Optical fiber microphone single-channel speech enhancement method fusing OMLSA and TQWT

Abstract

An optical fiber microphone single-channel speech enhancement method fusing OMLSA and TQWT comprises the steps that framing processing is firstly carried out on collected speech signals, OMLSA speech enhancement is carried out on each frame, and whether TQWT residual noise removal needs to be carried out or not is determined by using a segmented signal-to-noise ratio as a judgment condition. According to the method, the optimal values of a parameter Q factor and redundancy r required by TQWT decomposition are selected by using a reconstruction error, it is ensured that a group of most suitable wavelet coefficients are obtained through speech signal decomposition, a threshold value is adaptively selected according to each wavelet coefficient, the wavelet coefficients are sparsified by using basis tracking, and a finally enhanced speech signal is obtained through reconstruction of the sparsified wavelet coefficients. Compared with de-noising only by using OMLSA, the method of the invention can further effectively suppress residual noise on this basis, and improves the speech enhancement effect.

Description

Technical Field

[0001] The present invention relates to the technical field of microphone speech signal processing, and in particular to a single-channel speech enhancement method for an optical fiber microphone integrating OMLSA and TQWT. Background Art

[0002] Fiber optic microphones are acoustic sensors that use optical fiber as a sensing medium. They sense the phase change of the optical fiber caused by sound waves and then restore the sound wave information. They are passive, resistant to electromagnetic interference, and have a wide dynamic range. Their high sensitivity can achieve high-fidelity detection of sound waves and have been widely used in the fields of sound recognition and sound source localization. However, due to the presence of interference such as environmental noise and system background noise, the clarity and recognizability of speech signals are inevitably affected. Therefore, speech enhancement processing of speech signals collected by fiber optic microphones is of great significance for further improving speech quality.

[0003] Depending on whether the sound signal is collected by a single microphone or multiple microphones, it can correspond to a single-channel speech enhancement algorithm and a multi-channel speech enhancement algorithm. Compared with multiple microphones, a single microphone has attracted much attention due to its small size, low cost, and low power consumption. The single-channel speech enhancement algorithm mainly uses the characteristics of speech itself for speech enhancement. By extracting the difference in the characteristics of speech and noise signals in the time domain and transform domain, the noise component is suppressed to achieve the purpose of speech enhancement. Speech enhancement algorithms can be divided into traditional speech enhancement algorithms and deep learning-based speech enhancement algorithms. Deep learning-based speech enhancement requires a large amount of data as a training set to train the model, which takes a long time and cannot be deployed immediately. In addition, the performance depends on the distribution of the training set to a certain extent. Traditional single-channel speech enhancement includes a variety of methods, among which the minimum mean square error class can further reduce residual noise without losing speech quality, such as the optimally modified log-spectral amplitude estimation (OMLSA) based on the probability of speech existence, which achieves noise suppression by optimizing the speech amplitude spectrum. In addition, the tunable-Q wavelet transform (TQWT) decomposes the signal at multiple scales and removes interference noise by sparsely processing the decomposed wavelet coefficients, thereby achieving speech enhancement.

[0004] Defects and shortcomings of existing technology:

[0005] (1) Although some existing single-channel speech enhancement algorithms have excellent noise suppression effects, their implementation requires a clean speech signal as a reference, such as the Wiener filter and the adaptive filter, which is not advisable in practical applications. Therefore, it is necessary to implement a high-performance algorithm that can achieve denoising based only on the noisy speech itself.

[0006] (2) Due to the diversity of application scenarios of fiber optic microphones, speech signals may have low signal-to-noise ratio and non-stationary noise. The noise estimation error will significantly affect the speech enhancement performance of OMLSA, resulting in a large amount of residual noise in the signal after OMLSA processing, which seriously affects the quality of the speech signal. The algorithm needs to be improved to be suitable for low signal-to-noise ratio situations.

[0007] (3) The performance of TQWT depends on the selection of parameters, such as the Q factor, oversampling rate and number of decomposition layers, as well as the threshold for sparse wavelet coefficients. Improper parameter settings may result in the decomposed subband signal failing to accurately capture the main features of speech, thus affecting the enhancement effect. In speech signal processing, low-frequency components are usually important speech information, while the wavelet coefficients decomposed by TQWT are mainly composed of high-frequency components, which may result in the inability to completely suppress low-frequency noise. Summary of the invention

[0008] To solve the above technical problems, the present invention provides a single-channel speech enhancement method for a fiber optic microphone integrating OMLSA and TQWT. Compared with traditional electronic microphones, the present invention uses a high-sensitivity fiber optic microphone to collect signals and provides high-quality original data for subsequent speech enhancement processing. Moreover, the present invention only enhances the signal collected by the fiber optic microphone itself without providing additional clean speech signals, thereby achieving good speech enhancement effects and further improving language quality. The method can be widely used in practice.

[0009] The technical solution adopted by the present invention is:

[0010] A single-channel speech enhancement method for a fiber optic microphone integrating OMLSA and TQWT comprises the following steps:

[0011] Step 1: Assume that the speech signal detected by the optical fiber microphone is y(n), where n is the sampling point index, 0≤n≤N-1, and N is the total number of sampling points, which is guaranteed to be an even number;

[0012] Step 2: Select the Hanning window to frame the speech signal y(n) to obtain the time domain signal y(n,l), where l is the frame index, 0≤l≤L-1, and L is the total number of frames;

[0013] Step 3: Use discrete Fourier transform to convert the time domain signal y(n,l) into the frequency domain signal Y(k,l), where k is the frequency index, k = 0, 1, 2, ..., N-1;

[0014] Step 4: Enhance the speech based on the OMLSA algorithm and obtain the enhanced spectrum amplitude Calculated by the following formula:

[0015]

[0016] Where: G min is the gain function when speech does not exist; p(k,l) is the posterior probability of speech existence in the lth frame at frequency k; G H1 (k,l) is the gain function when the lth frame of speech exists at frequency k; (G min |Y(k,l)|) (1-p(k,l)) Indicates that when the l-th frame of speech does not exist at frequency k, a gain G is applied to |Y(k,l)| min To suppress noise, (G H1 (k,l)·|Y(k,l)|) p(k,l) Indicates that when the l-th frame of speech exists at frequency k, a gain G is applied to |Y(k,l)| H1 (k, l) emphasizes the speech component, and the symbol |·| indicates taking the absolute value.

[0017] Step 5: Perform inverse Fourier transform on the enhanced signal spectrum amplitude of all frames, and restore the speech signal x1(n) enhanced by the OMLSA algorithm according to the Hanning window function;

[0018] Step 6: Determine whether the segmented signal-to-noise ratio segSNR of the speech signal x1(n) satisfies the following conditions:

[0019] segSNR ≥ 5dB;

[0020] If satisfied, skip step 7, step 8, and step 9. At this time, x1(n) is the final enhanced speech signal x(n). Otherwise, proceed to step 7.

[0021] Step 7: Perform TQWT decomposition on x1(n) to obtain J+1 wavelet coefficients w{j}, where J is the optimal total decomposition level, j is the level index, 1≤j≤J+1;

[0022] Step 8: Use basis pursuit to sparse the wavelet coefficients w{j} to obtain the sparse wavelet coefficients w1{j};

[0023] Step 9: Perform inverse TQWT on the sparse wavelet coefficients w1{j} to obtain the final enhanced speech signal x(n).

