A method for identifying acoustic environmental noise based on fuzzy multi-Gaussian clustering
Through the fuzzy multi-Gaussian clustering method, the problems of instability and low efficiency of traditional Gaussian hybrid model training are solved, and the efficient identification and separation of audio signals are realized, which improves the stability and computing efficiency of the model.
Patent Information
- Application Number
- CN202210207683.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-04
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2042-03-04
AI Technical Summary
The traditional acoustic environmental noise recognition method based on Gaussian hybrid model has the problem of instability in the training process and low computational efficiency, especially due to the heuristic EM training algorithm and additional computational burden, resulting in low model training efficiency.
Using the method based on fuzzy multi-Gaussian clustering, the parameters are initialized by fuzzy mean clustering, and iteratively trained the fuzzy multi-Gaussian model, and iteratively loops through fuzzy membership, and finally converted into a Gaussian mixed distribution model, using posterior probability to realize the audio signal efficient identification and separation of foreground signals and background noise.
The stability and efficiency of model training are improved, and the efficient identification and separation of audio signals is achieved, which is more stable and efficient than traditional methods.
Smart Images

Figure CN114613386B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of acoustics, and in particular to a method for identifying acoustic environmental noise based on fuzzy multi-Gaussian clustering. Background Art
[0002] The sounds in the acoustic environment can be roughly divided into three categories, namely, speech, music and noise. For the user's auditory perception, any intentional or unintentional sound interference can be called noise. For example, there are various types of acoustic noise in mobile voice calls or multimedia conferences. These noises may be the background noise of canteens, supermarkets, etc., or the environmental noise on the streets where vehicles are driving. Even in acoustic echo cancellation, when the near-end user is speaking, the far-end user is speaking at the same time. For the far-end echo canceller, the speech of the far-end user is noise, which will interfere with the adaptive echo estimation. This is why the acoustic echo cancellation technology has the so-called duplex detection (double talk detection, DTD) problem. Another type of noise common in electroacoustic equipment is the HISS sound caused by the thermal noise of the circuit hardware, that is, the background sound when the circuit transmitting the sound signal is silent. Usually, background noise will greatly interfere with the user's recognition of the foreground voice.
[0003] In traditional pattern recognition applications based on Gaussian mixture models, k-means clustering is often used as an initialization method. The centers of each category output by the k-means clustering method are used as the mean centers of the Gaussian mixture distribution. The variances of each category are calculated based on the current classification status of each category as the variances of each branch of the Gaussian mixture distribution. Then, iterative training based on the expectation maximization method (EM) is started, and finally a converged Gaussian mixture model is obtained. The defect of the EM-based training mechanism of the Gaussian mixture model is that its mathematical logic is not very strict, its EM training algorithm is essentially heuristic, and the process of training the model is not stable. It is not a solid industrial process method, so it is called "alchemy" rather than "metallurgy". In addition, the training step of the Gaussian mixture model requires the arithmetic average of the posterior probability of each category to obtain the weight of each branch. This process is actually an additional computational burden, which reduces the efficiency of model training. Therefore, a stable and efficient environmental noise identification and separation method is needed. Summary of the invention
[0004] The purpose of the present invention is to solve the deficiencies of the prior art and provide a method for identifying acoustic environmental noise based on fuzzy multi-Gaussian clustering.
