A voice noise reduction method and system for emergency rescue communication equipment

Abstract

The application relates to the technical field of voice noise reduction, in particular to a voice noise reduction method and system of an emergency rescue communication device, which comprises the following steps: collecting audio signals in an emergency rescue communication environment through the emergency rescue communication device; analyzing the energy distribution difference of each frame of audio signals and adjacent frames of audio signals in a frequency domain, determining a burst non-stationary noise coefficient of each frame of audio signals; extracting the periodicity characteristics of peak signals in each frame of audio signals in a time domain, obtaining the amplitude fundamental frequency periodicity of each frame of audio signals, and obtaining a continuous periodic noise coefficient of each frame of audio signals; using the burst non-stationary noise coefficient and the continuous periodic noise coefficient to evaluate a signal retention factor of each frame of audio signals, which is used for correcting the forgetting gate calculation in an LSTM neural network containing gated learning, and using the corrected neural network to perform voice noise reduction. Therefore, the voice noise reduction processing effect of the emergency rescue communication device is improved.

Description

Technical Field

[0001] This application relates to the field of speech noise reduction technology, specifically to a speech noise reduction method and system for emergency rescue communication equipment. Background Technology

[0002] Emergency rescue communications are the "nerve center" of a modern emergency command system. Their quality directly impacts the success or failure of missions and the safety of rescue personnel. However, the harsh environments at rescue sites are filled with various noisy sources, placing higher demands on voice noise reduction technology. Advanced emergency rescue communication equipment must be able to filter out sudden, unstable background noise at the rescue site, effectively highlighting the voice information of rescue personnel, ensuring that the command center can clearly receive feedback from frontline rescuers, and enhancing communication reliability in extreme rescue environments.

[0003] This paper proposes a recurrent neural network (RNN) with LSTM gated learning to implement voice denoising for emergency rescue communication equipment. This effectively captures the temporal dependencies of voice signals and greatly alleviates the gradient vanishing or exploding problem that occurs in traditional RNNs for long-duration voice signals in emergency rescue scenarios. According to existing technology, while the forget gate calculation in the traditional LSTM neural network with gated learning provides flexibility in retaining the degree of information for each segment, it still suffers from insufficient adaptability to dynamic noise in high-real-time, sudden scenarios in emergency rescue environments. Specifically, it cannot quickly cut off the transmission of contaminated historical states when there is sudden strong noise, while excessively retaining voice features from continuous background noise, thus reducing the voice denoising effect of emergency rescue communication equipment and threatening the quality of voice interaction in high-risk emergency rescue environments. Summary of the Invention

[0004] To address the aforementioned technical problems, the purpose of this application is to provide a voice noise reduction method and system for emergency rescue communication equipment. The specific technical solution adopted is as follows: In a first aspect, embodiments of this application provide a voice noise reduction method for emergency rescue communication equipment, the method comprising the following steps: Audio signals are collected in the emergency rescue communication environment using emergency rescue communication equipment; The frequency domain transformation of each frame audio signal is performed to analyze the difference in energy distribution between each frame audio signal and its adjacent frame audio signals in the frequency domain, as well as the energy distribution of harmonic frequencies of each frame audio signal in the frequency domain. The fluctuation degree of amplitude of each frame audio signal in the time domain and the autocorrelation of amplitude in a short time are also analyzed to determine the burst non-stationary noise coefficient of each frame audio signal. The periodicity of the peak signal in each frame of audio signal in the time domain and the energy variation trend of the low frequency component of each frame of audio signal in the frequency domain are extracted to obtain the amplitude fundamental frequency periodicity of each frame of audio signal. Combined with the degree of energy enhancement of each frame of audio signal in the low frequency relative to the historical adjacent frame of audio signal, the continuous periodic noise figure of each frame of audio signal is obtained. The signal retention factor of each frame of audio signal is evaluated using the burst non-stationary noise coefficient and the continuous periodic noise coefficient. This is used to correct the forget gate calculation in the LSTM neural network that includes gating learning. The corrected neural network is then used for speech denoising.