[0024] In the step 2, Indicates rounding up, N f is the Hanning window frame length, N m For frame shift.

[0025] The step 4 comprises the following steps:

[0026] Step 4.1: Initialize the following parameters:

[0027] The initial value of the gain function G when speech exists H1 (k,0)=1, the initial value of the posterior signal-to-noise ratio of the speech γ(k,0)=1, the initial noise power spectrum λ d (k,0)=|Y(k,0)| 2 , initial noise power spectrum before offset compensation

[0028] Where Y(k,0) is the spectrum information of the initial frame, k represents the frequency index; the initial value of the first power spectrum time smoothing S(k,0) = S f (k,0), the initial value S of the minimum value of the first power spectrum time smoothing min (k,0)=S f (k,0), the initial value of the second power spectrum time smoothing Initial value of the second power spectrum time smoothing minimum S f (k,0) is the initial value of the speech frequency smoothed power spectrum; l is initialized to 1.

[0029] Step 4.2: Calculate the noise power spectrum of the lth frame before offset compensation at frequency k The calculation formula is:

[0030]

[0031] in: is the noise power spectrum of the l-1th frame before offset compensation at frequency k, Y(k,l-1) is the frequency domain signal of the l-1th frame at frequency k, is the time-varying smoothing parameter of the l-1th frame at frequency k, and its calculation formula is:

[0032]

[0033] Where: α d Represents the smoothing parameter before bias compensation, 0<α d <1; p(k,l-1) is the posterior probability of the presence of speech in the l-1th frame at frequency k;

[0034] Step 4.3: Calculate the noise power spectrum λ of the lth frame at frequency k d The calculation formula for (k,l) is as follows:

[0035]

[0036] Where: β is the bias factor, and the calculation formula is:

[0037]

[0038] Among them: γ1 is the regulatory factor; It represents the base of natural logarithm, e, raised to the power of -γ1;

[0039] Step 4.4: Calculate the posterior signal-to-noise ratio γ(k,l) of the lth frame of speech at frequency k. The calculation formula is:

[0040]

[0041] Step 4.5: Calculate the prior signal-to-noise ratio ξ(k,l) of the lth frame of speech at frequency k, and the calculation formula is:

[0042]

[0043] Where: α is the weight factor for controlling the trade-off between noise reduction and speech distortion; G H1 (k,l-1) is the gain function when the l-1th frame of speech exists at frequency k; γ(k,l-1) is the posterior signal-to-noise ratio of the l-1th frame of speech at frequency k; max{γ(k,l)-1,0} means taking the maximum value between γ(k,l)-1 and 0;

[0044] Step 4.6: Calculate the gain function G when the lth frame of speech exists at frequency k H1 (k,l), the calculation formula is:

[0045]

[0046] in: represents the balance value of the priori signal-to-noise ratio and the a posteriori signal-to-noise ratio of the lth frame of speech at frequency k;

[0047] Step 4.7: Calculate the first power spectrum time smoothing value S(k,l) of the lth frame of speech at frequency k, and the calculation formula is:

[0048] S(k,l)=α s ·S(k,l-1)+(1-α s )·S f (k,l);

[0049] Where: α s is the time smoothing parameter of the first power spectrum, S(k,l-1) is the time smoothing value of the first power spectrum of the l-1 frame speech at frequency k, S f (k, l) is the first power spectrum frequency smoothing value of the l-th frame speech at frequency k, and its calculation formula is:

[0050]

[0051] Where: b(i) is the normalized window function, the window length is 2w+1, w represents the half length of the window function, and i is the window function index, that is, Y(ki,l) is the frequency domain signal of the lth frame at frequency ki;

[0052] Step 4.8: Calculate the second power spectrum frequency smoothing of the l-th frame signal at frequency k and the second power spectrum time smoothing The calculation formulas are:

[0053]

[0054] in: is the second power spectrum time smoothing of the l-1 frame signal at frequency k, and the value of I(ki,l) is given by the following formula:

[0055]

[0056] Among them: γ0 and ζ0 are two thresholds for speech presence judgment, and “&&” represents the logical operation “and”;

[0057]

[0058] γ min (ki,l) represents the minimum signal-to-noise ratio of the l-th frame of speech at frequency ki, ζ(ki,l) represents the short-term stability index of the l-th frame of speech at frequency ki;

[0059] S(ki,l) is the first power spectrum time smoothing of the lth frame of speech at frequency ki, B min is the bias compensation factor for estimating the minimum value of the noise spectrum, and the minimum value S of the first power spectrum time smoothing of the l-th frame speech at frequency ki min (ki,l)=min{S min (ki,l-1),S(ki,l)},S min (ki,l-1) is the minimum value of the first power spectrum time smoothing of the l-1 frame of speech at frequency ki;

[0060] Step 4.9: Calculate the prior probability q(k,l) that the lth frame of speech does not exist at frequency k. The calculation formula is:

[0061]

[0062] in: S min (k,l-1) is the minimum value of the first power spectrum time smoothing of the l-1 frame of speech at frequency k; γ1 represents the adjustment factor;

[0063] Step 4.10: Calculate p(k,l), the calculation formula is:

[0064]

[0065] Step 4.11: Change the current G H1 Substitute (k, l) and p(k, l) into the calculation formula of step 4, as follows:

[0066]

[0067] Output frequency k at the enhanced spectrum amplitude of the lth frame

[0068] Step 4.12: Determine whether the following loop conditions are met:

[0069] l <L;

[0070] If satisfied, increase the value of l by 1 and repeat steps 4.2 to 4.11; otherwise proceed to step 5.

[0071] In step 5, the enhanced signal spectrum amplitudes of all frames are subjected to inverse Fourier transform, and the calculation formula is:

[0072]

[0073] in: is the time domain signal of the lth frame of speech, n1 is the sampling point index of the lth frame of time domain signal, 0≤n1≤N f -1, e is the base of natural logarithm, i1 is the imaginary unit;

[0074] And according to the Hanning window function, the speech signal x1(n) enhanced by the OMLSA algorithm is restored. The specific steps are:

[0075] right Apply the Hanning window to get the time domain signal a(n1,l) after windowing the l-th frame of speech:

[0076]

[0077] Where: hanning(n1) is the Hanning window formula, cos is the cosine trigonometric function; x1(n) is obtained by overlapping and adding a(n1,l) in frames, and the calculation formula is:

[0078]

[0079] Among them: hanning_overlap(n1) Hanning window overlap frame weight, its calculation formula is:

[0080]

[0081] In step 6, calculating segSNR includes the following steps:

[0082] Step 6.1: Divide x1(n) into M segments, where the value of M is:

[0083]

[0084] in: Indicates rounding down, S m is the size of each segment, usually 256, 512 or 1024, etc., m is the index of the segment, 1≤m≤M.

[0085] Step 6.2: For the t-th sampling point in the m-th segment of the residual noise in x1(n), 1_m The value of (t) is estimated by:

[0086]

[0087] Where: x 1_m (t) is the signal of the tth sampling point in the mth segment of x1(n), and th is the energy threshold of the speech silence segment.

