and MAX with the absolute value |s(i)| of the near-end input data. If the condition |s(i)|≥u*MAX is satisfied, it indicates that there is voice at the near-end enter; Step 6, updating the subframe counter, adding 1 to the subframe counter value; Step 7, if the subframe counter value is equal to M, then reset the subframe counter value to 0; Step eight, skip to step two, and calculate the next sampling point data of the subframe. 3. The method for realizing near-end speech detection in the echo suppressor according to claim 1 or 2 is characterized in that, when the method is applied to the application of the speech frame length concept itself, M should be the same time as the The common divisor of the speech frame length. 4. The implementation method of near-end voice detection in the echo suppressor according to claim 3, characterized in that, for a fixed system, according to different design requirements, there are one or two optimal M values, which are used in The optimal M value consumes the smallest number of processor operation instruction cycles. 5. The implementation method of near-end voice detection in the echo suppressor according to claim 4, characterized in that the optimal M value is obtained through actual testing. 6. The implementation method of near-end voice detection in the echo suppressor according to claim 4, characterized in that, in the method, the calculation amount of all subframes is the same, and the calculation amount of each sampling point in each subframe is The amount of operation is inconsistent. 7. The implementation method of near-end voice detection in the echo suppressor according to claim 3, characterized in that referring to the subscriber circuit echo attenuates at least 6dB relative to speech, the echo amplitude attenuation factor u is set to 0.5.

Images