[0005] In one embodiment, determining the burst non-stationary noise coefficient of each frame of audio signal includes: Extract the characteristic frequency components of each frame audio signal in the frequency domain, and form an energy sequence by combining all the energies corresponding to the characteristic frequency components; perform forward fusion on the metric distance of the energy sequences of each frame audio signal and all its adjacent frame audio signals, and calculate the degree of dispersion of the energy corresponding to all harmonic frequencies of each frame audio signal in the frequency domain. By combining the results of the forward fusion with the degree of dispersion, the resonant harmonic offset of each frame of audio signal is obtained; The amplitude mutation correlation of each frame audio signal is calculated by using the fluctuation degree and the autocorrelation. Combined with the resonant harmonic offset, the sudden non-stationary noise coefficient is obtained. The sudden non-stationary noise coefficient is positively correlated with both the amplitude mutation correlation and the resonant harmonic offset.

[0006] In one embodiment, calculating the amplitude abrupt change correlation of each frame of audio signal includes: Identify each peak of the audio signal in each frame of the time domain, form an amplitude change sequence by combining the amplitudes of each peak in the audio signal in each frame with the previous peak, determine the correlation coefficient of each amplitude change sequence, and aggregate and calculate all autocorrelation coefficients of the amplitude change sequences of all peaks in the audio signal in each frame. The numerical difference between the maximum and minimum peak values ​​of each frame of audio signal is measured. The absolute value of the aggregated calculation result is multiplied by the numerical difference and then normalized to obtain the amplitude abrupt change correlation of each frame of audio signal.

[0007] In one embodiment, extracting the characteristic frequency components of each frame of audio signal in the frequency domain includes: Signal processing techniques are used to obtain the frequency components of each frame of audio signal in the frequency domain. The energy of all frequency components is thresholded, and the frequency components corresponding to all energies greater than or equal to the threshold are taken as the characteristic frequency components of each frame of audio signal in the frequency domain.

[0008] In one embodiment, the process for determining the periodicity of the fundamental frequency of the amplitude of each frame of audio signal is as follows: The periodic terms of each peak are obtained by using a sequence decomposition algorithm. The frequencies of all frequency components corresponding to each frame of audio signal in the frequency domain are thresholded and the frequency components corresponding to all frequencies less than the segmentation threshold are recorded as low frequency components. Linear fitting is performed on all energies between the maximum and minimum energy values ​​in each low-frequency component to obtain the slope of the fitted line; the sum of the absolute values ​​of the slopes corresponding to all low-frequency components in the frequency domain of each frame of audio signal is calculated; the fundamental frequency periodicity of the amplitude is calculated by using the sum of the sums and the average value of the periodic terms of all peaks in each frame of audio signal.

[0009] In one embodiment, the periodicity of the fundamental frequency of the amplitude is positively correlated with the average value and negatively correlated with the accumulated result.

[0010] In one embodiment, obtaining the continuous periodic noise figure of each frame of audio signal includes: Mel spectrum analysis is performed on each frame of audio signal to extract the lowest-order Mel cepstral coefficients. The difference between the lowest-order Mel cepstral coefficients of each frame of audio signal and the previous frame of audio signal is calculated. The difference is then subjected to exponential operation. The normalized value of the product of the result of the exponential operation and the periodicity of the amplitude fundamental frequency is used as the continuous periodic noise coefficient of each frame of audio signal.

[0011] In one embodiment, the evaluation of the signal preservation factor of each frame of audio signal includes: The burst non-stationary noise coefficient and the continuous periodic noise coefficient of each frame of audio signal are used as inputs to the CIM composite index method, and the output composite score is used as the signal retention factor. The signal retention factor is negatively correlated with both the burst non-stationary noise coefficient and the continuous periodic noise coefficient.

[0012] In one embodiment, the correction of the forget gate calculation in the LSTM neural network involving gated learning is specifically expressed as follows: In the formula, Let t be the forgetting threshold for the audio signal of the t-th frame in an emergency rescue communication environment. , These represent the current input state and the hidden state of the previous frame, respectively. This is the original matrix containing the forget gate weights from the LSTM neural network. The signal preservation adjustment matrix for the t-th frame audio signal in an emergency rescue environment is determined by the signal preservation factor of the t-th frame audio signal. Let be the bias term of the audio signal in frame t. Let be the signal preservation factor of the audio signal in frame t. This is the sigmoid function.