[0088] Step 6.3: Calculate segSNR:

[0089]

[0090] In step 7, decomposing x1(n)TQWT to obtain J+1 wavelet coefficients w{j} includes the following steps:

[0091] Step 7.1: Use the reconstruction error E to find the optimal value of the quality factor Q opt and the optimal value r of redundancy r opt , the specific steps are:

[0092] S7.1.1: Set the maximum value that Q can take to be Q max , the maximum value that r can take is r max , initialize Q = 1, r is [3,r max ] any value in the interval;

[0093] S7.1.2: Calculate the number of decomposition levels J1 from Q and r using the formula:

[0094]

[0095] The Q value at this time is stored in the sequence Q_cell, and steps 7.2 to 7.8 are performed using the current three parameters Q, r, and J1, and step 8 is skipped. At this time, w{j} is w1{j} in step 9, and step 9 is performed.

[0096] S7.1.3: Calculate the reconstruction error E = max(|x1(n) - x(n)|) and store it in the sequence E_cell.

[0097] S7.1.4: Determine whether the following loop conditions are met:

[0098] Q max ;

[0099] If satisfied, add 0.2 to the value of Q and repeat S7.1.2 to S7.1.3, otherwise proceed to S7.1.5;

[0100] S7.1.5: The minimum reconstruction error min(E_cell) corresponds to the value in Q_cell, which is Q opt .

[0101] S7.1.6: Initialize r = 3, Q to [1, Q max ] any value in the interval;

[0102] S7.1.7: Calculate the decomposition level J1, store the r value at this time into the sequence r_cell, use the current parameters to perform steps 7.2 to 7.8, skip step 8, and w{j} is now w1{j} in step 9, and proceed to step 9.

[0103] S7.1.8: Calculate the reconstruction error E = max(|x1(n) - x(n)|) and store it in the sequence E_cell_1.

[0104] S7.1.9: Determine whether the following loop conditions are met:

[0105] r <r max ;

[0106] If satisfied, add 0.2 to the value of r and repeat S7.1.7 to S7.1.8, otherwise proceed to S7.1.10;

[0107] S7.1.10: The minimum reconstruction error min(E_cell_1) corresponds to the value in r_cell, which is r opt .

[0108] S7.1.11: Use Q opt and r opt Find the optimal decomposition level J.

[0109] Step 7.2: Perform unitary discrete Fourier transform uDFT(x1(n)) on x1(n) to obtain X1(k), whose formula is:

[0110]

[0111] Where: X_1(k) is the discrete Fourier transform of x1(n), and j=1 is initialized.

[0112] ​Step 7.3: Calculate the length of the low-pass subband of the jth layer and the high-pass subband length

[0113]

[0114] Among them: "round" means rounding operation,

[0115] Step 7.4: Calculate the passband length P of the j-th low-pass layer j , high-pass passband length S j , the length of the transition zone T j :

[0116]

[0117] Step 7.5: Obtain the j-th layer low-pass sub-band sequence and high-pass sub-band sequence. The specific steps are as follows:

[0118] S7.5.1: Get the j-th layer low-pass subband sequence k_low is the low-pass subband frequency index,

[0119]

[0120] Where: represents the initial value of the j-th layer low-pass subband sequence, X1(0) represents the initial value of X1(k) obtained by performing unitary discrete Fourier transform on x1(n);

[0121]

[0122] Where: Represents the j-th layer low-pass subband sequence from frequency 1 to frequency P j The value of X1(1:P j ) indicates that X1(k) goes from frequency 1 to frequency P j The value of

[0123]

[0124] Where: Represents the j-th layer low-pass subband sequence from frequency P j +1 to frequency P j +T j The value of

[0125] X1(P j +(1:T j )) indicates that X1(k) is from frequency P j +1 to frequency P j +Tj The value of θ(1:T j ) means to take values ​​from 1 to T in sequence j Perform the following theta operation:

[0126]

[0127] Where: D represents the parameter value for θ operation;

[0128]

[0129] Where: Indicates that the j-th layer low-pass subband sequence is at frequency The value of ;

[0130]

[0131] Where: Represents the j-th layer low-pass subband sequence from frequency To frequency The value of X1(NP j -(1:T j )) indicates that X1(k) is from frequency NP j -1 to frequency NP j -T j The value of

[0132]

[0133] Where: A1:B1 means the sequence is indexed from A1 to B1, and “.*” means multiplying the corresponding elements of the sequence;

[0134] S7.5.2: Get the j-th layer high-pass subband sequence V1 j (k_high), k_high is the high-pass subband frequency index,

[0135] V1 j (0) = 0;

[0136] Where: V1 j (0) represents the initial value of the j-th layer high-pass subband sequence;

[0137] V1 j (1:T j )=X1(P j +(1:T j )).*θ(T j -(1:T j )+1);

[0138] Where: V1 j (1:Tj ) represents the j-th layer high-pass subband sequence from frequency 1 to frequency T j The value of X1(P j +(1:T j )) indicates that X1(k) is from frequency P j +1 to frequency P j +T j The value of θ(T j -(1:T j )+1) means taking the value T in turn j To 1, perform theta operation;

[0139] V1 j (T j +(1:S j ))=X1(P j +T j +(1:S j ));

[0140] Where: V1 j (T j +(1:S j )) represents the j-th layer high-pass subband sequence from frequency T j +1 to frequency T j +S j The value of

[0141] X1(P j +T j +(1:S j )) indicates that X1(k) is from frequency P j +T j +1 to frequency P j +T j +S j The value of

[0142]

[0143] Where: Indicates that the j-th layer high-pass subband sequence is at frequency X1(N / 2) represents the value of X1(k) at k=N / 2;

[0144]

[0145] Where: Indicates that the j-th layer high-pass subband sequence is from frequency To frequency The value of X1(NP j -T j -(1:S j )) indicates that X1(k) is from frequency NP j-T j -1 to frequency NP j -T j -S j The value of

[0146]

[0147] Where: Indicates that the j-th layer high-pass subband sequence is from frequency To frequency The value of

[0148] X1(NP j -(1:T j )) indicates that X1(k) is from frequency NP j -1 to frequency NP j -T j The value of

[0149] Step 7.6: V1 j (k_high) performs inverse unitary discrete Fourier transform to obtain the jth wavelet coefficient w{j} of TQWT decomposition:

[0150] w{j}=uDFTinv(V1 j (k_high));

[0151] Where: uDFTinv represents the unitary inverse discrete Fourier transform, and its relationship with the inverse discrete Fourier transform is:

[0152]

[0153] Where: w_inv{j} is V1 j Inverse discrete Fourier transform of (k_high);

[0154] Step 7.7: Determine whether the following loop conditions are met:

[0155] j <J;

[0156] If satisfied, the value of j is increased by 1 and steps 7.3 to 7.6 are repeated; otherwise, proceed to step 7.8.

[0157] Step 7.8: Calculate the J+1th wavelet coefficient of the TQWT decomposition Represents the J-th layer low-pass subband sequence; so far, all the J+1 wavelet coefficients decomposed have been calculated.

[0158] In step 8, the basis pursuit is used to sparse w{j}, including the following steps:

[0159] Step 8.1: Initialize j=1;

[0160] Step 8.2: Generate an l j × j The identity matrix A j , l j is the length of the jth wavelet coefficient; vector b j =A j T *w{j} represents b j Projection onto the identity matrix, where A j T A j The transpose of ; matrix B j =[A j T *A j ,-A j T *A j ;-A j T *A j ,A j T *A j ] is used to construct quadratic terms in optimization problems;

[0161] Step 8.3: Calculate the regularization parameter τ of the jth layer j , and its calculation formula is:

[0162]

[0163] in: is a regulation factor that reflects the change of sparsity with the number of layers j. is the coefficient for adjusting the amplitude of w{j}, ln() is the natural logarithm, and median() is the median operation;

[0164] Step 8.4: Construct a j The column vector c j , as the driving vector for sparse optimization, is generated by:

[0165]

[0166] Where: C j is of length 2·l j And a column vector of all 1s;

[0167] Step 8.5: Solve the quadratic optimization problem:

[0168]

[0169] Among them: z is the intermediate quantity when solving, and its length is 2·l jColumn vector of z T represents the transpose of z; c j T Indicates c j The transpose of represents the optimal solution of z when minimizing the objective function, and the optimal solution is z j .

