Frequency domain noise reduction and reduction method and system for synchronously collecting double groups of human voices

Abstract

The invention relates to the technical field of voice communication, and discloses a frequency domain noise reduction and reduction method and system for double-group human voice synchronous acquisition, and the method comprises the steps: obtaining a synchronous signal through double groups of acquisition points, constructing a double-source frequency domain coordinate system, and mapping the coordinate system into an energy distribution point; calculating a spatial relationship to identify an emotional distortion state, and dividing four quadrants to generate a distortion type mark; dividing a frequency domain signal and executing targeted reconstruction and correction, and finally dividing a frequency band enhanced energy remarkably-reduced part; according to the method, the voice composite distortion under the extreme emotion can be accurately processed, and reliable voice quality guarantee is provided for accurate transmission of key information in an emergency communication scene.

Description

Technical Field

[0001] The present invention relates to the field of voice communication technology, and more specifically, to a method and system for frequency domain noise reduction and restoration for synchronously collecting two groups of human voices. Background Art

[0002] In the field of voice communications, especially in critical scenarios like emergency rescue and medical first aid, the quality of the caller's voice signal directly impacts the accuracy of information transmission and the efficiency of response. Under extreme emotional states, voice signals are often accompanied by complex physiological distortions. How to effectively reduce noise while preserving critical information becomes a core challenge in improving the reliability of voice communications.

[0003] In the prior art, Chinese patent application publication number CN114550752A discloses a method and device for analyzing waveform nodes for automatic noise reduction in two-way speech signals. This method achieves speech emotion classification by intercepting the sound waveform signal in time segments and comparing it with a pre-generated emotion training set. The core of this method is to associate emotional states with waveform features, while integrating automatic noise reduction and waveform modeling functions. Chinese patent application publication number CN119964537A discloses a mixed-audio noise reduction method and system. This method uses emotion recognition to determine whether the occupants of the vehicle are in a negative emotional state. If they are in an anxious state, the system collects sound signals, extracts noise feature vectors, and generates mixed-audio sound wave bands to neutralize the noise and reduce its impact on emotions.

[0004] However, these existing technologies still have specific shortcomings. The emotion analysis described in Chinese patent application publication number CN114550752A relies on waveform comparison, enabling only emotion classification and basic noise reduction. It fails to identify the physiological compound distortions associated with extreme emotions, such as frequency domain compression caused by vocal cord tension and periodic frequency shifts caused by emotional tremors. This can easily misclassify the compressed high-frequency enhancement components as noise. Chinese patent application publication number CN119964537A, while incorporating emotion-adjusted noise reduction strategies, focuses on the interplay between ambient noise and emotion, failing to distinguish the essential differences between physiological distortion and real noise. This can lead to frequency fluctuations caused by emotional tremors being misclassified as environmental interference and suppressed. More critically, the compound distortion frequencies associated with extreme emotions often overlap with the fundamental frequencies of key information, such as direction and number. The indiscriminate processing used in existing technologies results in "over-noise reduction," resulting in the loss of important information. In emergency communication scenarios, this can delay rescue decisions and even lead to serious consequences. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a frequency domain noise reduction and restoration method and system for the synchronous collection of two groups of human voices. Through the dual-group collection, a dual-source frequency domain coordinate system is constructed, and the four-quadrant partitioning is combined to identify the type of emotional distortion. The frequency domain reconstruction and correction are performed in a targeted manner to effectively distinguish between physiological distortion and environmental noise. The noise reduction and restoration accuracy of extreme emotional speech are significantly improved, providing reliable technical support for the accurate transmission of key information in emergency communications.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] The frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices includes:

[0008] Acquire two sets of synchronous human voice signals through two sets of acquisition points, construct a dual-source frequency domain coordinate system, and map the two sets of human voice signals into the constructed dual-source frequency domain coordinate system to form two sets of energy distribution points;

[0009] In the dual-source frequency domain coordinate system, the spatial relationship between the two sets of energy distribution points is calculated to obtain an index to measure the degree of distortion of the dual-set vocal synchronization signal, and to identify whether the current speech state of the dual-set vocal synchronization signal is in an emotional distortion state;

[0010] Divide the distortion recognition area into four quadrants in the dual-source frequency domain coordinate system, determine the quadrants to which the two groups of energy distribution points corresponding to the two groups of human voice synchronization signals identified as being in an emotional distortion state belong, and generate corresponding distortion type labels;

[0011] Segmenting the first set of frequency domain signals FreA and the second set of frequency domain signals FreB according to the generated distortion type labels to generate different types of frequency domain sub-signals, performing corresponding frequency domain reconstruction and correction on the different types of frequency domain sub-signals to obtain corrected first set of frequency domain signals Fre"A and second set of frequency domain signals Fre"B;

[0012] The corrected first group of frequency domain signals Fre'A and the second group of frequency domain signals Fre'B are divided into frequency bands, the frequency bands with significantly reduced energy are identified, and the frequency bands with significantly reduced energy are enhanced.

[0013] Furthermore, the two groups of human voice synchronization signals include a first group of human voice signals and a second group of human voice signals;

[0014] The method for forming two groups of energy distribution points includes: performing fast Fourier transform on the first group of human voice signals and the second group of human voice signals respectively to obtain a first group of frequency domain signals and a second group of frequency domain signals, using the amplitude of each frequency component in the first group of frequency domain signals as the Y coordinate and the frequency value as the X coordinate to form the first group of energy distribution points; and using the amplitude of each frequency component in the second group of frequency domain signals as the Y coordinate and the frequency value as the X coordinate to form the second group of energy distribution points.

[0015] Furthermore, the method of calculating the spatial relationship between the two groups of energy distribution points to obtain an index for measuring the degree of distortion of the two groups of synchronous vocal signals includes:

[0016] Based on the statistical analysis of historical normal speech data, a standard elliptical area is set in the dual-source frequency domain coordinate system; the shortest distances from the two groups of energy distribution points to the boundary of the standard elliptical area are calculated respectively, recorded as the first deviation value and the second deviation value; the difference Diff between the first deviation value and the second deviation value is calculated AB and the mean value AB , the difference Diff AB and mean AB As an indicator to measure the degree of distortion of two groups of vocal synchronization signals.

[0017] Furthermore, the method for identifying whether the voice state of the current two groups of human voice synchronization signals is in an emotional distortion state includes:

[0018] Set the number of continuous time windows Cou C , difference threshold Thr D and average value threshold Thr M ; When continuous Cou C The difference Diff within the time window AB Both are greater than the difference threshold Thr D And the average value AB Both are greater than the average threshold Thr M , the speech state of the two groups of voice synchronization signals is marked as the emotional distortion state.

[0019] Furthermore, the method for generating the distortion type mark includes:

[0020] In the dual-source frequency domain coordinate system, the origin O is used as the center to divide the image into four quadrants. The first quadrant is defined as the frequency domain compression distortion area, the second quadrant is defined as the mixed distortion area, the third quadrant is defined as the trembling distortion area, and the fourth quadrant is defined as the normal area.

[0021] Statistically determining the distribution ratios of the first group of energy distribution points and the second group of energy distribution points in four quadrants corresponding to the two groups of synchronized vocal signals identified as being in an emotionally distorted state;

[0022] According to the distribution ratio of the first group of energy distribution points and the second group of energy distribution points in the four quadrants, corresponding distortion type marks are automatically generated.

[0023] Furthermore, the method for automatically generating corresponding distortion type labels according to the distribution ratio of the first group of energy distribution points and the second group of energy distribution points in the four quadrants includes:

[0024] Determine the key distribution quadrants of the first group of energy distribution points and the second group of energy distribution points; generate distortion type marks based on the key distribution quadrants of the first group of energy distribution points and the second group of energy distribution points, wherein the distortion type marks include frequency domain compression distortion marks, tremor distortion marks, mixed distortion marks and normal speech marks.

[0025] Furthermore, the method for determining the key distribution quadrants of the first group of energy distribution points and the second group of energy distribution points includes: setting a distribution threshold T ratio , if the proportion of the first group of energy distribution points falling into the i-th quadrant area to the total number of points exceeds the distribution threshold T ratio , it means that the i-th quadrant is the key distribution quadrant of the first group of energy distribution points; otherwise, it means that the first group of energy distribution points has no key distribution quadrant; if the proportion of the second group of energy distribution points falling into the j-th quadrant to the total number of points exceeds the distribution threshold T ratio , it means that the j-th quadrant area is the key distribution quadrant of the second group of energy distribution points; otherwise, it means that the second group of energy distribution points has no key distribution quadrant; where i and j represent the quadrant numbers, i = 1, 2, 3, 4, j = 1, 2, 3, 4.

[0026] Furthermore, the first group of frequency domain signals FreA and the second group of frequency domain signals FreB are segmented according to the generated distortion type mark to generate different types of frequency domain sub-signals, including: if the distortion type mark is a frequency domain compression distortion mark, a compression distortion sub-signal is generated; if the distortion type mark is a trembling distortion mark, a trembling distortion sub-signal is generated; if the distortion type mark is a mixed distortion mark, a mixed distortion sub-signal is generated; if the distortion type mark is a normal speech mark, a normal sub-signal is generated.

