A method and system for detecting voice call status

Abstract

The present invention relates to a voice call state detection method and system in the field of voice detection technology, comprising the following steps: when far-end voice exists in the voice call, a near-end mixed signal is obtained, and the near-end mixed signal is pre-processed to obtain a high-frequency signal; near-end voice existence estimation is performed for each frequency point of each frame of the high-frequency signal, and the near-end voice existence estimation probability is calculated based on the estimation result of the near-end voice existence estimation for each frequency point of each frame; when the near-end voice existence estimation probability is that the near-end voice may exist, the noise existence probability is calculated based on the estimation result, and the actual near-end voice existence probability is calculated based on the noise existence probability; based on the actual near-end voice existence probability, the voice call state is judged to be a far-end single-talk state or a double-talk state, thereby solving the problem that the existing double-end simultaneous talk detection algorithm is sensitive to background noise and cannot detect nonlinear correlation, thereby misjudging the situation.

Description

Technical Field

[0001] The present invention relates to the technical field of voice detection, and in particular to a method and system for detecting a voice call state. Background Art

[0002] During a voice call, echoes are inevitable due to the influence of the environment and hardware equipment. Acoustic Echo Cancellation (AEC) can effectively suppress echo signals during a call and improve the user's call experience. In existing echo cancellation solutions, adaptive filtering is mainly used for echo path estimation. In order to improve the stability and reliability of the adaptive filter, a double-talk detection method is often used for pre-processing. By detecting whether the voice is in a far-end single talk or a far-end and near-end simultaneous talk state, combined with the echo cancellation algorithm, the echo that occurs during double talk can be effectively eliminated to ensure the effective conduct of the call.

[0003] In the relevant technical solutions, the double-talk detection algorithm mainly judges the status by detecting the signal characteristics such as the energy or correlation of the near-end signal and the far-end signal. The existing double-talk detection algorithm relies more on good call scenes and acoustic equipment. However, when faced with more complex actual application scenarios, due to the sensitivity to background noise and the inability to detect nonlinear correlation, the existing double-end simultaneous talk detection algorithm often makes misjudgments, resulting in echo signal leakage and key voice error elimination, which in turn causes audio signal distortion and reduces call quality. Summary of the invention

[0004] The present invention aims at the shortcomings of the prior art and provides a voice call status detection method and system, which solves the problem that the existing two-end simultaneous talk detection algorithm is sensitive to background noise and cannot detect nonlinear correlation, resulting in misjudgment.

[0005] In order to solve the above technical problems, the present invention is solved by the following technical solutions:

[0006] A method for detecting a voice call state comprises the following steps:

[0007] When there is far-end voice in the voice call, a near-end mixed signal is obtained, and the near-end mixed signal is pre-processed to obtain a high-frequency signal;

[0008] Performing near-end speech presence estimation for each frequency point of each frame of the high-frequency signal, and calculating a near-end speech presence estimation probability based on an estimation result of the near-end speech presence estimation for each frequency point of each frame;

[0009] When the estimated probability of the presence of the near-end speech is that the near-end speech may exist, calculating the probability of the presence of noise based on the estimation result, and calculating the actual probability of the presence of the near-end speech based on the probability of the presence of noise;

[0010] Based on the actual probability of the near-end voice being present, it is determined whether the voice call state is a far-end single-talk state or a far-end dual-talk state.

[0011] Optionally, performing near-end speech presence estimation for each frequency point of each frame of the high-frequency signal includes the following steps:

[0012] Convert each frequency point of each frame of the high-frequency signal into a frequency domain signal;

[0013] Calculating an adaptive smoothing factor 1, and performing a smoothing process on the frequency domain signal based on the adaptive smoothing factor 1 to obtain a frequency domain smoothed power spectrum;

[0014] Calculate a first near-end speech presence estimation parameter based on the frequency domain smoothed power spectrum, and calculate a second near-end speech presence estimation parameter based on the frequency domain signal;

[0015] Based on the first near-end speech presence estimation parameter and the second near-end speech presence estimation parameter, an estimation result of near-end speech presence estimation is obtained.

[0016] Optionally, the updating formula of the adaptive smoothing factor 1 is:

[0017] ,in, Indicates the frequency domain smoothed power spectrum of the kth frequency point in the n-1th frame during a smoothing process. The initial value of the frequency domain smoothed power spectrum is 0. Represents the minimum noise estimation deviation value; represents the minimum value of the frequency domain smoothed power spectrum tracked up to the kth frequency point in the n-1th frame during a smoothing process, and The initial value is set to greater than 0.