[0170] Step 8.6: By z j The sparse wavelet coefficient w1{j} is obtained, and its calculation formula is:

[0171] w1{j}=z j (1:l j )-z j (l j +1:2·l j );

[0172] Where: z j (1:l j ) represents the column vector z j Line 1 to line 1 j The value of the row, z j (l j +1:2·l j ) represents the column vector z j No.1 j +1 row to 2.1 j The value of the row;

[0173] Step 8.7: Determine whether the following loop conditions are met:

[0174] j≤J;

[0175] If satisfied, the value of j is increased by 1 and steps 8.2 to 8.6 are repeated; otherwise, proceed to step 9.

[0176] In step 9, performing inverse TQWT on w1{j} includes the following steps:

[0177] Step 9.1: Find the unitary discrete Fourier transform of the J+1th sparse wavelet coefficient w1{J+1} as Y J+1 (k_2), where k_2 is its frequency index, Represents Y J+1 The length of (k_2); initialize j=J.

[0178] Step 9.2: Calculate the passband length P of the j-th low-pass layer after sparseness j _1. High-pass passband length S j _1. Length of transition zone T j _1 are:

[0179]

[0180]

[0181] Where: N j _1 is the length of the unitary discrete Fourier transform of the jth sparse wavelet coefficient w1{j}, N j+1 _1 is the length of the unitary discrete Fourier transform of the j+1th sparse wavelet coefficient w1{j+1}.

[0182] Step 9.3: Calculate the j-th layer reconstructed signal Y j (k1), 0≤k1≤M j -1, k1 represents the frequency index of the j-th layer reconstructed signal; comprising the following steps:

[0183] S9.3.1: Perform unitary discrete Fourier transform on the jth sparse wavelet coefficient w1{j} to W j (k_1), k_1 is its frequency index, 0≤k_1≤N j _1-1;

[0184] S9.3.2: Signal Y reconstructed from layer j+1 j+1 (k1) Find the j-th layer sparse low-pass subband sequence

[0185]

[0186] Where: represents the initial value of the j-th layer sparse low-pass subband sequence, Y j+1 (0) represents the j+1th layer reconstructed signal Y j+1 The initial value of (k1);

[0187]

[0188] Where: Represents the j-th layer of sparse low-pass subband sequence from frequency 1 to frequency P j The value of _1, Y j+1 (1:P j _1) indicates Y j+1 (k1) from frequency 1 to frequency P j The value of _1;

[0189]

[0190] Where: Represents the j-th layer of sparse low-pass subband sequence from frequency P j _1+1 to frequency P j _1+T j _1 value; Y j+1 (Pj _1+(1:T j _1)) indicates Y j+1 (k1) from frequency P j _1+1 to frequency P j _1+T j _1 value;θ(1:T j _1) means to take values ​​from 1 to T in sequence j _1 performs theta operation;

[0191]

[0192] Where: Represents the j-th layer of sparse low-pass subband sequence from frequency P j _1+T j _1+1 to frequency P j _1+T j _1+S j The value of _1;

[0193]

[0194] Where: Represents the j-th layer of sparse low-pass subband sequence at frequency M j The value of ;

[0195]

[0196] Where: Denotes the j-th layer of sparse low-pass subband sequence from frequency M j -P j _1-T j _1-1 to frequency M j -P j _1-T j _1-S j The value of _1;

[0197]

[0198] Where: Represents the j-th layer of sparse low-pass subband sequence from frequency M j -P j _1-1 to frequency M j -P j _1-T j _1 value; Y j+1 (N j+1 _1-P j _1-(1:T j _1)) indicates Y j+1 (k1) from frequency N j+1 _1-Pj _1-1 to frequency N j+1 _1-P j _1-T j The value of _1,

[0199]

[0200] Where: Represents the j-th layer of sparse low-pass subband sequence from frequency M j -1 to frequency M j -P j _1 value; Y j+1 (N j+1 _1-(1:P j _1)) indicates Y j+1 (k1) from frequency N j+1 _1-1 to frequency N j+1 _1-P j The value of _1;

[0201] S9.3.3: Find the j-th layer sparse high-pass subband sequence Y1 j (k1):

[0202] Y1 j (0) = 0;

[0203] Where: Y1 j (0) represents the initial value of the j-th layer sparse high-pass subband sequence;

[0204] Y1 j (1:P j _1) = 0;

[0205] Where: Y1 j (1:P j _1) represents the j-th layer sparse high-pass subband sequence from frequency 1 to frequency P j The value of _1;

[0206] Y1 j (P j _1+(1:T j _1))=W j (1:T j _1).*θ(T j _1-(1:T j _1)+1);

[0207] Where: Y1 j (P j _1+(1:T j _1)) represents the j-th layer sparse high-pass subband sequence from frequency P j _1+1 to frequency P j_1+T j _1 value; W j (1:T j _1) represents the unitary discrete Fourier transform of w1{j} to W j (k_1) from frequency 1 to frequency T j _1 value;θ(T j _1-(1:T j _1)+1) means taking the value T in turn j _1 to 1 performs theta operation;

[0208] Y1 j (P j _1+T j _1+(1:S j _1))=W j (T j _1+(1:S j _1));

[0209] Where: Y1 j (P j _1+T j _1+(1:S j _1)) represents the j-th layer sparse high-pass subband sequence from frequency P j _1+T j _1+1 to frequency P j _1+T j _1+S j _1 value; W j (T j _1+(1:S j _1)) indicates W j (k_1) from frequency T j _1+1 to frequency T j _1+S j The value of _1;

[0210] Y1 j (M j / 2)=W j (N j _1 / 2);

[0211] Where: Y1 j (M j / 2) represents the j-th layer sparse high-pass subband sequence at frequency M j / 2; W j (N j _1 / 2) indicates W j (k_1) when k_1=N j _The value at 1 / 2;

[0212] Y1j (M j -P j _1-T j _1-(1:S j _1))=W j (N j _1-T j _1-(1:S j _1));

[0213] Where: Y1 j (M j -P j _1-T j _1-(1:S j _1)) represents the j-th layer sparse high-pass subband sequence from frequency M j -P j _1-T j _1-1 to frequency M j -P j _1-T j _1-S j _1 value; W j (N j _1-T j _1-(1:S j _1)) indicates W j (k_1) from frequency N j _1-T j _1-1 to frequency N j _1-T j _1-S j The value of _1;

[0214] Y1 j (M j -P j _1-(1:T j _1))=W j (N j _1-(1:T j _1)).*θ(T j _1-(1:T j _1)+1);

[0215] Where: Y1 j (M j -P j _1-(1:T j _1)) represents the j-th layer sparse high-pass subband sequence from frequency M j -P j _1-1 to frequency M j -P j _1-T j _1 value; Wj (N j _1-(1:T j _1)) indicates W j (k_1) from frequency N j _1-1 to frequency N j _1-T j The value of _1;

[0216] Y1 j (M j -(1:P j _1))=0;

[0217] Where: Y1 j (M j -(1:P j _1)) represents the j-th layer sparse high-pass subband sequence from frequency M j -1 to frequency M j -P j The value of _1;

[0218] S9.3.4: At this time, the j-th layer reconstructs the signal

[0219] Step 9.4: Determine whether the following loop conditions are met:

[0220] j>1

[0221] If satisfied, the value of j is reduced by 1 and steps 9.2 to 9.3 are repeated, otherwise proceed to step 9.5;

[0222] Step 9.5: After the step 9.4 cycle is completed, the final reconstructed signal Y 1 (k) Perform inverse unitary discrete Fourier transform to obtain the final enhanced language signal x(n), x(n) = uDFTinv(Y 1 (k)).

