Method for multi-speaker recognition based on fmcw radar

Abstract

The application discloses a multi-speaker recognition method based on FMCW radar, and the method comprises the following steps: positioning multi-speakers by using a DOA algorithm; weighting the received signals based on adaptive digital beamforming of DOA to obtain echo signals corresponding to angles; calculating distance dimension FFT on each echo signal, selecting a distance gate where a target is located, and extracting radar echo signals of the target; filtering out interference from the signals by using a band-pass filter; eliminating the influence caused by body movement by using a curve fitting method; dividing voiced sections by using a two-level threshold judgment method; processing and extracting a fundamental frequency of a vocal cord vibration signal by using an improved VMD algorithm; extracting speech parameters from the filtered radar echo signals; fusing the extracted fundamental frequency signals and the speech parameters of the radar echo signals into a GRU neural network to obtain a speaker recognition result. The method is effective and feasible, reliable in performance, and can accurately realize multi-speaker recognition.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular to a multi-speaker recognition method based on FMCW radar. Background Art

[0002] Language, with its rich connotations, is the most important medium of communication between people and the most direct way to convey information. Traditionally, acoustic sensors, such as microphones or microphone arrays, have been used to obtain speech information. However, their high sensitivity to acoustic noise prevents them from providing high-quality reproduction of the original sound in many scenarios with high background noise.

[0003] In 1953, Cherry proposed the cocktail party problem, which is a classic problem in the field of speech processing. It refers to the fact that when people speak at a cocktail party, the sound signals will overlap. When other applications such as speech recognition and speaker verification are required, speech separation algorithms are needed. In 1991, C. Jutten proposed the famous HJ algorithm to solve the problem of blind source separation. In 1994, Common P proposed the independent component analysis method. In 2000, Lewicki used the over-complete sparse representation based on the maximum a posteriori probability to obtain the sparse features of the signal and proposed the sparse component analysis (SCA) theory. In 2015, Cai et al. proposed a CMNN network structure. In 2019, Sun et al. proposed a two-stage algorithm using a single-output DNN to separate single-channel speech in a reverberant environment. Compared with traditional separation methods, the use of deep learning methods for speech separation has greatly improved the accuracy, but it has high requirements for the environment and cannot achieve a high accuracy in a noisy background.

[0004] In 1976, Atal et al. proposed linear prediction cepstral coefficients by analyzing and modeling the vocal tract, applying them to speaker verification. In 1994, to further improve the accuracy of speaker verification, Matsui applied the hidden Markov model to speaker verification. In 1995, Reynolds proposed a method for speaker identification using a Gaussian mixture model. In 2018, Ravanellil, Bengio et al. proposed an interpretable neural network structure for directly processing speech time series signals - SincNet. Speaker identification based on deep learning and neural networks can achieve a high accuracy rate, but microphone-based speaker recognition systems are vulnerable to replay attacks using the target speaker's voice, compromising the security of some speech recognition devices. Summary of the Invention

[0005] The object of the present invention is to provide a multi-speaker recognition method based on FMCW radar to address the problems existing in the above-mentioned prior art.

[0006] The technical solution for achieving the purpose of the present invention is: a multi-speaker recognition method based on FMCW radar, the method comprising the following steps:

[0007] Step 1: Receive the radar echo signal, use the DOA algorithm to locate the angle of the speaker, and use the MUSIC algorithm to obtain the angle θ of the speaker;

[0008] Step 2: Based on the speaker angle θ obtained in step 1, the weight vector W corresponding to the desired signal is obtained using the adaptive digital beamforming algorithm based on the LCMV criterion. LCMV , after weighting, the radar echo signal y1(t) in the direction of θ is obtained;

[0009] Step 3: Perform a short-time Fourier transform of the distance dimension on the radar echo signal y1(t) in the θ direction obtained in step 2 to obtain the signal y2(t) of the desired target;

[0010] Step 4: Perform motion compensation on the desired target signal y2(t) obtained in step 3 to obtain a signal y3(t) that eliminates the influence of body motion.