[0013] Secondly, embodiments of this application also provide a voice noise reduction system for emergency rescue communication equipment, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0014] This application has at least the following beneficial effects: This application modifies the forget gate calculation in an LSTM neural network that incorporates gated learning and trains an LSTM speech denoising neural network for emergency rescue scenarios. This allows for the acquisition of high-quality communication audio signals in emergency rescue scenarios. It solves the technical drawbacks of traditional forget gate learning, which cannot quickly forget historical noise in the face of sudden strong noise and retains excessive speech features of continuous background noise. While ensuring the temporal dependencies of the input communication audio signals to the speech denoising neural network, it effectively avoids the cumulative pollution of key interactive audio signals by complex environmental noise, thereby improving the quality of voice interaction in high-risk emergency rescue environments. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the steps of a voice noise reduction method for an emergency rescue communication device, as provided in one embodiment of this application. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by this application to achieve the intended purpose of the invention, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a voice noise reduction method and system for emergency rescue communication equipment proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0019] The following description, in conjunction with the accompanying drawings, details the specific scheme of the voice noise reduction method and system for emergency rescue communication equipment provided in this application.

[0020] Please see Figure 1 The document illustrates a flowchart of a voice noise reduction method for an emergency rescue communication device according to an embodiment of this application. The method includes the following steps: S1 collects audio signals in the emergency rescue communication environment through emergency rescue communication equipment.

[0021] This application utilizes Bluetooth smart microphones, smart communication hosts, and head-mounted speaker components deployed in emergency rescue equipment to collect audio signals and facilitate voice interaction in emergency rescue communication environments. The smart Bluetooth microphones collect communication audio signals in emergency rescue environments at a sampling frequency of 16kHz and a precision of 16bit. A bandpass filter is used to filter out high-frequency noise exceeding half of the sampling frequency to avoid signal aliasing.

[0022] The acquired communication audio signal was processed in frames of 25ms each. To avoid inter-frame breaks, the overlap rate between frames was set to 50%. The Hamming window function was used to window each frame of the communication audio signal. The communication audio signal in the emergency rescue environment was converted into a digital signal by an ADC analog-to-digital converter. Since the frame processing and windowing operation of the audio signal are existing known technologies, the specific acquisition process will not be described in detail.

[0023] S2. Perform frequency domain transformation on each frame of audio signal, analyze the difference in energy distribution between each frame of audio signal and its adjacent frame of audio signal in the frequency domain, as well as the energy distribution of harmonic frequencies of each frame of audio signal in the frequency domain, coordinate the fluctuation degree of amplitude of each frame of audio signal in the time domain, and the autocorrelation of amplitude in a short time, and determine the burst non-stationary noise coefficient of each frame of audio signal.

[0024] In the communication environment of emergency rescue equipment, there are various noise sources, including the operating noise of rescue equipment, sudden and unstable impact / explosion noise, personnel breathing noise, and noise from severe weather conditions. These noise audios can interfere with the stability of the gating mechanism in the LSTM neural network's denoising process for communication audio, causing abrupt changes in the input signal. This leads to abnormal fluctuations in the output of the forget gate and masks the true characteristics of the personnel's communication voice signals in the communication audio. It may cause the forget gate to be unable to accurately determine which information in the communication audio signal in the emergency rescue environment should be retained or forgotten, ultimately reducing the quality of communication audio interaction.

[0025] Specifically, during the training process of an LSTM speech denoising neural network incorporating gating learning, when any frame of communication audio signal in an emergency rescue environment should be forgotten during the forget gate calculation, the frame of communication audio signal is more likely to contain sudden, non-stationary, strong noise, and the lower its correlation with the contextual information of historical frame audio signals. That is, the more pronounced the sudden increase in the amplitude of the communication audio signal caused by sudden critical events in the emergency rescue communication audio, the lower the short-time autocorrelation of the amplitude of the communication audio signal within the frame; simultaneously, non-stationary noise and sudden strong noise make the difference in contextual information between the current frame and historical frames stronger, that is, the more severe the shift of specific frequency resonant peaks between the current frame and historical frames, and the higher the energy difference in the harmonic frequency structure.

[0026] Based on the above analysis, this embodiment divides each 1 second into a communication frame segment and constructs a burst non-stationary noise coefficient to characterize the amplitude abrupt change correlation and resonant harmonic shift degree in each frame of the communication audio signal. Specifically: Each frame of the communication audio signal is used as input. The AMPD (Automatic Multiscale-based Peak Detection) algorithm is used to obtain all peak signals in each frame of the communication audio signal. The sequence of audio signal amplitudes between each peak signal and its predecessor, arranged in temporal order, is taken as the amplitude change sequence of each peak signal. For the first peak of each frame of the communication audio signal, the sequence of the minimum amplitude of the signal closest to the first peak and all signal amplitudes between the first peak signal, arranged in temporal order, is taken as the amplitude change sequence of the first peak signal. The correlation coefficient of the amplitude change sequence of each peak signal in each frame of the communication audio signal is obtained through the ACF function. The mean of all autocorrelation coefficients of the amplitude change sequence corresponding to each peak signal is calculated and denoted as the amplitude autocorrelation intensity of each peak signal. The lower the amplitude autocorrelation intensity, the higher the probability of the presence of sudden non-stationary noise in the current frame of the communication audio signal.