[0027] Furthermore, the method for obtaining the corrected first set of frequency domain signals Fre'A and second set of frequency domain signals Fre'B includes:

[0028] Performing corresponding frequency domain reconstruction processing on different types of frequency domain sub-signals to obtain processed first group of frequency domain signals Fre'A and second group of frequency domain signals Fre'B; correcting the processed first group of frequency domain signals Fre'A and second group of frequency domain signals Fre'B to obtain corrected first group of frequency domain signals Fre"A and second group of frequency domain signals Fre"B;

[0029] The method for correcting the processed first set of frequency domain signals Fre'A and the second set of frequency domain signals Fre'B includes: marking processing error points in the processed first set of frequency domain signals Fre'A and the second set of frequency domain signals Fre'B, and adjusting relevant parameters of frequency domain reconstruction for the frequency points marked as processing error points.

[0030] A frequency domain noise reduction and restoration system for synchronously collecting two groups of human voices is used to implement the above-mentioned frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices. The system includes:

[0031] Mapping module: obtains two sets of synchronous human voice signals through two sets of acquisition points, constructs a dual-source frequency domain coordinate system, and maps the two sets of human voice signals into the constructed dual-source frequency domain coordinate system to form two sets of energy distribution points;

[0032] Distortion Identification Module: This module calculates the spatial relationship between two sets of energy distribution points in the dual-source frequency domain coordinate system, obtains an indicator to measure the degree of distortion of the dual-set vocal synchronization signal, and identifies whether the current speech state of the dual-set vocal synchronization signal is in an emotional distortion state.

[0033] Quadrant marking module: used to divide the four-quadrant distortion recognition area in the dual-source frequency domain coordinate system, determine the quadrant belonging of the two groups of energy distribution points corresponding to the two groups of human voice synchronization signals identified as emotional distortion states, and generate corresponding distortion type labels;

[0034] Frequency domain correction module: Based on the generated distortion type mark, the first set of frequency domain signals FreA and the second set of frequency domain signals FreB are segmented to generate different types of frequency domain sub-signals. The corresponding frequency domain reconstruction and correction are performed on the different types of frequency domain sub-signals to obtain the corrected first set of frequency domain signals Fre"A and second set of frequency domain signals Fre"B.

[0035] Frequency band enhancement module: used to divide the corrected first group of frequency domain signals Fre'A and the second group of frequency domain signals Fre'B into frequency bands, identify the frequency bands with significantly reduced energy, and enhance the frequency bands with significantly reduced energy.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention realizes the accurate identification and targeted processing of voice composite distortion under extreme emotions through the frequency domain-space fusion analysis framework constructed by dual-group synchronous acquisition. With the help of the spatial relationship analysis of energy distribution points in the dual-source frequency domain coordinate system, it can effectively distinguish physiological composite distortion from environmental noise, and avoid misjudging the high-frequency enhancement components caused by frequency domain compression and the periodic frequency offset caused by emotional tremors as noise for suppression; through the distortion type label generated by four-quadrant partitioning, adaptive frequency domain reconstruction and correction can be performed for different composite distortion features. Combined with frequency band enhancement processing, it can effectively retain and restore frequency components related to key information, prevent key information from being lost or blurred due to excessive or insufficient noise reduction, and significantly improve the noise reduction and restoration accuracy of voice signals under extreme emotional states, providing stable and reliable technical support for scenarios such as emergency communications that rely on clear voice to convey key information. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0039] Figure 1 This is a principle flow chart of the frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices in the present invention;

[0040] Figure 2 A schematic diagram of the quadrant distortion recognition area division of the present invention;

[0041] Figure 3 This is a flow chart of the frequency domain reconstruction processing and correction principle of the present invention;

[0042] Figure 4 This is a functional module diagram of the frequency domain noise reduction and restoration system for synchronously collecting two groups of human voices in the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] Example 1

[0045] See also Figure 1 As shown, this embodiment provides a frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices, including:

[0046] Step S10, obtaining two groups of human voice synchronization signals through two groups of acquisition points, constructing a dual-source frequency domain coordinate system, and mapping the two groups of human voice signals into the constructed dual-source frequency domain coordinate system to form two groups of energy distribution points;

[0047] Furthermore, step S10 includes:

[0048] Step S11, obtaining two groups of voice synchronization signals through two groups of collection points, the two groups of collection points including the first group of collection points Pos A and the second set of collection points Pos B The two groups of voice synchronization signals include the first group of voice signals Sig obtained by the first group of acquisition points A The second group of human voice signals Sig obtained from the second group of collection points B ;

[0049] Step S12: The first set of collection points Pos A Establish a dual-source frequency domain coordinate system for the origin O, and start from the first set of acquisition points Pos A Point to the second set of collection points Pos B The direction of is set as the positive direction of X axis, and the direction passing through the origin O and perpendicular to the X axis is set as Y axis;

[0050] Step S13 , performing frequency domain transformation on the two groups of human voice synchronization signals and mapping them into the constructed dual-source frequency domain coordinate system to form two groups of energy distribution points.

[0051] Specifically, step S10 aims to provide a basic framework for the identification and processing of complex speech distortion under extreme emotions through the combination of spatial information and frequency domain features, and to solve the problem that the traditional single-group acquisition system is difficult to distinguish between distortion and noise due to the lack of spatial reference. The two groups of human voice synchronization signals are collected by dual-group acquisition equipment. The traditional single-group acquisition system cannot distinguish physiological distortion by spatial differences due to the lack of reference of the other group of signals. The dual-group synchronous acquisition provides a basis for distinction for subsequent analysis through the spatial correlation of the two groups of signals. If step S11 is missing, the subsequent dual-source frequency domain coordinate system will not be established due to the lack of valid input signals, and the spatial information utilization basis of the entire method will not exist, resulting in the inability to identify distortion features from the spatial dimension. The establishment of the dual-source frequency domain coordinate system is based on the need for spatial position quantification, and the origin is selected as Pos A In order to use one of the signal groups as a benchmark, it is convenient to calculate the relative spatial relationship between the two groups of signals. The setting of the X-axis direction is directly related to the physical position connection line of the two groups of acquisition points, so that the X-axis coordinate can reflect the frequency value. In the subsequent mapping, the X-axis represents the frequency and can also imply spatial position information. The Y-axis is set to a direction perpendicular to the X-axis to characterize the amplitude of the signal, that is, the energy intensity. This setting organically integrates the frequency domain characteristics, namely frequency and amplitude, with the spatial position characteristics, namely relative distance, solving the defect that traditional frequency domain analysis only focuses on a single dimension, namely frequency-amplitude, while ignoring spatial information.

[0052] The method for forming two groups of energy distribution points specifically includes: A And the second group of human voice signals Sig BFast Fourier transform is performed respectively to obtain a first group of frequency domain signals FreA and a second group of frequency domain signals FreB. The amplitude of each frequency component in the first group of frequency domain signals FreA is used as the Y coordinate, and the frequency value is used as the X coordinate to form a first group of energy distribution points. Similarly, the amplitude of each frequency component in the second group of frequency domain signals FreB is used as the Y coordinate, and the frequency value is used as the X coordinate to form a second group of energy distribution points. In the process of forming energy distribution points, each frequency component corresponds to a point. For example, FreA contains frequency f1, such as 500 Hz, and amplitude A1, such as 0.8 V. Then, this component corresponds to the point (500, 0.8) in the coordinate system. All such points constitute the first set of energy distribution points. This mapping converts abstract frequency domain data into a geometric point set that can be intuitively analyzed, so that the frequency-energy distribution difference between the two groups of signals can be reflected through the spatial distribution pattern of the point set. For example, when emotional tremors occur, the periodic fluctuations of the two groups of point sets in the X-axis direction are consistent, while the fluctuations of environmental noise do not have such consistency. This solves the problem of difficult visualization and spatial correlation analysis of signal features in traditional frequency domain analysis. If step S13 is missing, the frequency domain features of the two groups of signals will still exist in the form of abstract data, and subsequent deviation calculations cannot be performed through spatial geometric relationships, such as the distance to the standard elliptical area in step S20, resulting in the loss of quantitative basis for distortion identification.

[0053] Step S10 realizes the transformation of speech signals from the time domain-space domain to the frequency domain-space fusion domain through the coordination of dual-group synchronous acquisition, coordinate system construction and frequency domain mapping. The dual-group synchronous acquisition provides a spatial reference basis, the dual-source frequency domain coordinate system provides a fusion analysis framework, and the frequency domain mapping converts signal features into geometric features. The combination of the three enables the complex distortion of speech under extreme emotions, namely frequency domain compression and periodic frequency offset, to be identified through the spatial pattern of energy distribution points. Traditional methods cannot achieve this due to the lack of spatial information and fusion framework.