[0011] Step 5: High-pass filter the motion-compensated signal y3(t) obtained in step 4 to obtain an output signal y4(t);

[0012] Step 6: Perform voice segment cutting on the signal y4(t) obtained in step 5 after removing interference, to obtain the voice segment signal y5(t);

[0013] Step 7: Based on the voiced segment signal y5(t) obtained in step 6, the baseband signal y6(t) of the radar signal is extracted by using an improved VMD transform;

[0014] Step 8: extracting speech feature parameters based on the voiced segment signal y5(t) obtained in step 6;

[0015] In step 9, the fundamental frequency signal y6(t) obtained in step 7 and the speech feature parameters obtained in step 8 are fed into the GRU neural network to realize speaker recognition.

[0016] Compared with the existing technology, the present invention has the following significant advantages: 1) a two-transmitter, four-receiver FMCW radar is used in combination with digital beamforming technology to obtain the speaker's vocal cord vibration signal, which is not affected by background noise; 2) an improved VMD algorithm is proposed, which has adaptive characteristics for extracting the fundamental frequency and can avoid the influence of over-segmentation; 3) a speaker recognition method based on vocal cord vibration signals is proposed, which is not affected by background noise. Compared with microphone-based speaker recognition, speaker recognition using vocal cord vibration has higher security.

[0017] The present invention is further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of the multi-speaker recognition method based on FMCW radar of the present invention.

[0019] Figure 2 A schematic diagram of DOA estimation based on the MUSIC algorithm.

[0020] Figure 3 is an adaptively formed beam pattern.

[0021] Figure 4 (a) is the spectrum diagram of the signal frequency shift caused by body movement. Figure 4 (b) is the spectrum diagram after body motion compensation.

[0022] Figure 5 This is the time domain signal diagram after passing through the high-pass filter.

[0023] Figure 6 Schematic diagram of voice segment cutting obtained using the two-level threshold decision method.

[0024] Figure 7 Signal diagrams of different modes using improved VMD transform.

[0025] Figure 8 This is the time-frequency diagram of the fundamental frequency signal after mode selection.

[0026] Figure 9 This is the fundamental frequency diagram after the envelope is obtained through Hilbert transform.

[0027] Figure 10 This is the confusion matrix output by the GRU neural network. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0029] In addition, if the present invention has descriptions involving "first", "second", etc., the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the fact that they can be implemented by ordinary technicians in this field. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0030] In one embodiment, combined Figure 1 , provides a multi-speaker recognition method based on FMCW radar, the method comprising the following steps:

[0031] Step 1: Receive the radar echo signal, use the DOA algorithm to locate the angle of the speaker, and use the MUSIC algorithm to obtain the angle θ of the speaker;

[0032] Step 2: Based on the speaker angle θ obtained in step 1, the weight vector W corresponding to the desired signal is obtained using the adaptive digital beamforming algorithm based on the LCMV criterion. LCMV , after weighting, the radar echo signal y1(t) in the direction of θ is obtained;

[0033] Step 3: Perform a short-time Fourier transform of the distance dimension on the radar echo signal y1(t) in the θ direction obtained in step 2 to obtain the signal y2(t) of the desired target;

[0034] Step 4: Perform motion compensation on the desired target signal y2(t) obtained in step 3 to obtain a signal y3(t) that eliminates the influence of body motion.

[0035] Step 5: High-pass filter the motion-compensated signal y3(t) obtained in step 4 to obtain an output signal y4(t);

[0036] Step 6: Perform voice segment cutting on the signal y4(t) obtained in step 5 after removing interference, to obtain the voice segment signal y5(t);

[0037] Step 7: Based on the voiced segment signal y5(t) obtained in step 6, the baseband signal y6(t) of the radar signal is extracted by using an improved VMD transform;

[0038] Step 8: extracting speech feature parameters based on the voiced segment signal y5(t) obtained in step 6;

[0039] In step 9, the fundamental frequency signal y6(t) obtained in step 7 and the speech feature parameters obtained in step 8 are fed into the GRU neural network to realize speaker recognition.

[0040] Furthermore, in one embodiment, the step 1 of obtaining the angle θ of the speaker using the MUSIC algorithm is specifically performed by using the MUSIC algorithm to estimate the direction of arrival and obtain the angle θ of the speaker. The process includes:

[0041] Step 1-1, obtain the covariance matrix based on N received signal vectors, which is:

[0042] R=E[XX H ]

[0043] Where R is the covariance matrix and X is the received signal matrix;

[0044] Step 1-2, perform eigenvalue decomposition on the covariance matrix, expressed as:

[0045]

[0046] Among them, Λ S and σ 2 They are respectively represented as the eigenvalue matrix of the signal and the eigenvalue matrix of the noise, U S Λ S The characteristic vector of U N is σ 2 The features are sorted by eigenvalue size. The first M values, which are the same as the number of signals M, represent the eigenvalues ​​of the signal, and the other NM values ​​represent the eigenvalues ​​of the noise.