[0018] Optionally, a calculation formula for the first near-end speech presence estimation parameter is:

[0019] , It represents the frequency domain smoothed power spectrum of the kth frequency point in the nth frame during a smoothing process. Indicates the minimum value of the frequency domain smoothed power spectrum tracked up to the kth frequency point of the nth frame during a smoothing process. ;

[0020] The calculation formula of the second near-end speech presence estimation parameter is:

[0021] , Represents the frequency domain signal of the kth frequency point in the nth frame.

[0022] Optionally, obtaining an estimation result of near-end speech presence estimation includes the following steps:

[0023] Setting a near-end speech presence estimation threshold 1 and a near-end speech presence estimation threshold 2;

[0024] When the first near-end speech presence estimation parameter is greater than near-end speech presence estimation threshold one, and the second near-end speech presence estimation parameter is greater than near-end speech presence estimation threshold two, the estimation result is that near-end speech exists; otherwise, the estimation result is that near-end speech does not exist.

[0025] Optionally, calculating the near-end speech presence estimation probability based on the estimation result of the near-end speech presence estimation at each frequency point of each frame includes the following steps:

[0026] The estimated probability of the presence of near-end speech in the 2 / 3~3 / 4 frequency band in the high-frequency signal is calculated, and the calculation formula is as follows:

[0027] ,in, It represents the estimation result of the near-end speech presence estimation of the k-th frequency point in the n-th frame, and M represents the frame length.

[0028] Optionally, calculating the noise existence probability based on the estimation result comprises the following steps:

[0029] transforming each frequency point of each frame of the high-frequency signal into a frequency domain signal, and performing secondary smoothing processing on the frequency domain signal based on the estimation result;

[0030] Calculating an adaptive smoothing factor 2, and performing time-dimensional smoothing processing on the power spectrum after the secondary smoothing processing based on the adaptive smoothing factor 2 to obtain a time-domain smoothed power spectrum;

[0031] Calculate a first noise presence parameter based on the time domain smoothed power spectrum, and calculate a second noise presence parameter based on the frequency domain signal;

[0032] A noise existence probability is obtained based on the first noise existence parameter and the second noise existence parameter.

[0033] Optionally, the updating formula of the adaptive smoothing factor 2 is:

[0034] ,in, It represents the time domain smoothed power spectrum of the kth frequency point in the n-1th frame. The initial value of the time domain smoothed power spectrum is 0. Represents the minimum noise estimation deviation value; represents the tracking minimum value of the time domain smoothed power spectrum up to the kth frequency point in the n-1th frame, and The initial value is set to greater than 0.

[0035] Optionally, calculating the actual probability of near-end speech existence based on the noise existence probability comprises the following steps:

[0036] Based on the noise existence probability, the near-end speech existence probability of each frequency point in each frame is calculated, wherein the calculation formula of the near-end speech existence probability is: ,in, represents the probability of noise existence; represents the posterior signal-to-noise ratio of the signal; represents the signal-to-noise ratio a priori;

[0037] Based on the near-end speech existence probability of each frequency point in each frame, the actual probability of the near-end speech existence is calculated, where the calculation formula for the actual probability of the near-end speech existence is: , M represents the frame length.

[0038] Optionally, based on the actual probability of the presence of the near-end voice, determining whether the voice call state is a far-end single-talk state or a far-end dual-talk state comprises the following steps:

[0039] A near-end speech presence threshold 2 is set, and the near-end speech presence threshold 2 is set to: 0.8<near-end speech presence threshold 2<1;

[0040] When the actual probability of the near-end speech existence is less than the near-end speech existence threshold 2, the speech double-talk state is determined to be the far-end single-talk state; when the actual probability of the near-end speech existence is greater than the near-end speech existence threshold 2, the speech double-talk state is determined to be the double-talk state.

[0041] Optionally, before detecting the presence of far-end voice in the voice call, the following steps are further included:

[0042] After up-sampling the far-end reference signal, the near-end mixed signal is collected by the near-end device;

[0043] The short-time energy of the far-end reference signal and the near-end mixed signal of each frame is calculated, and a first voice call state judgment is performed based on the calculation result, wherein the result of the first voice call state judgment includes near-end single talk, silence, and the presence of far-end single talk.

