Intelligent voice separation method and system based on subway carriage operation environment

Abstract

The invention discloses an intelligent voice separation method and system based on a subway carriage operation environment. The method belongs to the technical field of voice processing, and comprises the following operation steps: acquiring a mixed voice signal in a subway carriage by using a multi-channel microphone array; carrying out band-pass filtering on the acquired mixed signal, removing unnecessary frequency band noise by using a Butterworth filter, and reserving a core frequency band of voice; separating the mixed signal by adopting an independent component analysis model, establishing a linear instantaneous mixed model, assuming that noise sources are independent from each other, updating a separation matrix through a natural gradient algorithm, and finally separating the mixed signal into an independent voice signal and an independent noise signal; in addition, the convergence speed and the separation precision are improved by adopting an adaptive step length and momentum factor optimization algorithm. The system part comprises a signal acquisition module, a signal processing module and an output module. According to the method, the signal-to-noise ratio and the separation effect of the voice signals can be remarkably improved in the complex noise environment of the subway carriage.

Description

Technical Field

[0001] The present invention belongs to the technical field of speech processing and relates to an intelligent speech separation method; namely, an intelligent speech separation method and system based on the operation environment of a subway car; specifically, it relates to an intelligent speech separation method and system based on the complex noise environment of a subway car. Background Art

[0002] In the complex operating environment of subway cars, there are a large number of noise sources, including mechanical noise, environmental noise and passenger noise. These noises seriously affect the quality of voice communication between subway dispatching center staff and train drivers and passengers. Existing speech separation technologies, such as the FastICA (Fast Independent Component Analysis) algorithm, have limited separation effects in high-noise environments and are difficult to meet real-time communication needs. Therefore, it is necessary to develop an intelligent speech separation method suitable for subway car environments to ensure the clarity and reliability of voice communication. Summary of the invention

[0003] In response to the above problems, the purpose of the present invention is to propose an intelligent speech separation method based on an improved natural gradient algorithm and independent component analysis (ICA) (i.e., an intelligent speech separation method and system based on the complex noise environment of a subway car), which can effectively cope with the complex noise environment in a subway car and improve the signal-to-noise ratio and separation accuracy of voice communication.

[0004] The technical solution of the present invention is: the intelligent speech separation method and system based on the subway car operation environment of the present invention comprises the following steps:

[0005] S1. Signal acquisition: using a multi-channel microphone array to collect mixed voice signals in the subway car, the signal includes mechanical noise, environmental noise and passenger noise;

[0006] S2, signal preprocessing: band-pass filtering is performed on the collected mixed signal, using a Butterworth filter with a passband frequency of 300 Hz to 3400 Hz, retaining the effective frequency band of the speech and removing low-frequency and high-frequency noise;

[0007] S3, signal separation: The mixed signal is processed by independent component analysis (ICA) to establish a linear instantaneous mixed model:

[0008] X(t)=A·S(t)

[0009] Where X(t) is the observed signal, A is the mixing matrix, and S(t) is the source signal. The goal of signal separation is to find the separation matrix W so that the estimated signal Y(t) satisfies:

[0010] Y(t)=W·X(t)

[0011] S4, adaptive step size and momentum factor optimization: In the natural gradient algorithm, the adaptive step size factor η(t) and momentum factor β are introduced, and the weight update formula is:

[0012]

[0013] In the formula, η(t) is dynamically adjusted according to the change of the signal:

[0014]

[0015] Where ρ(t) is the correlation coefficient in the current iteration, α is the control factor, and ∈ is a small positive number to prevent division by zero;

[0016] S5. Signal reconstruction: After the separation matrix W converges, the mixed signal is reconstructed through the separation matrix to obtain independent source signal S(t) and noise signal.

[0017] Furthermore, in S3, the signal separation adopts the adaptive step size natural gradient method, and iterates through the following formula:

[0018] W(k+1)=W(k)+μ·[If(y(k))·y T (k)]·W(k)

[0019] Where μ is the adaptive step size factor and f(y(k)) is the nonlinear function of the source signal.

[0020] Furthermore, in S3, the separation matrix W is updated according to the following optimization formula:

[0021]

[0022] In the formula, α is the learning rate, is the gradient of the objective function.

[0023] Furthermore, the further step of the adaptive step size and momentum factor is: optimizing the step size using the Spearman rank correlation coefficient, and the change of the correlation coefficient is calculated by the formula:

[0024]

[0025] Dynamically control the step size to accelerate convergence and improve separation accuracy.

