Beauty resonance feature extraction method

Abstract

The invention discloses a beautiful sound resonance feature extraction method, and relates to a beautiful singing method identification method, and the method comprises the steps: firstly, respectively collecting standard and non-standard beautiful sound high pitch sounding samples, and carrying out the pre-emphasis, framing and windowing of sound signals; secondly, converting a time domain signal into a frequency domain through discrete Fourier transform, retaining an amplitude spectrum, using a plurality of Mel filters formed by triangular band-pass filters, converting linear frequency into non-linearly distributed Mel frequency, calculating logarithmic energy output by each filter bank, and performing dimensionality reduction on an obtained Mel frequency cepstrum coefficient by using a ReliefF algorithm; secondly, performing feature classification by using a support vector machine; by comparing the sound production signals of the beauty sound teacher and the beauty sound beginner, scientific sound production and wrong sound production are effectively distinguished, the recognition performance of the system is improved, and an auxiliary teaching effect on beauty sound production training is achieved.

Description

Technical Field

[0001] The invention relates to a bel canto singing recognition method, and in particular to a bel canto resonance feature extraction method. Background Art

[0002] Bel Canto is a beautiful resonant singing method that is different from other music genres. Singers are required to combine the resonance of the head cavity, chest cavity, oral cavity and nasal cavity as much as possible when vocalizing, so as to achieve a balance between the upper and lower resonances. The combination of these cavities makes Bel Canto round and full. At the same time, it is also the key factor that distinguishes Bel Canto from other music genres such as pop and folk.

[0003] In the bel canto vocal training, beginners often cannot use each resonance cavity well, resulting in vocal errors. Without the guidance of a teacher, it is difficult for students to accurately judge vocal errors through their own singing, and they cannot adjust the wrong resonance position in time. At present, there is no vocal teaching system that can judge the resonance cavity errors.

[0004] Therefore, studying the use of resonance chambers in bel canto can not only improve the efficiency of vocal training, but also has important significance for promoting the cross-development of singing training and signal processing technology. Summary of the invention

[0005] The purpose of the present invention is to provide a method for extracting bel canto resonance features. The method extracts the Mel-frequency cepstrum coefficients by comparing the standard and non-standard bel canto soprano vocalization signals, effectively distinguishes scientific vocalization from incorrect vocalization, and plays an auxiliary teaching role in bel canto vocalization training.

[0006] The objective of the present invention is achieved through the following technical solutions:

[0007] A method for extracting beautiful voice resonance features, the method comprising the following steps:

[0008] Step 1: Collect high-pitched signals from a certain bel canto beginner and teacher, perform endpoint cutting on the collected sound signals, remove the noise and silence segments in the sound samples, and obtain interference-free bel canto signals.

[0009] Step 2: The time domain signal is converted into the frequency domain by discrete Fourier transform and the amplitude spectrum is preserved. The linear frequency is converted into nonlinearly distributed Mel frequency using Mel filter. Discrete cosine transform (DCT) is applied to calculate the logarithmic energy of each filter group output.

[0010] Step 3: Use the ReliefF algorithm to reduce the dimension of the extracted Mel frequency cepstral coefficient feature parameters

[0011] Step 4: Use cross-validation and test set partitioning methods in the support vector machine to classify the data.

[0012] The step 2 further comprises:

[0013] 2.1 Discrete Fourier Transform:

[0014]

[0015] Where N is the number of fast Fourier transform points, S(n) is the framed signal, and the obtained P(n) is the signal energy spectrum.

[0016] 2.2 After the speech signal is transformed by fast Fourier transform, an energy spectrum composed of a complex matrix will be obtained. The amplitude spectrum is retained by squaring the energy spectrum and the amplitude spectrum is calculated:

[0017] P(n)=P 2 (n)

[0018] 2.3 Calculate the Mel filter. The Mel filter consists of M triangular bandpass filters with a center frequency of f m ,m=1,2,...,M usually takes the value of 22-26. This set of filters can convert linear frequency into nonlinearly distributed Mel frequency. The frequency response of the triangular filter is defined as:

[0019]

[0020] Compute the log energy of each filter bank output:

[0021]

[0022] 2.4 Obtained energy E m By performing discrete cosine transform, we can get the characteristic parameters of MFCC:

[0023]

[0024] The step 3 further comprises:

[0025] First, the prediction weight W of all factors is set to 0. Next, the algorithm randomly selects a sample R from the sample set through iteration, and then finds the k neighboring sample sets of R from the samples of the same class as R, set as H; finds the k neighboring sample sets from the samples of different classes from R, and defines them as M(C) according to the different classes C to which they belong, and updates the weight value of each feature:

[0026]

[0027]

[0028] Among them, diff(A,R1,R2) represents the difference between samples R1 and R2 on feature A, Mj (C) represents the jth nearest neighbor sample in the sample of different class from R, and m is the number of iterations.