[0027] Secondly, the numerical difference between the maximum and minimum peaks of each frame of audio signal is measured. The inverse proportional mapping result of the absolute value of the mean of the amplitude autocorrelation intensity of all peaks in each frame of communication audio signal is multiplied by the numerical difference and then normalized to obtain the amplitude abrupt correlation of each frame of audio signal.

[0028] It should be noted that difference represents the degree of difference between two variables, which can be calculated using methods such as difference, absolute value of difference, square of difference, ratio, etc. Inverse proportional mapping represents a relationship between two variables, specifically the property that an increase in one variable leads to a decrease in the other. In this embodiment, the mean of the amplitude autocorrelation intensity of all peaks in each frame of the communication audio signal is the result of aggregating all autocorrelation coefficients of the amplitude change sequence of all peaks in each frame of the audio signal. Aggregation calculation can specifically employ methods such as summation or averaging.

[0029] In this embodiment, the difference between the maximum and minimum peak values ​​of each frame's audio signal is calculated and multiplied by the reciprocal of the absolute value of the mean of the amplitude autocorrelation intensity of all peaks in each frame's communication audio signal. The multiplication result is then normalized to obtain the amplitude abrupt change correlation of each frame's audio signal (the reciprocal is the inverse proportional mapping result of the absolute value). This embodiment calculates this for all frame audio signals within the communication frame segment and uses a maximum-minimum value normalization method for normalization.

[0030] Furthermore, each frame of the communication audio signal is used as input, and the Fast Fourier Transform (FFT) is used to obtain the frequency domain representation of each frame of the communication audio signal. The Hilbert Transform is used to obtain all frequency components in each frame of the communication audio signal. The response energy of all frequency components in each frame of the communication audio signal is used as input, and the Otsu method is used to obtain the segmentation threshold. The frequency components corresponding to all response energies greater than or equal to the segmentation threshold are recorded as the characteristic frequency components of each frame of the communication audio signal. The frequency corresponding to the maximum response energy in each frame of the communication audio signal is taken as the fundamental frequency, and the harmonic frequencies corresponding to the fundamental frequency are obtained.

[0031] The energy sequence is formed by arranging all the energies corresponding to the characteristic frequency components of each frame audio signal in the frequency domain in ascending order of frequency; the metric distance of the energy sequence of each frame audio signal and all other frame audio signals in the communication frame segment is positively fused, and the dispersion of the energy corresponding to all harmonic frequencies of each frame audio signal in the frequency domain is calculated. By combining the result of the forward fusion with the degree of dispersion, the resonant harmonic shift of each frame of audio signal is obtained. Forward fusion refers to combining multiple variables, which can be calculated using methods such as addition, multiplication, a combination of addition and multiplication, or averaging. The degree of dispersion can be calculated using methods such as variance, standard deviation, and coefficient of variation.

[0032] In this embodiment, the cumulative result of the DTW distance of the energy sequence of each frame audio signal and all other frame audio signals within its communication frame segment is calculated, and then multiplied by the standard deviation of the energy corresponding to all harmonic frequencies of each frame audio signal in the frequency domain. The result of the multiplication is used as the resonant harmonic offset of each frame audio signal. In another embodiment, the distance can be calculated using Euclidean distance.

[0033] The burst non-stationary noise figure of each frame of audio signal is calculated using the following expression: In the formula, Let be the burst non-stationary noise figure of the i-th frame of communication audio signal in an emergency rescue environment; The correlation of amplitude abrupt changes in the i-th frame of the communication audio signal in an emergency rescue environment. Let be the resonant harmonic offset of the i-th frame of the communication audio signal in the emergency rescue environment.

