A method for extracting characteristic wavelet coefficients of English long vowel u based on spline interpolation wavelet neural network

Abstract

The present invention discloses a method for extracting characteristic wavelet coefficients of the English long vowel u based on a spline interpolation wavelet neural network. The method first normalizes and thresholds the audio signal to effectively remove environmental noise interference. Secondly, a three-layer neural network feedback matrix based on a sixth-order spline wavelet function is constructed, and the feedback matrix is generated by transposing the global matrix Ψ and calculating the inverse matrix. Then, a criterion function containing frequency domain errors is constructed using an inverse discrete Fourier transform. Finally, the output layer weights are dynamically adjusted through iterative training. When the criterion function modulus is less than the training error iteration end condition, the training is terminated to obtain a set of wavelet coefficients representing the characteristics of the long vowel u. The present invention innovatively combines the frequency domain localization characteristics of the sixth-order spline wavelet with the adaptive learning of the neural network, improving the noise resistance performance while ensuring feature extraction accuracy, and solving the problem of individual pronunciation differences.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal analysis, and in particular relates to a method for extracting characteristic wavelet coefficients of an English long vowel u based on a spline interpolation wavelet neural network. Background Art

[0002] The English long vowel "U" is an important component of English pronunciation, so extracting the audio signal features of "U" is a key step in developing English speech recognition technology. Because wavelet energy is highly concentrated in the time-frequency plane, using wavelet signal processing technology to extract the signal features of "U" has become one of the most important research directions in speech recognition.

[0003] However, traditional wavelet U sound extraction technology has the following two shortcomings: 1. The U audio signal usually exists in the form of discrete data, and the classic wavelet transform requires the application of continuous integration, which greatly limits the resolution of the extracted wavelet coefficients (the distribution characteristics of the wavelet coefficients represent the signal characteristics of the U sound); 2. The pronunciation of the U sound varies for different individuals, which makes the classic wavelet extraction of the U sound signal characteristics unstable. Summary of the Invention

[0004] The present invention aims to provide a method for recognizing and synthesizing the English long vowel u based on spline interpolation wavelet and multi-resolution analysis with low computational complexity, high transformation accuracy, high resolution and high stability, and to solve the problem of individual pronunciation differences by using the adaptive technology of neural networks.

[0005] To achieve the purpose of the present invention, the present invention provides a method for extracting characteristic wavelet coefficients of English long vowel u based on spline interpolation wavelet neural network, comprising the following steps:

[0006] Step 1: Preprocessing of audio signal array: obtain the analog audio signal array of long vowel u by sampling, normalize it, and then perform threshold processing to obtain the preprocessed audio signal array ;

[0007] Step 1.1: Sample the analog audio signal of the long vowel u to obtain a quantized audio signal array. ;in, Represents the nth quantized audio signal array element, N represents the length of the quantized audio signal array;

[0008] Step 1.2: quantize the audio signal array Perform normalization processing to obtain a normalized audio signal array ;in, Represents the normalized nth audio signal array element;

[0009] Step 1.3: If the modulus of the nth normalized audio signal array element is , then let ,otherwise, Remain unchanged, thus obtaining the audio signal array after threshold processing ;in, represents the kth audio signal array element after threshold processing, and T represents the threshold.

[0010] Step 2: Construct feedback matrix: Use sixth-order spline wavelet function As the hidden neurons of the three-layer wavelet neural network, Construct the row vector global matrix of the sixth-order spline wavelet function as vector-valued matrix elements , by global matrix row vector Transpose and perform matrix multiplication with itself to obtain a square matrix , then the square matrix Inverse, and finally obtain the wavelet neural network feedback matrix through matrix composite operation ;

[0011] Step 2.1: Set up a three-layer wavelet neural network. The wavelet function space of the audio signal features to be extracted is ,in, For the spatial layer number of wavelet function decomposition in the process of signal feature extraction, set one by one , and determine the input layer weight coefficient as ,in, and is the time shift of the wavelet function, which is determined by the length of the recorded audio time: set the recorded audio time length to ,but and The value of is: , ;