[0029] Feature A is divided into two cases. If feature A is discrete, then:

[0030]

[0031] If feature A is continuous, then:

[0032]

[0033] Where max(F j ) and min(F j ) represent the maximum and minimum values ​​of feature A respectively.

[0034] The above weight updating process is iterated m times to obtain the weights of n features W = {w1, w2, ..., w n}, and arrange the features from large to small according to the weights, and we can get the order of the features’ ability to classify samples from strong to weak.

[0035] The step 4 further includes: using a support vector machine to classify the data. This study adopts a cross-validation method, using 80% of the data samples as a training set and the remaining 20% ​​as a test set. After 5-fold repeated cross-validation, the recognition rate and average value of all test samples are obtained.

[0036] The beneficial effects of the present invention are:

[0037] The present invention can clearly distinguish scientific and incorrect phonation by comparing standard and non-standard bel canto soprano vocalization signals and extracting Mel frequency cepstrum coefficients. This enables teachers to provide targeted guidance to students based on accurate judgment results during the bel canto vocalization training and teaching process, correct incorrect phonation habits, help students improve their phonation level faster and more effectively, and make teaching more scientific and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a flow chart of the present invention;

[0039] Figure 2 It is the time domain diagram of the original soprano signal in the present invention;

[0040] Figure 3 It is a time domain diagram of the soprano signal after preprocessing in the present invention;

[0041] Figure 4 It is the MFCC three-dimensional view of the vowel in the present invention;

[0042] Figure 5It is a line graph of component weights calculated by the ReliefF algorithm using the MFCC features obtained in the present invention. DETAILED DESCRIPTION

[0043] The present invention will be described in detail below with reference to the embodiments shown in the accompanying drawings.

[0044] A method for extracting beautiful voice resonance features, comprising the following steps:

[0045] Step 1: The first step is to collect soprano vocal signals from bel canto beginners who are in good health, have no cold symptoms, and have normal voices, as well as bel canto teachers with rich teaching experience in a music school. Through a strict screening process and scientific classification method, a series of unprocessed original soprano vocal signals are initially obtained, such as Figure 2 As shown in the figure, on this basis, the collected sound signals are preprocessed. For each sound sample, professional audio processing technology is used to accurately remove the noise and silence segments, so as to obtain interference-free beautiful sound signals, such as Figure 3 shown.

[0046] Step 2: The time domain signal is converted into the frequency domain by discrete Fourier transform and the amplitude spectrum is preserved. The linear frequency is converted into nonlinearly distributed Mel frequency using Mel filter. Discrete cosine transform (DCT) is applied to calculate the logarithmic energy of each filter group output.

[0047] 2.1 Discrete Fourier Transform:

[0048]

[0049] Where N is the number of fast Fourier transform points, S(n) is the framed signal, and the obtained P(n) is the signal energy spectrum.

[0050] 2.2 After the speech signal is transformed by fast Fourier transform, an energy spectrum composed of a complex matrix will be obtained. The amplitude spectrum is retained by squaring the energy spectrum and the amplitude spectrum is calculated:

[0051] P(n)=P 2 (n)

[0052] 2.3 Calculate the Mel filter. The Mel filter consists of M triangular bandpass filters with a center frequency of f m ,m=1,2,...,M usually takes the value of 22-26. This set of filters can convert linear frequency into nonlinearly distributed Mel frequency. The frequency response of the triangular filter is defined as:

[0053]

[0054] Compute the log energy of each filter bank output:

[0055]

[0056] 2.4 Obtained energy E m By performing discrete cosine transform, we can get the characteristic parameters of MFCC:

[0057]

[0058] The MFCC three-dimensional view of the soprano vocal signal is obtained as follows Figure 4 shown.

[0059] Step 3: First, the prediction weights W of all factors are set to 0. Next, the algorithm randomly selects a sample R from the sample set through iteration, and then finds the k neighboring sample sets of R from the samples of the same class as R, set as H; finds the k neighboring sample sets from the samples of different classes from R, and defines them as M(C) according to the different classes C to which they belong, and updates the weight value of each feature:

[0060]

[0061] Among them, diff(A,R1,R2) represents the difference between samples R1 and R2 on feature A, M j (C) represents the jth nearest neighbor sample in the sample of different class from R, and m is the number of iterations.