[0044] A voice call status detection system, the voice call status detection system executing any of the above-mentioned voice call status detection methods, comprising a signal acquisition unit, a first judgment unit, a signal processing unit, a near-end voice presence estimation unit, a near-end voice presence calculation unit, and a second judgment unit;

[0045] The signal acquisition unit is used to acquire a near-end mixed signal through a near-end device after up-sampling the far-end reference signal;

[0046] The first judgment unit is used to calculate the short-time signal energy of the far-end reference signal and the near-end mixed signal of each frame, and perform a first voice call state judgment based on the calculation result, wherein the result of the first voice call state judgment includes near-end single talk, silence, and existence of far-end single talk;

[0047] The signal processing unit is used to pre-process the near-end mixed signal to obtain a high-frequency signal when there is a far-end voice in the voice call;

[0048] The near-end speech presence estimation unit is used to perform near-end speech presence estimation for each frequency point of each frame of the high-frequency signal, and calculate the near-end speech presence estimation probability based on the estimation result of the near-end speech presence estimation for each frequency point of each frame;

[0049] The near-end speech existence calculation unit is used to calculate the noise existence probability based on the estimation result when the near-end speech existence estimation probability is that the near-end speech may exist, and calculate the near-end speech existence actual probability based on the noise existence probability;

[0050] The second judgment unit is used to judge whether the voice call state is a far-end single-talk state or a far-end dual-talk state based on the actual probability of the near-end voice existence.

[0051] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects:

[0052] By calculating the probability of noise existence and calculating the actual probability of near-end speech existence based on the probability of noise existence, the influence of background noise on the result of voice call status judgment is eliminated, thereby ensuring the accuracy of judgment. At the same time, through the estimated probability of near-end speech existence and the actual probability of near-end speech existence, the near-end speech is judged twice, thereby improving the reliability of the result judgment when the voice call status is the far-end single-talk state or the dual-talk state. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0054] Figure 1 This is a flow chart of a voice call status detection method proposed in the first embodiment. DETAILED DESCRIPTION

[0055] The present invention is further described in detail below in conjunction with embodiments. The following embodiments are for explanation of the present invention but the present invention is not limited to the following embodiments.

[0056] Before performing voice call status detection, the voice call status and each voice signal during the process are explained as follows:

[0057] For the voice call status, during a voice call, the voice call status is mainly divided into four situations: near-end single talk, far-end single talk, both ends talking at the same time, and silence. In the voice call status detection of this application, it is mainly used to distinguish the above four situations;

[0058] For the explanation of each voice signal, during a voice call, take the call between a smart device and a mobile phone as an example. When the voice call status is detected on the mobile phone, for the mobile phone, the mobile phone is the near end and the smart device is the far end. The sound emitted by the user on the mobile phone and the background sound are called near-end voice signals, the sound recorded by the microphone on the smart device is called the far-end reference signal, the voice emitted by the user on the smart device is called the far-end voice signal, and the microphone on the mobile phone records both the far-end reference signal and the near-end voice signal. The two signals are mixed together and are called near-end mixed signals.

[0059] Embodiment 1

[0060] like Figure 1 As shown, a voice call status detection method includes the following steps: S1, after upsampling the far-end reference signal, the near-end mixed signal is collected by the near-end device, specifically, the near-end mixed signal d(n) is sampled at a high sampling rate (for example, at a sampling rate of 16 Hz), and the far-end reference signal x(n) is sampled at a low sampling rate (for example, at a sampling rate of 8 Hz). The far-end reference signal is received at the receiving end, upsampled, and then played by the near-end speaker. At this time, since the far-end reference signal is collected at a different sampling rate, there is no energy in the high-frequency region. Therefore, in the near-end mixed signal d(n), only the near-end voice signal s(n) has energy in the high-frequency region, so that the far-end voice signal and the near-end voice signal have differences in the high-frequency band, thereby improving the reliability and stability of the detection result when the signal short-time energy is used for status detection in the subsequent process.

[0061] S2. Calculate the short-time energy of the far-end reference signal and the near-end mixed signal of each frame, and make a first voice call status judgment based on the calculation result, wherein the result of the first voice call status judgment includes near-end single talk, silence, and the presence of far-end single talk.

[0062] Specifically, the first voice call state is determined by using the short-time energy of the signal to distinguish between the near-end single talk and silence states.