[0047] Steps 1-3 calculate the spectrum function and obtain the estimated value of the direction of arrival by finding the peak. The spatial spectrum estimation of the MUSIC algorithm is defined as:

[0048]

[0049] Where, P M (θ) is the spectral power in the θ direction, a(θ) is the direction vector in the θ direction, U N is the eigenvector of the noise eigenvalue matrix.

[0050] Furthermore, in one embodiment, in step 2, the speaker angle θ obtained in step 1 is used to obtain a weight vector W corresponding to the desired signal using an adaptive digital beamforming algorithm based on the LCMV criterion. LCMV , after weighting, the radar echo signal y1(t) in the direction of θ is obtained, which specifically includes:

[0051] Step 2-1, calculate the weight vector under the LCMV rule, expressed as:

[0052] W LCMV =R -1 C(CH R -1 C) -1 f

[0053] Where W LCMV is the weight vector based on the LCMV criterion, C=[a(θ1),a(θ2),…,a(θ N )] is the constraint matrix, f=(0,0,…,1,…,0) T is a response vector, R is the covariance matrix of the received signal;

[0054] Step 2-2, the weight vector W LCMV Multiplying the received signal X(t) with the radar echo signal y1(t) in the θ direction is as shown in the following formula:

[0055]

[0056] Furthermore, in one embodiment, step 3 performs a short-time Fourier transform in the range dimension on the radar echo signal y1(t) in the θ direction obtained in step 2 to obtain a signal y2(t) of the desired target. Specifically, the vocal cord vibration signal y2(t) of the desired target is obtained by selecting a range gate where the desired target is located. The process includes:

[0057] Step 3-1, perform distance dimension FFT on y1(t) in the fast time dimension and extract the peak value of the spectrum;

[0058] In step 3-2, after accumulating multiple chirps, the units at the same distance are reorganized in the slow time dimension, the Doppler frequency is obtained in the slow time dimension, and the vocal cord vibration signal y2(t) of the desired target is obtained.

[0059] Furthermore, in one embodiment, step 4 performs motion compensation on the desired target signal y2(t) obtained in step 3 to obtain a signal y3(t) that eliminates the effect of body motion, specifically including:

[0060] Step 4-1, curve fitting is performed on the desired target signal y2(t), i.e., the phase angle of the body motion signal;

[0061] Step 4-2: Perform phase compensation on the body motion signal using the fitted phase angle to obtain a compensated signal y3(t).

[0062] Furthermore, in one embodiment, step 5 performs high-pass filtering on the motion-compensated signal y3(t) obtained in step 4 (a high-pass filter is selected to remove noise and other interference, considering that the fundamental frequency is mostly above 50 Hz) to obtain the output signal y4(t), which specifically includes:

[0063] Step 5-1, construct a high-pass filter with a cutoff frequency of 50 Hz;

[0064] Step 5-2: Pass the body motion compensated signal y3(t) through the high-pass filter to obtain the output signal y4(t).

[0065] Furthermore, in one embodiment, in step 6, the signal y4(t) obtained in step 5 after the interference is removed is subjected to voice segment segmentation to obtain the voice segment signal y5(t). Specifically, a two-level decision method is used, and the process includes:

[0066] Step 6-1, frame the signal y4(t), extract the zero-crossing rate and short-time energy of each frame of data, and calculate the average zero-crossing rate and short-time average energy;

[0067] Step 6-2: Determine a threshold T1 for the short-time average energy and perform a rough judgment. The signal portion with short-time average energy exceeding the threshold T1 is determined as a speech segment. The points where the first and last short-time average energy exceeds the threshold T1 are set as endpoints, namely points A and B.

[0068] Step 6-3, determine a threshold T2 on the average zero-crossing rate, search from point A to the left and from point B to the right, find two points C and D whose short-time average zero-crossing rate is lower than T2, as the starting and ending points of the speech segment, thereby obtaining the voiced segment signal y5(t).