[0223] uDFTinv stands for unitary inverse discrete Fourier transform, which is different from the ordinary inverse discrete Fourier transform.

[0224] The present invention provides a single-channel speech enhancement method for an optical fiber microphone integrating OMLSA and TQWT, and the technical effects are as follows:

[0225] 1) The present invention makes full use of the characteristics of OMLSA and TQWT, and the two are combined to achieve complementarity. OMLSA has a strong denoising ability when there is speech, but there is still a lot of residual noise when the signal-to-noise ratio is low, and TQWT obtains wavelet coefficients by decomposition, which further suppresses the residual noise of OMLSA when the signal-to-noise ratio is low.

[0226] 2) The present invention uses the reconstruction error to provide a reference for the Q factor and oversampling rate required for TQWT denoising, and by minimizing the reconstruction error, the optimal parameters are selected to achieve the best speech enhancement performance of TQWT.

[0227] 3) In the present invention, when the segmented signal-to-noise ratio of the signal after OMLSA processing is high, it means that the residual noise of the speech signal is very small. If TQWT processing is still performed, it will cause part of the useful signal to be lost. Therefore, the present invention uses the segmented signal-to-noise ratio as the decision condition for the signal after OMLSA processing to decide whether further TQWT processing is needed, thereby saving computing resources while ensuring the best voice quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0228] Figure 1 It is a flow chart of the present invention.

[0229] Figure 2 This is how the reconstruction error changes with the Q value in Example 1.

[0230] Figure 3 This is how the reconstruction error changes with the r value in Example 1.

[0231] Figure 4 This is a comparison chart of the results of performing OMLSA on the signal in Example 1 and the speech enhancement results of the present invention. DETAILED DESCRIPTION

[0232] Embodiment 1:

[0233] The present invention performs speech enhancement on the signal collected by the fiber optic microphone, the sampling frequency is 10kHz, and the collected signal has a total of 32812 sampling points. The parameters are set as follows:

[0234] Hanning window frame length and frame shift settings for framing:

[0235] Hanning window frame length N f Take 512, frame shift N m Take 128.

[0236] Gain function G when speech is absent min Value: G min =0.1257.

[0237] Smoothing parameter α before bias compensation d The value of: α d =0.85

[0238] The value of the adjustment factor γ1 is: γ1=3.

[0239] The value of the weight factor α that controls the trade-off between noise reduction and speech distortion is: α=0.92.

[0240] The first power spectrum time smoothing parameter α s The value of: α s =0.9.

[0241] The value of the half-length w of the normalized window function is: w=1.

[0242] The two thresholds γ0 and ζ0 for language existence judgment are set as follows:

[0243] Take γ0=4.6, ζ0=1.67.

[0244] Noise spectrum minimum estimation bias compensation factor B min Value: B min =1.66.

[0245] Segmented signal-to-noise ratio Size of each segment: S m =256.

[0246] The value of the speech silence segment energy threshold th is: th=0.02.

[0247] Maximum value of quality factor Q max and the maximum redundancy r max set up:

[0248] Get Q max =10, r max =6.

[0249] The signal is enhanced using the present invention and compared with the original noisy signal and the result of only OMLSA enhancement:

[0250] The calculated signal-to-noise ratio of the signal after OMLSA enhancement is 4.25 dB, which is less than 5 dB. Therefore, TQWT is required for further denoising. Figure 2 and Figure 3 It can be seen that the best Q value for this speech is 1.8, the best r value is 3, and the corresponding decomposition layer number J is 11. The enhancement effect of the algorithm on the speech is observed by the time domain waveform of the signal. Figure 4 It can be seen that after OMLSA processing, a lot of noise in the speech signal is removed, but there is still some residual noise. Compared with the processing results of the present invention, the present invention can further reduce the residual noise on the basis of OMLSA, especially the blank part of the speech can see an obvious denoising effect, and the partial waveform of the speech is very similar to the original speech, which shows that the present invention can reduce the interference noise to the greatest extent while ensuring the fidelity of the speech, thereby achieving speech enhancement.

[0251] Embodiment 2:

[0252] 8 sentences of speech were selected from the TIMIT speech database, and four noise types, F16, Pink, Tank, and White, were selected from the Noisex-92 noise database, and mixed with speech signals at signal-to-noise ratios of -5dB, 0dB, 5dB, and 10dB, respectively, to obtain speech signals with different signal-to-noise ratios. OMLSA and the present invention were used to enhance the noisy speech signals, respectively, and the speech enhancement algorithm was evaluated using a comprehensive evaluation index, which includes three parameters: signal distortion Csig, background noise Cbak, and overall quality Covl of the speech. The higher the score, the better the speech quality. The parameter settings for speech enhancement are the same as those in Example 1. The scores of the comprehensive evaluation indexes under different noise types are shown in Table 1:

[0253] Table 1 Comparison of comprehensive evaluation indicators under different noise types

[0254]

[0255]

[0256] As can be seen from Table 1, except for a few indicators with shading, the speech quality processed by the speech enhancement method of the present invention is greatly improved compared with only OMLSA processing, and the indicator Cbak of the present invention in all cases is no worse than that of only OMLSA processing, indicating that the method of the present invention has excellent performance in suppressing background noise. Moreover, when the speech signal evaluation index is high, the index data after OMLSA processing and the processing of the present invention are the same, such as the bold part of the index data, indicating that the segmented signal-to-noise ratio of the speech after OMLSA processing is greater than 5dB at this time, and the speech can already meet the application requirements, and does not need to be further enhanced using TQWT.

[0257] The present invention first performs frame processing on the collected speech signal, performs OMLSA speech enhancement on each frame, and uses the segmented signal-to-noise ratio as a judgment condition to decide whether TQWT is needed to remove residual noise. The reconstruction error is used to select the optimal values ​​of the parameters Q factor and redundancy r required for TQWT decomposition to ensure that the most suitable set of wavelet coefficients is obtained for the decomposition of the speech signal, and the threshold is adaptively selected according to each wavelet coefficient. The wavelet coefficients are sparsely processed using basis pursuit, and the final enhanced speech signal is obtained by reconstructing the sparse wavelet coefficients. Compared with only using OMLSA for denoising, the method of the present invention can further effectively suppress residual noise on this basis, thereby improving the speech enhancement effect.