[0063] First, the short-time energy of the remote reference signal of each frame is calculated. The calculation formula is:

[0064] , where M is the frame length, is the short-time energy of the nth frame of the far-end reference signal x(n).

[0065] Then the short-time energy of the near-end mixed signal of each frame is calculated, and the calculation formula is:

[0066] , is the short-time energy of the nth frame of the near-end mixed signal d(n).

[0067] Next, set the far-end signal voice presence threshold T1 and the near-end signal voice presence threshold T2, where T1>0, T2>0, when <T1, and When >T2, it indicates that there is no far-end voice signal and there is a near-end voice signal, then the result of the first voice call state judgment is near-end single talk, and the state judgment result is output.

[0068] when <T1, and When <T2, it indicates that there is no far-end voice signal and no near-end voice signal, then the result of the first voice call state judgment is silence, and the state judgment result is output.

[0069] when >T1, and When >T2, it indicates that there must be a far-end voice signal. Since the near-end mixed signal contains the far-end voice signal and the near-end voice signal, >T2 cannot accurately determine whether there is a near-end voice signal, but can only determine whether there is a far-end single talk, so it is necessary to enter step S3 for further determination.

[0070] S3. When there is far-end voice in the voice call, preprocess the near-end mixed signal to obtain a high-frequency signal. Specifically, high-pass filter the near-end mixed signal to obtain a high-frequency signal d of the near-end mixed signal. h (n) To prevent the far-end voice signal from introducing signal distortion during the playback of the near-end speaker and affecting the voice call status detection effect, the cutoff frequency of the high-pass filter is set to 2 / 3 of the near-end mixed signal sampling rate.

[0071] On the other hand, since the existence of the far-end voice signal has been determined before step S3, it has been determined that there must be far-end speech. Therefore, after step S3, it is only necessary to detect whether there is near-end speech. Since there is a difference between the far-end voice signal and the near-end mixed signal in the high-frequency band, after high-pass filtering, only the high-frequency band of the near-end mixed signal is detected. Since the far-end voice signal has no energy in the high-frequency band, the far-end voice signal has no effect on the detection result during detection, thereby improving the accuracy and reliability of the detection result.

[0072] After completing step S3, execute step S4 to perform near-end speech presence estimation for each frequency point of each frame of the high-frequency signal, and calculate the near-end speech presence estimation probability based on the estimation result of the near-end speech presence estimation for each frequency point of each frame.

[0073] Specifically, the near-end speech presence estimation is performed for each frequency point of each frame of the high-frequency signal, including the following steps: first, each frequency point of each frame of the high-frequency signal is transformed into a frequency domain signal, and the transformation formula is:

[0074] ,in, represents the high-frequency signal of the nM+mth frame, j represents an imaginary number, which is specifically set to , represents the frequency domain signal of the kth frequency point in the nth frame, w represents the window function, and in this embodiment, the window function is a Hamming window.

[0075] Then, the adaptive smoothing factor 1 is calculated, and the frequency domain signal is smoothed in the frequency dimension based on the adaptive smoothing factor 1 to obtain a frequency domain smoothed power spectrum.

[0076] Specifically, the calculation formula of the frequency domain smoothed power spectrum is:

[0077] ,in, is a normalized window function with a length of 2l+1, l>0, l represents half of the window length, and , represents the adaptive smoothing factor 1. It should be noted that The initial value of is 0, so when calculating When the frequency domain point of the previous frame is a known value.

[0078] In order to track the non-stationary segment in the speech signal more stably, the adaptive smoothing factor is set to 1, and its update formula is:

[0079] ,in, Indicates that during one smoothing process, the frequency domain smoothed power spectrum of the kth frequency point in the n-1th frame is a known value, and the initial value of the frequency domain smoothed power spectrum is 0; represents the minimum noise estimation deviation value, which is 1.66 in this embodiment; Indicates the tracking minimum value of the frequency domain smoothed power spectrum ending at the kth frequency point of the n-1th frame during a smoothing process, which is a known value, where The initial value of is greater than 0, and in this embodiment can be set to 32768, which can be calculated by the following formula:

[0080] .

[0081] In this embodiment, an adaptive smoothing factor of one is used to prevent the tracking minimum value from entering a wider resonance peak region, so as to better track the non-stationary region of the signal.