[0069] Furthermore, in one embodiment, step 7 extracts the radar signal's baseband signal y6(t) by using an improved VMD transform based on the voice segment signal y5(t) obtained in step 6, and specifically includes:

[0070] Step 7-1, perform fast FFT on the voiced signal y5(t) to obtain the Fourier spectrum of the signal;

[0071] Step 7-2, using the subdivided cubic spline interpolation method to extract the Fourier spectrum envelope of the voiced segment signal y5(t);

[0072] Step 7-3: Calculate the number of sampling points of the spectrum envelope and traverse from the second sampling point to the second-to-last sampling point. When a sampling point is greater than both the previous and next sampling points, determine that sampling point as a maximum point. This determines the number of maximum values ​​K of the spectrum envelope and sets it as the mode number of VMD.

[0073] Step 7-4, normalize the horizontal coordinate of the spectrum trend maximum to x m , and set it as the initial center frequency of VMD

[0074] Step 7-5, based on the initial center frequency calculated in steps 7-3 and 7-4 and the number of modes K, perform variational mode decomposition;

[0075] Step 7-6: Calculate the energy of different modes, determine the mode of the fundamental frequency signal based on the spectrum energy and permutation entropy, and extract the fundamental frequency signal.

[0076] Furthermore, in one embodiment, step 8 of extracting speech feature parameters based on the voiced segment signal y5(t) obtained in step 6 specifically includes:

[0077] Step 8-1, extract MFCC parameters:

[0078]

[0079] Where C mel (n) is the Mel cepstral coefficient, M is the number of MEL filters, N is the order of MFCC parameters, L is the m represents the logarithmic power spectrum of the mth MEL filter extracted;

[0080] L m =ln(P m )(0<m<M)

[0081] Where, P m is the power spectrum of the mth MEL filter;

[0082] Step 8-2, extract LPC parameters:

[0083]

[0084] E n (i) =(1-k i 2 )E n (i-1)

[0085] Where, is the jth prediction coefficient of the i-th order predictor in the system, E n ( i ) is the prediction residual energy obtained after the i-th order predictor, R n (i) is the autocorrelation coefficient,

[0086] Step 8-3, extract LPCC parameters:

[0087] c0=a1

[0088]

[0089] Where a n represents the LPC coefficient, c0 represents the DC component of the signal, c n represents the LPCC coefficient, and N' represents the LPCC order.

[0090] Example

[0091] As a specific example, the present invention is further verified and explained. Figure 1 The present invention provides a multi-speaker recognition method based on FMCW radar, comprising the following contents:

[0092] 1. In this embodiment, a two-transmitter, four-receiver FMCW radar is used, with a radar carrier frequency of 77 GHz, a bandwidth of 5000 MHz, and a sampling frequency of 4 kHz.

[0093] 2. Place a single-transmitter, multi-receiver FMCW radar in front of multiple human targets to be tested. Two experimenters sit quietly in front of the radar and use the DOA estimation based on the MUSIC algorithm to obtain the direction of arrival of the signal. The DOA results are as follows: Figure 2 As shown, they are -12° and 18° respectively.

[0094] 3. After obtaining the angle position through DOA, adaptive digital beam forming is performed to suppress interference signals so that the signal in the target direction is optimally output. The adaptive digital beam pattern is as follows: Figure 3 shown.

[0095] 4. After multi-target signal separation, select the range gate where the vibration source is located to obtain the vocal cord vibration signal of each object.

[0096] 5. Perform phase compensation on the signal to compensate for the influence of body movement. The effect before and after processing is as follows: Figure 4 As shown in (a)(b).

[0097] 6. Use a high-pass filter to filter out noise and other interference signals. The time domain signal after the signal passes through the filter is as follows: Figure 5 shown.

[0098] 7. Cut the voiced segment and use the two-level judgment method. Figure 6 shown.

[0099] 8. By improving VMD to extract the fundamental frequency, different modal signals such as Figure 7 shown.

[0100] 9. Combine different modal signals to obtain the fundamental frequency signal. The combined fundamental frequency signal time spectrum is as follows: Figure 8 shown.

[0101] 10. Perform Hilbert transform on the extracted baseband signal to get the envelope, such as Figure 9 shown.

[0102] 11. Extract MFCC, LPCC parameters and base frequency time domain features and frequency domain features by extracting radar echo signals, and send them into GRU network for speaker recognition. The recognition effect is as follows: Figure 10 shown.

