Jiangsan silk bamboo flute technique identification method

Abstract

The invention discloses a Jiangsan silk bamboo flute technique identification method. The method comprises the following steps: receiving an audio signal; generating a time-frequency representation; extracting features from the time-frequency representation; inputting the extracted features into a pre-trained convolutional neural network; inputting the extracted features into a pre-trained recurrent neural network; inputting the output of the convolutional neural network and the recurrent neural network into a pre-trained hidden Markov model; and integrating the outputs of the convolutional neural network, the recurrent neural network and the hidden Markov model to generate a final technique recognition result. According to the method, the influence of different playing styles and environmental noise can be better adapted by using short-time Fourier transform and Mel-frequency cepstrum coefficients, and the adaptability and robustness of the model are enhanced by capturing the subtle changes of the bamboo flute technique, including the frequency fluctuation of tremolo, the pitch track of gliding and the strength change of beating sound, so that the method is more suitable for popularization and application. And the fineness and the accuracy of feature extraction are further improved.

Description

Technical Field

[0001] The invention relates to the field of audio signal processing, and in particular to a method for identifying Jiangnan silk and bamboo flute techniques. Background Art

[0002] Jiangnan Sizhu music is an important part of traditional Chinese music. As one of the main instruments, bamboo flute has rich and diverse playing techniques, including vibrato, glissando, and striking.

[0003] Traditional technique recognition mainly relies on manual listening and empirical judgment, which is not only inefficient but also difficult to guarantee accuracy. With the development of audio processing technology and machine learning algorithms, it has become possible to use computers to automatically recognize and analyze music techniques.

[0004] However, existing technologies still have certain limitations when dealing with the complex and varied Jiangnan silk and bamboo flute techniques, such as insufficient feature extraction and insufficient model generalization capabilities. Summary of the invention

[0005] The purpose of the present invention is to provide a method for identifying Jiangnan Sizhu bamboo flute techniques to solve the problems raised in the above background technology.

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

[0007] A method for identifying Jiangnan Sizhu bamboo flute techniques, comprising the following steps:

[0008] Step S1, receiving an audio signal of a Jiangnan Sizhu bamboo flute performance;

[0009] Step S2, performing short-time Fourier transform on the audio signal and generating a time-frequency representation;

[0010] Step S3, extracting features reflecting the rhythm, melody and harmony characteristics of Jiangnan Sizhu music, as well as features reflecting bamboo flute playing techniques, from the time-frequency representation, including frequency fluctuations of vibrato, pitch trajectory of glissando and force changes of striking notes;

[0011] Step S4, inputting the extracted features into a pre-trained convolutional neural network to identify melody and rhythm features in the audio signal;

[0012] Step S5, inputting the extracted features into a pre-trained recurrent neural network to identify the temporal dynamic characteristics in the audio signal;

[0013] Step S6, inputting the outputs of the convolutional neural network and the recurrent neural network into a pre-trained hidden Markov model to identify the conversion and dependency relationships between the Jiangnan silk and bamboo flute techniques;

[0014] Step S7, integrating the outputs of the convolutional neural network, the recurrent neural network and the hidden Markov model to generate the final technique recognition result.

[0015] Furthermore, in step S1, the audio signal of the Jiangnan Sizhu bamboo flute performance is captured by an audio acquisition device, and the collected audio signal is preprocessed, including but not limited to noise reduction and gain adjustment.

[0016] Furthermore, in step S2, the audio signal is processed using a short-time Fourier transform algorithm to generate a time-frequency diagram. The short-time Fourier transform algorithm converts the audio signal from the time domain to the time-frequency domain to provide distribution information of the signal in time and frequency;

[0017] Short-time Fourier transform converts the signal from the time domain to the time-frequency domain. The formula is as follows:

[0018] X(m,ω)=∑_{n=-∞}^{∞}x(n)w(n-mR)e^{-jωn}

[0019] Among them, X(m,ω) represents the time-frequency domain representation, x(n) represents the input signal, w(n) represents the window function, m represents the frame index, R represents the frame shift, and ω represents the angular frequency.