[0082] Next, a first near-end speech presence estimation parameter is calculated based on the frequency domain smoothed power spectrum, and a second near-end speech presence estimation parameter is calculated based on the frequency domain signal, wherein the calculation formula of the first near-end speech presence estimation parameter is:

[0083] , It represents the frequency domain smoothed power spectrum of the kth frequency point in the nth frame during a smoothing process. Indicates the minimum value of the frequency domain smoothed power spectrum tracked up to the kth frequency point of the nth frame during a smoothing process. ;

[0084] The calculation formula of the second near-end speech presence estimation parameter is:

[0085] , Represents the frequency domain signal of the kth frequency point in the nth frame.

[0086] Finally, based on the first near-end speech presence estimation parameter and the second near-end speech presence estimation parameter, an estimation result of the near-end speech presence estimation is obtained, which specifically includes the following steps: setting a near-end speech presence estimation threshold value 1 and near-end speech presence estimation threshold 2 ; When the first near-end speech presence estimation parameter is greater than the near-end speech presence estimation threshold 1, and the second near-end speech presence estimation parameter is greater than the near-end speech presence estimation threshold 2, the estimation result is that the near-end speech exists, otherwise the estimation result is that the near-end speech does not exist, which can be expressed by the formula: , The estimated result of the near-end speech presence estimation of the k-th frequency point in the n-th frame. When it is 1, it means that the k-th frequency point in the n-th frame may have a near-end speech. When it is 0, it means that the k-th frequency point in the n-th frame does not have a near-end speech and the probability of existence is 0. In this embodiment, the near-end speech presence estimation threshold is set to 1. The value is 4.6, which is the near-end speech presence estimation threshold 2. The value is 1.67.

[0087] In this embodiment, the low-frequency band of the high-frequency region of the near-end mixed signal is used to estimate the presence of near-end speech, thereby reducing the influence of high-frequency band noise on the judgment result. By setting the first near-end speech presence estimation parameter and the second near-end speech presence estimation parameter, the weak sound part can be better detected to prevent the weak sound area from being missed.

[0088] Furthermore, since the high-frequency signal is the high-frequency region of the near-end mixed signal, and the speech signal is mainly concentrated in the low-frequency band of the high-frequency signal, when calculating the near-end speech presence estimation probability based on the estimation result of the near-end speech presence estimation at each frequency point of each frame, the near-end speech presence estimation probability of the 2 / 3~3 / 4 frequency band in the high-frequency signal is calculated, and the calculation formula is as follows: .

[0089] Next, set the near-end voice presence threshold , and 0.6< <1, if , it is determined that there is no voice activity in this frequency band, and the output judgment result is far-end single talk; if , it is determined that there may be voice activity in this frequency band, that is, there may be near-end voice, and the operation of step S5 is performed.

[0090] S5. When the estimated probability of the presence of the near-end speech is that the near-end speech may exist, the probability of the presence of noise is calculated based on the estimation result, and the actual probability of the presence of the near-end speech is calculated based on the probability of the presence of noise.

[0091] The calculation of the noise existence probability based on the estimation result includes the following steps: first, each frequency point of each frame of the high-frequency signal is transformed into a frequency domain signal, and the frequency domain signal is subjected to secondary smoothing in the frequency domain dimension based on the estimation result. The calculation formula of the secondary smoothing is as follows:

[0092] ;in, The initial value of is 0.

[0093] Secondly, the adaptive smoothing factor 2 is calculated, and the power spectrum after the secondary smoothing is smoothed in the time dimension based on the adaptive smoothing factor 2 to obtain the time domain smoothed power spectrum, where the calculation formula of the time domain smoothed power spectrum is:

[0094] , represents the adaptive smoothing factor 2. It should be noted that The initial value of is 0, so when calculating When the frequency domain point of the previous frame is a known value.

[0095] In order to track the non-stationary segment in the speech signal more stably, the adaptive smoothing factor 2 is set, and its update formula is:

[0096] ,in, It represents the time domain smoothed power spectrum of the kth frequency point in the n-1th frame. The initial value of the time domain smoothed power spectrum is 0. Represents the minimum noise estimation deviation value; Indicates the tracking minimum value of the time domain smoothed power spectrum ending at the kth frequency point in the n-1th frame.

[0097] Minimum tracking of a time-smoothed power spectrum: ,in, The initial value of is greater than 0, and can be set to 32768 in this embodiment.

[0098] In this embodiment, an adaptive smoothing factor of two is used to prevent the tracking minimum value from entering a wider resonance peak region, so as to better track the non-stationary region of the signal.