Claims

1. A single-channel speech enhancement method for fiber microphone integrating OMLSA and TQWT, characterized in that The following steps are involved: Step 1: Assume that the speech signal detected by the optical fiber microphone is y(n), where n is the sampling point index, 0≤n≤N-1, and N is the total number of sampling points; Step 2: Select the Hanning window to frame the speech signal y(n) to obtain the time domain signal y(n,l), where l is the frame index, 0≤l≤L-1, and L is the total number of frames; Step 3: Use discrete Fourier transform to convert the time domain signal y(n,l) into the frequency domain signal Y(k,l), where k is the frequency index, k = 0, 1, 2, ..., N-1; Step 4: Enhance the speech based on the OMLSA algorithm and obtain the enhanced spectrum amplitude Calculated by the following formula: Where: G min is the gain function when speech does not exist; p(k,l) is the posterior probability of speech existence in the lth frame at frequency k; G H1 (k,l) is the gain function when the lth frame of speech exists at frequency k; (G min |Y(k,l)|) (1-p(k,l)) Indicates that when the l-th frame of speech does not exist at frequency k, a gain G is applied to |Y(k,l)| min To suppress noise, (G H1 (k,l)·|Y(k,l)|) p(k,l) Indicates that when the l-th frame of speech exists at frequency k, a gain G is applied to |Y(k,l)| H1 (k, l) emphasizes the phonetic component, and the symbol |·| indicates taking the absolute value; Step 5: Perform inverse Fourier transform on the enhanced signal spectrum amplitude of all frames, and restore the speech signal x1(n) enhanced by the OMLSA algorithm according to the Hanning window function; Step 6: Determine whether the segmented signal-to-noise ratio segSNR of the speech signal x1(n) satisfies the following conditions: segSNR ≥ 5dB; If satisfied, skip step 7, step 8, and step 9. At this time, x1(n) is the final enhanced speech signal x(n). Otherwise, proceed to step 7. Step 7: Perform TQWT decomposition on x1(n) to obtain J+1 wavelet coefficients w{j}, where J is the optimal total decomposition level, j is the level index, 1≤j≤J+1; Step 8: Use basis pursuit to sparse the wavelet coefficients w{j} to obtain the sparse wavelet coefficients w1{j}; Step 9: Perform inverse TQWT on the sparse wavelet coefficients w1{j} to obtain the final enhanced speech signal x(n).

2. According to claim 1, a fiber microphone single-channel speech enhancement method integrating OMLSA and TQWT is characterized in that: In the step 2, Indicates rounding up, N f is the Hanning window frame length, N m For frame shift.

3. According to claim 1, a fiber microphone single-channel speech enhancement method integrating OMLSA and TQWT is characterized in that: The step 4 comprises the following steps: Step 4.1: Initialize the following parameters: Initial value of gain function G when speech exists H1 (k,0)=1, the initial value of the posterior signal-to-noise ratio of the speech γ(k,0)=1, the initial noise power spectrum λ d (k,0)=|Y(k,0)| 2 , initial noise power spectrum before offset compensation Where Y(k,0) is the spectrum information of the initial frame, k represents the frequency index; the initial value of the first power spectrum time smoothing S(k,0) = S f (k,0), the initial value S of the minimum value of the first power spectrum time smoothing min (k,0)=S f (k,0), the initial value of the second power spectrum time smoothing Initial value of the minimum value of the second power spectrum time smoothing S f (k,0) is the initial value of the smoothed power spectrum of speech frequency; l is initialized to 1; Step 4.2: Calculate the noise power spectrum of the lth frame before offset compensation at frequency k The calculation formula is: in: is the noise power spectrum of the l-1th frame before offset compensation at frequency k, Y(k,l-1) is the frequency domain signal of the l-1th frame at frequency k, is the time-varying smoothing parameter of the l-1th frame at frequency k, and its calculation formula is: Where: α d Represents the smoothing parameter before bias compensation, 0<α d <1; p(k,l-1) is the posterior probability of the presence of speech in the l-1th frame at frequency k; Step 4.3: Calculate the noise power spectrum λ of the lth frame at frequency k d The calculation formula for (k,l) is as follows: Where: β is the bias factor, and the calculation formula is: Among them: γ1 is the regulatory factor; It represents the base of natural logarithm, e, raised to the power of -γ1; Step 4.4: Calculate the posterior signal-to-noise ratio γ(k,l) of the lth frame of speech at frequency k, and the calculation formula is: Step 4.5: Calculate the prior signal-to-noise ratio ξ(k,l) of the lth frame of speech at frequency k, and the calculation formula is: Where: α is the weight factor for controlling the trade-off between noise reduction and speech distortion; G H1 (k,l-1) is the gain function when the l-1th frame of speech exists at frequency k; γ(k,l-1) is the posterior signal-to-noise ratio of the l-1th frame of speech at frequency k; max{γ(k,l)-1,0} means taking the maximum value between γ(k,l)-1 and 0; Step 4.6: Calculate the gain function G when the lth frame of speech exists at frequency k H1 (k,l), the calculation formula is: in: represents the balance value of the priori signal-to-noise ratio and the a posteriori signal-to-noise ratio of the lth frame of speech at frequency k; Step 4.7: Calculate the first power spectrum time smoothing value S(k,l) of the lth frame of speech at frequency k, and the calculation formula is: S(k,l)=a s ·S(k,l-1)+(1-α s )·S f (k,l); Where: α s is the time smoothing parameter of the first power spectrum, S(k,l-1) is the time smoothing value of the first power spectrum of the l-1 frame speech at frequency k, S f (k, l) is the first power spectrum frequency smoothing value of the l-th frame speech at frequency k, and its calculation formula is: Where: b(i) is the normalized window function, the window length is 2w+1, w represents the half length of the window function, and i is the window function index, that is, Y(ki,l) is the frequency domain signal of the lth frame at frequency ki; Step 4.8: Calculate the second power spectrum frequency smoothing of the l-th frame signal at frequency k and the second power spectrum time smoothing The calculation formulas are: in: is the second power spectrum time smoothing of the l-1 frame signal at frequency k, and the value of I(ki,l) is given by the following formula: Among them: γ0 and ζ0 are two thresholds for speech presence judgment, "&&" represents the logical operation "and"; γ min (ki,l) represents the minimum signal-to-noise ratio of the l-th frame of speech at frequency ki, ζ(ki,l) represents the short-term stability index of the l-th frame of speech at frequency ki; S(ki,l) is the first power spectrum time smoothing of the l-th frame of speech at frequency ki, B min is the bias compensation factor for estimating the minimum value of the noise spectrum, and the minimum value S of the first power spectrum time smoothing of the l-th frame speech at frequency ki min (ki,l)=min{S min (ki,l-1),S(ki,l)},S min (ki,l-1) is the minimum value of the first power spectrum time smoothing of the l-1 frame of speech at frequency ki; Step 4.9: Calculate the prior probability q(k,l) that the lth frame of speech does not exist at frequency k. The calculation formula is: in: S min (k,l-1) is the minimum value of the first power spectrum time smoothing of the l-1 frame of speech at frequency k; γ1 represents the adjustment factor; Step 4.10: Calculate p(k,l), the calculation formula is: Step 4.11: Change the current G H1 Substitute (k, l) and p(k, l) into the calculation formula of step 4, as follows: Output frequency k at the enhanced spectrum amplitude of the lth frame Step 4.12: Determine whether the following loop conditions are met: l<L； If satisfied, increase the value of l by 1 and repeat steps 4.2 to 4.11; otherwise proceed to step 5.