[0054] Step S20, calculating the spatial relationship between the two groups of energy distribution points in the dual-source frequency domain coordinate system, obtaining an index for measuring the degree of distortion of the dual-group vocal synchronization signal, and identifying whether the current speech state of the dual-group vocal synchronization signal is in an emotional distortion state;

[0055] Furthermore, step S20 includes:

[0056] Step S21, based on the statistical analysis of historical normal speech data, setting a standard elliptical area in the dual-source frequency domain coordinate system;

[0057] Step S22: Calculate the shortest distance from the two groups of energy distribution points to the boundary of the standard ellipse area, and record it as the first deviation value Dev. A and the second deviation value Dev B ;

[0058] Step S23, calculate the first deviation value Dev A and the second deviation value Dev B Diff AB and the mean value AB , the difference Diff AB and mean AB As an indicator to measure the degree of distortion of the synchronous signal of two groups of human voices;

[0059] Step S24, set the number of continuous time windows Cou C , difference threshold Thr D and average value threshold Thr M ; When continuous Cou C The difference Diff within the time window AB Both are greater than the difference threshold Thr D And the average value AB Both are greater than the average threshold Thr M , the speech state of the two groups of voice synchronization signals is marked as the emotional distortion state.

[0060] Specifically, step S20 aims to establish an objective distortion identification standard by quantitatively analyzing the degree of deviation between energy distribution points and normal speech features, thereby solving the problem of distortion misjudgment caused by traditional algorithms relying on a single feature threshold. Step S21 sets a standard elliptical area in a dual-source frequency domain coordinate system based on statistical analysis of historical normal speech data. The specific implementation process is as follows: normal speech samples of different genders, ages, and speaking speeds are collected, with a sample volume of no less than 1,000 hours, covering non-extreme emotional states such as calmness and slight excitement. These samples are converted into two groups of energy distribution points using the method of step S10. The energy distribution points of all normal samples are spatially clustered using an ellipse fitting algorithm. The ellipse obtained by fitting is the standard elliptical area. The ellipse's major axis is along the X-axis, representing frequency. Its length is determined by the frequency range of normal speech, taking the 95th percentile of the difference between the maximum and minimum frequency values ​​of all samples as its major axis length. The minor axis is along the Y-axis, representing amplitude. Its length is determined by the amplitude fluctuation range of normal speech, taking three times the standard deviation of the amplitude values ​​of all samples as its minor axis length. The ellipse's center coordinates are the intersection of the frequency mean and amplitude mean of all normal sample energy distribution points. The ellipse was chosen as the standard region because the frequency and amplitude distribution of normal speech exhibits a continuous distribution characteristic similar to an ellipse, with a wide frequency range and a smooth amplitude variation with frequency. However, the energy distribution points of distorted signals under extreme emotions deviate from this region. The geometric properties of the ellipse better align with the natural distribution patterns of normal speech. Using a rectangular or circular region would reduce subsequent recognition accuracy by excessively including unnaturally distributed edge points or omitting the normal fluctuation range.

[0061] Step S22 calculates the shortest distances from the two groups of energy distribution points to the boundary of the standard ellipse area, which is recorded as the first deviation value Dev. A and the second deviation value Dev B The calculation method is as follows: for each point in the first set of energy distribution points, the shortest distance from each point to the ellipse boundary is calculated using the standard equation of the ellipse; a positive result indicates that the point is outside the ellipse, and a negative result indicates that it is inside. The larger the absolute value, the greater the degree of deviation. The arithmetic average of the distances of all energy distribution points in the first set is taken as Dev A Similarly, calculate the average value of the second group of energy distribution points as Dev B This process converts abstract signal features into comparable numerical indicators by quantifying the degree of deviation of each energy point from the normal area. This solves the problem of traditional methods that judge distortion only through subjective experience. If the deviation value is not calculated, it will be impossible to objectively measure the degree of signal distortion, resulting in a lack of quantitative basis for subsequent indicators.

[0062] In step S23, Diff AB For Dev A Subtract Dev B The absolute value of Mean AB For (Dev A +Dev B ) / 2. Diff AB Reflects the consistency of the distortion patterns of the two groups of signals. Under normal speech conditions, the deviation of the two groups of signals due to the difference in spatial position is small and Diff AB Stable in a smaller range, while physiological distortion under extreme emotions will make the two sets of signals show similar deviation patterns. AB Keep the value small, the environmental noise may cause a group of signal deviations to increase significantly, making Diff AB Get bigger; Mean AB Reflects the overall deviation between the two sets of signals. The greater the deviation, the more likely the signal is distorted. The combination of the two can distinguish physiological distortion from environmental noise. For example, the frequency fluctuation caused by emotional tremors will cause both sets of signals to deviate from the elliptical area at the same time. AB Increase and Diff AB Small, but the wind noise near one side of the microphone will cause one group of signals to deviate significantly, resulting in Diff AB If only a single indicator is used, it will be impossible to distinguish between the two situations, resulting in misjudgment.

[0063] In step S24, the length of the time window is determined based on the average duration of human voice syllables to ensure that each window contains a complete speech fundamental frequency cycle, which is usually between 10 and 30 milliseconds; C The value range of Thr is 2 to 5, which is determined according to the stability requirements of the speech signal;D and Thr M Determined through historical data statistics: Collect at least 300 hours of speech samples containing extreme emotional distortion, and calculate the Diff of these samples in the emotional distortion state. AB and Mean AB Distribution, take Diff AB The 20% to 30% percentile of the distribution is taken as Thr D , ensuring more than 90% physiological distortion sample Diff AB Less than this value; take the Mean AB The 40% to 60% percentile of the distribution is taken as Thr M , ensuring that the MeanAB of more than 90% of normal speech samples is less than this value. The emotion distortion state determination mechanism, through the requirement of temporal continuity, filters out accidental deviations caused by transient interference, improving recognition stability. Without continuous window judgment, brief signal fluctuations can be misinterpreted as emotion distortion, affecting the accuracy of subsequent processing.

[0064] Step S20 realizes the objective quantitative identification of the emotional distortion state through the coordination of the setting of the standard elliptical area, the calculation of the deviation value, the analysis of the comprehensive index and the verification of time continuity. The standard elliptical area provides a benchmark framework for distortion judgment based on a large amount of historical data, the deviation value calculation converts the spatial distribution characteristics into numerical indicators, the combined analysis of the difference and the average value distinguishes physiological distortion from environmental noise, and the continuous time window verification filters out instantaneous interference. The quantitative index of step S20 converts the spatial features constructed in step S10 into a determinate state label. The combination of step S10 and step S20 upgrades the distortion identification from qualitative description to quantitative analysis. If step S20 is missing, the energy distribution points generated by step S10 will not be able to be converted into clear distortion state labels, resulting in the loss of the premise of the quadrant attribution analysis of step S30. The entire processing flow will not be able to determine whether distortion restoration is required, and ultimately the noise reduction processing of extreme emotional speech will still remain at the traditional indifferent suppression level.

[0065] Step S30, dividing the four-quadrant distortion identification area in the dual-source frequency domain coordinate system, determining the quadrants to which the two groups of energy distribution points corresponding to the two groups of human voice synchronization signals identified as being in an emotional distortion state belong, and generating corresponding distortion type labels;

[0066] Furthermore, step S30 includes:

[0067] Step S31, see Figure 2 As shown, in the dual-source frequency domain coordinate system, the origin O is used as the center to divide the area into four quadrants. The first quadrant is defined as the frequency domain compression distortion area, the second quadrant is defined as the mixed distortion area, the third quadrant is defined as the trembling distortion area, and the fourth quadrant is defined as the normal area.

[0068] Step S32, counting the distribution ratios of the first group of energy distribution points and the second group of energy distribution points corresponding to the two groups of human voice synchronization signals identified as being in an emotionally distorted state in four quadrants;

[0069] Step S33, automatically generating corresponding distortion type labels according to the distribution ratio of the first group of energy distribution points and the second group of energy distribution points in the four quadrants;

[0070] Specifically, step S30 aims to achieve accurate classification of different types of composite distortions through geometric partitioning of spatial quadrants, and solve the problem of insufficient targeted processing caused by the confusion of composite distortion features by traditional algorithms. Step S31 divides the dual-source frequency domain coordinate system into four quadrants with the origin O as the center. The first quadrant is defined as the frequency domain compression distortion area, the second quadrant is defined as the mixed distortion area, the third quadrant is defined as the trembling distortion area, and the fourth quadrant is defined as the normal area. The specific basis for the quadrant division is: the positive direction of the X-axis corresponds to the frequency range higher than the center of the standard ellipse, and the negative direction corresponds to the frequency range lower than the center of the standard ellipse; the positive direction of the Y-axis corresponds to the amplitude range higher than the center of the standard ellipse, and the negative direction corresponds to the amplitude range lower than the center of the standard ellipse. Frequency domain compression distortion manifests as a narrowing of the frequency range and an abnormal increase in high-frequency amplitude, with energy distribution points concentrated in the first quadrant bounded by the positive x- and y-axes. Periodic frequency shifts caused by emotional tremor manifest as low-frequency fluctuations and unstable amplitudes, with energy distribution points concentrated in the third quadrant bounded by the negative x- and y-axes. Mixed distortion, due to the combination of compression and tremor, has points distributed across multiple quadrants, with at least one quadrant in the second quadrant (i.e., the negative x- and positive y-axes), corresponding to a mixed characteristic with increased low-frequency amplitude. Energy distribution points for normal speech are primarily concentrated in the fourth quadrant (i.e., the positive x- and negative y-axes), with both frequency and amplitude within normal fluctuations. A four-quadrant division was chosen over other partitioning methods because the three typical states of extreme emotion exhibit orthogonal characteristics in frequency and amplitude deviations from normal speech: compression, tremor, and mixed. This orthogonality maximizes the ability to distinguish between the spatial distribution differences of different distortion types. Using a sector-shaped or multi-region division would result in blurred classification due to overlapping boundaries.