[0099] Next, the first noise presence parameter is calculated based on the time domain smoothed power spectrum, and the second noise presence parameter is calculated based on the frequency domain signal, wherein the calculation formula of the first noise presence parameter is:

[0100] , represents the time domain smoothed power spectrum of the kth frequency point in the nth frame, It represents the tracking minimum value of the time domain smoothed power spectrum ending at the kth frequency point of the nth frame. The calculation formula of the second noise existence parameter is: .

[0101] Finally, based on the first noise existence parameter and the second noise existence parameter, the noise existence probability is obtained , specifically including the following steps: setting the noise existence threshold and noise presence threshold 2 ,in, >0, so the probability of noise existence is:

[0102] , Indicates the probability value of noise at the kth frequency point in the nth frame. When it is 1, it means that the probability of noise at the kth frequency point in the nth frame is 100%. When it is 0, it means that there is no noise at the kth frequency point in the nth frame, and the probability value of noise is 0. When The calculation result of is used to obtain the noise existence probability value. In this embodiment, Set to equal.

[0103] Further, calculating the actual probability of near-end speech existence based on the noise existence probability includes the following steps:

[0104] Based on the noise existence probability, the near-end speech existence probability of each frequency point in each frame is calculated, where the calculation formula for the near-end speech existence probability is: ,in, represents the probability of noise existence; represents the posterior signal-to-noise ratio of the signal; represents the signal-to-noise ratio.

[0105] Among them, the calculation formula of the signal prior signal-to-noise ratio is:

[0106] ,in, is the weight factor, and its value range is 0.3< <1, is the Wiener gain of the previous frame signal.

[0107] The calculation formula of the signal posterior signal-to-noise ratio is:

[0108] ,

[0109] in, represents the estimated noise spectrum and is calculated as follows:

[0110] .

[0111] At the same time, in the calculation Then set the noise spectrum over-estimation protection mechanism:

[0112] , where 1<μ<10, thus the probability of near-end speech existence at each frequency point in each frame can be calculated. , calculate the actual probability of near-end speech , where the calculation formula for the actual probability of near-end speech existence is: , M represents the frame length.

[0113] S6: Based on the actual probability of the presence of near-end voice, determine whether the voice call state is a far-end single-talk state or a far-end dual-talk state, specifically including the following steps: setting a near-end voice presence threshold value 2 , and the near-end speech presence threshold 2 is set to: 0.8< <1; when the actual probability of the presence of near-end speech is less than the near-end speech existence threshold 2, it indicates that there is no near-end single talk, and the speech double talk state is determined to be far-end single talk; when the actual probability of the presence of near-end speech is greater than the near-end speech existence threshold 2, it indicates that there is near-end single talk, and the speech double talk state is determined to be the double talk state.

[0114] Finally, the detection result of the voice call status is obtained It is expressed as:

[0115] .

[0116] When step S4 determines that there may be voice activity in this frequency band, the frequency domain signal is secondary smoothed in the frequency domain dimension based on the estimation result, and the posterior signal-to-noise ratio and the prior signal-to-noise ratio of the signal are used to perform a second near-end voice presence judgment in the entire frequency band of the signal, thereby increasing the reliability of the detection result and avoiding nonlinear echo signal interference and echo leakage.

[0117] At this point, the near-end mixed signal is detected. In this stage, the signal features are used to determine the near-end speech twice, thereby increasing the reliability of the determination result.

[0118] Embodiment 2

[0119] A voice call status detection system comprises a signal acquisition unit, a first judgment unit, a signal processing unit, a near-end voice presence estimation unit, a near-end voice presence calculation unit, and a second judgment unit;

[0120] A signal acquisition unit, configured to acquire a near-end mixed signal through a near-end device after up-sampling the far-end reference signal;

[0121] A first judgment unit is used to calculate the short-time energy of the far-end reference signal and the near-end mixed signal of each frame, and perform a first voice call state judgment based on the calculation result, wherein the result of the first voice call state judgment includes near-end single talk, silence, and existence of far-end single talk;

[0122] A signal processing unit, used for preprocessing the near-end mixed signal to obtain a high-frequency signal when there is a far-end voice in the voice call;

[0123] A near-end speech presence estimation unit, configured to perform near-end speech presence estimation for each frequency point of each frame of the high-frequency signal, and calculate a near-end speech presence estimation probability based on an estimation result of the near-end speech presence estimation for each frequency point of each frame;

[0124] A near-end speech existence calculation unit, configured to calculate the noise existence probability based on the estimation result when the near-end speech existence estimation probability is that the near-end speech may exist, and calculate the near-end speech existence actual probability based on the noise existence probability;

[0125] The second judgment unit is used to judge whether the voice call state is a far-end single-talk state or a far-end dual-talk state based on the actual probability of the near-end voice being present.