[0034] The burst non-stationary noise figure reflects the short-time correlation of amplitude abrupt changes and the degree of resonant harmonic frequency shift within each frame of communication audio signal in an emergency rescue environment. Amplitude abrupt change correlation reflects the maximum difference in peak signal amplitude and the short-time autocorrelation strength of all peak signals within each frame of communication audio signal in an emergency rescue environment. The resonant harmonic shift characterizes the difference in characteristic peaks between each frame of communication audio signal and its adjacent frames, as well as the fluctuation in the response energy of harmonic frequency components. During the training process using an LSTM speech denoising neural network with gated learning, when any frame of communication audio signal in an emergency rescue environment is more likely to contain burst non-stationary strong noise with low correlation to historical frame context information, the maximum amplitude difference of that frame's communication audio signal is stronger, and the short-time autocorrelation of the intra-frame peak signals is lower. Simultaneously, the greater the difference in characteristic peak response energy between it and historical frames, the higher the difference in response energy between harmonic frequency components. In this case, the calculation results of the LSTM forget gate should be reduced to decrease the pollution of key speech information by burst strong noise features and reduce dependence on burst strong noise states.

[0035] S3 extracts the periodicity of the peak signal in each frame of audio signal in the time domain, as well as the energy change trend of the low-frequency components of each frame of audio signal in the frequency domain, to obtain the amplitude fundamental frequency periodicity of each frame of audio signal. Combined with the degree of energy enhancement of each frame of audio signal in the low frequency relative to the historical adjacent frame of audio signal, the continuous periodic noise coefficient of each frame of audio signal is obtained.

[0036] In the training process of the LSTM speech denoising neural network with gating learning, when relying solely on the burst non-stationary noise coefficients in the communication audio signal as the basis for improving the forget gate calculation, the degree of preservation of stable and continuous background noise such as the strong breathing noise of rescuers in the emergency rescue environment is not considered in the forget gate calculation. The analysis of the low-frequency continuous periodic characteristics in the communication audio signal is lacking, which leads to the inability of the LSTM speech denoising neural network with gating learning to effectively suppress stable background noise and reduce the quality of speech denoising.

[0037] Specifically, during the training process of an LSTM speech denoising neural network incorporating gating learning, the more pronounced the continuous periodic noise of human breathing in any frame of the communication audio signal in an emergency rescue environment, the higher the periodicity of the amplitude change of the communication audio signal, and the slower the fundamental frequency energy attenuation of low-frequency components. Simultaneously, the increase in low-order Mel-frequency cepstral coefficients within a certain time range caused by continuous periodic noise such as human breathing is more pronounced. In this case, the retention rate of that frame of communication audio signal during gating learning should be reduced to prevent the accumulation of continuous periodic noise in the emergency rescue environment from contaminating the critical communication audio signal and reducing denoising quality.

[0038] Based on the above analysis, this embodiment calculates the continuous periodic noise figure of each frame of audio signal to evaluate the amplitude fundamental frequency periodicity and low-frequency energy enhancement degree of each frame of communication audio signal in an emergency rescue environment. Specifically: The sequence of all peak signals in each frame of the communication audio signal, arranged in temporal order, is used as input. The STL (Seasonal and Trend decomposition using Loess) algorithm is employed to obtain the periodic term of each peak signal within each frame of the communication audio signal. The frequency components of each frame of the communication audio signal are then segmented using the Otsu's method to obtain a segmentation threshold. Frequency components corresponding to frequencies below the segmentation threshold are designated as low-frequency components. A sequence of all response energies from the maximum to the minimum response energy value within each low-frequency component of the communication audio signal is obtained. This sequence is then linearly fitted using the least squares method to obtain the slope of the fitted line. The least squares method is a well-known existing technique. In another embodiment, the implementer may choose other existing feasible linear fitting algorithms.

[0039] Secondly, the amplitude spectrum of each frame of communication audio signal is obtained by using FFT (Fast Fourier Transform). The linear frequency is converted to Mel frequency by Mel filter bank and logarithmic compression and DCT transformation are performed to obtain the Mel cepstral coefficients of each frame of communication audio signal. The low-order Mel cepstral coefficients of each frame of communication audio signal are taken. In this embodiment, the lowest-order Mel cepstral coefficient in each frame of communication audio signal is taken.

[0040] The periodicity of the amplitude fundamental frequency of each frame of the audio signal is calculated using the following expression: In the formula, For the fundamental frequency periodicity of the amplitude of the i-th frame of communication audio signal in an emergency rescue environment, Let be the average of the periodic terms of all peak signals in the i-th frame of the communication audio signal. The slope of the i-th frame of the communication audio signal corresponding to each low-frequency component in the frequency domain is given. Let e ​​represent the sum of the absolute values ​​of the slopes corresponding to all low-frequency components of the i-th frame of the communication audio signal in the frequency domain, where e is a natural constant.