[0020] Further, in step S3, based on the frequency fluctuation of the vibrato, the pitch trajectory of the glissando, and the change in the intensity of the striking sound in the bamboo flute technique, a feature extraction algorithm is used, including Mel-frequency cepstral coefficients;

[0021] The Fourier transform converts the signal from the time domain to the frequency domain;

[0022] The Mel frequency cepstral coefficient converts the linear frequency domain to the Mel frequency domain. The formula is as follows:

[0023] m=2595log_{10}(1+f / 700)

[0024] Where m represents the Mel frequency and f represents the frequency;

[0025] The discrete cosine transform converts the energy spectrum in the Mel frequency domain to the cepstrum domain. The formula is as follows:

[0026] c_i=∑_{j=1}^{N}log(S_j)cos(i(j-0.5)π / N)

[0027] Among them, c_i represents the MFCC coefficient, S_j represents the output of the Mel filter bank, N represents the number of filters, and i represents the MFCC coefficient index.

[0028] Furthermore, in step S4, a pre-trained convolutional neural network model is used, the extracted features are used as input, and forward propagation is performed to obtain melody and rhythm recognition results, and a large amount of labeled audio data is used to train the convolutional neural network model;

[0029] The convolutional neural network model is used to identify the melody and rhythm features in the audio signal, including convolutional layers and pooling layers;

[0030] The formula for the convolutional layer is as follows:

[0031] y_{ij}=∑_{m}∑_{n}x_{i+m,j+n}w_{mn}+b

[0032] Among them, y_{ij} represents the output feature map, x_{i+m,j+n} represents the input feature map, w_{mn} represents the convolution kernel, and b represents the bias term;

[0033] The formula for maximum pooling is as follows:

[0034] y_{ij}=max_{m,n}(x_{i+m,j+n})

[0035] Among them, y_{ij} represents the output feature map after pooling, x_{i+m,j+n} represents the local area in the input feature map, and m,n represent the index in the pooling window;

[0036] The formula for average pooling is as follows:

[0037] y_{ij}=\frac{1}{K}∑_{m}∑_{n}x_{i+m,j+n}

[0038] Among them, y_{ij} represents the output feature map after pooling, x_{i+m,j+n} represents the local area in the input feature map, m, n represent the index in the pooling window, and K represents the size of the pooling window, that is, the number of elements in the window.

[0039] Furthermore, in step S5, a pre-trained recurrent neural network model is used, the extracted features are used as input, forward propagation is performed, and recognition results of temporal dynamic characteristics are obtained, and a large amount of labeled audio data is used to train the recurrent neural network model;

[0040] The recurrent neural network model is used to identify the temporal dynamic characteristics of audio signals. The formula of the recurrent layer is:

[0041] h_t=f(W_{hh}h_{t-1}+W_{xh}x_t+b_h)

[0042] Among them, h_t represents the hidden state at the current moment, h_{t-1} represents the hidden state at the previous moment, x_t represents the input at the current moment, W_{hh} represents the weight matrix from hidden state to hidden state, W_{xh} represents the weight matrix from input to hidden state, b_h represents the bias term, and f represents the activation function.

[0043] Furthermore, in step S6, a pre-trained hidden Markov model is used, and the outputs of the convolutional neural network model and the recurrent neural network model are used as inputs to perform a forward-backward algorithm to obtain the recognition results of the conversion and dependency relationship between the techniques;

[0044] Forward algorithm formula:

[0045] α_t(i)=∑_{j=1}^{N}α_{t-1}(j)a_{ji}b_i(o_t)

[0046] Among them, α_t(i) represents the forward probability of being in state i at time t, α_{t-1}(j) represents the forward probability of being in state j at time t-1, a_{ji} represents the state transition probability, b_i(o_t) represents the observation probability, and N represents the number of states;

[0047] Backward algorithm formula:

[0048] β_t(i)=∑_{j=1}^{N}a_{ij}b_j(o_{t+1})β_{t+1}(j)

[0049] Where β_t(i) is the backward probability of being in state i at time t and the observation sequence after time t+1, a_{ij} is the state transition probability from state i to state j, b_j(o_{t+1}) is the observation probability of observing o_{t+1} in state j, and β_{t+1}(j) is the backward probability of being in state j at time t+1.

[0050] Furthermore, in step S7, the outputs of multiple models are integrated, and the outputs of all models are integrated through weighted fusion to generate a final technique recognition result.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. The present invention can better adapt to the influence of different playing styles and environmental noise by using short-time Fourier transform and Mel-frequency cepstral coefficients. It captures the subtle changes in bamboo flute techniques, including the frequency fluctuation of vibrato, the pitch trajectory of glissando and the change in the force of striking the notes, thereby enhancing the adaptability and robustness of the model and further improving the refinement and accuracy of feature extraction.