[0026] Furthermore, an intelligent voice separation system based on a subway car operation environment includes:

[0027] (1) Signal acquisition module, used to collect mixed voice signals in subway cars;

[0028] (2) Signal processing module, including independent component analysis (ICA) model, natural gradient algorithm, bandpass filter and adaptive step size optimization algorithm, used for preprocessing, separation and reconstruction of mixed signals;

[0029] (3) Output module, used to output the separated speech signal and noise.

[0030] The beneficial effects of the present invention are as follows: the present invention obtains the first biometric information of the new order based on historical order analysis and performs fuzzy verification, which effectively avoids the security risks caused by forged identity information in traditional identity verification, and at the same time improves the reliability of verification by utilizing the first attribute feature; generates a verification logistics track according to the real-time verification relationship, so that the verification process can adapt in real time with the topology changes of the transportation device, significantly improves the verification efficiency and reduces the verification resource requirements; the real-time verification relationship is also based on the real-time adjustment of the log record priority value, comprehensively considers the verification resource requirements and credit evaluation, and ensures the optimality of the verification logistics track; based on watermark key verification and consumer user feature credit evaluation, the new order verification is enhanced to resist quantum computing attacks and malicious order disguises; the risk monitoring mechanism of the real-time verification relationship can monitor order consumer users in real time, quickly mark risky orders and execute re-verification steps, thereby improving the robustness of the verification device and the overall safety of the transportation device; in addition, the present invention does not need to rely on the preliminary calculations of transportation supervision personnel, and can accurately and objectively judge the accurate safety situation of the merchant in the later stage, which has a good effect on maintaining user stickiness afterwards. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a diagram of a speech separation system in a subway car operation environment of the present invention;

[0032] Figure 2 is a signal-to-noise ratio diagram of the speech signal separation result in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The specific technical scheme of the present invention is further described in detail below with reference to specific examples.

[0034] As shown in the figure, the intelligent speech separation method or system based on the complex noise environment of a subway car described in the present invention specifically includes the following steps:

[0035] S1. Signal acquisition and preprocessing: A microphone array is used to collect mixed speech signals in the subway car. The collected signals contain mixed signals from multiple different noise sources.

[0036] S2. Independent Component Analysis (ICA) Model: Using the independent component analysis (ICA) technology, the mixed signal in the subway car environment is regarded as a linear instantaneous mixed model, and it is assumed that the noise sources are independent of each other. The model decomposes the mixed signal into several independent source signals through matrix operations.

[0037] Adaptive step size and momentum factor optimization: The adaptive step size and momentum factor mechanism are introduced to enhance the robustness and separation effect of the algorithm in complex environments by adjusting the convergence speed of the algorithm. The adaptive step size is dynamically adjusted according to the correlation coefficient between the separation signal and the source signal to ensure the stability and efficiency of the signal separation process.

[0038] Iterative update of separation matrix: Use the natural gradient algorithm to iteratively update the separation matrix until it converges to the optimal solution. Each iteration optimizes the independence of the separated signals by calculating the non-Gaussianity of the signal.

[0039] Signal separation and reconstruction: After the algorithm converges, the mixed signal is processed through the separation matrix to reconstruct each independent source signal, including the speech signal and the background noise signal. The low-frequency and high-frequency noise are suppressed through the filter to further improve the clarity of the speech signal.

[0040] Output separation signal: Finally, the separated voice signal and noise signal are output separately for real-time communication between the staff of the subway dispatching center and the train drivers and passengers in the subway car.

[0041] Embodiment: An intelligent speech separation method and system based on a subway car operation environment comprises the following steps:

[0042] We selected a pure human voice dialogue dataset, and selected NoiseX-92 and TUT Acoustic scenes noise libraries as experimental data. The noise dataset mainly consists of mechanical noise, environmental noise, and passenger noise to simulate the real subway car environment.

[0043] refer to Figure 1 As shown, it is a speech separation system in the subway car operation environment, wherein the speech separation process includes: speech signal acquisition; speech signal preprocessing; blind source separation of the signal; signal enhancement; judging whether the signal-to-noise ratio (SNR) of the signal to be output is higher than 25dB, if not, reprocessing the signal, and if so, outputting the audio signal.

[0044] Setting parameters: First, you need to determine the number of signals, that is, the number of independent speech sources to be processed; in this example, setting K=4 means that four source signals will be used for processing and mixing; the setting of the number of signals will determine the dimensions in subsequent matrix operations, ensuring that the size of the matrix is ​​consistent with the number of processed signals.