[0012] Step 2.2, using the sixth-order spline wavelet function As the hidden neurons of the wavelet neural network, for The inverse Fourier transform of the frequency domain expression, Indicates frequency;

[0013] Step 2.3: Construct the sixth-order spline wavelet function No. vector-valued matrices :

[0014] ;

[0015] in, is the number of spatial layers of wavelet function decomposition, That is, the length of the audio signal array; here the vector-valued matrix element ,in, is the input variable factor number, is the time shift of the wavelet function;

[0016] Step 2.4: Using the vector value matrix , construct the sixth-order spline wavelet function global matrix ;

[0017] Step 2.5, according to the sixth-order spline wavelet function global matrix , by the sixth-order spline wavelet function global matrix Transpose to get , after transposition, Matrix multiplication with itself , and then invert the result ,at last Then with the global matrix Perform matrix multiplication and multiply by the coefficient 0.5 to construct the wavelet neural network training feedback matrix .

[0018] Step 3: Construct the criterion function: Use the inverse discrete Fourier transform to calculate the frequency domain expression of the sixth-order spline wavelet function Perform inverse discrete Fourier transform to obtain coefficient vector , the audio signal array Construct audio signal array vector by ordered arrangement , set the output layer weight matrix , by the audio signal array vector , output layer weight matrix and the sixth-order spline wavelet function global matrix Perform matrix multiplication and subtraction to get the error on the training set , the error Transpose With coefficient vector Perform convolution to obtain the frequency domain error matrix of the criterion function , and finally the frequency domain error matrix of the criterion function Perform matrix transposition and matrix multiplication to obtain the criterion function ;

[0019] Step 3.1: Use the sixth-order spline wavelet function Frequency domain expression of Through the expression With coefficient The discrete Fourier transform expression of The relationship between the wavelet frequency domain is:

[0020] ;

[0021] in, is the sixth-order spline wavelet function The frequency domain expression of the Fourier transform is: is the imaginary unit, Indicates frequency, is an integer,

[0022] Step 3.2: Calculate the wavelet frequency domain coefficients using inverse discrete Fourier transform :

[0023] ;

[0024] in, , here take , and obtain the coefficient , and construct the coefficient vector , , is an integer;

[0025] Step 3.3: Using the audio signal array , construct the audio signal array vector , Indicates the length of the audio signal array;

[0026] Step 3.4: Determine the output layer weights The matrix expression of is:

[0027] ;

[0028] in, is the output layer weight matrix The transpose of is the output layer weight matrix, For the The output layer weight at this moment, For the The output layer weight at this moment, for Output layer weight at each moment; and is the time shift of the wavelet function;

[0029] Step 3.5: Use the audio signal array vector Transpose Subtract the sixth-order spline wavelet function global matrix With the output layer weight matrix The vector product of , obtains the error on the training set ;

[0030] Step 3.6: The error on the training set Transpose and coefficient vector Convolution operation , and obtain the criterion function frequency domain error matrix , and then pass the criterion function frequency domain error matrix Transpose The product of Get the criterion function :

[0031] ;

[0032] .

[0033] Step 4: Train the output layer weight vector: through the criterion function Training wavelet neural network training feedback matrix The final output layer weight is determined by the training error iteration end condition , complete the extraction of characteristic wavelet coefficients of English long vowel u;

[0034] Step 4.1: Convert the audio signal array into a vector and wavelet neural network training feedback matrix Substitute into the training iteration process: pass the audio signal array vector Transpose and the global matrix of the sixth-order spline wavelet function Hedi Step output layer weight matrix The product of the difference is the first Step error , among which Step output layer weight matrix By Step output layer weight matrix and wavelet neural network training feedback matrix With the Step error The product of is obtained by summing, as shown in the following formula:

[0035] ;

[0036] in, Represents the error matrix on the training set In the The value in the training step iteration, Represents the output layer weight The output layer weight matrix is expressed in the matrix expression In the The value in the step iteration training;

[0037] Step 4.2, using The training set error in the step-iterative training Calculate the frequency domain error matrix of the criterion function through convolution operation ,Right now ;

[0038] Step 4.3: Use the criterion function frequency domain error matrix Calculation criteria function ,Right now ;

[0039] Step 4.4: If the modulus of the criterion function , the training stops;

[0040] Step 4.5: When training stops, obtain the output layer weights of the wavelet neural network As the final training result , and Expressed as:

[0041] ;

[0042] in The first characteristic wavelet coefficients, The first characteristic wavelet coefficients, The first characteristic wavelet coefficients.

[0043] Compared with the existing technology, the significant progress of the present invention is that: the present invention solves the shortcomings of traditional wavelet U sound extraction technology, uses interpolation wavelet technology to break through the limitation of wavelet coefficient resolution, and uses neural network adaptive technology to solve the problem of individual pronunciation differences, thereby reducing the amount of calculation, improving transformation accuracy, improving resolution, and improving stability.

[0044] In order to more clearly illustrate the functional characteristics and structural parameters of the present invention, further description is given below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0046] Figure 1 is a flow chart of the steps of the present invention;

[0047] Figure 2It is a three-layer wavelet neural network structure diagram of the present invention;

[0048] Figure 3 It is a graph of characteristic wavelet coefficients of long vowel u extracted by wavelet neural network of the present invention;

[0049] FIG4 (a) is a diagram of an analog signal of the English long vowel u collected by the present invention;

[0050] FIG4( b ) is a signal diagram reconstructed from characteristic wavelet coefficients of the English long vowel u according to the present invention. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0052] The invention discloses a method for extracting characteristic wavelet coefficients of English long vowel u based on spline interpolation wavelet neural network, combined with Figure 1 , including the following steps:

[0053] Step 1: Preprocessing of audio signal array: obtain the analog audio signal array of long vowel u by sampling, normalize it, and then perform threshold processing to obtain the preprocessed audio signal array ;

[0054] Step 1.1: Sample the analog audio signal of the long vowel u to obtain a quantized audio signal array. ;in, Represents the nth quantized audio signal array element, N represents the length of the quantized audio signal array;

[0055] Step 1.2: quantize the audio signal array Perform normalization processing to obtain a normalized audio signal array ;in, Represents the normalized nth audio signal array element;

[0056] Step 1.3: If the modulus of the nth normalized audio signal array element is , then let ,otherwise, Remain unchanged, thus obtaining the audio signal array after threshold processing ;in, represents the kth audio signal array element after threshold processing, and T represents the threshold.

[0057] Step 2: Combine Figure 2 , construct the feedback matrix: using the sixth-order spline wavelet function As the hidden neurons of the three-layer wavelet neural network, Construct the row vector global matrix of the sixth-order spline wavelet function as vector-valued matrix elements , by global matrix row vector Transpose and perform matrix multiplication with itself to obtain a square matrix , then the square matrix Inverse, and finally obtain the wavelet neural network feedback matrix through matrix composite operation ;

[0058] Step 2.1: Set up a three-layer wavelet neural network. The wavelet function space of the audio signal features to be extracted is ,in, For the spatial layer number of wavelet function decomposition in the process of signal feature extraction, set one by one , and determine the input layer weight coefficient as ,in, and is the time shift of the wavelet function, which is determined by the length of the recorded audio time: set the recorded audio time length to ,but and The value of is: , ;

[0059] Step 2.2, using the sixth-order spline wavelet function As the hidden neurons of the wavelet neural network, for The inverse Fourier transform of the frequency domain expression, Indicates frequency;