[0053] 2. The present invention adopts convolutional neural network and recurrent neural network to effectively capture the spatial and temporal features in audio signals, while the hidden Markov model can better handle the conversion and dependency between techniques, thereby enhancing the generalization ability of the model and enabling it to adapt to the technique recognition needs of different performers and different repertoires. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 The present invention is a flow chart of a method for identifying Jiangnan Sizhu bamboo flute techniques. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] See also Figure 1 , the present invention provides a technical solution:

[0057] A method for identifying Jiangnan Sizhu bamboo flute techniques, comprising the following steps:

[0058] Step S1, receiving an audio signal of a Jiangnan Sizhu bamboo flute performance;

[0059] Preferably, the audio signal of the Jiangnan Sizhu bamboo flute performance is captured by an audio acquisition device, and the collected audio signal is preprocessed, including but not limited to noise reduction and gain adjustment.

[0060] Step S2, performing short-time Fourier transform on the audio signal and generating a time-frequency representation;

[0061] Preferably, the audio signal is processed using a short-time Fourier transform algorithm to generate a time-frequency diagram, and the short-time Fourier transform algorithm converts the audio signal from the time domain to the time-frequency domain to provide distribution information of the signal in time and frequency;

[0062] Short-time Fourier transform converts the signal from the time domain to the time-frequency domain. The formula is as follows:

[0063] X(m,ω)=∑_{n=-∞}^{∞}x(n)w(n-mR)e^{-jωn}

[0064] Among them, X(m,ω) represents the time-frequency domain representation, x(n) represents the input signal, w(n) represents the window function, m represents the frame index, R represents the frame shift, and ω represents the angular frequency.

[0065] Step S3, extracting features reflecting the rhythm, melody and harmony characteristics of Jiangnan Sizhu music, as well as features reflecting bamboo flute playing techniques, from the time-frequency representation, including frequency fluctuations of vibrato, pitch trajectory of glissando and force changes of striking notes;

[0066] Preferably, a feature extraction algorithm is used based on the frequency fluctuation of the vibrato, the pitch trajectory of the glissando, and the force change of the striking sound in the bamboo flute technique, including Mel-frequency cepstral coefficients;

[0067] The Fourier transform converts the signal from the time domain to the frequency domain;

[0068] The Mel frequency cepstral coefficient converts the linear frequency domain to the Mel frequency domain. The formula is as follows:

[0069] m=2595log_{10}(1+f / 700)

[0070] Where m represents the Mel frequency and f represents the frequency;

[0071] The discrete cosine transform converts the energy spectrum in the Mel frequency domain to the cepstrum domain. The formula is as follows:

[0072] c_i=∑_{j=1}^{N}log(S_j)cos(i(j-0.5)π / N)

[0073] Among them, c_i represents the MFCC coefficient, S_j represents the output of the Mel filter bank, N represents the number of filters, and i represents the MFCC coefficient index.

[0074] Step S4, inputting the extracted features into a pre-trained convolutional neural network to identify melody and rhythm features in the audio signal;

[0075] Preferably, a pre-trained convolutional neural network model is used, the extracted features are used as input, forward propagation is performed, melody and rhythm recognition results are obtained, and a large amount of labeled audio data is used to train the convolutional neural network model;

[0076] The convolutional neural network model is used to identify the melody and rhythm features in the audio signal, including convolutional layers and pooling layers;

[0077] The formula for the convolutional layer is as follows:

[0078] y_{ij}=∑_{m}∑_{n}x_{i+m,j+n}w_{mn}+b

[0079] Among them, y_{ij} represents the output feature map, x_{i+m,j+n} represents the input feature map, w_{mn} represents the convolution kernel, and b represents the bias term;

[0080] The formula for maximum pooling is as follows:

[0081] y_{ij}=max_{m,n}(x_{i+m,j+n})

[0082] Among them, y_{ij} represents the output feature map after pooling, x_{i+m,j+n} represents the local area in the input feature map, and m,n represent the index in the pooling window;

[0083] The formula for average pooling is as follows:

[0084] y_{ij}=\frac{1}{K}∑_{m}∑_{n}x_{i+m,j+n}

[0085] Among them, y_{ij} represents the output feature map after pooling, x_{i+m,j+n} represents the local area in the input feature map, m, n represent the index in the pooling window, and K represents the size of the pooling window, that is, the number of elements in the window.