[0071] Step S32 counts the distribution ratios of the first and second energy distribution points corresponding to the two groups of human voice synchronization signals identified as being in an emotionally distorted state in the four quadrants. The specific implementation process is as follows: for the two groups of human voice synchronization signals marked as being in an emotionally distorted state, extract the corresponding first and second energy distribution points, and determine the quadrant to which the coordinates (f, a) of each energy distribution point belong one by one, where f is the abscissa of the energy distribution point and a is the ordinate of the energy distribution point; when f is greater than the ellipse center frequency and a is greater than the ellipse center amplitude, it is determined to be in the first quadrant; when f is less than the ellipse center frequency and a is greater than the ellipse center amplitude, it is determined to be in the second quadrant; when f is less than the ellipse center frequency and a is less than the ellipse center amplitude, it is determined to be in the third quadrant; when f is greater than the ellipse center frequency and a is less than the ellipse center amplitude, it is determined to be in the fourth quadrant. Count the number of energy distribution points in each quadrant respectively, divide the number of energy distribution points in each quadrant by the total number of energy distribution points in the group, and obtain the distribution ratios of the four quadrants. For example, if the first set of energy distribution points consists of 100 points, 60 of which are in the first quadrant, 20 in the second quadrant, 10 in the third quadrant, and 10 in the fourth quadrant, the distribution ratios for the first, second, and fourth quadrants are 60%, 20%, 10%, and 10%, respectively. We use percentages rather than single points because speech signals under extreme emotions may experience brief, normal fluctuations. The percentage reflects the overall distribution trend and avoids misclassification due to individual outliers. Relying solely on single points can reduce classification stability due to transient fluctuations.

[0072] The method for generating the corresponding distortion type mark in step S33 is to determine the key distribution quadrants of the first group of energy distribution points and the second group of energy distribution points; and generate the distortion type mark according to the key distribution quadrants of the first group of energy distribution points and the second group of energy distribution points. The method for determining the key distribution quadrants of the first group of energy distribution points and the second group of energy distribution points includes: setting a distribution threshold T ratio The distribution threshold is determined by the statistics of historical distortion samples. 300 hours of speech samples are collected, including pure frequency domain compression, pure tremor, and mixed distortion. The proportion of energy distribution points in the corresponding characteristic quadrant in each sample is calculated. The minimum value of the characteristic quadrant ratio among all samples is taken as T ratio , usually ranging from 50% to 70%, ensuring that more than 95% of pure distortion samples can be correctly identified. If the proportion of the first set of energy distribution points falling into the i-th quadrant area to the total number of points exceeds the distribution threshold T ratio , it means that the i-th quadrant is the key distribution quadrant of the first group of energy distribution points; otherwise, it means that the first group of energy distribution points has no key distribution quadrant; if the proportion of the second group of energy distribution points falling into the j-th quadrant to the total number of points exceeds the distribution threshold T ratio, it means that the j-th quadrant area is the key distribution quadrant of the second group of energy distribution points; otherwise, it means that the second group of energy distribution points has no key distribution quadrant; where i and j represent the quadrant numbers, i = 1, 2, 3, 4, j = 1, 2, 3, 4.

[0073] The method for generating a distortion type label based on the focused distribution quadrants of the first and second energy distribution points includes: if the focused distribution quadrants of the first and second energy distribution points are both in the first quadrant, generating a frequency domain compression distortion label; if the focused distribution quadrants of the first and second energy distribution points are both in the third quadrant, generating a tremor distortion label; if the focused distribution quadrants of the first and second energy distribution points are different, and at least one of the focused distribution quadrants is in the second quadrant, generating a mixed distortion label; if neither the first and second energy distribution points have a focused distribution quadrant, or if the focused distribution quadrants of the first and second energy distribution points are both in the fourth quadrant, generating a normal speech label. If only one group has a focused distribution quadrant and the other group does not, calculating the difference between the characteristic quadrant ratio of the group with the focused distribution quadrant and the highest ratio of all quadrants in the other group, and if the difference exceeds 20%, the type of the group with the focused distribution quadrant prevails; otherwise, labeling it as mixed distortion.

[0074] For example, in a certain emotional distortion signal, if the first group of energy distribution points has a 65% distribution ratio in the first quadrant and the second group has a 70% distribution ratio, and the Tratio is set to 60%, both groups will be focused in the first quadrant, generating a frequency domain compression distortion mark. In another signal, if the first group has a 62% distribution ratio in the third quadrant and the second group has a 58% distribution ratio in the third quadrant, and the Tratio is set to 55%, a tremor distortion mark will be generated. If the first group focuses on the first quadrant and the second group focuses on the second quadrant, a mixed distortion mark will be generated. This judgment mechanism verifies the distribution consistency of the two groups of signals, reducing misjudgments caused by noise interference from a single signal. If the distribution of only a single group of signals is relied upon, the type of the signal may be misjudged due to local distortion in that group.

[0075] Step S30 achieves accurate classification from emotional distortion state to specific distortion type through the coordination of four-quadrant partitioning, distribution ratio statistics, and multi-condition judgment. Among them, the four-quadrant partitioning provides spatial feature templates for different distortion types, the distribution ratio statistics quantifies the overall distribution trend, and the multi-condition judgment combines the consistency of the two groups of signals to improve the classification reliability. Step S30 cooperates with the energy distribution points generated by step S10 and the emotional distortion state identified by step S20. The spatial mapping of step S10 provides a specific analysis object for the quadrant partitioning, and the state identification of step S20 limits the classification range of the emotional distortion signal. The combination of the three forms a progressive analysis chain of "feature mapping-state identification-type subdivision". If step S30 is missing, step S40 will not be able to determine the specific distortion type, and can only use a general processing method, resulting in a decrease in the processing effect of mixed distortion. For example, when compression and tremor are present at the same time, a single processing method will lose sight of one and fail to take into account the repair needs of both distortions. Step S30 provides precise type guidance for subsequent processing, enabling dynamic adaptation of processing strategies according to different distortion characteristics, thus avoiding the limitations of "one-size-fits-all" processing.

[0076] Step S40: Segment the first set of frequency domain signals FreA and the second set of frequency domain signals FreB according to the generated distortion type labels to generate different types of frequency domain sub-signals, and perform corresponding frequency domain reconstruction and correction on the different types of frequency domain sub-signals to obtain corrected first set of frequency domain signals Fre"A and second set of frequency domain signals Fre"B.

[0077] Step S40 aims to achieve accurate restoration of extreme emotional speech signals by targeted processing of frequency domain features of different distortion types, and to solve the problem of key information loss or increased distortion caused by the traditional noise reduction algorithm using a single processing strategy for all signals.

[0078] See also Figure 3 As shown, further, step S40 includes:

[0079] Step S41: Segment the first set of frequency domain signals FreA and the second set of frequency domain signals FreB according to the generated distortion type labels to generate different types of frequency domain sub-signals, and perform corresponding frequency domain reconstruction processing on the different types of frequency domain sub-signals to obtain processed first set of frequency domain signals Fre'A and second set of frequency domain signals Fre'B.

[0080] Specifically, based on the generated distortion type labels, the first frequency domain signal FreA and the second frequency domain signal FreB are segmented to generate different types of frequency domain sub-signals. These include: if the distortion type is labeled as frequency domain compression distortion, a compression distortion sub-signal is generated; if the distortion type is labeled as trembling distortion, a trembling distortion sub-signal is generated; if the distortion type is labeled as mixed distortion, a mixed distortion sub-signal is generated; and if the distortion type is labeled as normal speech, a normal sub-signal is generated. This targeted segmentation and processing approach stems from the fundamental differences in the frequency domain characteristics of different distortion types: frequency domain compression distortion manifests as concentrated energy in a specific high-frequency band, trembling distortion manifests as periodic frequency fluctuations in a low-frequency band, mixed distortion exhibits both characteristics, and normal speech maintains a stable distribution across the entire frequency band. Traditional algorithms apply a uniform filtering or enhancement strategy to all frequency bands, resulting in oversuppression of key compression distortion bands and misclassification of trembling distortion fluctuations as noise. However, sub-signal segmentation guided by the type label ensures that processing is limited to the distortion components, preventing interference with normal signals.