[0060] Step 2.3: Construct the sixth-order spline wavelet function No. vector-valued matrices :

[0061] ;

[0062] in, is the number of spatial layers of wavelet function decomposition, That is, the length of the audio signal array; here the vector-valued matrix element ,in, is the input variable factor number, is the time shift of the wavelet function;

[0063] Step 2.4: Using the vector value matrix , construct the sixth-order spline wavelet function global matrix ;

[0064] Step 2.5, according to the sixth-order spline wavelet function global matrix , by the sixth-order spline wavelet function global matrix Transpose to get , after transposition, Matrix multiplication with itself , and then invert the result ,at last Then with the global matrix Perform matrix multiplication and multiply by the coefficient 0.5 to construct the wavelet neural network training feedback matrix .

[0065] Step 3: Construct the criterion function: Use the inverse discrete Fourier transform to calculate the frequency domain expression of the sixth-order spline wavelet function Perform inverse discrete Fourier transform to obtain coefficient vector , the audio signal array Construct audio signal array vector by ordered arrangement , set the output layer weight matrix , by the audio signal array vector , output layer weight matrix and the sixth-order spline wavelet function global matrix Perform matrix multiplication and subtraction to get the error on the training set , the error Transpose With coefficient vector Perform convolution to obtain the frequency domain error matrix of the criterion function , and finally the frequency domain error matrix of the criterion function Perform matrix transposition and matrix multiplication to obtain the criterion function ;

[0066] Step 3.1: Use the sixth-order spline wavelet function Frequency domain expression of Through the expression With coefficient The discrete Fourier transform expression of The relationship between the wavelet frequency domain is:

[0067] ;

[0068] in, is the sixth-order spline wavelet function The frequency domain expression of the Fourier transform is: is the imaginary unit, Indicates frequency, is an integer,

[0069] Step 3.2: Calculate the wavelet frequency domain coefficients using inverse discrete Fourier transform :

[0070] ;

[0071] in, , here take , and obtain the coefficient , and construct the coefficient vector , , is an integer;

[0072] Step 3.3: Using the audio signal array , construct the audio signal array vector , Indicates the length of the audio signal array;

[0073] Step 3.4: Determine the output layer weights The matrix expression of is:

[0074] ;

[0075] in, is the output layer weight matrix The transpose of is the output layer weight matrix, For the The output layer weight at this moment, For the The output layer weight at this moment, for Output layer weight at each moment; and is the time shift of the wavelet function;

[0076] Step 3.5: Use the audio signal array vector Transpose Subtract the sixth-order spline wavelet function global matrix With the output layer weight matrix The vector product of , obtains the error on the training set ;

[0077] Step 3.6: The error on the training set Transpose and coefficient vector Convolution operation , and obtain the criterion function frequency domain error matrix , and then pass the criterion function frequency domain error matrix Transpose The product of Get the criterion function :

[0078] ;

[0079] .

[0080] Step 4: Train the output layer weight vector: through the criterion function Training wavelet neural network training feedback matrix The final output layer weight is determined by the training error iteration end condition , complete the extraction of characteristic wavelet coefficients of English long vowel u;

[0081] Step 4.1: Convert the audio signal array into a vector and wavelet neural network training feedback matrix Substitute into the training iteration process: pass the audio signal array vector Transpose and the global matrix of the sixth-order spline wavelet function Hedi Step output layer weight matrix The product of the difference is the first Step error , among which Step output layer weight matrix By Step output layer weight matrix and wavelet neural network training feedback matrix With the Step error The product of is obtained by summing, as shown in the following formula:

[0082] ;

[0083] in, Represents the error matrix on the training set In the The value in the training step iteration, Represents the output layer weight The output layer weight matrix is expressed in the matrix expression In the The value in the step iteration training;

[0084] Step 4.2, using The training set error in the step-iterative training Calculate the frequency domain error matrix of the criterion function through convolution operation ,Right now ;

[0085] Step 4.3: Use the criterion function frequency domain error matrix Calculation criteria function ,Right now ;

[0086] Step 4.4: If the modulus of the criterion function , the training stops;

[0087] Step 4.5: When training stops, obtain the output layer weights of the wavelet neural network As the final training result , and Expressed as:

[0088] ;

[0089] in The first characteristic wavelet coefficients, The first characteristic wavelet coefficients, The first characteristic wavelet coefficients.