[0086] Step S5, inputting the extracted features into a pre-trained recurrent neural network to identify the temporal dynamic characteristics in the audio signal;

[0087] Preferably, a pre-trained recurrent neural network model is used, with the extracted features as input.

[0088] Perform forward propagation to obtain the recognition results of temporal dynamic characteristics, and use a large amount of labeled audio data to train the recurrent neural network model;

[0089] The recurrent neural network model is used to identify the temporal dynamic characteristics of audio signals. The formula of the recurrent layer is:

[0090] h_t=f(W_{hh}h_{t-1}+W_{xh}x_t+b_h)

[0091] Among them, h_t represents the hidden state at the current moment, h_{t-1} represents the hidden state at the previous moment, x_t represents the input at the current moment, W_{hh} represents the weight matrix from hidden state to hidden state, W_{xh} represents the weight matrix from input to hidden state, b_h represents the bias term, and f represents the activation function.

[0092] Step S6, inputting the outputs of the convolutional neural network and the recurrent neural network into a pre-trained hidden Markov model to identify the conversion and dependency relationships between the Jiangnan silk and bamboo flute techniques;

[0093] Preferably, a pre-trained hidden Markov model is used, and the outputs of the convolutional neural network model and the recurrent neural network model are used as inputs to perform a forward-backward algorithm to obtain the recognition results of the conversion and dependency relationships between the techniques;

[0094] Forward algorithm formula:

[0095] α_t(i)=∑_{j=1}^{N}α_{t-1}(j)a_{ji}b_i(o_t)

[0096] Among them, α_t(i) represents the forward probability of being in state i at time t, α_{t-1}(j) represents the forward probability of being in state j at time t-1, a_{ji} represents the state transition probability, b_i(o_t) represents the observation probability, and N represents the number of states;

[0097] Backward algorithm formula:

[0098] β_t(i)=∑_{j=1}^{N}a_{ij}b_j(o_{t+1})β_{t+1}(j)

[0099] Where β_t(i) is the backward probability of being in state i at time t and the observation sequence after time t+1, a_{ij} is the state transition probability from state i to state j, b_j(o_{t+1}) is the observation probability of observing o_{t+1} in state j, and β_{t+1}(j) is the backward probability of being in state j at time t+1.

[0100] Step S7, integrating the outputs of the convolutional neural network, the recurrent neural network and the hidden Markov model to generate a final technique recognition result;

[0101] Preferably, the outputs of multiple models are integrated, and the outputs of all models are combined through weighted fusion to generate the final technique recognition result.

[0102] The specific implementation steps of the above embodiment are as follows:

[0103] Suppose that you need to identify a 2-second bamboo flute audio signal with a sampling frequency of 16kHz, and use STFT to analyze the time-frequency characteristics of this signal.

[0104] Select a window function, for example, a Hamming window, whose length is 512 samples, and calculate the Fourier transform of the window function:

[0105] w(n)=0.54-0.46\cos\left(\frac{2\pi n}{N-1}\right)

[0106] Where N is the window length, n is the sampling point index, the signal is framed, each frame length is 512 sampling points, the frame shift is 256 sampling points, the window function is applied to each frame, and then the Fourier transform is performed:

[0107] X(m,\omega)=\sum_{n=0}^{N-1}x(n+mR)w(n)e^{-j\omega n}

[0108] Where m is the frame index and R is the frame shift.

[0109] Based on the above, the time-frequency representation has been obtained through STFT, and then the MFCC features are extracted.

[0110] Compute the power spectrum:

[0111] S(k)=|X(m,k)|^2

[0112] Where k is the frequency index.

[0113] Apply a Mel filter bank with each filter center frequency evenly spaced on the Mel scale:

[0114] H_m(k)=\begin{cases}

[0115] 0&k <f(m-1)\\

[0116] \frac{kf(m-1)}{f(m)-f(m-1)}&f(m-1)\leq k\leq f(m)\\

[0117] \frac{f(m+1)-k}{f(m+1)-f(m)}&f(m)\leq k\leq f(m+1)\\

[0118] 0&k>f(m+1)

[0119] \end{cases}

[0120] Where f(m) is the center frequency of the mth filter.

[0121] Compute the energy of each filter:

[0122] E_m=\sum_{k=0}^{N-1}S(k)H_m(k)

[0123] Logarithmic compression:

[0124] \log E_m=\log(E_m)

[0125] Apply the discrete cosine transform:

[0126] c_i=\sum_{m=1}^{M}\log E_m\cos\left(\frac{i(m-0.5)\pi}{M}\right)

[0127] Where M is the number of filters and c_i is the i-th MFCC coefficient.