[0041] Furthermore, the difference between the lowest-order Mel-frequency cepstral coefficients of each frame of audio signal and the previous frame is calculated. This difference is then subjected to an exponential operation. The normalized value of the product of the exponential operation result and the periodicity of its amplitude fundamental frequency is used as the continuous periodic noise figure of each frame of audio signal. In this embodiment, the exponential operation on this difference specifically involves using the difference as the exponent of an exponential function with the natural constant as its base, and using the calculated result of this exponential function as the result of the exponential operation. In this embodiment, the normalization method involves analyzing and calculating the continuous periodic noise figure of all frame audio signals within the communication frame segment, and then using a maximum-minimum value normalization method for normalization.

[0042] The continuous periodic noise figure reflects the periodicity of the fundamental frequency amplitude of each frame of communication audio signal in an emergency rescue environment and the degree of low-frequency energy accumulation. The periodicity of the fundamental frequency amplitude reflects the intensity of the periodic amplitude change of each frame of communication audio signal in an emergency rescue environment and the slow attenuation of the fundamental frequency energy of the low-frequency components. The degree of low-frequency energy enhancement characterizes the enhancement of the low-order Mel-frequency cepstral coefficients of each frame of communication audio signal in an emergency rescue environment, that is, the enhancement of the continuous periodic noise energy intensity in each frame of communication audio signal. During the training process of the LSTM speech denoising neural network with gating learning, when the continuous periodic noise of human breathing in any frame of communication audio signal is more obvious, the intensity of the periodic amplitude change of the communication audio signal is higher, and the attenuation of the fundamental frequency energy of the low-frequency components is slower. At the same time, the enhancement of the low-order Mel-frequency cepstral coefficients of the communication audio signal within the communication frame segment is more obvious. At this time, the retention degree of the forget gate in the gating learning of the communication audio signal should be reduced to avoid excessive accumulation of continuous periodic noise contaminating the key communication audio signal and reducing the denoising quality.

[0043] S4. Using the burst non-stationary noise coefficient and the continuous periodic noise coefficient, the signal retention factor of each frame of audio signal is evaluated, which is used to correct the forget gate calculation in the LSTM neural network containing gating learning, and the corrected neural network is used for speech denoising.

[0044] During the training process of an LSTM speech denoising neural network with gating learning, the more severe the sudden non-stationary strong noise and continuous periodic noise in any frame of communication audio signal in the emergency rescue environment, the more the retention of that frame of communication audio signal in the forget gate should be reduced to avoid the real rescue personnel's interactive audio signal being contaminated by noise signals.

[0045] Therefore, this application constructs a signal retention factor to characterize the degree of audio feature retention of each frame of communication audio signal in the training process of an LSTM speech denoising neural network including gating learning in an emergency rescue environment. This factor can be obtained through burst non-stationary noise coefficients and continuous periodic noise coefficients. Specifically, in one implementation, the burst non-stationary noise coefficients and continuous periodic noise coefficients of each frame of communication audio signal are used as inputs, and the CIM comprehensive index method is used to obtain a comprehensive score, which is recorded as the signal retention factor. The burst non-stationary noise coefficients and continuous periodic noise coefficients are negatively correlated with the signal retention factor, and their calculation weights can be determined by the entropy weight method. When the signal retention factor is larger, it indicates that the intra-frame communication audio signal should be retained more in the forget gate calculation, so as to avoid the accumulation of complex noise in the emergency rescue environment, which pollutes the audio signal of key personnel interaction and reduces the quality of voice interaction. Since the CIM comprehensive index method and the entropy weight method are known technologies, the specific acquisition process will not be described in detail.

