A Speech Signal Filtering Iterative Method

By combining the 3rd and 7th order Newtonian iteration methods and the FastICA algorithm, the iterative processing of speech signals is optimized, and the disorder and amplitude uncertainty in the blind source separation of speech signals is solved, and the separation effect and iteration of speech signals are improved. speed.

CN114242107BActive Publication Date: 2024-12-03NANJING METRO CONSTRUCTION CO LTD +1
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202111438778.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-30
Publication Date
2024-12-03
Estimated Expiration
2041-11-30

AI Technical Summary

Technical Problem

Existing speech signal blind source separation algorithms such as FastICA have disorder and amplitude uncertainty of the separation signal, making it difficult to effectively separate without understanding the relationship between the number of sensors and the number of source signals, affecting the accuracy of human-computer interaction and Validity.

Method used

The 3rd and 7th order Newton's iterative method is used to combine with the FastICA algorithm to construct a new iterative formula. By centralizing and whitening the source signal during the iteration process, the estimation of the demix matrix is ​​optimized and the separation effect is improved.

Benefits of technology

The average signal-to-noise ratio and separation speed of the voice signal are significantly improved, the number of iterations is reduced, and the separation effect and iteration speed are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114242107B_ABST
    Figure CN114242107B_ABST
Patent Text Reader

Abstract

The present invention discloses a speech signal filtering iterative method, comprising step 1: selecting two Newton iteration methods from 3rd order, 5th order and 7th order Newton iteration methods; step 2: combining with FastICA algorithm, using the Newton iteration method selected in step 1 to construct an iteration formula; step 3: using the iteration formula in step 2 to filter the speech signal. The present invention can effectively improve the average signal-to-noise ratio, reduce the number of iterations, and improve the speech separation effect and separation speed of blind source speech signals.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of speech signal processing, and in particular relates to a speech signal filtering iterative method. Background Art

[0002] As an important communication medium in human daily life, voice carries a lot of information and is one of the important ways of communication for humans. In recent years, with the continuous development of computer science and technology and the popularization of personal computers, the ways of interaction between people and computers have gradually increased. Voice interaction is the most convenient way of interaction between people and computers. Its core technology is speech recognition technology, which involves the cross-integration of multiple disciplines such as signal processing, pattern recognition, probability theory, and artificial intelligence.

[0003] How to process and analyze these speech signals and extract useful information from them is a very important technology. Independent Component Correlation Algorithm (ICA) is an algorithm that can achieve blind source separation in signal processing and has been widely used in speech recognition. ICA is fundamentally a statistical and computational technology that is often used to reveal hidden components in random variables, measurement data or signals. It can separate and extract independent signal sources when both the source signal and the mixing mode are unknown. Among the many ICA algorithms, the FastICA algorithm stands out for its fast convergence speed, high robustness, and good separation effect. It was first introduced into the field of neural network modeling in the early 1980s. After more than 30 years of application in different fields, its reliability has been proven. But at the same time, FastICA also has certain shortcomings. For example, due to the disorder and amplitude uncertainty of the separation signal in the blind source separation algorithm itself, the relationship between the number of sensors and the number of source signals needs to be known during calculation, otherwise blind source separation cannot be performed.

[0004] In actual environments, speech signals are always mixed with a variety of different sound signals and noises. Therefore, research on blind source separation methods for speech signals to separate and extract useful speech signals from mixed audio signals is of great significance for improving the accuracy and effectiveness of human-computer interaction. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a speech signal filtering iterative method in view of the above-mentioned deficiencies in the prior art.

[0006] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0007] A speech signal filtering iterative method, comprising:

[0008] Step 1: Select two Newton iteration methods from the 3rd, 5th and 7th order Newton iteration methods;

[0009] Step 2: Combined with the FastICA algorithm, the Newton iteration method selected in step 1 is used to construct the iteration formula;

[0010] Step 3: Use the iterative formula in step 2 to filter the speech signal.

[0011] To optimize the above technical solutions, the specific measures taken also include:

[0012] In the above step 1, the third-order Newton iteration method and the fifth-order Newton iteration method are selected, or the third-order Newton iteration method and the seventh-order Newton iteration method are selected.