[0090] After steps 1 to 4, we can get the result graph, as shown in the following figure: Figure 3 As shown, the pre-processed English long vowel u phoneme signal is processed by a spline wavelet neural network, and the wavelet coefficients of the phoneme signal can be extracted, as shown in Figure 4 (b). After the wavelet coefficients are reconstructed in the scale space, an accurate approximation result can be obtained. The original signal in Figure 4 (a) is the analog signal of the English long vowel u collected by the present invention, and the reconstructed signal in Figure 4 (b) is the signal reconstructed by the characteristic wavelet coefficients of the English long vowel u of the present invention, which proves that the characteristic wavelet coefficients of the English long vowel u can completely reproduce the analog signal of the collected English long vowel u.

[0091] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0092] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for extracting characteristic wavelet coefficients of English long vowel u based on spline interpolation wavelet neural network, characterized in that: The following steps are involved: Step 1: Preprocessing of audio signal array: obtain the analog audio signal array of long vowel u by sampling, normalize it, and then perform threshold processing to obtain the preprocessed audio signal array ; Step 2: Construct feedback matrix: Use sixth-order spline wavelet function As the hidden neurons of the three-layer wavelet neural network, Construct the row vector global matrix of the sixth-order spline wavelet function as vector-valued matrix elements , by global matrix row vector Transpose and perform matrix multiplication with itself to obtain a square matrix , then the square matrix Inverse, and finally obtain the wavelet neural network feedback matrix through matrix composite operation ; Step 3: Construct the criterion function: Use the inverse discrete Fourier transform to calculate the frequency domain expression of the sixth-order spline wavelet function Perform inverse discrete Fourier transform to obtain coefficient vector , the audio signal array Construct audio signal array vector by ordered arrangement , set the output layer weight matrix , by the audio signal array vector , output layer weight matrix and the global matrix of the sixth-order spline wavelet function Perform matrix multiplication and subtraction to get the error on the training set , the error Transpose With coefficient vector Perform convolution to obtain the frequency domain error matrix of the criterion function , and finally the frequency domain error matrix of the criterion function Perform matrix transposition and matrix multiplication to obtain the criterion function ; Step 4: Train the output layer weight vector: through the criterion function Training wavelet neural network training feedback matrix The final output layer weight is determined by the training error iteration end condition , complete the extraction of characteristic wavelet coefficients of English long vowel u.

2. The method for extracting characteristic wavelet coefficients of English long vowel u based on spline interpolation wavelet neural network according to claim 1, characterized in that: The step 1 comprises the following steps: Step 1.1: Sample the analog audio signal of the long vowel u to obtain a quantized audio signal array. ;in, Represents the nth quantized audio signal array element, N represents the length of the quantized audio signal array; Step 1.2: quantize the audio signal array Perform normalization processing to obtain a normalized audio signal array ;in, Represents the normalized nth audio signal array element; Step 1.3: If the modulus of the nth normalized audio signal array element is , then let ,otherwise, Remain unchanged, thus obtaining the audio signal array after threshold processing ;in, represents the kth audio signal array element after threshold processing, and T represents the threshold.