[0045] Read audio files: Use the audioread function in MATLAB to read four audio files separately. These audio files will be used as input for source signals. Each audio file represents an independent source speech signal, assuming that they are all recorded at the same sampling rate. The audio sampling rate Fs is the number of samples collected per second, which determines the temporal resolution of the signal. Assume that all audio signals have the same sampling rate, which facilitates subsequent signal processing.

[0046] Signal preprocessing: In order to remove unnecessary frequency components in the audio signal, a Butterworth bandpass filter is designed; the function of the bandpass filter is to retain only the frequency components in the range of 300Hz to 3400Hz, because this frequency range is the effective frequency band of the speech signal, and other frequency components (such as low-frequency background noise and high-frequency mechanical noise) will be filtered out; the main parameters of the filter design include the filter order and half-power cutoff frequency, the formula is as follows:

[0047]

[0048] Where H(s) is the filter transfer function, ωc is the cutoff frequency, n is the filter order, and G0 is the gain;

[0049] Apply the designed filter to each audio signal; the filter function is used to implement the filtering operation on the signal; after filtering, the unnecessary frequency components in each signal will be effectively removed, leaving mainly the effective frequency part of the speech signal; for the convenience of subsequent processing, it is necessary to ensure that all filtered audio signals have the same length; therefore, use the min function to find the shortest audio length and truncate all audio signals to the same length;

[0050] The four filtered audio signals are stored in a matrix S, where each row of the matrix S represents an independent source signal and each column represents a sampling point. This is done to facilitate subsequent matrix operations, such as signal mixing and separation. By constructing signal data in matrix form, multiple signals can be processed uniformly.

[0051] The source signals are mixed together using a randomly generated mixing matrix A. The signal mixing process can be expressed as a matrix operation: X = A·S, where X is the mixed signal matrix, A is a randomly generated mixing matrix, and S is the source signal matrix. Through this formula, the obtained X is a linear combination of several signals, and each mixed signal contains the components of multiple source signals.

[0052] The signal data is centralized and whitened to remove the DC offset of the signal, that is, the mean of each signal becomes zero; then whitening is performed. This step makes the covariance matrix of the signal become the unit matrix through the decorrelation operation, thereby eliminating the correlation between different signals; the mathematical expression of whitening is:

[0053]

[0054] Where E and D are the eigenvector and eigenvalue matrices of the covariance matrix, respectively.

[0055] In the signal separation process, the natural gradient algorithm is combined with the adaptive step size, and the Spearman rank correlation coefficient is introduced to dynamically adjust the step size to further optimize the convergence speed and stability of the algorithm; the natural gradient algorithm is an optimization method, especially suitable for signal separation problems in independent component analysis; compared with ordinary gradient algorithms, the natural gradient algorithm is more suitable for processing parameter optimization problems in information geometry; its core idea is to multiply the gradient vector by the inverse Fisher information matrix to obtain a faster convergence speed.

[0056] In the signal separation process, the natural gradient algorithm is used to update the separation matrix W, and its update formula is:

[0057]

[0058] In the formula, is the gradient of the separation matrix, η(t) is the step size; to improve the efficiency and stability of the algorithm, the momentum factor and adaptive step size are combined to dynamically adjust the step size of the weight update.

[0059] By introducing momentum, we can keep a certain update direction based on the previous iterations, thus avoiding over-reliance on the current gradient direction. The formula for momentum update is:

[0060]

[0061] Where β is the momentum factor (usually close to 1, such as 0.95), which represents the trade-off between the current gradient and the previous gradient update. The introduction of momentum can make the weight update smoother and avoid falling into the local minimum.

[0062] In the natural gradient algorithm, the step size η(t) has an important influence on the convergence speed of the algorithm; a larger step size can speed up the convergence speed, but may cause oscillation or instability; while a step size that is too small will lead to slow convergence; therefore, adaptive step size is a strategy for dynamically adjusting the step size, which automatically adjusts the step size at each iteration according to the changes in the signal.

[0063] The effect of signal separation is judged by the Spearman rank correlation coefficient, and the step size is adjusted dynamically based on this; the Spearman correlation coefficient ρ measures the nonlinear correlation of two signals by comparing their ranking differences, which is more suitable for evaluating the nonlinear relationship between the source signal and the separated signal than the Pearson correlation coefficient.