[0013] The above step 2 combines the third-order Newton iteration formula and the fifth-order Newton iteration formula to construct an iteration formula, and obtains a new iteration formula:

[0014]

[0015]

[0016]

[0017] Among them, x n is the sound signal to be separated, f(x n ) is x n is the functional relationship of the variables, f′(x n ) is the function f(x n ) for variable x n The derivative of n ,y n is the intermediate variable in the iteration process, x n+1 is the sound signal after one Newton iteration;

[0018] Substitute the new iterative formula into the FastICA algorithm to obtain the final iterative formula:

[0019]

[0020] in:

[0021]

[0022]

[0023]

[0024]

[0025] Wherein, W is the unmixing matrix. When the number of source signals is known, the mixed sound signal is considered to be a mixture of several source signals according to their respective weights. The matrix composed of the weights is the unmixing matrix W. By finding the unmixing matrix W, we can get the proportion of each source signal in the process of forming the mixed sound signal, and thus separate the source signal.

[0026] W0 is the initial value of W, W k+1 is the unmixing matrix after one iteration, W1 and W2 are intermediate variables in the iteration process, g(y) is any second-order continuous and differentiable non-quadratic function, which will directly affect the final separation effect according to the different functions, and E{} represents the mathematical expectation of the function inside the brackets.

[0027] Step 3 above includes:

[0028] Step 3.1.1 Centralize the source signal to make it a random variable with zero mean;

[0029] Step 3.1.2 performs whitening on the signal after the centralization process, thereby obtaining the whitened data z;

[0030] Step 3.1.3 sets the maximum number of iterations to no more than 200;

[0031] Step 3.1.4 randomly selects an initial weight vector W0 as the initial matrix;

[0032] Step 3.1.5: normalize the initial weight vector W0;

[0033] Step 3.1.6 Application Iterate W0;

[0034] Step 3.1.7 Pass Standardize it;

[0035] Step 3.1.8 If Then it is judged to be convergent, otherwise, return to step 3.1.6;

[0036] Step 3.1.9 Output vector W n+1 .

[0037] The above step 2 combines the third-order Newton iteration method and the seventh-order Newton iteration method to construct an iteration formula and obtain a new iteration formula:

[0038]

[0039]

[0040]

[0041]

[0042] Among them, x n is the sound signal to be separated, f(x n ) is x n is the functional relationship of the variables, f′(x n ) is the function f(x n ) for variable x n The derivative of n ,y n is the intermediate variable in the iteration process, x n+1 is the sound signal after one Newton iteration;

[0043] Substitute the new iterative formula into the FastICA algorithm to obtain the final iterative formula:

[0044]

[0045] in:

[0046]

[0047]

[0048]

[0049]

[0050]

[0051] Where W is the unmixing matrix; W0 is the initial value of W, W k+1 is the unmixing matrix after one iteration, W1 and W2 are intermediate variables in the iteration process, g(y) is any second-order continuous and differentiable non-quadratic function, which will directly affect the final separation effect according to the different functions, and E{} represents the mathematical expectation of the function inside the brackets.

[0052] Step 3 above includes:

[0053] Step 3.2.1 Centralize the source signal to make its mean 0, that is, convert the source signal into a random variable with zero mean;

[0054] Step 3.2.2 performs whitening on the signal after the centralization process, thereby obtaining the whitened data z;

[0055] Step 3.2.3 sets the maximum number of iterations to no more than 200;

[0056] Step 3.2.4 randomly selects an initial weight vector W0 as the initial matrix;

[0057] Step 3.2.5: normalize the initial weight vector W0;

[0058] Step 3.2.6 Application Iterate W0;

[0059] Step 3.2.7 Pass Standardize it;

[0060] Step 3.2.8 If Then it is judged to be convergent, otherwise, return to step 3.2.6;

[0061] Step 3.2.9 Output vector W n+1 .