[0005] In order to solve the above problems, the present invention adopts the following technical solutions:
[0006] A method for identifying acoustic environmental noise based on fuzzy multi-Gaussian clustering comprises the following steps:
[0007] Step 1: Collect audio data in the time domain and initialize parameter data through fuzzy mean clustering;
[0008] Step 2: Increase the training output of the fuzzy multi-Gaussian model through repeated iterative operations
[0009] Where K represents the number of Gaussian models in the fuzzy multi-Gaussian model; N(x j ,μ k ,σ k ) represents a normal distribution, x j represents the jth observation data, j∈(1,N), μ k represents data expectation, σ k represents the data standard deviation; Γ kj Represents x j Belongs to the kth class N(x,μ k ,σ k )’s fuzzy membership, Ν(x,μ k ,σ k ) represents normal distribution; the training samples of the model are observed data, and there are N training samples in total;
[0010] Step 3: Measure N training samples separately to maximize and update represents the data expectation at time t, Represents the standard deviation of the data at time t;
[0011] Step 4: Based on the maximum obtained in step 3 Determine the convergence characteristics of the value over time; if the set convergence conditions are met, it means that the training of the fuzzy multi-Gaussian model is completed and proceed to the next step; otherwise, return to step 2;
[0012] Step 5: After completing the training of the fuzzy multi-Gaussian model, convert it to obtain the Gaussian mixture distribution;
[0013] Step 6: Convert the audio data to be identified from the time domain to the frequency domain, and calculate the posterior probability that the audio data belongs to each category;
[0014] Step 7: Classify the audio data into the type with the highest posterior probability and end the step.
[0015] Furthermore, in step 1, during the parameter data initialization process, Γ is set t kj represents Γ at the tth step kj Value, Γ kj Represents data x j Belongs to the kth class N(x,μk ,σ k )’s fuzzy membership; is μ at the tth step k , μ k Express data expectations; is σ at the tth step k , σ k represents the data standard deviation; sets the number of Gaussian distribution types to K; sets the collected audio data set to x j represents the jth audio data, N represents the number of data; the audio data set is clustered with a known number of clusters of 2; after the mean clustering algorithm converges, the output membership coefficient is ω ij ,ω ij Represents x j Fuzzy membership of the i-th class in c-mean, 0≤ω ij ≤1, where membership is an existing concept in fuzzy mathematics; the initialization formula is as follows;
[0016] Γ 0 kj =ω kj
[0017]
[0018]
[0019] Among them, Γ 0 kj represents the initial step Γ kj Numerical value; ω kj Represents x j The fuzzy membership of the kth class in the Gaussian distribution; μ represents the initial step k ; represents the initial step σ k .
[0020] Furthermore, the number of types of the Gaussian distribution is K, which has a value of 2, and the two types correspond to the acoustic background noise and the foreground sound, respectively.
[0021] Furthermore, in step 2, N(x, μ k ,σ k ) is a normal distribution, expressed as:
[0022]
[0023] Among them, μ k represents data expectation; σ k Represents the standard deviation of the data.
[0024] Furthermore, in step 2, by updating Γ kj , to increase the training output of the fuzzy multi-Gaussian model Update Γ by kj :
[0025]
[0026] Update Γ by timing kj , increase The value of .
[0027] Furthermore, in step 3, update The process includes:
[0028]
[0029]
[0030] According to the Γ obtained in step 2 t kj , update the normal distribution and Parameters; by updating The value of is maximized at step t
[0031] In step 4, the following formula is used to determine The convergence characteristics of the value of over time:
[0032] Δ<ε
[0033] If Δ satisfies the condition, the algorithm meets the convergence characteristic, otherwise return to step 2.
[0034] Furthermore, the value of Δ is as follows:
[0035]
[0036] Where ε is the set threshold; K represents the number of Gaussian distribution types; and N represents the number of data.
[0037] Furthermore, in step 5, by using the Gaussian mixture distribution weight coefficient π k Substitute the fuzzy multi-Gaussian model obtained in step 4 to obtain the Gaussian mixture distribution model Weight coefficient π k The conversion process is expressed as:
[0038]
[0039] Among them, π k is a non-negative weight coefficient,
[0040] Furthermore, the model number K′ of the Gaussian mixture distribution model is consistent with the number K in the fuzzy multi-Gaussian model, both of which are 2.
[0041] The beneficial effects of the present invention are:
[0042] By directly performing iterative cycles based on fuzzy membership, and converting the weights of each branch through the arithmetic mean of the fuzzy membership after convergence or iteration, the fuzzy multi-Gaussian is converted into a Gaussian mixture model in one step. Finally, the Bayesian calculation step of the posterior probability of the Gaussian mixture model is used to achieve efficient recognition and separation of foreground signals and background noise of online audio. Compared with traditional methods, it has higher stability and the model training efficiency is also improved. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 This is the overall process of the first embodiment of the present invention. DETAILED DESCRIPTION
[0044] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.