[0103] It can be seen from the above that the method of the present invention is effective and feasible, has reliable performance, and can accurately separate multi-person speech and extract fundamental frequency.

[0104] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-speaker recognition method based on FMCW radar, characterized in that: The method comprises the following steps: Step 1: Receive the radar echo signal, use the DOA algorithm to locate the angle of the speaker, and use the MUSIC algorithm to obtain the angle θ of the speaker; Step 2: Based on the speaker angle θ obtained in step 1, the weight vector W corresponding to the desired signal is obtained using the adaptive digital beamforming algorithm based on the LCMV criterion. LCMV , after weighting, the radar echo signal y1(t) in the direction of θ is obtained; Step 3: Perform a short-time Fourier transform of the distance dimension on the radar echo signal y1(t) in the θ direction obtained in step 2 to obtain the signal y2(t) of the desired target; Step 4: Perform motion compensation on the desired target signal y2(t) obtained in step 3 to obtain a signal y3(t) that eliminates the influence of body motion. Step 5: High-pass filter the motion-compensated signal y3(t) obtained in step 4 to obtain an output signal y4(t); Step 6: Perform voice segment cutting on the signal y4(t) obtained in step 5 after removing interference, to obtain the voice segment signal y5(t); Step 7: Based on the voiced segment signal y5(t) obtained in step 6, the baseband signal y6(t) of the radar signal is extracted by using an improved VMD transform; Step 8: extracting speech feature parameters based on the voiced segment signal y5(t) obtained in step 6; Step 9: Send the fundamental frequency signal y6(t) obtained in step 7 and the speech feature parameters obtained in step 8 to the GRU neural network to realize speaker recognition; Step 7 extracts the radar signal's baseband signal y6(t) based on the voiced segment signal y5(t) obtained in step 6 by using an improved VMD transform, specifically including: Step 7-1, perform fast FFT on the voiced signal y5(t) to obtain the Fourier spectrum of the signal; Step 7-2, using the subdivided cubic spline interpolation method to extract the Fourier spectrum envelope of the voiced segment signal y5(t); Step 7-3: Calculate the number of sampling points of the spectrum envelope and traverse from the second sampling point to the second-to-last sampling point. When a sampling point is greater than both the previous and next sampling points, determine that sampling point as a maximum point. This determines the number of maximum values ​​K of the spectrum envelope and sets it as the mode number of VMD. Step 7-4, normalize the horizontal coordinate of the spectrum trend maximum to x m , and set it as the initial center frequency of VMD Step 7-5, based on the initial center frequency calculated in steps 7-3 and 7-4 and the number of modes K, perform variational mode decomposition; Step 7-6: Calculate the energy of different modes, determine the mode of the fundamental frequency signal based on the spectrum energy and permutation entropy, and extract the fundamental frequency signal.

2. The multi-speaker recognition method based on FMCW radar according to claim 1, characterized in that: The method described in step 1 uses the MUSIC algorithm to obtain the angle θ of the speaker. Specifically, the method uses the MUSIC algorithm to estimate the direction of arrival and obtain the angle θ of the speaker. The process includes: Step 1-1, obtain the covariance matrix based on N received signal vectors, which is: R=E[XX H ] Where R is the covariance matrix and X is the received signal matrix; Step 1-2, perform eigenvalue decomposition on the covariance matrix, expressed as: Among them, Λ S and σ 2 They are respectively represented as the eigenvalue matrix of the signal and the eigenvalue matrix of the noise, U S Λ S The characteristic vector of U N is σ 2 The features are sorted by eigenvalue size. The first M values, which are the same as the number of signals M, represent the eigenvalues ​​of the signal, and the other NM values ​​represent the eigenvalues ​​of the noise. Steps 1-3 calculate the spectrum function and obtain the estimated value of the direction of arrival by finding the peak. The spatial spectrum estimation of the MUSIC algorithm is defined as: Where, P M (θ) is the spectral power in the θ direction, a(θ) is the direction vector in the θ direction, U N is the eigenvector of the noise eigenvalue matrix.