[0062] The present invention has the following beneficial effects:

[0063] Compared with the traditional algorithm, the speech filtering signal iterative method based on the third-order Newton iterative method and the fifth-order Newton iterative method has an average signal-to-noise ratio improved by 4.2103, and the average total number of iterations reduced by 3.9 times. The separation effect and iteration speed are significantly improved.

[0064] Compared with traditional algorithms, the speech filtering signal iterative method based on the third-order Newton iteration method and the seventh-order Newton iteration method can also effectively improve the signal-to-noise ratio and zone speed, thereby improving the separation speed and effect of speech blind source signals. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is the separation effect of the traditional algorithm;

[0066] Figure 2 It is the comparison between the source signal and the separated signal of the traditional algorithm;

[0067] Figure 3 It is the separation effect of the speech signal filtering iterative method based on the third-order Newton iterative method and the fifth-order Newton iterative method of the present invention;

[0068] Figure 4 The present invention is based on the third-order Newton iteration method and the fifth-order Newton iteration method of the speech signal filtering iterative method and the comparison between the source signal and the separated signal;

[0069] Figure 5 It is the separation effect of a seventh-order convergent Newton iterative FastICA separation algorithm;

[0070] Figure 6 It is a comparison between the source sound and the demixed sound of the Newton iterative FastICA separation algorithm with seventh-order convergence;

[0071] Figure 7It is the separation effect of the speech signal filtering iterative method based on the third-order Newton iterative method and the seventh-order Newton iterative method of the present invention;

[0072] Figure 8 The present invention compares the source signal and the separated signal of the speech signal filtering iterative method based on the 3rd order Newton iterative method and the 7th order Newton iterative method;

[0073] Fig. 9 A is a comparison diagram of the number of iterations of the speech signal filtering iterative method based on the third-order Newton iterative method and the fifth-order Newton iterative method of the present invention and the traditional filtering iterative method, each running 10 times;

[0074] Fig. 9 B is a comparison diagram of the number of iterations of the speech signal filtering iterative method based on the third-order Newton iteration method and the seventh-order Newton iteration method according to the present invention and the speech signal filtering iterative method based on the seventh-order Newton iteration, each of which is run 10 times;

[0075] Fig.10 The present invention is a flow chart of a speech signal filtering iterative method. DETAILED DESCRIPTION

[0076] The embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0077] See also Fig.10 The present invention provides a speech signal filtering iterative method, comprising:

[0078] Step 1: Select two Newton iteration methods from the 3rd, 5th and 7th order Newton iteration methods;

[0079] Step 2: Combined with the FastICA algorithm, the Newton iteration method selected in step 1 is used to construct the iteration formula;

[0080] Step 3: Use the iterative formula in step 2 to filter the speech signal.

[0081] In the embodiment, the higher the order of Newton iteration is, the faster the convergence speed is, but the amount of calculation will increase accordingly, and it should be selected according to the specific situation.

[0082] The step 1 selects the 3rd order Newton iteration method and the 5th order Newton iteration method, or selects the 3rd order Newton iteration method and the 7th order Newton iteration method.

[0083] Embodiment 1:

[0084] The step 2 combines the third-order Newton iteration formula and the fifth-order Newton iteration formula to construct an iteration formula, and obtains a new iteration formula:

[0085]

[0086]

[0087]

[0088] Among them, x n is the sound signal to be separated, f(x n ) is x n is the functional relationship of the variables, f′(x n ) is the function f(x n ) for variable x n The derivative of n ,y n is the intermediate variable in the iteration process, x n+1 is the sound signal after one Newton iteration;

[0089] Substitute the new iterative formula into the FastICA algorithm to obtain the final iterative formula:

[0090]

[0091] in:

[0092]

[0093]

[0094]

[0095]

[0096] Wherein, W is the unmixing matrix. When the number of source signals is known, the mixed sound signal is considered to be a mixture of several source signals according to their respective weights. The matrix composed of the weights is the unmixing matrix W. By finding the unmixing matrix W, we can get the proportion of each source signal in the process of forming the mixed sound signal, and thus separate the source signal.