[0045] It should be noted that the illustrations provided in the following embodiments are only used to schematically illustrate the basic concept of the present invention. The figures only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.
[0046] Embodiment 1:
[0047] like Figure 1 As shown, a method for identifying acoustic environmental noise based on fuzzy multi-Gaussian clustering includes the following steps:
[0048] Step 1: Collect audio data in the time domain and initialize parameter data through fuzzy mean clustering;
[0049] Step 2: Increase the training output of the fuzzy multi-Gaussian model through repeated iterative operations
[0050] Where K represents the number of Gaussian models in the fuzzy multi-Gaussian model; N(x j ,μ k ,σ k) represents a normal distribution, x j represents the jth observation data, j∈(1,N), μ k represents data expectation, σ k represents the data standard deviation; Γ kj Represents x j Belongs to the kth class N(x,μ k ,σ k )’s fuzzy membership, Ν(x,μ k ,σ k ) represents normal distribution; in this example, the training samples of the model are observation data, including N training samples;
[0051] Step 3: Measure N training samples separately to maximize and update represents the data expectation at time t, Represents the standard deviation of the data at time t;
[0052] Step 4: Based on the maximum obtained in step 3 Determine the convergence characteristics of the value over time; if the set convergence conditions are met, it means that the training of the fuzzy multi-Gaussian model is completed and proceed to the next step; otherwise, return to step 2;
[0053] Step 5: After completing the training of the fuzzy multi-Gaussian model, convert it to obtain the Gaussian mixture distribution model;
[0054] Step 6: Convert the audio data to be identified from the time domain to the frequency domain, and calculate the posterior probability that the audio data belongs to each class in the Gaussian mixture distribution model;
[0055] Step 7: Classify the audio data into the type with the highest posterior probability and end the step.
[0056] In step 1, during the parameter data initialization process, set Γ t kj represents Γ at the tth step kj Value, Γ kj Represents data x j Belongs to the kth class N(x,μ k ,σ k )’s fuzzy membership; is μ at the tth step k , μ k Express data expectations; is σ at the tth step k , σ k represents the data standard deviation; the number of types of fuzzy Gaussian distribution is set to K. In this example, there are two situations: acoustic background noise and foreground sound, so K is set to 2; the collected audio data set is set to x j Represents the jth audio data, and N represents the number of data. Perform fuzzy c-means clustering on the audio data set with a known cluster number of 2. At this time, c = 2. After the mean clustering algorithm converges, the output membership coefficient is ω ij ,ω ij Represents x j Fuzzy membership of the i-th class in c-mean, 0≤ω ij ≤1, where membership is an existing concept in fuzzy mathematics. The initialization formula is as follows;
[0057] Γ 0 kj =ω kj
[0058]
[0059]
[0060] Among them, Γ 0 kj represents the initial step Γ kj Numerical value; ω kj Represents x j The fuzzy membership of the kth class in the Gaussian distribution; represents the initial step μ k ; represents the initial step σ k .
[0061] In step 2, N(x, μ k ,σ k ) is a normal distribution, x, μ k , σ k is the parameter in the normal distribution, expressed as:
[0062]
[0063] Among them, μ k represents data expectation; σ k represents the standard deviation of the data;
[0064] In step 2, by updating Γ kj , to increase the training output of the fuzzy multi-Gaussian model In this example, Γ is updated by kj :
[0065]
[0066] Among them, Γ t-1 ij Indicates that at the t-1th step, the jth data belongs to the i-th category N(x,μi ,σ i ) is the fuzzy membership of . According to the known mathematical proof process, as the fuzzy multi-Gaussian model is trained, Γ is updated by the above formula kj , can be increased The value of makes the model in this example stable.