4. According to claim 1, a fiber microphone single-channel speech enhancement method integrating OMLSA and TQWT is characterized in that: In step 5, the enhanced signal spectrum amplitudes of all frames are subjected to inverse Fourier transform, and the calculation formula is: in: is the time domain signal of the lth frame of speech, n1 is the sampling point index of the lth frame of time domain signal, 0≤n1≤N f -1, e is the base of natural logarithm, i1 is the imaginary unit; And according to the Hanning window function, the speech signal x1(n) enhanced by the OMLSA algorithm is restored. The specific steps are: Apply the Hanning window to get the time domain signal a(n1,l) after windowing the l-th frame of speech: Where: hanning(n1) is the Hanning window formula, cos is the cosine trigonometric function; x1(n) is obtained by overlapping and adding a(n1,l) in frames, and the calculation formula is: Among them: hanning_overlap(n1) Hanning window overlap frame weight, its calculation formula is: 。 5. According to claim 1, a fiber microphone single-channel speech enhancement method integrating OMLSA and TQWT is characterized in that: In step 6, calculating segSNR includes the following steps: Step 6.1: Divide x1(n) into M segments, where the value of M is: in: Indicates rounding down, S m is the size of each segment, m is the segment index, 1≤m≤M; Step 6.2: For the t-th sampling point in the m-th segment of the residual noise in x1(n), 1_m The value of (t) is estimated by: Where: x 1_m (t) is the signal of the t-th sampling point in the m-th segment of x1(n), and th is the energy threshold of the speech silence segment; Step 6.3: Calculate segSNR:

6. According to claim 1, a fiber microphone single-channel speech enhancement method integrating OMLSA and TQWT is characterized in that: In step 7, decomposing x1(n)TQWT to obtain J+1 wavelet coefficients w{j} includes the following steps: Step 7.1: Use the reconstruction error E to find the optimal value of the quality factor Q opt and the optimal value r of redundancy r opt ; Step 7.2: Perform unitary discrete Fourier transform uDFT(x1(n)) on x1(n) to obtain X1(k), whose formula is: Where: X_1(k) is the discrete Fourier transform of x1(n), initialized with j=1; Step 7.3: Calculate the length of the low-pass subband of the jth layer and the high-pass subband length Among them: "round" means rounding operation, Step 7.4: Calculate the passband length P of the j-th low-pass layer j , high-pass passband length S j , the length of the transition zone T j : Step 7.5: Obtain the j-th layer low-pass sub-band sequence and high-pass sub-band sequence; Step 7.6: Perform inverse unitary discrete Fourier transform to obtain the j-th wavelet coefficient w{j} of TQWT decomposition: Where: uDFTinv represents the unitary inverse discrete Fourier transform, and its relationship with the inverse discrete Fourier transform is: Where: w_inv{j} is The inverse discrete Fourier transform of Step 7.7: Determine whether the following loop conditions are met: j <J; If satisfied, the value of j is increased by 1 and steps 7.3 to 7.6 are repeated, otherwise proceed to step 7.8; Step 7.8: Calculate the J+1th wavelet coefficient of the TQWT decomposition Represents the J-th layer low-pass subband sequence; so far, all the J+1 wavelet coefficients decomposed have been calculated.

7. According to claim 6, a fiber optic microphone single-channel speech enhancement method integrating OMLSA and TQWT is characterized in that: The step 7.1 comprises the following steps: S7.1.1: Set the maximum value that Q can take to be Q max , the maximum value that r can take is r max , initialize Q = 1, r is [3,r max ] any value in the interval; S7.1.2: Calculate the number of decomposition levels J1 from Q and r using the formula: The Q value at this time is stored in the sequence Q_cell, and steps 7.2 to 7.8 are performed using the current three parameters Q, r, and J1, and step 8 is skipped. At this time, w{j} is w1{j} in step 9, and step 9 is performed; S7.1.3: Calculate the reconstruction error E = max(|x1(n) - x(n)|) and store it in the sequence E_cell; S7.1.4: Determine whether the following loop conditions are met: Q<Q max ; If satisfied, add 0.2 to the value of Q and repeat S7.1.2 to S7.1.3, otherwise proceed to S7.1.5; S7.1.5: The minimum reconstruction error min(E_cell) corresponds to the value in Q_cell, which is Q opt ; S7.1.6: Initialize r = 3, Q to [1, Q max ] any value in the interval; S7.1.7: Calculate the decomposition level J1, store the current r value in the sequence r_cell, use the current parameters to perform steps 7.2 to 7.8, skip step 8, and w{j} is now w1{j} in step 9, and proceed to step 9; S7.1.8: Calculate the reconstruction error E = max(|x1(n) - x(n)|) and store it in sequence E_cell_1; S7.1.9: Determine whether the following loop conditions are met: r<r max ； If satisfied, add 0.2 to the value of r and repeat S7.1.7 to S7.1.8, otherwise proceed to S7.1.10; S7.1.10: The minimum reconstruction error min(E_cell_1) corresponds to the value in r_cell, which is r opt ; S7.1.11: Use Q opt and r opt Find the optimal decomposition level J.

8. According to claim 6, a fiber optic microphone single-channel speech enhancement method integrating OMLSA and TQWT is characterized in that: The step 7.5 comprises the following steps: S7.5.1: Get the j-th layer low-pass subband sequence k_low is the low-pass subband frequency index, Where: represents the initial value of the j-th layer low-pass subband sequence, X1(0) represents the initial value of X1(k) obtained by performing unitary discrete Fourier transform on x1(n); Where: Represents the j-th layer low-pass subband sequence from frequency 1 to frequency P j The value of X1(1:P j ) indicates that X1(k) goes from frequency 1 to frequency P j The value of Where: Represents the j-th layer low-pass subband sequence from frequency P j +1 to frequency P j +T j The value of X1(P j +(1:T j )) indicates that X1(k) is from frequency P j +1 to frequency P j +T j The value of θ(1:T j ) means to take values ​​from 1 to T in sequence j Perform the following theta operation: Where: D represents the parameter value for θ operation; Where: Indicates that the j-th layer low-pass subband sequence is at frequency The value of ; Where: Indicates that the j-th layer low-pass subband sequence is from frequency To frequency The value of X1(NP j -(1:T j )) indicates that X1(k) is from frequency NP j -1 to frequency NP j -T j The value of Where: A1:B1 means the sequence is indexed from A1 to B1, ".*" means the corresponding elements of the sequence are multiplied; S7.5.2: Get the j-th layer high-pass subband sequence k_high is the high-pass subband frequency index, Where: represents the initial value of the j-th layer high-pass subband sequence; Where: Represents the j-th layer high-pass subband sequence from frequency 1 to frequency T j The value of X1(P j +(1:T j )) indicates that X1(k) is from frequency P j +1 to frequency P j +T j The value of θ(T j -(1:T j )+1) means taking the value T in turn j To 1, perform theta operation; Where: Represents the j-th layer high-pass subband sequence from frequency T j +1 to frequency T j +S j The value of X1(P j +T j +(1:S j )) indicates that X1(k) is from frequency P j +T j +1 to frequency P j +T j +S j The value of Where: Indicates that the j-th layer high-pass subband sequence is at frequency X1(N / 2) represents the value of X1(k) at k=N / 2; Where: Indicates that the j-th layer high-pass subband sequence is from frequency To frequency The value of X1(NP j -T j -(1:S j )) indicates that X1(k) is from frequency NP j -T j -1 to frequency NP j -T j -S j The value of Where: Indicates that the j-th layer high-pass subband sequence is from frequency To frequency The value of X1(NP j -(1:T j )) indicates that X1(k) is from frequency NP j -1 to frequency NP j -T j The value of .