[0064] In each iteration, the current separation signal W(:,p) is calculated T The Spearman correlation coefficient ρ(t) between X and the true source signal S(p,:), and then adjust the step size according to the change of the correlation coefficient; if the Spearman correlation coefficient of the current iteration changes greatly, it indicates that the signal separation effect has been significantly improved, and the step size can be appropriately increased; if the change is small or no change, the step size can be appropriately reduced to ensure the stability of the separation process;

[0065] The step size adjustment formula is:

[0066]

[0067] Where ρ(t) is the Spearman correlation coefficient of the current iteration; ρ(t-1) is the Spearman correlation coefficient of the previous iteration; factor_up and factor_down are factors for increasing and decreasing the step size, usually 1.05 and 0.95; threshold is the threshold of the Spearman correlation coefficient change. If the change exceeds the threshold, it means that the current step size needs to be adjusted.

[0068] After signal separation, the Spearman correlation coefficient between the separated signal and the source signal is calculated. This coefficient is used to measure the nonlinear correlation between the two signals. The final separated signal can be matched with the source signal to calculate the signal-to-noise ratio (SNR), as shown in the following formula:

[0069]

[0070] Signal power and noise power are measures of the error between the source signal and the separated signal, respectively.

Claims

1. An intelligent speech separation method based on the subway car operation environment is characterized by: The operation steps are as follows: Step (1): signal acquisition; Step (2): signal preprocessing; Step (3): signal separation; Step (4): Adaptive step size and momentum factor optimization; Step (5): Signal reconstruction.

2. The intelligent speech separation method based on the subway car operation environment according to claim 1 is characterized in that: In step (1), the signal collection specifically includes: using a multi-channel microphone array to collect mixed voice signals in the subway car; The signals include mechanical noise, environmental noise and passenger noise.

3. The intelligent speech separation method based on the subway car operation environment according to claim 1 is characterized in that: In step (2), the signal preprocessing is specifically: bandpass filtering the collected mixed signal, using a Butterworth filter, the passband frequency of the filter is 300 Hz to 3400 Hz, retaining the effective frequency band of the speech, and removing low-frequency and high-frequency noise.

4. The intelligent speech separation method based on the subway car operation environment according to claim 1 is characterized by: In step (3), the signal separation is specifically: The mixed signal is processed by independent component analysis to establish a linear instantaneous mixed model: X(t)=A·S(t) In the formula, X(t) represents the observed signal, A represents the mixing matrix, and S(t) represents the source signal; The goal of signal separation is to find the separation matrix W so that the estimated signal Y(t) satisfies: Y(t)=W·X(t).

5. The intelligent speech separation method based on the subway car operation environment according to claim 4 is characterized in that: The signal separation adopts the adaptive step size natural gradient method, which is iterated by the following formula: W(k+1)=W(k)+μ·[I-f(y(k))·y T (k)]·W(k) Where μ is the adaptive step size factor and f(y(k)) is the nonlinear function of the source signal.

6. The intelligent speech separation method based on the subway car operation environment according to claim 1 is characterized in that: In step (4), the adaptive step size and momentum factor optimization refers to: in the natural gradient algorithm, introducing the adaptive step size factor η(t) and the momentum factor β, and the weight update formula is: In the formula, η(t) is dynamically adjusted according to the change of the signal: Where ρ(t) is the correlation coefficient in the current iteration, α is the control factor, and ∈ is a small positive number to prevent division by zero.

7. The intelligent speech separation method based on the subway car operation environment according to claim 6 is characterized by: The further step of the adaptive step size and momentum factor optimization is to adjust the step size using the Spearman rank correlation coefficient, and the change of the correlation coefficient is calculated by the formula: Dynamically control the step size to accelerate convergence and improve separation accuracy.

8. The intelligent speech separation method based on the subway car operation environment according to claim 1 is characterized by: In step (5), the signal reconstruction is specifically: after the separation matrix W converges, the mixed signal is reconstructed by the separation matrix to obtain independent source signals S(t) and noise signals.

9. The intelligent speech separation method based on the subway car operation environment according to claim 8 is characterized in that: The update of the separation matrix W is based on the following optimization formula: In the formula, α is the learning rate, is the gradient of the objective function.

10. An intelligent voice separation system based on the subway car operation environment, characterized by: include: A signal acquisition module is used to collect mixed voice signals in subway cars; Signal processing module, including independent component analysis model, natural gradient algorithm, bandpass filter and adaptive step size optimization algorithm, used for preprocessing, separation and reconstruction of mixed signals; The output module is used to output the separated speech signal and noise signal.

Images