[0067] Update in step 3 The process includes:
[0068]
[0069]
[0070] According to the Γ obtained in step 2 t kj , update the normal distribution and According to known mathematical proofs, it can be achieved by updating The value of is maximized at step t
[0071] In step 4, the following formula is used to determine The convergence characteristics of the value of with time:
[0072] Δ<ε
[0073]
[0074] Among them, ε is a preset threshold; K represents the number of fuzzy Gaussian distribution types; N represents the number of data; if Δ meets the conditions, the algorithm meets the convergence characteristics, otherwise return to step 2.
[0075] In step 5, the Gaussian mixture distribution weight coefficient π is k Substitute the fuzzy multi-Gaussian model obtained in step 4 to obtain the Gaussian mixture distribution model In this example, the number of models K′ in the Gaussian mixture distribution model is the same as the number K in the fuzzy multi-Gaussian model, both of which are 2; the weight coefficient π k The conversion process is expressed as:
[0076]
[0077] Among them, π k is a non-negative weight coefficient, Represents the membership degree Γ of all data points involved in training kj Average.
[0078] In step 6, the posterior probability p is calculated ij The process is as follows:
[0079]
[0080] Among them, p ij express The probability of belonging to type i in the Gaussian mixture distribution model, i∈(1,K′); Ν(x j ,μ i t ,σ i t ) and N(x j ,μ k t ,σ k t ) represents the normal distribution of Gaussian mixture distribution model type i and type k;
[0081] Then according to The data x j Identified as type i0 in the Gaussian mixture distribution model, in this example, the data x is divided into two types according to the Gaussian mixture distribution model. j Identify as acoustic background noise or foreground sound.
[0082] In step 7, by identifying and classifying the audio data, the noise in the data can be identified, which facilitates subsequent noise processing and other processes.
[0083] In step 1 and step 5, the method of converting data from the time domain to the frequency domain is Fourier transform.
[0084] In the traditional Gaussian distribution model training process, the initial value of the training uses the output result of k-means clustering, as shown in the following formula:
[0085]
[0086]
[0087]
[0088] Among them, N i Represents the number of vector data of the i-th category; represents the jth data of the i-th category; the initial value π 0 i , and The value is different from that in this embodiment; secondly, in the traditional training of Gaussian distribution model training, step E is also included:
[0089]
[0090] According to known mathematical proofs, it can be known that the traditional E step is difficult to ensure that If the value of increases, it may decrease, so the traditional Gaussian distribution model is not stable.
[0091] During the implementation process, the data samples collected are first trained using fuzzy mean clustering, and the output is used as the input of the clustering of the fuzzy multi-Gaussian algorithm for further training to generate a noise feature model based on the fuzzy multi-Gaussian mechanism, thereby establishing a noise recognition framework; after the data training is completed, for each data based on the size of the membership value of the generated multi-Gaussian, it is classified into the corresponding Gaussian distribution type; the membership of all data of the type belonging to the strong and weak is arithmetic averaged to obtain the weight of the type, thereby generating a mixed Gaussian distribution by the fuzzy multi-Gaussian model; finally, for the input data to be identified, the posterior probability of each category is calculated according to the mixed Gaussian distribution, and it is classified into the corresponding type with the largest probability value. Compared with the traditional distribution model based on k-means clustering, this embodiment has the characteristics of high efficiency and stability.
[0092] The above description is only a specific example of the present invention and does not constitute any limitation to the present invention. It is obvious that for professionals in this field, after understanding the content and principle of the present invention, various modifications and changes in form and details may be made without departing from the principle and structure of the present invention, but these modifications and changes based on the idea of the present invention are still within the scope of protection of the claims of the present invention.