[0097] W0 is the initial value of W, W k+1 is the unmixing matrix after one iteration, W1 and W2 are intermediate variables in the iteration process, g(y) is any second-order continuous and differentiable non-quadratic function, which will directly affect the final separation effect according to the different functions, and E{} represents the mathematical expectation of the function inside the brackets.

[0098] Accordingly, step 3 includes:

[0099] Step 3.1.1 Centralize the source signal to make it a random variable with zero mean;

[0100] Its main purpose is to transform the source signal into a random variable with zero mean, so that it meets the assumptions of the ICA mathematical model and does not affect the estimation of the unmixing matrix.

[0101] Step 3.1.2 performs whitening on the signal after the centralization process, thereby obtaining the whitened data z;

[0102] After the signal is whitened, the correlation between signals can be removed and the redundancy of the output signal can be reduced, thereby reducing the complexity of the problem, improving the convergence speed of the algorithm, and reducing the steady-state error;

[0103] Step 3.1.3 sets the maximum number of iterations to no more than 200;

[0104] Step 3.1.4 randomly selects an initial weight vector W0 as the initial matrix;

[0105] Step 3.1.5: normalize the initial weight vector W0;

[0106] Step 3.1.6 Application Iterate W0;

[0107] Step 3.1.7 Pass Standardize it;

[0108] Step 3.1.8 If Then it is judged to be convergent. At this time, it can be considered that the unmixing matrix has been obtained. Otherwise, return to step 3.1.6;

[0109] Step 3.1.9 Output vector W n+1 .

[0110] Embodiment 2:

[0111] The step 2 combines the third-order Newton iteration method and the seventh-order Newton iteration method to construct an iteration formula, and obtains a new iteration formula:

[0112]

[0113]

[0114]

[0115]

[0116] Among them, x n is the sound signal to be separated, f(x n ) is x n is the functional relationship of the variables, f′(x n ) is the function f(x n ) for variable x n The derivative of n ,y n is the intermediate variable in the iteration process, xn+1 is the sound signal after one Newton iteration;

[0117] Substitute the new iterative formula into the FastICA algorithm to obtain the final iterative formula:

[0118]

[0119] in:

[0120]

[0121]

[0122]

[0123]

[0124]

[0125] Where W is the unmixing matrix; W0 is the initial value of W, W k+1 is the unmixing matrix after one iteration, W1 and W2 are intermediate variables in the iteration process, g(y) is any second-order continuous and differentiable non-quadratic function, which will directly affect the final separation effect according to the different functions, and E{} represents the mathematical expectation of the function inside the brackets.

[0126] Accordingly, step 3 includes:

[0127] Step 3.2.1 Centralize the source signal to make its mean 0, that is, convert the source signal into a random variable with zero mean;

[0128] Its main purpose is to transform the source signal into a random variable with zero mean, so that it meets the assumptions of the ICA mathematical model and does not affect the estimation of the unmixing matrix.

[0129] Step 3.2.2 performs whitening on the signal after the centralization process, thereby obtaining the whitened data z;

[0130] After the signal is whitened, the correlation between signals can be removed and the redundancy of the output signal can be reduced, thereby reducing the complexity of the problem, improving the convergence speed of the algorithm, and reducing the steady-state error;

[0131] Step 3.2.3 sets the maximum number of iterations to no more than 200;

[0132] Step 3.2.4 randomly selects an initial weight vector W0 as the initial matrix;

[0133] Step 3.2.5: normalize the initial weight vector W0;

[0134] Step 3.2.6 Application Iterate W0;

[0135] Step 3.2.7 Pass Standardize it;

[0136] Step 3.2.8 If Then it is judged to be convergent. At this time, it can be considered that the unmixing matrix has been obtained. Otherwise, return to step 3.2.6;

[0137] Step 3.2.9 Output vector W n+1 .

[0138] The iterative method of speech signal filtering based on the third-order Newton iteration method and the fifth-order Newton iteration method in the present invention is compared with the operating effect of the traditional FastICA algorithm. The main method is to use Matlab to generate a random matrix to mix the same source signal, and then use the two algorithms to separate the mixed signals respectively, so as to compare the separation effects of the two algorithms.