[0081] The method for generating a compression distortion sub-signal includes: counting the frequency range FA1 = [FA1, min, FA1, max] of the first group of energy distribution points in the first quadrant, and the frequency range FB1 = [FB1, min, FB1, max] of the second group of energy distribution points in the first quadrant; taking the intersection frequency range of FA1 and FB1 as the compression distortion frequency range F comp ; Extract the frequency range F from the first set of frequency domain signals FreA comp The corresponding part is used as the first group of compression distortion sub-signals; the frequency range F is extracted from the second group of frequency domain signals FreB comp The corresponding part is used as the second group of compression distortion sub-signals. The method for generating the trembling distortion sub-signal is as follows: the frequency range FA3 of the first group of energy distribution points in the third quadrant, that is, the interval formed by the minimum and maximum frequencies of all points in the quadrant, and the frequency range FB3 of the second group of energy distribution points in the third quadrant are counted, and the intersection of the two is taken as the trembling distortion frequency range F trem In the specific implementation, the frequency values ​​of each point in the third quadrant are sorted, the minimum value is taken as the lower limit of the interval, and the maximum value is taken as the upper limit of the interval. If a group of signals has no point distribution in the third quadrant, the range of the other group is used as the benchmark. Extract the frequency range F from the first group of frequency domain signals FreA trem The corresponding part is taken as the first group of jitter distortion sub-signals, and the corresponding part is extracted from the second group of frequency domain signals FreB as the second group of jitter distortion sub-signals. For example, the frequency range of the first group in the third quadrant is 100Hz to 300Hz, and the frequency range of the second group is 150Hz to 350Hz. The intersection of 150Hz to 300Hz is taken as F tremThe signal components within this range are extracted as the tremor distortion sub-signal. This quadrant-based range determination is directly related to the correspondence between the third quadrant and the tremor distortion in step S30, ensuring that the sub-signal contains only the fluctuation components caused by emotional tremor.

[0082] The method for generating the mixed distorted sub-signal is as follows: Count the frequency range FA1 of the first quadrant and the frequency range FA3 of the third quadrant of the first group of energy distribution points, and take the union of the two as the first group of mixed frequency range FA mix ; Count the frequency range FB1 of the first quadrant and the frequency range FB3 of the third quadrant of the second group of energy distribution points, and take the union of the two as the second group of mixed frequency range FB mix ; Then take FA mix With Facebook mix The intersection of the mixed distortion frequency range F mix Extract the parts within the range from the two sets of frequency domain signals as mixed distortion sub-signals. For example, the first quadrant of the first group ranges from 500Hz to 1000Hz, and the third quadrant ranges from 100Hz to 300Hz. mix The first quadrant of the second group is 600Hz to 1100Hz, the third quadrant is 150Hz to 250Hz, and the FB mix 150Hz to 1100Hz; intersection F mix The frequency range of the mixed distortion sub-signal is 150Hz to 1000Hz. This method ensures that both components of the mixed distortion are included in the processing range by covering the frequency band corresponding to compression and tremor.

[0083] The method for generating a normal sub-signal is: counting the frequency range FA4 of the first group of energy distribution points in the fourth quadrant and the frequency range FB4 of the second group of energy distribution points in the fourth quadrant, and taking the union of the two as the normal frequency range F norm Extract the part within the range from each of the two frequency domain signals as the normal sub-signal. If one group of signals has no point distribution in the fourth quadrant, the range of the other group is used as the benchmark; if both groups have no fourth quadrant points, the normal sub-signal is determined to be empty. For example, the fourth quadrant range of the first group is 300Hz to 500Hz, and the fourth quadrant range of the second group is 250Hz to 600Hz, F norm The frequency range is 250Hz to 600Hz, and the signal within this range is extracted as the normal sub-signal. The fourth quadrant is selected because the signal components corresponding to this area are within the normal fluctuation range in frequency and amplitude, and do not require complex processing.

[0084] The corresponding frequency domain reconstruction processing is performed for different types of frequency domain sub-signals, including the frequency expansion mapping method for compression distortion sub-signals, the center locking stabilization method for trembling distortion sub-signals, the layered collaborative processing method for mixed distortion sub-signals, and the amplitude fine-tuning method for normal sub-signals. The frequency expansion mapping method is used for compression distortion sub-signals, specifically to calculate the center frequency Center of the compression distortion sub-signal. Freq , Center Freq is the compression distortion frequency range F comp The midpoint of (F comp Minimum value of +F comp The maximum value of the value) is divided by 2; the preset expansion ratio Ratio R , Ratio R The ratio is determined by historical compression distortion sample experiments. 500 hours of frequency domain compression distortion speech is collected. The ratio of the frequency range of the corresponding speech in the normal state to the frequency range after compression is calculated. The average of all ratios is taken as the Ratio R , usually ranging from 1.5 to 2.5, to ensure that the frequency range after expansion is close to the natural distribution of normal speech; Center Freq As a benchmark, it will be lower than Center Freq The frequency component of the center frequency component expands toward the low frequency direction, and the expansion amplitude is Center Freq With F comp The difference between the minimum values ​​is multiplied by (Ratio R -1), will be higher than Center Freq The frequency component of the scalar expands toward the high frequency direction, and the expansion amplitude is F comp Maximum and Center Freq The difference is multiplied by (Ratio R -1), redistributing the compressed frequency components across a wider frequency band. This method exploits the reversibility of frequency domain compression distortion—vocal cord tension during extreme emotion only narrows the frequency range, not the components themselves. Symmetrical expansion restores their original distribution. Compared to traditional amplitude enhancement methods, this method can restore frequency richness without introducing noise.

[0085] The center-locked stabilization method is used for the tremor distortion sub-signal. Specifically, the frequency variation pattern of the tremor distortion sub-signal on the time axis is analyzed. The average frequency in each window is calculated by sliding the window. The window length involved here is consistent with the time window in step S24. The average frequencies of all windows are linearly fitted to obtain the trend line of frequency variation. The midpoint of the trend line is taken as the center frequency of the fluctuation, that is, the anchor point position Anchor. Pos ; For each frequency component in the time window, calculate its relationship with Anchor PosThe deviation value is subtracted from the original frequency to realign all frequency components to the Anchor Pos , eliminating periodic fluctuations. We chose linear fitting rather than arithmetic averaging to calculate the center frequency because the frequency fluctuations of emotional tremors often show a trend of gradual strengthening or weakening. Linear fitting can more accurately reflect the overall center of fluctuations. If simple averaging is used, extreme fluctuations will cause the anchor point to shift, and it will not be able to completely eliminate instability.

[0086] The mixed distorted sub-signals are processed by layered collaborative processing. Specifically, the mixed distorted sub-signals are split into high frequency band and low frequency band according to the frequency range. The high frequency band is the frequency higher than the anchor. Pos The mixed distorted sub-signal, the low frequency band is lower than or equal to the Anchor Pos The mixed distorted sub-signal uses the frequency expansion mapping method in the high frequency band, and the parameters are the same as those for the compression distortion processing. The center locking stabilization method is used in the low frequency band, and the parameters are the same as those for the trembling distortion processing. After processing, the two signals are placed in the Anchor Pos The amplitudes at the splicing points are smoothly spliced, and a linear transition is used to ensure smooth amplitude transitions. This layered processing is based on the physical properties of mixed distortion. Speech under extreme emotion often exhibits a coexistence of high-frequency compression and low-frequency tremors. Layered processing optimizes both characteristics separately. Using a single method would result in reduced processing effectiveness due to the mutual interference between high-frequency compression and low-frequency fluctuations. An amplitude fine-tuning method is applied to the normal sub-signal. Specifically, the ratio of the normal sub-signal's amplitude to the average amplitude of the corresponding frequency within the standard elliptical region is calculated. Lower and upper ratio thresholds, along with corresponding adjustment coefficients, are set. These thresholds are determined based on historical normal speech sample statistics to cover the fluctuation range of most normal samples. For example, the lower ratio threshold is set to 0.8, and the upper ratio threshold is set to 1.2. If the ratio is below the lower ratio threshold, the amplitude is multiplied by the lower adjustment coefficient, which does not exceed the amplitude corresponding to the upper ratio threshold. If the ratio is above the upper ratio threshold, the amplitude is multiplied by the upper adjustment coefficient, which does not exceed the amplitude corresponding to the lower ratio threshold. The intermediate range remains unchanged. This fine-tuning avoids over-processing of normal signals by traditional algorithms, compensates for amplitude attenuation caused by slight ambient noise, and ensures natural speech.

[0087] The first set of processed frequency domain signals, Fre'A, and the second set of processed frequency domain signals, Fre'B, are generated by placing each processed sub-signal back into the original frequency domain signal within its original frequency range. For blank frequency bands (i.e., portions not covered by any sub-signal), the amplitude of the original signal is retained but multiplied by 0.5 to treat them as weak noise or ineffective components. For overlapping frequency bands (i.e., portions belonging to multiple sub-signals), the average of the processed amplitudes of each sub-signal is taken as the final amplitude to ensure complete and conflict-free frequency band coverage. For example, in the first set of frequency domain signals, the sub-signals from 150Hz to 300Hz are the dithered sub-signals, the compressed sub-signals from 500Hz to 1000Hz, and the normal sub-signals from 300Hz to 500Hz are spliced ​​in this order after processing to form the complete Fre'A.