[0126] Since the voice call status detection system executes the voice call status detection method described in the first embodiment, it will not be repeated in this embodiment.

[0127] The above is only a preferred embodiment of the present invention, and is not any formal or substantial limitation of the present invention. It should be pointed out that ordinary technicians in this technical field can make several improvements and supplements without departing from the method of the present invention, and these improvements and supplements should also be regarded as the protection scope of the present invention. Any technician familiar with this profession, without departing from the spirit and scope of the present invention, can make some changes, modifications and evolutions of the technical content disclosed above, which are equivalent embodiments of the present invention; at the same time, any changes, modifications and evolutions of any equivalent changes made to the above embodiments based on the essential technology of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A method for detecting a voice call status, characterized in that: The following steps are involved: When there is far-end voice in the voice call, a near-end mixed signal is obtained, and the near-end mixed signal is pre-processed to obtain a high-frequency signal; Performing near-end speech presence estimation for each frequency point of each frame of the high-frequency signal, and calculating a near-end speech presence estimation probability based on an estimation result of the near-end speech presence estimation for each frequency point of each frame; When the estimated probability of the presence of the near-end speech is that the near-end speech may exist, calculating the probability of the presence of noise based on the estimation result, and calculating the actual probability of the presence of the near-end speech based on the probability of the presence of noise; Based on the actual probability of the near-end voice being present, determining whether the voice call state is a far-end single-talk state or a far-end dual-talk state; Calculating the noise existence probability based on the estimation result includes the following steps: Each frequency point of each frame of the high-frequency signal is transformed into a frequency domain signal, and the frequency domain signal is subjected to secondary smoothing processing based on the estimation result; an adaptive smoothing factor 2 is calculated, and the power spectrum after the secondary smoothing processing is subjected to time dimension smoothing processing based on the adaptive smoothing factor 2 to obtain a time domain smoothed power spectrum; a first noise existence parameter is calculated based on the time domain smoothed power spectrum, and a second noise existence parameter is calculated based on the frequency domain signal; a noise existence probability is obtained based on the first noise existence parameter and the second noise existence parameter; Calculating the actual probability of near-end speech existence based on the noise existence probability includes the following steps: Based on the noise existence probability, the near-end speech existence probability of each frequency point in each frame is calculated, wherein the calculation formula of the near-end speech existence probability is: ,in, represents the probability of noise existence; represents the posterior signal-to-noise ratio of the signal; represents the signal-to-noise ratio a priori; Based on the near-end speech existence probability of each frequency point in each frame, the actual probability of the near-end speech existence is calculated, where the calculation formula for the actual probability of the near-end speech existence is: , M represents the frame length.

2. A method for detecting a voice call status according to claim 1, characterized in that: The near-end speech presence estimation is performed for each frequency point of each frame of the high-frequency signal, including the following steps: Convert each frequency point of each frame of the high-frequency signal into a frequency domain signal; Calculating an adaptive smoothing factor 1, and performing a smoothing process on the frequency domain signal based on the adaptive smoothing factor 1 to obtain a frequency domain smoothed power spectrum; Calculate a first near-end speech presence estimation parameter based on the frequency domain smoothed power spectrum, and calculate a second near-end speech presence estimation parameter based on the frequency domain signal; Based on the first near-end speech presence estimation parameter and the second near-end speech presence estimation parameter, an estimation result of near-end speech presence estimation is obtained.

3. A method for detecting a voice call status according to claim 2, characterized in that: The updating formula of the adaptive smoothing factor 1 is: ,in, Indicates the frequency domain smoothed power spectrum of the kth frequency point in the n-1th frame during a smoothing process. The initial value of the frequency domain smoothed power spectrum is 0. Represents the minimum noise estimation deviation value; represents the minimum value of the frequency domain smoothed power spectrum tracked up to the kth frequency point in the n-1th frame during a smoothing process, and The initial value is set to greater than 0.