[0128] Based on the MFCC features extracted above, CNN is used to identify melody and rhythm features.

[0129] Design the convolution layer, for example, use 32 convolution kernels, each of size 3x3.

[0130] Convolution operation:

[0131] y_{ij}=\sum_{m}\sum_{n}x_{i+m,j+n}w_{mn}+b

[0132] Among them, x is the input feature map, w is the convolution kernel, and b is the bias term.

[0133] Apply the activation function ReLU:

[0134] y_{ij}=\max(0,y_{ij})

[0135] Pooling operation, maximum pooling:

[0136] y_{ij}=\max(x_{i:i+k,j:j+k})

[0137] Here, k is the pooling window size.

[0138] Based on the feature representation obtained by CNN, RNN is used to identify temporal dynamic characteristics.

[0139] Design the recurrent layer and use LSTM units.

[0140] Forward propagation:

[0141] h_t=\text{LSTM}(h_{t-1},x_t)

[0142] Among them, h_t is the hidden state at the current moment, and x_t is the input at the current moment.

[0143] Apply output layer, fully connected layer:

[0144] y_t=W_{hy}h_t+b_y

[0145] Among them, W_{hy} is the weight matrix and b_y is the bias term.

[0146] Based on the feature representation obtained through CNN and RNN, HMM is now used to identify the conversion and dependency between techniques.

[0147] Define the state set and observation set.

[0148] Initialize the forward probability:

[0149] \alpha_1(i)=\pi_i b_i(o_1)

[0150] Among them, \pi_i is the initial state probability and b_i(o_1) is the initial observation probability.

[0151] Recursively calculate the forward probability:

[0152] \alpha_t(i)=\sum_{j=1}^{N}\alpha_{t-1}(j)a_{ji}b_i(o_t)

[0153] Among them, a_{ji} is the state transition probability and b_i(o_t) is the observation probability.

[0154] Calculate the final probability:

[0155] P(O|\lambda)=\sum_{i=1}^{N}\alpha_T(i)

[0156] Where O is the observation sequence and \lambda is the model parameter.

[0157] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0158] Although 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 the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying Jiangnan Sizhu bamboo flute techniques, characterized in that: The following steps are involved: Step S1, receiving an audio signal of a Jiangnan Sizhu bamboo flute performance; Step S2, performing short-time Fourier transform on the audio signal and generating a time-frequency representation; Step S3, extracting features reflecting the rhythm, melody and harmony characteristics of Jiangnan Sizhu music, as well as features reflecting bamboo flute playing techniques, from the time-frequency representation, including frequency fluctuations of vibrato, pitch trajectory of glissando and force changes of striking notes; Step S4, inputting the extracted features into a pre-trained convolutional neural network to identify melody and rhythm features in the audio signal; Step S5, inputting the extracted features into a pre-trained recurrent neural network to identify the temporal dynamic characteristics in the audio signal; Step S6, inputting the outputs of the convolutional neural network and the recurrent neural network into a pre-trained hidden Markov model to identify the conversion and dependency relationships between the Jiangnan silk and bamboo flute techniques; Step S7, integrating the outputs of the convolutional neural network, the recurrent neural network and the hidden Markov model to generate the final technique recognition result.

2. The method for identifying Jiangnan Sizhu bamboo flute techniques according to claim 1, characterized in that: In step S1, the audio signal of the Jiangnan Sizhu bamboo flute performance is captured by an audio acquisition device, and the collected audio signal is preprocessed, including but not limited to noise reduction and gain adjustment.

3. The method for identifying Jiangnan Sizhu bamboo flute techniques according to claim 1, characterized in that: In step S2, the audio signal is processed using a short-time Fourier transform algorithm to generate a time-frequency diagram. The short-time Fourier transform algorithm converts the audio signal from the time domain to the time-frequency domain to provide distribution information of the signal in time and frequency; Short-time Fourier transform converts the signal from the time domain to the time-frequency domain. The formula is as follows: X(m,ω)=∑_{n=-∞}^{∞}x(n)w(n-mR)e^{-jωn} Among them, X(m,ω) represents the time-frequency domain representation, x(n) represents the input signal, w(n) represents the window function, m represents the frame index, R represents the frame shift, and ω represents the angular frequency.