[0139] The separation effect of traditional algorithms is as follows Figure 1 As shown, the contrast between the source signal and the separated signal is Figure 2 As shown, the iterative method for speech signal filtering based on the third-order Newton iterative method and the fifth-order Newton iterative method in the present invention is as follows Figure 3 As shown, the contrast between the source signal and the separated signal is Figure 4 As shown in , the two algorithms are run ten times on the same set of source signals, and the number of iterations of each run of the two algorithms is obtained, and a line graph is made for comparison, as shown in Fig. 9 As shown in A.

[0140] The calculation results of the two separation algorithms are quantified, and the calculation results are shown in Table 1 and Table 2.

[0141] Table 1 Results of ten operations of the traditional FastICA algorithm

[0142]

[0143]

[0144] It can be calculated from Table 1 that the average signal-to-noise ratio of source signal 1 after ten operations is: 17.6552, the average signal-to-noise ratio of source signal 2 after ten operations is: 22.3749, and the average total number of iterations is: 10.1 times.

[0145] Table 2 Results of ten operations of the iterative method for speech signal filtering based on the third-order Newton iterative method and the fifth-order Newton iterative method

[0146] Source signal 1 signal-to-noise ratio Source signal 2 signal-to-noise ratio Total number of iterations 19.1669 21.2671 6 19.1669 22.6519 6 17.0073 22.6519 6 19.1669 22.6519 6 19.1669 22.6519 6 19.1669 22.6519 6 19.1669 22.6519 6 17.0073 22.6519 7 17.0073 22.6519 6 17.0073 22.6519 7

[0147] It can be calculated from Table 2 that the average signal-to-noise ratio of source signal 1 after ten operations is: 18.3031, the average signal-to-noise ratio of source signal 2 after ten operations is: 22.5134, and the average total number of iterations is: 6.2 times.

[0148] The performance comparison of the two algorithms is shown in Table 3.

[0149] Table 3 Performance comparison of two algorithms

[0150]

[0151] It can be seen that the iterative method of speech signal filtering based on the third-order Newton iteration method and the fifth-order Newton iteration method improves the average signal-to-noise ratio of source signal 1 by 0.6479 and the average signal-to-noise ratio of source signal 2 by 0.1385, compared with the traditional FastICA algorithm, and the average total number of iterations is reduced by 3.9 times.

[0152] The iterative method of speech signal filtering based on the third-order Newton iteration method and the seventh-order Newton iteration method in the present invention is compared with the operating effect of the modified FastICA algorithm with a seventh-order convergent Newton iteration. The main method is to use Matlab to generate a random matrix to mix the same source signal, and then use the two algorithms to separate the mixed signals respectively, so as to compare the separation effects of the two algorithms.

[0153] The separation effect of a seventh-order Newton iterative FastICA algorithm is shown in Figure 5 As shown, the comparison between the separated source sound and the unmixed sound is as follows Figure 6 As shown, the separation effect of the iterative method of speech signal filtering based on the third-order Newton iterative method and the seventh-order Newton iterative method in the present invention is as follows Figure 7 As shown in the figure, the contrast between the separated source signal and the separated signal is Figure 8 Run the two algorithms on the same set of source signals ten times each, obtain the number of iterations of each run of the two algorithms, and make a line graph for comparison, as shown in Fig. 9 As shown in B.

[0154] The calculation results of the two separation algorithms are quantified, and the calculation results are shown in Table 4 and Table 5.

[0155] Table 4 Results of ten operations of a seventh-order convergent Newton iteration algorithm

[0156] Source sound 1 signal-to-noise ratio Source Sound 2 Signal-to-Noise Ratio Total number of iterations 17.0078 22.2518 15 19.1668 21.2672 17 17.0078 21.2672 12 17.0078 21.2672 20 17.0078 22.6518 13 19.1668 21.2672 18 17.0078 21.2674 15 19.1668 21.2672 12 19.1668 21.2672 13 17.0078 21.2672 16

[0157] From Table 4, it can be calculated that the average signal-to-noise ratio of source sound 1 after ten operations is: 17.8714, the average signal-to-noise ratio of source sound 2 after ten operations is: 21.4641, and the average total number of iterations is: 15.1 times