[0046] The forget gate calculation in an LSTM neural network incorporating gated learning is corrected using a signal retention factor. The specific calculation method is as follows: ;in, This is the forgetting threshold for the t-th frame of audio signal in an emergency rescue communication environment. , These represent the current input state and the hidden state of the previous frame, respectively. This means concatenating the previous hidden state and the current input vector into a single vector. The original matrix containing the forget gate weights of the LSTM neural network can be determined using the backpropagation algorithm and gradient descent optimization method. It is responsible for processing the current input... Map the hidden state from the previous time step to the unit state dimension; The signal preservation adjustment matrix for the t-th frame audio signal in an emergency rescue environment can be determined by the signal preservation factor of the t-th frame communication audio signal. Specifically, a... The matrix is ​​a set of data elements, where each element is a signal preservation factor of the t-th frame of the communication audio signal, and n is a variable. The dimension of the forget gate weight matrix will be adjusted as the signal retention matrix using the matrix obtained in the above manner; This is the bias term for the audio signal in frame t, used to increase model flexibility and can be automatically learned and updated through backpropagation; Let be the signal preservation factor of the audio signal in frame t; For the sigmoid function, such that The value range is within the range of [0,1].

[0047] The reason why this application improves the current input state, the hidden state of the previous frame, and the bias term in the forget gate calculation is to avoid the fact that improving only one of the weight matrix or the bias term would disrupt the overall distribution offset of the gated signal, thereby breaking the temporal dependency between the input communication audio signals.

[0048] This embodiment obtains N frames of noisy communication audio signals and corresponding frames of clean speech signals as training pairs to train an LSTM speech denoising neural network with gated learning. A multi-layer LSTM network is constructed; in this embodiment, the LSTM network has three layers. Each LSTM unit includes an input gate, a forget gate, and an output gate to learn long-term dependencies in the time-series communication audio signals. A fully connected layer is connected after the LSTM network layers to map the hidden states to an output with the same dimension as the target. Mean squared error is used as the loss function. The original gradients of the loss with respect to the gate weights, input weights, and bias weights are calculated through backpropagation. The forget gate output is calculated using the improved gated learning described above. The Adam optimizer is used to adaptively adjust the learning rate and accelerate the convergence and stabilization training process. The loss and speech quality metrics are monitored on the retained validation set. In this application, POLQA score is used as the speech quality metric. Since the training of the LSTM speech denoising neural network with gated learning is a well-known technique, the specific process will not be described in detail. In this embodiment, the size of the training dataset is N = 18,000,000 frames. Implementers can set this size according to their actual situation; this embodiment does not impose any restrictions.

[0049] The system acquires each frame of communication audio signal in the emergency rescue communication environment and uses it as input. It then uses the pre-trained LSTM speech denoising neural network with gating learning to output the denoised communication audio signal, which is transmitted to the dispatch center via the intelligent communication host for the purpose of assessing and evaluating high-quality communication audio signals in emergency rescue scenarios.

[0050] Based on the same inventive concept as the above method, this application embodiment also provides a voice noise reduction system for an emergency rescue communication device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described voice noise reduction methods for an emergency rescue communication device.

[0051] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments of this specification have been described above. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0052] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0053] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the principles of this application should be included within the protection scope of this application.

Claims

1. A voice noise reduction method for emergency rescue communication equipment, characterized in that, The method includes the following steps: Audio signals are collected in the emergency rescue communication environment using emergency rescue communication equipment; The frequency domain transformation of each frame audio signal is performed to analyze the difference in energy distribution between each frame audio signal and its adjacent frame audio signals in the frequency domain, as well as the energy distribution of harmonic frequencies of each frame audio signal in the frequency domain. The fluctuation degree of amplitude of each frame audio signal in the time domain and the autocorrelation of amplitude in a short time are also analyzed to determine the burst non-stationary noise coefficient of each frame audio signal. The periodicity of the peak signal in each frame of audio signal in the time domain and the energy change trend of the low frequency component in the frequency domain of each frame of audio signal are extracted to obtain the amplitude fundamental frequency periodicity of each frame of audio signal. Combined with the degree of energy enhancement of each frame of audio signal in the low frequency relative to the historical adjacent frame of audio signal, the continuous periodic noise figure of each frame of audio signal is obtained. The signal retention factor of each frame of audio signal is evaluated by using the burst non-stationary noise coefficient and the continuous periodic noise coefficient, which is used to correct the forget gate calculation in the LSTM neural network containing gating learning, and the corrected neural network is used for speech denoising. The evaluation of the signal retention factor of each frame of audio signal includes: taking the burst non-stationary noise coefficient and the continuous periodic noise coefficient of each frame of audio signal as input to the CIM comprehensive index method, and taking the output comprehensive score as the signal retention factor, wherein the signal retention factor is negatively correlated with both the burst non-stationary noise coefficient and the continuous periodic noise coefficient. The correction includes the forget gate calculation in the gated learning LSTM neural network, and the specific expression is as follows: In the formula, Let t be the forgetting threshold for the audio signal of the t-th frame in an emergency rescue communication environment. , These represent the current input state and the hidden state of the previous frame, respectively. This is the original matrix containing the forget gate weights from the LSTM neural network. The signal preservation adjustment matrix for the t-th frame audio signal in an emergency rescue environment is determined by the signal preservation factor of the t-th frame audio signal. Let be the bias term of the audio signal in frame t. Let be the signal preservation factor of the audio signal in frame t. This is the sigmoid function.