3. The method for extracting characteristic wavelet coefficients of English long vowel u based on spline interpolation wavelet neural network according to claim 2, characterized in that: The step 2 comprises the following steps: Step 2.1: Set up a three-layer wavelet neural network. The wavelet function space of the audio signal features to be extracted is ,in, For the spatial layer number of wavelet function decomposition in the process of signal feature extraction, set one by one , and determine the input layer weight coefficient as ,in, and is the time shift of the wavelet function, which is determined by the length of the recorded audio time: set the recorded audio time length to ,but and The value of is: , ; Step 2.2, using the sixth-order spline wavelet function As the hidden neurons of the wavelet neural network, for The inverse Fourier transform of the frequency domain expression, Indicates frequency; Step 2.3: Construct the sixth-order spline wavelet function No. vector-valued matrices : ； in, is the number of spatial layers of wavelet function decomposition, That is, the length of the audio signal array; here the vector-valued matrix element ,in, is the input variable factor number, is the time shift of the wavelet function; Step 2.4: Using the vector value matrix , construct the sixth-order spline wavelet function global matrix ; Step 2.5, according to the sixth-order spline wavelet function global matrix , by the sixth-order spline wavelet function global matrix Transpose to get , after transposition, Matrix multiplication with itself , and then invert the result ,at last Then with the global matrix Perform matrix multiplication and multiply by the coefficient 0.5 to construct the wavelet neural network training feedback matrix .

4. The method for extracting characteristic wavelet coefficients of English long vowel u based on spline interpolation wavelet neural network according to claim 3, characterized in that: The step 3 comprises the following steps: Step 3.1: Use the sixth-order spline wavelet function Frequency domain expression of Through the expression With coefficient The discrete Fourier transform expression of The relationship between the wavelet frequency domain is: ； in, is the sixth-order spline wavelet function The frequency domain expression of the Fourier transform is: is the imaginary unit, Indicates frequency, is an integer, Step 3.2: Calculate the wavelet frequency domain coefficients using inverse discrete Fourier transform : ； in, , here take , and obtain the coefficient , and construct the coefficient vector , , is an integer; Step 3.3: Using the audio signal array , construct the audio signal array vector , Indicates the length of the audio signal array; Step 3.4: Determine the output layer weights The matrix expression of is: ； in, is the output layer weight matrix The transpose of is the output layer weight matrix, For the The output layer weight at this moment, For the The output layer weight at this moment, for Output layer weight at each moment; and is the time shift of the wavelet function; Step 3.5: Use the audio signal array vector Transpose Subtract the sixth-order spline wavelet function global matrix With the output layer weight matrix The vector product of , obtains the error on the training set ; Step 3.6: The error on the training set Transpose and coefficient vector Convolution operation , and obtain the criterion function frequency domain error matrix , and then pass the criterion function frequency domain error matrix Transpose The product of Get the criterion function : ； 。 5. The method for extracting characteristic wavelet coefficients of English long vowel u based on spline interpolation wavelet neural network according to claim 4, characterized in that: The step 4 specifically includes the following steps: Step 4.1: Convert the audio signal array into a vector and wavelet neural network training feedback matrix Substitute into the training iteration process: pass the audio signal array vector Transpose and the global matrix of the sixth-order spline wavelet function Hedi Step output layer weight matrix The product of the difference is the first Step error , among which Step output layer weight matrix By Step output layer weight matrix and wavelet neural network training feedback matrix With the Step error The product of is obtained by summing, as shown in the following formula: ； in, Represents the error matrix on the training set In the The value in the training step iteration, Represents the output layer weight The output layer weight matrix is expressed in the matrix expression In the The value in the step iteration training; Step 4.2, using The training set error in the step-iterative training Calculate the frequency domain error matrix of the criterion function through convolution operation ,Right now ; Step 4.3: Use the criterion function frequency domain error matrix Calculation criteria function ,Right now ; Step 4.4: If the modulus of the criterion function , the training stops; Step 4.5: When training stops, obtain the output layer weights of the wavelet neural network As the final training result , and Expressed as: ； in The first characteristic wavelet coefficients, The first characteristic wavelet coefficients, The first characteristic wavelet coefficients.

Images