[0158] Table 5 Results of ten operations of iterative method for speech signal filtering based on 3rd order Newton iterative method and 7th order Newton iterative method

[0159] Source sound 1 signal-to-noise ratio Source Sound 2 Signal-to-Noise Ratio Total number of iterations 19.1668 22.6519 6 17.0073 21.2671 6 19.1668 22.6519 8 17.0073 22.6519 8 19.1668 21.2671 8 17.0073 21.2671 7 19.1668 22.6519 6 17.0073 22.6519 7 19.1668 21.2671 6 19.1668 22.6519 8

[0160] It can be calculated from Table 5 that the average signal-to-noise ratio of source sound 1 after ten operations is: 18.3030, the average signal-to-noise ratio of source sound 2 after ten operations is: 22.0980, and the average total number of iterations is: 7 times.

[0161] The performance comparison of the two algorithms is shown in Table 6:

[0162] Table 6 Performance comparison of two algorithms

[0163]

[0164] The higher the signal-to-noise ratio, the better the separation effect. From Table 6, it can be clearly seen that the speech signal filtering iterative method based on the 3rd-order Newton iterative method and the 7th-order Newton iterative method has a better separation effect for both source sound 1 and source sound 2. Compared with the 7th-order convergent Newton iterative method, the average signal-to-noise ratio of source sound 1 is improved by 0.4316, the average signal-to-noise ratio of source sound 2 is improved by 0.6339, and the average number of iterations is reduced by 8.1 times.

[0165] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A speech signal filtering iterative method, characterized in that: include: Step 1: Select the 3rd order Newton iteration method and the 5th order Newton iteration method; Step 2: Combined with the FastICA algorithm, the Newton iteration method selected in step 1 is used to construct the iteration formula; Step 3: Use the iterative formula in step 2 to filter the speech signal; Step 2: When the iterative formula is constructed by combining the third-order Newton iterative formula and the fifth-order Newton iterative formula, a new iterative formula is obtained: Among them, x n is the sound signal to be separated, f(x n ) is x n is the functional relationship of the variables, f′(x n ) is the function f(x n ) for variable x n The derivative of n ,y n is the intermediate variable in the iteration process, x n+1 is the sound signal after one Newton iteration; Substitute the new iterative formula into the FastICA algorithm to obtain the final iterative formula: in: Wherein, W is the unmixing matrix. When the number of source signals is known, the mixed sound signal is considered to be a mixture of several source signals according to their respective weights. The matrix composed of the weights is the unmixing matrix W. By finding the unmixing matrix W, we can get the proportion of each source signal in the process of forming the mixed sound signal, and thus separate the source signal. W0 is the initial value of W, W k+1 is the unmixing matrix after one iteration, W1 and W2 are intermediate variables in the iteration process, g(y) is any second-order continuous and differentiable non-quadratic function, which will directly affect the final separation effect according to the different functions, and E{} represents the mathematical expectation of the function inside the brackets; The step 3 comprises: Step 3.1.1 performs central processing on the speech source signal to make the source signal a random variable with zero mean; Step 3.1.2 performs whitening on the signal after the centralization process, thereby obtaining the whitened data z; Step 3.1.3 sets the maximum number of iterations to no more than 200; Step 3.1.4 randomly selects an initial weight vector W0 as the initial matrix; Step 3.1.5: normalize the initial weight vector W0; Step 3.1.6 Application Iterate W0; Step 3.1.7 Pass Standardize it; Step 3.1.8 If Then it is judged to be convergent, otherwise, return to step 3.1.6; Step 3.1.9 Output vector W n+1 .

Citation Information

Patent Citations

  • Blind voice separation method and structure, voice control system and electric appliance assembly

    CN106356075A

  • Multi-person speech separation method and system for chatting robot

    CN108922557A

  • Noisy signal-containing blind source separation algorithm

    CN109584900A

  • Source separation by independent component analysis in conjunction with source direction information

    US20130297296A1

  • Blind signal separation method and structure, voice control system, and electrical appliance assembly

    US20200027473A1