3. The multi-speaker recognition method based on FMCW radar according to claim 2, characterized in that: In step 2, the speaker angle θ obtained in step 1 is used to obtain the weight vector W corresponding to the desired signal using the adaptive digital beamforming algorithm based on the LCMV criterion. LCMV , after weighting, the radar echo signal y1(t) in the direction of θ is obtained, which specifically includes: Step 2-1, calculate the weight vector under the LCMV rule, expressed as: W LCMV =R -1 C(C H R -1 C) -1 f Where W LCMV is the weight vector based on the LCMV criterion, C=[a(θ1),a(θ2),…,a(θ N )] is the constraint matrix, f=(0,0,…,1,…,0) T is a response vector, R is the covariance matrix of the received signal; Step 2-2, the weight vector W LCMV Multiplying the received signal X(t) with the radar echo signal y1(t) in the θ direction is as shown in the following formula:

4. The multi-speaker recognition method based on FMCW radar according to claim 3, characterized in that: Step 3 performs a short-time Fourier transform of the distance dimension on the radar echo signal y1(t) in the θ direction obtained in step 2 to obtain the signal y2(t) of the desired target. Specifically, the vocal cord vibration signal y2(t) of the desired target is obtained by selecting the range gate where the desired target is located. The process includes: Step 3-1, perform distance dimension FFT on y1(t) in the fast time dimension and extract the peak value of the spectrum; In step 3-2, after accumulating multiple chirps, the units at the same distance are reorganized in the slow time dimension, the Doppler frequency is obtained in the slow time dimension, and the vocal cord vibration signal y2(t) of the desired target is obtained.

5. The multi-speaker recognition method based on FMCW radar according to claim 4, characterized in that: Step 4 performs motion compensation on the desired target signal y2(t) obtained in step 3 to obtain a signal y3(t) that eliminates the influence of body motion, specifically including: Step 4-1, curve fitting is performed on the desired target signal y2(t), i.e., the phase angle of the body motion signal; Step 4-2: Perform phase compensation on the body motion signal using the fitted phase angle to obtain a compensated signal y3(t).

6. The multi-speaker recognition method based on FMCW radar according to claim 5, characterized in that: Step 5 performs high-pass filtering on the motion-compensated signal y3(t) obtained in step 4 to obtain an output signal y4(t), which specifically includes: Step 5-1, construct a high-pass filter with a cutoff frequency of 50 Hz; Step 5-2: Pass the body motion compensated signal y3(t) through the high-pass filter to obtain the output signal y4(t).

7. The multi-speaker recognition method based on FMCW radar according to claim 6, characterized in that: In step 6, the signal y4(t) obtained in step 5 after interference removal is subjected to voice segment segmentation to obtain the voice segment signal y5(t). Specifically, a two-level decision method is used, and the process includes: Step 6-1, frame the signal y4(t), extract the zero-crossing rate and short-time energy of each frame of data, and calculate the average zero-crossing rate and short-time average energy; Step 6-2: Determine a threshold T1 for the short-time average energy and perform a rough judgment. The signal portion with short-time average energy exceeding the threshold T1 is determined as a speech segment. The points where the first and last short-time average energy exceeds the threshold T1 are set as endpoints, namely points A and B. Step 6-3, determine a threshold T2 on the average zero-crossing rate, search from point A to the left and from point B to the right, find two points C and D whose short-time average zero-crossing rate is lower than T2, as the starting and ending points of the speech segment, thereby obtaining the voiced segment signal y5(t).

8. The multi-speaker recognition method based on FMCW radar according to claim 7, characterized in that: Step 8 extracts speech feature parameters based on the voiced segment signal y5(t) obtained in step 6, specifically including: Step 8-1, extract MFCC parameters: Where C mel (n) is the Mel cepstral coefficient, M is the number of MEL filters, N is the order of MFCC parameters, L is the m represents the logarithmic power spectrum of the mth MEL filter extracted; L m =ln(P m )(0<m<M) Where, P m is the power spectrum of the mth MEL filter; Step 8-2, extract LPC parameters: E n (i) =(1-k i 2 )E n (i-1) Where, is the jth prediction coefficient of the i-th order predictor in the system, E n (i) is the prediction residual energy obtained after the i-th order predictor, R n (i) is the autocorrelation coefficient, Step 8-3, extract LPCC parameters: c0=a1 Where a n represents the LPC coefficient, c0 represents the DC component of the signal, c n represents the LPCC coefficient, and N' represents the LPCC order.

Images