[0062] Feature A is divided into two cases. If feature A is discrete, then:

[0063]

[0064] If feature A is continuous, then:

[0065]

[0066] Where max(F j ) and min(F j ) represent the maximum and minimum values ​​of feature A respectively.

[0067] The above weight updating process is iterated m times to obtain the weights of n features W = {w1, w2, ..., w n}, and arrange the features from large to small according to the weights, and we can get the order of the features’ ability to classify samples from strong to weak.

[0068] The MFCC features of the soprano voice signal are calculated by the ReliefF algorithm, and the component weights are shown in the line graph. Figure 5 shown.

[0069] Step 4: Use support vector machine to classify the data. This study uses the cross-validation method, taking 80% of the data samples as the training set and the remaining 20% ​​as the test set. After 5-fold repeated cross-validation, the recognition rate and average value of all test samples are obtained.

Claims

1. A method for extracting beautiful voice resonance features, characterized in that: The method comprises the following steps: Step 1: Collect high-pitched signals from a certain bel canto beginner and teacher, perform endpoint cutting on the collected sound signals, remove the noise and silence segments in the sound samples, and obtain interference-free bel canto signals; Step 2: Convert the time domain signal to the frequency domain by discrete Fourier transform and preserve the amplitude spectrum, use Mel filter to convert the linear frequency into nonlinearly distributed Mel frequency, and apply discrete cosine transform (DCT) to calculate the logarithmic energy of each filter group output; Step 3: Use the ReliefF algorithm to reduce the dimension of the extracted Mel frequency cepstral coefficient feature parameters; Step 4: Use cross-validation and test set partitioning methods in the support vector machine to classify the data.

2. A method for extracting beautiful voice resonance features according to claim 1, characterized in that: In step 2, the operation of squaring the energy spectrum and retaining the amplitude spectrum highlights the amplitude information of the signal in the frequency domain, calculates the logarithmic energy of each filter group output, and uses the obtained logarithmic energy as an effective feature representation to reflect the distribution of signal energy in different frequency bands and capture the frequency characteristics.

3. A method for extracting beautiful voice resonance features according to claim 1, characterized in that: The implementation process of step 2 is as follows: 2.1 Discrete Fourier Transform: Where N is the number of fast Fourier transform points, S(n) is the framed signal, and the obtained P(n) is the signal energy spectrum; 2.2 After the speech signal is fast Fourier transformed, an energy spectrum consisting of a complex matrix is ​​obtained. The amplitude spectrum is retained by taking the square of the energy spectrum and the amplitude spectrum is calculated: P(n)=P 2 (n) 2.3 Calculate the Mel filter. The Mel filter consists of M triangular bandpass filters with a center frequency of f m ,m=1,2,...,M usually takes the value of 22-26. This set of filters can convert linear frequency into nonlinearly distributed Mel frequency; the frequency response of the triangular filter is defined as: Compute the log energy of each filter bank output: 2.4 Obtained energy E m By performing discrete cosine transform, we can get the characteristic parameters of MFCC:

4. The method for extracting the characteristics of the beautiful voice resonance according to claim 1, characterized in that: In step 3, the ReliefF algorithm is used as a feature selection algorithm. By calculating the weights of feature parameters, features with large weights are selected from high-dimensional feature samples, and features with small weights are removed, thereby reducing the feature dimension. The process includes: First, the prediction weight W of all factors is set to 0. Next, the algorithm randomly selects a sample R from the sample set through iteration, and then finds the k neighboring sample sets of R from the samples of the same class as R, set as H; finds the k neighboring sample sets from the samples of different classes from R, and defines them as M(C) according to the different classes C to which they belong, and updates the weight value of each feature: Among them, diff(A,R1,R2) represents the difference between samples R1 and R2 on feature A, M j (C) represents the jth nearest neighbor sample among samples of different classes from R, where m is the number of iterations; Feature A is divided into two cases. If feature A is discrete, then: If feature A is continuous, then: Where max(F j ) and min(F j ) represent the maximum and minimum values ​​of feature A respectively; The above weight updating process is iterated m times to obtain the weights of n features W = {w1, w2, ..., w n }, and arrange the features from large to small according to the weights, and we can get the order of the features’ ability to classify samples from strong to weak.

5. The method for extracting the characteristics of the beautiful voice resonance according to claim 1, characterized in that: In step 4, the support vector machine classifies the data by using a cross-validation method, taking 80% of the data samples as a training set and the remaining 20% ​​as a test set, and performing 5-fold repeated cross-validation to obtain the recognition rate and average value of all test samples.

Images