4. A method for detecting a voice call status according to claim 3, characterized in that: The calculation formula of the first near-end speech presence estimation parameter is: , It represents the frequency domain smoothed power spectrum of the kth frequency point in the nth frame during a smoothing process. Indicates the minimum value of the frequency domain smoothed power spectrum tracked up to the kth frequency point of the nth frame during a smoothing process. ; The calculation formula of the second near-end speech presence estimation parameter is: , Represents the frequency domain signal of the kth frequency point in the nth frame.

5. A method for detecting a voice call status according to claim 2, characterized in that: Obtaining an estimation result of near-end speech presence estimation includes the following steps: Setting a near-end speech presence estimation threshold 1 and a near-end speech presence estimation threshold 2; When the first near-end speech presence estimation parameter is greater than near-end speech presence estimation threshold one, and the second near-end speech presence estimation parameter is greater than near-end speech presence estimation threshold two, the estimation result is that near-end speech exists; otherwise, the estimation result is that near-end speech does not exist.

6. A method for detecting a voice call status according to claim 1, characterized in that: Calculating the near-end speech presence estimation probability based on the estimation result of the near-end speech presence estimation at each frequency point of each frame includes the following steps: The estimated probability of the presence of near-end speech in the 2 / 3~3 / 4 frequency band in the high-frequency signal is calculated, and the calculation formula is as follows: ,in, It represents the estimation result of the near-end speech presence estimation of the k-th frequency point in the n-th frame, and M represents the frame length.

7. A method for detecting a voice call status according to claim 1, characterized in that: The updating formula of the adaptive smoothing factor 2 is: ,in, It represents the time domain smoothed power spectrum of the kth frequency point in the n-1th frame. The initial value of the time domain smoothed power spectrum is 0. Represents the minimum noise estimation deviation value; represents the tracking minimum value of the time domain smoothed power spectrum up to the kth frequency point in the n-1th frame, and The initial value is set to greater than 0.

8. A method for detecting a voice call status according to claim 1, characterized in that: Based on the actual probability of the near-end voice being present, determining whether the voice call state is a far-end single-talk state or a far-end dual-talk state includes the following steps: A near-end speech presence threshold 2 is set, and the near-end speech presence threshold 2 is set to: 0.8<near-end speech presence threshold 2<1; When the actual probability of the near-end speech existence is less than the near-end speech existence threshold 2, the speech double-talk state is determined to be the far-end single-talk state; when the actual probability of the near-end speech existence is greater than the near-end speech existence threshold 2, the speech double-talk state is determined to be the double-talk state.

9. A method for detecting a voice call status according to claim 1, characterized in that: Before detecting the presence of far-end voice in a voice call, the following steps are also included: After up-sampling the far-end reference signal, the near-end mixed signal is collected by the near-end device; The short-time energy of the far-end reference signal and the near-end mixed signal of each frame is calculated, and a first voice call state judgment is performed based on the calculation result, wherein the result of the first voice call state judgment includes near-end single talk, silence, and existence of far-end single talk.

10. A voice call status detection system, characterized in that: The voice call status detection system implements the voice call status detection method according to any one of claims 1 to 8, including a signal acquisition unit, a first judgment unit, a signal processing unit, a near-end voice presence estimation unit, a near-end voice presence calculation unit, and a second judgment unit; The signal acquisition unit is used to acquire a near-end mixed signal through a near-end device after up-sampling the far-end reference signal; The first judgment unit is used to calculate the short-time signal energy of the far-end reference signal and the near-end mixed signal of each frame, and perform a first voice call state judgment based on the calculation result, wherein the result of the first voice call state judgment includes near-end single talk, silence, and existence of far-end single talk; The signal processing unit is used to pre-process the near-end mixed signal to obtain a high-frequency signal when there is a far-end voice in the voice call; The near-end speech presence estimation unit is used to perform near-end speech presence estimation for each frequency point of each frame of the high-frequency signal, and calculate the near-end speech presence estimation probability based on the estimation result of the near-end speech presence estimation for each frequency point of each frame; The near-end speech existence calculation unit is used to calculate the noise existence probability based on the estimation result when the near-end speech existence estimation probability is that the near-end speech may exist, and calculate the near-end speech existence actual probability based on the noise existence probability; The second judgment unit is used to judge whether the voice call state is a far-end single-talk state or a far-end dual-talk state based on the actual probability of the near-end voice existence.

Images