4. The method for identifying Jiangnan Sizhu bamboo flute techniques according to claim 1, characterized in that: In step S3, based on the frequency fluctuation of the vibrato, the pitch trajectory of the glissando, and the change in the intensity of the striking sound in the bamboo flute technique, a feature extraction algorithm is used, including Mel-frequency cepstral coefficients; The Fourier transform converts the signal from the time domain to the frequency domain; The Mel frequency cepstral coefficient converts the linear frequency domain to the Mel frequency domain. The formula is as follows: m=2595log_{10}(1+f / 700) Where m represents the Mel frequency and f represents the frequency; The discrete cosine transform converts the energy spectrum in the Mel frequency domain to the cepstrum domain. The formula is as follows: c_i=∑_{j=1}^{N}log(S_j)cos(i(j-0.5)π / N) Among them, c_i represents the MFCC coefficient, S_j represents the output of the Mel filter bank, N represents the number of filters, and i represents the MFCC coefficient index.

5. The method for identifying Jiangnan Sizhu bamboo flute techniques according to claim 1, characterized in that: In step S4, a pre-trained convolutional neural network model is used, the extracted features are used as input, forward propagation is performed, melody and rhythm recognition results are obtained, and a large amount of labeled audio data is used to train the convolutional neural network model; The convolutional neural network model is used to identify the melody and rhythm features in the audio signal, including convolutional layers and pooling layers; The formula for the convolutional layer is as follows: y_{ij}=∑_{m}∑_{n}x_{i+m,j+n}w_{mn}+b Among them, y_{ij} represents the output feature map, x_{i+m,j+n} represents the input feature map, w_{mn} represents the convolution kernel, and b represents the bias term; The formula for maximum pooling is as follows: y_{ij}=max_{m,n}(x_{i+m,j+n}) Among them, y_{ij} represents the output feature map after pooling, x_{i+m,j+n} represents the local area in the input feature map, and m,n represent the index in the pooling window; The formula for average pooling is as follows: y_{ij}=\frac{1}{K}∑_{m}∑_{n}x_{i+m,j+n} Among them, y_{ij} represents the output feature map after pooling, x_{i+m,j+n} represents the local area in the input feature map, m, n represent the index in the pooling window, and K represents the size of the pooling window, that is, the number of elements in the window.

6. The method for identifying Jiangnan Sizhu bamboo flute techniques according to claim 1, characterized in that: In step S5, a pre-trained recurrent neural network model is used, the extracted features are taken as input, forward propagation is performed, and recognition results of temporal dynamic characteristics are obtained, and a large amount of labeled audio data is used to train the recurrent neural network model; The recurrent neural network model is used to identify the temporal dynamic characteristics of audio signals. The formula of the recurrent layer is: h_t=f(W_{hh}h_{t-1}+W_{xh}x_t+b_h) Among them, h_t represents the hidden state at the current moment, h_{t-1} represents the hidden state at the previous moment, x_t represents the input at the current moment, W_{hh} represents the weight matrix from hidden state to hidden state, W_{xh} represents the weight matrix from input to hidden state, b_h represents the bias term, and f represents the activation function.

7. The method for identifying Jiangnan Sizhu bamboo flute techniques according to claim 1, characterized in that: In step S6, a pre-trained hidden Markov model is used, and the outputs of the convolutional neural network model and the recurrent neural network model are used as inputs to perform a forward-backward algorithm to obtain the recognition results of the conversion and dependency relationship between the techniques; Forward algorithm formula: α_t(i)=∑_{j=1}^{N}α_{t-1}(j)a_{ji}b_i(o_t) Among them, α_t(i) represents the forward probability of being in state i at time t, α_{t-1}(j) represents the forward probability of being in state j at time t-1, a_{ji} represents the state transition probability, b_i(o_t) represents the observation probability, and N represents the number of states; Backward algorithm formula: β_t(i)=∑_{j=1}^{N}a_{ij}b_j(o_{t+1})β_{t+1}(j) Where β_t(i) is the backward probability of being in state i at time t and the observation sequence after time t+1, a_{ij} is the state transition probability from state i to state j, b_j(o_{t+1}) is the observation probability of observing o_{t+1} in state j, and β_{t+1}(j) is the backward probability of being in state j at time t+1.

8. The method for identifying Jiangnan Sizhu bamboo flute techniques according to claim 1, characterized in that: In step S7, the outputs of multiple models are integrated, and the outputs of all models are integrated through weighted fusion to generate a final technique recognition result.

Images