2. The voice noise reduction method for emergency rescue communication equipment as described in claim 1, characterized in that, The determination of the burst non-stationary noise coefficient of each frame of audio signal includes: Extract the characteristic frequency components of each frame audio signal in the frequency domain, and form an energy sequence by combining all the energies corresponding to the characteristic frequency components; perform forward fusion on the metric distance of the energy sequences of each frame audio signal and all its adjacent frame audio signals, and calculate the degree of dispersion of the energy corresponding to all harmonic frequencies of each frame audio signal in the frequency domain. By combining the results of the forward fusion with the degree of dispersion, the resonant harmonic offset of each frame of audio signal is obtained; The amplitude mutation correlation of each frame audio signal is calculated by using the fluctuation degree and the autocorrelation. Combined with the resonant harmonic offset, the sudden non-stationary noise coefficient is obtained. The sudden non-stationary noise coefficient is positively correlated with both the amplitude mutation correlation and the resonant harmonic offset.

3. The voice noise reduction method for emergency rescue communication equipment as described in claim 2, characterized in that, The calculation of the amplitude abrupt change correlation of each frame of audio signal includes: Identify each peak of the audio signal in each frame of the time domain, form an amplitude change sequence by combining the amplitudes of each peak in the audio signal in each frame with the previous peak, determine the correlation coefficient of each amplitude change sequence, and aggregate and calculate all autocorrelation coefficients of the amplitude change sequences of all peaks in the audio signal in each frame. The numerical difference between the maximum and minimum peak values ​​of each frame of audio signal is measured. The absolute value of the aggregated calculation result is multiplied by the numerical difference and then normalized to obtain the amplitude abrupt change correlation of each frame of audio signal.

4. The voice noise reduction method for an emergency rescue communication device as described in claim 3, characterized in that, The extraction of characteristic frequency components of each frame of audio signal in the frequency domain includes: Signal processing techniques are used to obtain the frequency components of each frame of audio signal in the frequency domain. The energy of all frequency components is thresholded, and the frequency components corresponding to all energies greater than or equal to the threshold are taken as the characteristic frequency components of each frame of audio signal in the frequency domain.

5. The voice noise reduction method for an emergency rescue communication device as described in claim 4, characterized in that, The process for determining the periodicity of the fundamental frequency of the amplitude of each frame of audio signal is as follows: The periodic terms of each peak are obtained by using a sequence decomposition algorithm. The frequencies of all frequency components corresponding to each frame of audio signal in the frequency domain are thresholded and the frequency components corresponding to all frequencies less than the segmentation threshold are recorded as low frequency components. Linear fitting is performed on all energies between the maximum and minimum energy values ​​in each low-frequency component to obtain the slope of the fitted line; The absolute values ​​of the slopes corresponding to all low-frequency components in the frequency domain of each frame of audio signal are accumulated. The fundamental frequency periodicity of the amplitude is calculated by combining the accumulated result with the average value of the periodic terms of all peaks in each frame of audio signal.

6. The voice noise reduction method for an emergency rescue communication device as described in claim 5, characterized in that, The periodicity of the fundamental frequency of the amplitude is positively correlated with the average value and negatively correlated with the cumulative result.

7. The voice noise reduction method for an emergency rescue communication device as described in claim 1, characterized in that, The acquisition of the continuous periodic noise figure of each frame of audio signal includes: Mel spectrum analysis is performed on each frame of audio signal to extract the lowest-order Mel cepstral coefficients. The difference between the lowest-order Mel cepstral coefficients of each frame of audio signal and the previous frame of audio signal is calculated. The difference is then subjected to exponential operation. The normalized value of the product of the result of the exponential operation and the periodicity of the amplitude fundamental frequency is used as the continuous periodic noise coefficient of each frame of audio signal.

8. A voice noise reduction system for an emergency rescue communication device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

Images