9. According to claim 1, a fiber optic microphone single-channel speech enhancement method integrating OMLSA and TQWT is characterized in that: In step 8, the basis pursuit is used to sparse w{j}, including the following steps: Step 8.1: Initialize j=1; Step 8.2: Generate an l j × j The identity matrix A j , l j is the length of the jth wavelet coefficient; vector b j =A j T *w{j} represents b j Projection onto the identity matrix, where A j T A j The transpose of ; matrix B j =[A j T *A j ,-A j T *A j ;-A j T *A j ,A j T *A j ] is used to construct quadratic terms in optimization problems; Step 8.3: Calculate the regularization parameter τ of the jth layer j , and its calculation formula is: in: is a regulation factor that reflects the change of sparsity with the number of layers j. is the coefficient for adjusting the amplitude of w{j}, ln() is the natural logarithm, and median() is the median operation; Step 8.4: Construct a j The column vector c j , as the driving vector for sparse optimization, is generated by: Where: C j is of length 2·l j And a column vector of all 1s; Step 8.5: Solve the quadratic optimization problem Among them: z is the intermediate quantity when solving, and its length is 2·l j Column vector of z T represents the transpose of z; c j T Indicates c j The transpose of represents the optimal solution of z when minimizing the objective function, and the optimal solution is z j ; Step 8.6: By z j The sparse wavelet coefficient w1{j} is obtained, and its calculation formula is: w1{j}=z j (1:l j )-z j (l j +1:2·l j ); Where: z j (1:l j ) represents the column vector z j Row 1 to row 1 j The value of the row, z j (l j +1:2·l j ) represents the column vector z j No.1 j +1 row to 2.1 j The value of the row; Step 8.7: Determine whether the following loop conditions are met: j≤J; If satisfied, the value of j is increased by 1 and steps 8.2 to 8.6 are repeated; otherwise, proceed to step 9.

10. The method for single-channel speech enhancement of a fiber microphone integrating OMLSA and TQWT according to claim 1, characterized in that: In step 9, performing inverse TQWT on w1{j} includes the following steps: Step 9.1: Find the unitary discrete Fourier transform of the J+1th sparse wavelet coefficient w1{J+1} as Y J+1 (k_2), where k_2 is its frequency index, Represents Y J+1 The length of (k_2); initialize j = J; Step 9.2: Calculate the passband length P of the j-th low-pass layer after sparseness j _1. High-pass passband length S j _1. Length of transition zone T j _1 are: Where: N j _1 is the length of the unitary discrete Fourier transform of the jth sparse wavelet coefficient w1{j}, N j+1 _1 is the length of the unitary discrete Fourier transform of the j+1th sparse wavelet coefficient w1{j+1}; Step 9.3: Calculate the j-th layer reconstructed signal Y j (k1), 0≤k1≤M j -1, k1 represents the frequency index of the j-th layer reconstructed signal; comprising the following steps: S9.3.1: Perform unitary discrete Fourier transform on the jth sparse wavelet coefficient w1{j} to W j (k_1), k_1 is its frequency index, 0≤k_1≤N j _1-1; S9.3.2: Signal Y reconstructed from layer j+1 j+1 (k1) Find the j-th layer sparse low-pass subband sequence Where: represents the initial value of the j-th layer sparse low-pass subband sequence, Y j+1 (0) represents the j+1th layer reconstructed signal Y j+1 The initial value of (k1); Where: Represents the j-th layer of sparse low-pass subband sequence from frequency 1 to frequency P j The value of _1, Y j+1 (1:P j _1) indicates Y j+1 (k1) from frequency 1 to frequency P j The value of _1; Where: Represents the j-th layer of sparse low-pass subband sequence from frequency P j _1+1 to frequency P j _1+T j _1 value; Y j+1 (P j _1+(1:T j _1)) indicates Y j+1 (k1) from frequency P j _1+1 to frequency P j _1+T j _1 value;θ(1:T j _1) means to take values ​​from 1 to T in sequence j _1 performs theta operation; Where: Represents the j-th layer of sparse low-pass subband sequence from frequency P j _1+T j _1+1 to frequency P j _1+T j _1+S j The value of _1; Where: Represents the j-th layer of sparse low-pass subband sequence at frequency M j The value of ; Where: Denotes the j-th layer sparse low-pass subband sequence from frequency M j -P j _1-T j _1-1 to frequency M j -P j _1-T j _1-S j The value of _1; Where: Represents the j-th layer of sparse low-pass subband sequence from frequency M j -P j _1-1 to frequency M j -P j _1-T j _1 value; Y j+1 (N j+1 _1-P j _1-(1:T j _1)) indicates Y j+1 (k1) from frequency N j+1 _1-P j _1-1 to frequency N j +1 _1-P j _1-T j The value of _1, Where: Represents the j-th layer of sparse low-pass subband sequence from frequency M j -1 to frequency M j -P j _1 value; Y j+1 (N j+1 _1-(1:P j _1)) indicates Y j+1 (k1) from frequency N j+1 _1-1 to frequency N j+1 _1-P j The value of _1; S9.3.3: Find the j-th layer of sparse high-pass subband sequence Where: represents the initial value of the j-th layer sparse high-pass subband sequence; Where: Represents the j-th layer of sparse high-pass subband sequence from frequency 1 to frequency P j The value of _1; Where: Represents the j-th layer of sparse high-pass subband sequence from frequency P j _1+1 to frequency P j _1+T j _1 value; W j (1:T j _1) represents the unitary discrete Fourier transform of w1{j} to W j (k_1) from frequency 1 to frequency T j _1 value;θ(T j _1-(1:T j _1)+1) means taking the value T in turn j _1 to 1 performs theta operation; Where: Represents the j-th layer of sparse high-pass subband sequence from frequency P j _1+T j _1+1 to frequency P j _1+T j _1+S j _1 value; W j (T j _1+(1:S j _1)) indicates W j (k_1) from frequency T j _1+1 to frequency T j _1+S j The value of _1; Where: Represents the j-th layer sparse high-pass subband sequence at frequency M j / 2; W j (N j _1 / 2) indicates W j (k_1) when k_1=N j _The value at 1 / 2; Where: Represents the j-th layer of sparse high-pass subband sequence from frequency M j -P j _1-T j _1-1 to frequency M j -P j _1-T j _1-S j _1 value; W j (N j _1-T j _1-(1:S j _1)) indicates W j (k_1) from frequency N j _1-T j _1-1 to frequency N j _1-T j _1-S j The value of _1; Where: Represents the j-th layer of sparse high-pass subband sequence from frequency M j -P j _1-1 to frequency M j -P j _1-T j _1 value; W j (N j _1-(1:T j _1)) indicates W j (k_1) from frequency N j _1-1 to frequency N j _1-T j The value of _1; Where: Represents the j-th layer of sparse high-pass subband sequence from frequency M j -1 to frequency M j -P j The value of _1; S9.3.4: At this time, the j-th layer reconstructs the signal Step 9.4: Determine whether the following loop conditions are met: j>1 If satisfied, the value of j is reduced by 1 and steps 9.2 to 9.3 are repeated, otherwise proceed to step 9.5; Step 9.5: After the step 9.4 cycle is completed, the final reconstructed signal Y 1 (k) Perform inverse unitary discrete Fourier transform to obtain the final enhanced language signal x(n), x(n) = uDFTinv(Y 1 (k)).

Images