Claims
1. A method for identifying acoustic environmental noise based on fuzzy multi-Gaussian clustering, characterized in that: The steps include: Step 1: Collect audio data in the time domain and initialize parameter data through fuzzy mean clustering; Step 2: Increase the training output of the fuzzy multi-Gaussian model through repeated iterative operations Where K represents the number of Gaussian models in the fuzzy multi-Gaussian model; N(x j , μ k , σ k ) represents a normal distribution, x j represents the jth observation data, j∈(1,N), μ k represents data expectation, σ k represents the data standard deviation; Γ kj express Belongs to the kth class N(x, μ k , σ k )’s fuzzy membership, N(x, μ k , σ k ) represents normal distribution; the training samples of the model are observed data, and there are N training samples in total; Step 3: Measure N training samples separately to maximize and update represents the data expectation at time t, Represents the standard deviation of the data at time t; Step 4: Based on the maximum obtained in step 3 Determine the convergence characteristics of the value over time; if the set convergence conditions are met, it means that the training of the fuzzy multi-Gaussian model is completed and proceed to the next step; otherwise, return to step 2; Step 5: After completing the training of the fuzzy multi-Gaussian model, convert it to obtain the Gaussian mixture distribution; Step 6: Convert the audio data to be identified from the time domain to the frequency domain, and calculate the posterior probability that the audio data belongs to each category; Step 7: Classify the audio data into the type with the largest posterior probability, and end the step; In step 1, during the parameter data initialization process, set Γ t kj represents Γ at the tth step kj Value, Γ kj Representation data Belongs to the kth class N(x, μ k , σ k )’s fuzzy membership; set the number of Gaussian distribution types to K; set the collected audio data set to x j represents the jth audio data, N represents the number of data; the audio data set is clustered with a known number of clusters of 2; after the mean clustering algorithm converges, the output membership coefficient is ω ij ,ω ij express Fuzzy membership of the i-th class in c-mean, 0≤ω ij ≤1, where membership is an existing concept in fuzzy mathematics; the initialization formula is as follows; C 0 kj =ω kj Among them, Γ 0 kj represents the initial step Γ kj Numerical value; ω kj express The fuzzy membership of the kth class in the Gaussian distribution; μ represents the initial step k ; represents the initial step σ k ; In step 5, the Gaussian mixture distribution weight coefficient π is k Substitute the fuzzy multi-Gaussian model obtained in step 4 to obtain the Gaussian mixture distribution model Weight coefficient π k The conversion process is expressed as: Among them, π k is a non-negative weight coefficient, 2. The method for identifying acoustic environmental noise based on fuzzy multi-Gaussian clustering according to claim 1, characterized in that: The number of types of the Gaussian distribution is K, which has a value of 2, and the two types correspond to the acoustic background noise and the foreground sound, respectively.
3. The method for identifying acoustic environmental noise based on fuzzy multi-Gaussian clustering according to claim 1, characterized in that: In step 2, N(x, μ k , σ k ) is a normal distribution, expressed as: Among them, μ k represents data expectation; σ k Represents the standard deviation of the data.
4. The method for identifying acoustic environmental noise based on fuzzy multi-Gaussian clustering according to claim 1, characterized in that: In step 2, by updating Γ kj , to increase the training output of the fuzzy multi-Gaussian model Update Γ by kj : Update Γ by timing kj , increase The value of .
5. The method for identifying acoustic environmental noise based on fuzzy multi-Gaussian clustering according to claim 1, characterized in that: Update in step 3 The process includes: According to the Γ obtained in step 2 t kj , update the normal distribution and Parameters; by updating The value of is maximized at step t 6. The method for identifying acoustic environmental noise based on fuzzy multi-Gaussian clustering according to claim 1, characterized in that: In step 4, the following formula is used to determine The convergence characteristics of the value of over time: Δ<ε If Δ satisfies the condition, the algorithm meets the convergence characteristics, otherwise returns to step 2; The value of Δ is as follows: Where ε is the set threshold; K represents the number of Gaussian distribution types; and N represents the number of data.
7. The method for identifying acoustic environmental noise based on fuzzy multi-Gaussian clustering according to claim 1, characterized in that: The model number K of the Gaussian mixture distribution model is consistent with the number K in the fuzzy multi-Gaussian model, both of which are 2.