[0088] Step S41 achieves accurate repair of different distortion types through type-guided sub-signal segmentation, physical property adaptation processing methods, and multi-band collaborative integration. In coordination with the distortion type labeling of step S30, it ensures that the processing method strictly matches the distortion characteristics, avoiding the loss of focus caused by traditional single processing; in coordination with the dual-source frequency domain coordinate system of step S10, the frequency range of each sub-signal is determined based on the quadrant distribution in the coordinate system to ensure the accuracy of segmentation. The independent processing of each frequency domain sub-signal enables subsequent error correction to adjust parameters for specific distortion types without affecting other processing, significantly improving the correction efficiency. If this step is missing, subsequent error correction cannot be effectively performed due to the lack of targeted processing objects, resulting in the key information of extreme emotional speech still being masked or over-suppressed by noise.

[0089] Step S42, correcting the processed first set of frequency domain signals Fre'A and the second set of frequency domain signals Fre'B: marking processing error points in the processed first set of frequency domain signals Fre'A and the second set of frequency domain signals Fre'B, adjusting relevant parameters of frequency domain reconstruction, and obtaining corrected first set of frequency domain signals Fre"A and second set of frequency domain signals Fre"B;

[0090] Furthermore, step S42 includes:

[0091] Step S421, respectively calculating the phase difference and amplitude ratio of the processed first set of frequency domain signals Fre'A and the second set of frequency domain signals Fre'B at the same frequency point;

[0092] Step S422: Setting a normal phase difference range and a normal amplitude ratio range. When the phase difference value exceeds the normal phase difference range, or the amplitude ratio value exceeds the normal amplitude ratio range, marking the corresponding frequency point as a processing error point.

[0093] Step S423 : adjusting the relevant parameters of frequency domain reconstruction for the frequency points marked as processing error points.

[0094] Specifically, step S42 aims to identify and correct the deviations introduced in the frequency domain reconstruction process by verifying the consistency of the two groups of signals, and solve the problem of distortion amplification that is difficult to detect in single signal processing. Step S421 calculates the phase difference and amplitude ratio of the processed first group of frequency domain signals Fre'A and the second group of frequency domain signals Fre'B at the same frequency point. The phase difference is the phase value of Fre'A minus the phase value of Fre'B at the same frequency point, and the amplitude ratio is the amplitude of Fre'A divided by the amplitude of Fre'B at the same frequency point. Phase and amplitude are the key correlation features of the two groups of synchronously collected signals. The phase difference of the signals collected by the same sound source at two spatially separated points is determined by the propagation distance difference and is stable as a constant. The amplitude ratio is determined by the distance attenuation characteristics and is stable within a fixed range. If the processed signal deviates from this inherent relationship, it means that the processing process has introduced errors. Step S422 sets the normal phase difference range and normal amplitude ratio range. The normal phase difference range is determined by statistically analyzing the phase differences of two groups of signals from 1000 hours of normal speech samples. The arithmetic mean of the phase differences at each frequency point is taken, and the standard deviation of the phase differences across all frequency points is calculated. The normal range is calculated by adding or subtracting three times the standard deviation from the mean. The normal amplitude ratio range is determined using the same method, analyzing the amplitude ratio distribution of normal samples and using a 95% confidence interval as the range. When the phase difference value or amplitude ratio value at a frequency point exceeds the normal phase difference range, the corresponding frequency point is marked as a processing error point. The choice of three times the standard deviation and a 95% confidence interval is based on the fact that the phase difference and amplitude ratio of two groups of signals for normal speech follow a normal distribution. This range covers the vast majority of normal fluctuations and reduces false positives caused by random noise. If the range is too narrow, invalid error points will be added, while if it is too wide, true errors will not be identified.

[0095] Step S423 adjusts the parameters related to frequency domain reconstruction for the frequency points marked as processing error points. The adjustment follows the principle of "error type-parameter association": for processing error points of compression distortion, the error is caused by the phase asynchrony of the two groups of signals due to excessive or insufficient frequency expansion. In this case, reduce or increase the expansion ratio Ratio R , each adjustment amplitude is 10% of the initial value, and the phase difference and amplitude ratio are calculated after re-execution of frequency expansion until the point returns to the normal range; for the processing error point of tremor distortion, the error comes from the anchor point position Anchor Pos The amplitude ratio is abnormal due to calculation deviation. In this case, the number of sliding windows is increased from 3 to 5, the frequency trend line is refitted, and a new anchor is determined. PosAfter re-performing frequency alignment, verify the performance until it meets the target. For errors in mixed distortion processing, which often arise from improper segmentation of the high and low frequency bands, resulting in a sudden phase shift at the splice, the segmentation threshold is shifted toward the frequency band where the error occurs, by 50Hz at a time. The layering process is then re-performed and a smooth transition is performed until the deviation is eliminated. The magnitude and method of parameter adjustment are determined based on historical error correction data to ensure that each adjustment significantly reduces the deviation without introducing new fluctuations.

[0096] Step S42 achieves precise correction of frequency domain reconstruction errors through dual-dimensional verification of phase difference and amplitude ratio and dynamic parameter adjustment. In coordination with the targeted processing of step S41, the sub-signal processing of S41 provides a specific object for error identification, and the correction of S42 feeds back to optimize the processing parameters of S41, forming a closed loop of "processing-verification-optimization"; the unified frequency reference provided by the dual-source frequency domain coordinate system of step S10 ensures the comparability of phase difference and amplitude ratio calculations, avoiding misjudgment of errors caused by inconsistent frequency axes. The distribution characteristics of the error points can be used to reversely locate the weak links of the previous processing. For example, if compression distortion error points continue to appear in a certain frequency band, it means that the expansion ratio of the frequency band needs to be optimized in a targeted manner, so that the processing parameters are upgraded from fixed values ​​to variables that dynamically adapt to different frequency bands, which significantly improves the robustness of processing complex composite distortions. If step S42 is omitted, the processing errors in step S41 will accumulate and be passed to subsequent steps, causing the frequency band enhancement in step S50 to mistakenly identify the error components as valid signals. The final output speech will still contain distortion or residual noise. At the same time, the consistency of the two sets of signals cannot be guaranteed, and subsequent applications based on the fusion of the two sets of signals, such as sound source localization, will fail due to signal deviation. Step S42 provides a quality control mechanism for the entire processing flow, ensuring that the physical characteristics of the pre-processed signals remain consistent, laying the foundation for the accuracy of the final speech restoration.

[0097] Step S50 , dividing the corrected first group of frequency domain signals Fre”A and the second group of frequency domain signals Fre”B into frequency bands, identifying frequency bands with significantly reduced energy, and performing enhancement processing on the frequency bands with significantly reduced energy.

[0098] Furthermore, step S50 includes:

[0099] Step S51, dividing the corrected first group of frequency domain signals Fre'A and the second group of frequency domain signals Fre'B into frequency bands;

[0100] Step S52: analyzing the energy loss value of the corrected first set of frequency domain signals Fre'A and the second set of frequency domain signals Fre'B on a frequency band by frequency band basis;

[0101] Calculate the first set of original energy values ​​Ori before each frequency band processing of the first set of frequency domain signals Fre'A and the second set of frequency domain signals Fre'B respectively A、The second set of original energy value Ori B And the first set of current energy values ​​Cur after processing A 、The second set of current energy value Cur B ; by Ori A Subtract Cur A , get the energy loss value Loss of each frequency band of the first set of frequency domain signals Fre”A A , by Ori B Subtract Cur B , get the energy loss value Loss of each frequency band of the second set of frequency domain signals Fre”B B .

[0102] Step S53: Loss A Greater than the first loss threshold Thr A The frequency band of Loss is marked as the first group of frequency bands with significantly reduced energy. B Greater than the second loss threshold Thr B The frequency bands are marked as the second group of frequency bands with significantly reduced energy;

[0103] Step S54: Calculate the first energy ratio Ene of the first group of frequency bands with significantly reduced energy and the second group of frequency bands with significantly reduced energy to the adjacent normal frequency bands. RA and the second set of energy ratios Ene RB ;

[0104] Step S55: Based on the first set of energy ratios Ene RA and the second set of energy ratios Ene RB , generating a first set of gain compensation coefficients and a second set of gain compensation coefficients, applying the first set of gain compensation coefficients to the first set of frequency bands with significantly reduced energy, and applying the second set of gain compensation coefficients to the second set of frequency bands with significantly reduced energy.

[0105] Specifically, step S50 aims to address the noise amplification problem caused by traditional full-band enhancement by specifically enhancing the energy in key frequency bands, compensating for any residual energy attenuation from previous processing. Step S51 then divides the two corrected frequency domain signals into frequency bands. The frequency resolution for the band divisions is determined based on the frequency characteristics of the speech signal, covering the human voice audio range of 20Hz to 20kHz. A value such as 128, 256, or 512 is selected so that each band contains a corresponding number of adjacent frequency points. This range is chosen because key speech information is primarily concentrated between 300Hz and 3400Hz. A higher resolution, such as 256, ensures that subdivisions within this frequency band are sufficient to capture subtle energy variations. Using a lower resolution, such as 128, for the low- and high-frequency bands reduces computational effort, balancing processing efficiency and accuracy. The division is performed by equally dividing the entire frequency range according to the resolution. For example, when the resolution is 256, the width of each band is the total frequency range divided by the number of bands. This ensures that adjacent bands are continuous and non-overlapping, providing independent units for subsequent energy analysis. Step S52 analyzes the energy loss value frequency band by frequency band, and calculates the original energy value before processing and the current energy value after processing for each frequency band of the two groups of signals. The original energy value is the sum of the energy of the frequency domain signal mapped in step S10 in the frequency band, which is obtained by summing the squares of the amplitudes of all frequency points in the frequency band; the current energy value is the sum of the energy of the corrected frequency domain signal in the same frequency band, and the calculation method is the same. The energy loss value is the original energy value minus the current energy value. A positive value indicates that the energy is reduced after processing, and a negative value or zero indicates that the energy has not decreased or increased. This calculation method directly quantifies the impact of the processing process on the energy of each frequency band, avoids misjudgment of energy changes due to subjective judgment, and provides an objective basis for subsequent enhancement.

[0106] Step S53 marks frequency bands with significantly reduced energy. A first loss threshold and a second loss threshold are set. The first loss threshold is determined for the corrected first set of frequency domain signals, Fre"A, and the second loss threshold is determined for the corrected second set of frequency domain signals, Fre"B. Both are independently calculated based on the energy loss distribution of their respective signals. Due to differences in the spatial locations of the two sets of acquisition points, such as distance from the sound source and the degree of environmental interference, the energy loss patterns of the two sets of signals naturally differ, requiring separate settings to ensure accurate marking. The specific process is as follows: 500 hours of the first and second sets of speech samples, corrected in step S42, are collected. For the first set of samples, energy loss values ​​are calculated for all frequency bands to form a first energy loss distribution. This distribution is sorted from smallest to largest, and the value corresponding to the 75th percentile of the sorted samples is taken as the first loss threshold. This means that only 25% of the samples have energy loss values ​​above the first loss threshold, ensuring that only frequency bands with significant energy loss in the first set of signals are marked for enhancement. The second loss threshold is set similarly. When the energy loss value of a frequency band exceeds the corresponding threshold, it is marked as a band with significantly reduced energy. The 75% percentile is chosen because most energy losses after normal processing are concentrated in the lower range. This threshold can filter out slight attenuation and focus on the significant loss frequency band that really affects speech clarity. If the threshold is too low, it will lead to over-enhancement, and if it is too high, the frequency band that needs to be repaired will be missed.

[0107] In step S54, the first energy ratio Ene of the first group of frequency bands with significantly reduced energy and the second group of frequency bands with significantly reduced energy to the adjacent normal frequency bands is calculated respectively. RA and the second set of energy ratios Ene RB For the first group of frequency bands with significantly reduced energy, select the normal frequency bands on both sides of each frequency band with significantly reduced energy as reference frequency bands, calculate the ratio of the current energy value of the frequency band with significantly reduced energy to the average energy value of the reference frequency band, and obtain the energy ratio Ene RA ;Ene RBCalculated using a similar method. The selection of the reference frequency band is based on the continuity of the speech energy distribution. The energy levels of adjacent frequency bands are usually correlated. This serves as a benchmark to ensure that the enhancement amplitude is coordinated with the surrounding energy and avoid abruptness. For example, if the current energy of a significantly reduced frequency band is 10, the average energy of the reference frequency band on the left is 20, and the average energy of the reference frequency band on the right is 25, then the energy ratio is 10 / (20+25) / 2≈0.44, reflecting the degree of energy attenuation of this frequency band relative to the surrounding energy. Step S55 generates a gain compensation coefficient and applies it to the significantly reduced frequency band. The gain compensation coefficient is inversely proportional to the energy ratio. The smaller the ratio, the larger the gain compensation coefficient, ensuring that the frequency band with more severe energy attenuation receives stronger compensation. The maximum value of the gain compensation coefficient is set to the ratio of the average energy value of the reference frequency band to the current energy value of the significantly reduced frequency band to avoid distortion caused by excessive enhancement. At the same time, progressive gain adjustment is adopted. At the edge of the significantly reduced frequency band, the gain compensation coefficient gradually transitions from 1 to the target value. The transition range is 10% of the width of the frequency band to ensure smooth energy changes.

[0108] Step S50 achieves accurate repair of the energy of key frequency bands through fine-grained frequency band division, energy loss quantification, and targeted gain compensation. The corrected signal of step S42 provides high-quality input for frequency band division, ensuring that the enhanced object is a valid signal with real attenuation rather than noise. By comparing the energy ratios of the two groups of signals, the unilateral energy attenuation caused by differences in the acquisition equipment can be identified, and only the signal on that side is targeted for enhancement, making the energy distribution of the two groups of signals more balanced and improving the stability of subsequent signal fusion. If step S50 is missing, the residual energy attenuation in the early processing will cause key information, such as numbers and directional words in emergency scenarios, to be blurred, affecting the accuracy of communication; at the same time, enhancement without transition will cause a sense of discontinuity in the voice and reduce naturalness. As the final optimization link of the entire processing flow, step S50 ensures that the voice signal remains clear and natural after distortion repair through the energy compensation mechanism, providing reliable voice quality assurance for emergency communication scenarios.

[0109] Example 2

[0110] This embodiment provides a frequency domain noise reduction and restoration system for synchronously collecting two groups of human voices based on embodiment 1, such as Figure 4 As shown, including:

[0111] Mapping module: obtains two sets of synchronous human voice signals through two sets of acquisition points, constructs a dual-source frequency domain coordinate system, and maps the two sets of human voice signals into the constructed dual-source frequency domain coordinate system to form two sets of energy distribution points.

[0112] Distortion Identification Module: This module is used to calculate the spatial relationship between two sets of energy distribution points in the dual-source frequency domain coordinate system, obtain an indicator to measure the degree of distortion of the dual-set vocal synchronization signal, and identify whether the current speech state of the dual-set vocal synchronization signal is in an emotionally distorted state.

[0113] Quadrant marking module: used to divide the four-quadrant distortion recognition area in the dual-source frequency domain coordinate system, determine the quadrant affiliation of the two groups of energy distribution points corresponding to the two groups of human voice synchronization signals identified as emotional distortion states, and generate corresponding distortion type labels.

[0114] Frequency domain correction module: According to the generated distortion type mark, the first group of frequency domain signals FreA and the second group of frequency domain signals FreB are divided to generate different types of frequency domain sub-signals, and corresponding frequency domain reconstruction processing and correction are performed on the different types of frequency domain sub-signals to obtain the corrected first group of frequency domain signals Fre”A and the second group of frequency domain signals Fre”B.

[0115] Frequency band enhancement module: used to divide the corrected first group of frequency domain signals Fre'A and the second group of frequency domain signals Fre'B into frequency bands, identify the frequency bands with significantly reduced energy, and enhance the frequency bands with significantly reduced energy.

[0116] In the frequency domain correction module, the method of performing corresponding frequency domain reconstruction processing and correction on different types of frequency domain sub-signals to obtain the corrected first set of frequency domain signals Fre"A and second set of frequency domain signals Fre"B includes:

[0117] Step S41: Segment the first set of frequency domain signals FreA and the second set of frequency domain signals FreB according to the generated distortion type labels to generate different types of frequency domain sub-signals, and perform corresponding frequency domain reconstruction processing on the different types of frequency domain sub-signals to obtain processed first set of frequency domain signals Fre'A and second set of frequency domain signals Fre'B.

[0118] Step S42, correcting the processed first group of frequency domain signals Fre'A and the second group of frequency domain signals Fre'B: marking the processing error points in the processed first group of frequency domain signals Fre'A and the second group of frequency domain signals Fre'B, adjusting the relevant parameters of frequency domain reconstruction, and obtaining the corrected first group of frequency domain signals Fre"A and the second group of frequency domain signals Fre"B.

[0119] Furthermore, step S42 includes:

[0120] Step S421, respectively calculating the phase difference and amplitude ratio of the processed first set of frequency domain signals Fre'A and the second set of frequency domain signals Fre'B at the same frequency point;

[0121] Step S422: Setting a normal phase difference range and a normal amplitude ratio range. When the phase difference value exceeds the normal phase difference range, or the amplitude ratio value exceeds the normal amplitude ratio range, marking the corresponding frequency point as a processing error point.

[0122] Step S423 : adjusting the relevant parameters of frequency domain reconstruction for the frequency points marked as processing error points.

[0123] In the frequency band enhancement module, the method of identifying the frequency band with significantly reduced energy and performing enhancement processing on the frequency band with significantly reduced energy includes:

[0124] Step S51, dividing the corrected first group of frequency domain signals Fre'A and the second group of frequency domain signals Fre'B into frequency bands;

[0125] Step S52: analyzing the energy loss value of the corrected first set of frequency domain signals Fre'A and the second set of frequency domain signals Fre'B on a frequency band by frequency band basis;

[0126] Step S53: Loss A Greater than the first loss threshold Thr A The frequency band of Loss is marked as the first group of frequency bands with significantly reduced energy. B Greater than the second loss threshold Thr B The frequency bands are marked as the second group of frequency bands with significantly reduced energy;

[0127] Step S54: Calculate the first energy ratio Ene of the first group of frequency bands with significantly reduced energy and the second group of frequency bands with significantly reduced energy to the adjacent normal frequency bands. RA and the second set of energy ratios Ene RB ;

[0128] Step S55: Based on the first set of energy ratios Ene RA and the second set of energy ratios Ene RB , generating a first set of gain compensation coefficients and a second set of gain compensation coefficients, applying the first set of gain compensation coefficients to the first set of frequency bands with significantly reduced energy, and applying the second set of gain compensation coefficients to the second set of frequency bands with significantly reduced energy.

[0129] The methods and systems of the present application may be implemented in many ways. For example, the methods and systems of the present application may be implemented using software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps used in the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified.

[0130] In addition, the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0131] The above-described specific embodiments further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices, characterized in that: The method comprises: Acquire two sets of synchronous human voice signals through two sets of acquisition points, construct a dual-source frequency domain coordinate system, and map the two sets of human voice signals into the constructed dual-source frequency domain coordinate system to form two sets of energy distribution points; In the dual-source frequency domain coordinate system, the spatial relationship between the two sets of energy distribution points is calculated to obtain an index to measure the degree of distortion of the dual-set vocal synchronization signal, and to identify whether the current speech state of the dual-set vocal synchronization signal is in an emotional distortion state; Divide the distortion recognition area into four quadrants in the dual-source frequency domain coordinate system, determine the quadrants to which the two groups of energy distribution points corresponding to the two groups of human voice synchronization signals identified as being in an emotional distortion state belong, and generate corresponding distortion type labels; Segmenting the first set of frequency domain signals FreA and the second set of frequency domain signals FreB according to the generated distortion type labels to generate different types of frequency domain sub-signals, performing corresponding frequency domain reconstruction and correction on the different types of frequency domain sub-signals to obtain corrected first set of frequency domain signals Fre"A and second set of frequency domain signals Fre"B; The corrected first group of frequency domain signals Fre'A and the second group of frequency domain signals Fre'B are divided into frequency bands, the frequency bands with significantly reduced energy are identified, and the frequency bands with significantly reduced energy are enhanced.

2. The frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices according to claim 1 is characterized in that: The two groups of human voice synchronization signals include a first group of human voice signals and a second group of human voice signals; The method for forming two groups of energy distribution points includes: performing fast Fourier transform on the first group of human voice signals and the second group of human voice signals respectively to obtain a first group of frequency domain signals and a second group of frequency domain signals, using the amplitude of each frequency component in the first group of frequency domain signals as the Y coordinate and the frequency value as the X coordinate to form the first group of energy distribution points; and using the amplitude of each frequency component in the second group of frequency domain signals as the Y coordinate and the frequency value as the X coordinate to form the second group of energy distribution points.

3. The frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices according to claim 2 is characterized in that: The method of calculating the spatial relationship between the two groups of energy distribution points to obtain an index for measuring the degree of distortion of the two groups of synchronous vocal signals includes: Based on the statistical analysis of historical normal speech data, a standard elliptical area is set in the dual-source frequency domain coordinate system; the shortest distances from the two groups of energy distribution points to the boundary of the standard elliptical area are calculated respectively, recorded as the first deviation value and the second deviation value; the difference Diff between the first deviation value and the second deviation value is calculated AB and the mean value AB , the difference Diff AB and mean AB As an indicator to measure the degree of distortion of two groups of vocal synchronization signals.

4. The frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices according to claim 3 is characterized in that: The method for identifying whether the voice state of the current two groups of human voice synchronization signals is an emotional distortion state includes: Set the number of continuous time windows Cou C , difference threshold Thr D and average value threshold Thr M ; When continuous Cou C The difference Diff within the time window AB Both are greater than the difference threshold Thr D And the average value AB Both are greater than the average threshold Thr M , the speech state of the two groups of voice synchronization signals is marked as the emotion distortion state.

5. The frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices according to claim 4 is characterized in that: The method for generating the distortion type mark includes: In the dual-source frequency domain coordinate system, the origin O is used as the center to divide the image into four quadrants. The first quadrant is defined as the frequency domain compression distortion area, the second quadrant is defined as the mixed distortion area, the third quadrant is defined as the trembling distortion area, and the fourth quadrant is defined as the normal area. Statistically determining the distribution ratios of the first group of energy distribution points and the second group of energy distribution points in four quadrants corresponding to the two groups of synchronized vocal signals identified as being in an emotionally distorted state; According to the distribution ratio of the first group of energy distribution points and the second group of energy distribution points in the four quadrants, corresponding distortion type marks are automatically generated.

6. The frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices according to claim 5, characterized in that: The method for automatically generating corresponding distortion type marks according to the distribution ratio of the first group of energy distribution points and the second group of energy distribution points in the four quadrants includes: Determine the key distribution quadrants of the first group of energy distribution points and the second group of energy distribution points; generate distortion type marks based on the key distribution quadrants of the first group of energy distribution points and the second group of energy distribution points, wherein the distortion type marks include frequency domain compression distortion marks, tremor distortion marks, mixed distortion marks and normal speech marks.

7. The frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices according to claim 6, characterized in that: The method for determining the key distribution quadrants of the first group of energy distribution points and the second group of energy distribution points includes: setting a distribution threshold, if the ratio of the number of points of the first group of energy distribution points falling into the i-th quadrant area to the total number of points exceeds the distribution threshold, it indicates that the i-th quadrant area is the key distribution quadrant of the first group of energy distribution points; otherwise, it indicates that the first group of energy distribution points has no key distribution quadrant; if the ratio of the number of points of the second group of energy distribution points falling into the j-th quadrant area to the total number of points exceeds the distribution threshold T ratio , it means that the j-th quadrant area is the key distribution quadrant of the second group of energy distribution points; otherwise, it means that the second group of energy distribution points has no key distribution quadrant; where i and j represent the quadrant numbers, i = 1, 2, 3, 4, j = 1, 2, 3, 4.

8. The frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices according to claim 7, characterized in that: The first group of frequency domain signals FreA and the second group of frequency domain signals FreB are segmented according to the generated distortion type mark to generate different types of frequency domain sub-signals, including: if the distortion type mark is a frequency domain compression distortion mark, a compression distortion sub-signal is generated; if the distortion type mark is a tremor distortion mark, a tremor distortion sub-signal is generated; if the distortion type mark is a mixed distortion mark, a mixed distortion sub-signal is generated; if the distortion type mark is a normal speech mark, a normal sub-signal is generated.

9. The frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices according to claim 8, characterized in that: The method for obtaining the corrected first set of frequency domain signals Fre'A and second set of frequency domain signals Fre'B includes: Performing corresponding frequency domain reconstruction processing on different types of frequency domain sub-signals to obtain processed first group of frequency domain signals Fre'A and second group of frequency domain signals Fre'B; correcting the processed first group of frequency domain signals Fre'A and second group of frequency domain signals Fre'B to obtain corrected first group of frequency domain signals Fre"A and second group of frequency domain signals Fre"B; The method for correcting the processed first set of frequency domain signals Fre'A and the second set of frequency domain signals Fre'B includes: marking processing error points in the processed first set of frequency domain signals Fre'A and the second set of frequency domain signals Fre'B, and adjusting relevant parameters of frequency domain reconstruction for the frequency points marked as processing error points.

10. A frequency domain noise reduction and restoration system for synchronously collecting two groups of human voices, which is used to implement the frequency domain noise reduction and restoration method for synchronously collecting two groups of human voices according to any one of claims 1 to 9, characterized in that: The system comprises: Mapping module: obtains two sets of synchronous human voice signals through two sets of acquisition points, constructs a dual-source frequency domain coordinate system, and maps the two sets of human voice signals into the constructed dual-source frequency domain coordinate system to form two sets of energy distribution points; Distortion Identification Module: This module calculates the spatial relationship between two sets of energy distribution points in the dual-source frequency domain coordinate system, obtains an indicator to measure the degree of distortion of the dual-set vocal synchronization signal, and identifies whether the current speech state of the dual-set vocal synchronization signal is in an emotional distortion state. Quadrant marking module: used to divide the four-quadrant distortion recognition area in the dual-source frequency domain coordinate system, determine the quadrant belonging of the two groups of energy distribution points corresponding to the two groups of human voice synchronization signals identified as emotional distortion states, and generate corresponding distortion type labels; Frequency domain correction module: Based on the generated distortion type mark, the first set of frequency domain signals FreA and the second set of frequency domain signals FreB are segmented to generate different types of frequency domain sub-signals. The corresponding frequency domain reconstruction and correction are performed on the different types of frequency domain sub-signals to obtain the corrected first set of frequency domain signals Fre"A and second set of frequency domain signals Fre"B. Frequency band enhancement module: used to divide the corrected first group of frequency domain signals Fre'A and the second group of frequency domain signals Fre'B into frequency bands, identify the frequency bands with significantly reduced energy, and enhance the frequency bands with significantly reduced energy.

Images