Sound analysis method and system based on neural network, terminal and storage medium
Through neural network-based analysis methods, the acoustic parameters of the sound signal are extracted and converted into a Mel spectrogram, solving the problems of the existing acoustic acoustic analysis time and single diagnostic standards, achieving faster and more accurate acoustic analysis results.
Patent Information
- Application Number
- CN202510056638.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-14
- Publication Date
- 2025-04-25
AI Technical Summary
The existing acoustic acoustic analysis methods are time-consuming, single diagnostic standards, complex and costly.
Using a neural network-based analysis method, by obtaining sound signals, calculating acoustic parameters, and converting them into a meer spectrogram, the features are extracted using CNN neural network, and the final spectrogram is generated in combination with acoustic parameters.
Faster sound analysis is achieved, the results obtained are more accurate and intuitive, and the defects of single parameter or feature extraction are overcome.
Smart Images
Figure CN119889358A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of sound analysis, and in particular to a sound analysis method, system, terminal and storage medium based on a neural network. Background Art
[0002] At present, voice assessment mainly includes five aspects: subjective scoring, perceived sound assessment, dynamic stroboscopic laryngoscopy, aerodynamics and acoustic assessment. Among them, acoustic testing is one of the preferred methods for objective voice assessment because it is non-invasive, easy to obtain, repeatable and can provide quantitative evidence related to laryngeal function.
[0003] Sound acoustic detection can be used in various occasions, such as analyzing sound to obtain acoustic parameters and spectrograms for crime analysis, and observing the sound acoustic parameters and spectrogram characteristics of various laryngeal disease patients to determine the laryngeal lesions of patients. However, the current sound acoustic analysis is time-consuming and has a single diagnostic standard. Traditional sound acoustic analysis determines acoustic characteristics by analyzing sound acoustic parameters. This method is relatively complex and requires multiple tests on the sound, which is time-consuming, labor-intensive, and costly. Summary of the invention
[0004] The technical problem to be solved by the present invention is that the currently used sound acoustic analysis is time-consuming and has a single diagnostic standard. Traditional sound acoustic analysis determines acoustic features by analyzing sound acoustic parameters. This method is relatively complex and requires multiple detections of the sound, which is time-consuming, labor-intensive, and costly. In view of the above-mentioned defects of the prior art, a sound analysis method, system, terminal, and storage medium based on a neural network are provided.
[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:
[0006] Construct a sound analysis method based on neural network, including the following steps:
[0007] Acquire sound signals and save the sound signals;
[0008] Calculate acoustic parameters of sound signals;
[0009] The sound signal is converted into a Mel spectrogram, and features in the Mel spectrogram are extracted through a CNN neural network, and the extracted features are combined to obtain a processed Mel spectrogram;
[0010] The mel-spectrogram and acoustic parameters are combined to obtain the processed spectrogram.
[0011] Preferably, in the process of extracting features from the Mel-spectrogram by using the CNN neural network, the process further includes:
[0012] Features in the mel spectrogram are extracted through at least one set of convolutional layers, and the extracted features are downsampled through a pooling layer. The extracted downsampled features are then combined by a fully connected layer to obtain a processed mel spectrogram.
[0013] Preferably, the CNN neural network includes a first convolutional layer and a second convolutional layer, the first convolutional layer extracts a first feature of the mel spectrogram, the second convolutional layer extracts a second feature of the mel spectrogram, the pooling layer downsamples the first feature and the second feature and extracts features from the first feature and the second feature, and the fully connected layer combines the first feature, the second feature and the features extracted by the pooling layer to obtain a processed mel spectrogram.
[0014] Preferably, the proportion of the second feature in the mel spectrogram is greater than the proportion of the first feature in the mel spectrogram;
[0015] After the first convolution layer extracts the first feature from the mel spectrogram, the pooling layer downsamples the first feature and pools the mel spectrogram, and then the second convolution layer extracts the second feature from the pooled mel spectrogram, and the pooling layer downsamples the second feature again.
[0016] Preferably, in the process of extracting the features of the mel-spectrogram, the convolution layer further includes:
[0017] Introduce a nonlinear function into the convolution layer so that the convolution layer can be used as a nonlinear function for feature extraction.
[0018] Preferably, the CNN neural network further includes an input layer and an output layer, and the CNN neural network includes:
[0019] Some neurons will be paralyzed with a defined probability, and these neurons will not propagate, while the number of neurons in the input and output layers will remain unchanged, and then the parameters will be updated.
[0020] Preferably, the CNN neural network randomly paralyzes some neurons in the next iteration so that they do not participate in training, until the training is completed to obtain the trained CNN neural network, and the trained neural network is used to process the Mel spectrogram.
[0021] Construct a sound analysis system, including:
[0022] A sound collection module, used to obtain and save sound signals;
[0023] Acoustic parameter analysis module, used to analyze acoustic parameters in sound signals;
[0024] The acoustic diagnosis module converts the sound signal into a Mel-type spectrogram, uses a CNN neural network to extract features from the converted Mel-type spectrogram, and combines the extracted features to obtain a processed spectrogram;
[0025] The report generation module combines the acoustic parameters with the processed spectrogram to obtain the sound frequency graph.
[0026] A smart terminal is constructed, which includes at least one memory, at least one processor, and a sound analysis program stored in the memory and executable on the processor. When the sound analysis program is executed by the processor, the steps of the sound analysis method described above are implemented.
[0027] A computer-readable storage medium is constructed, wherein a sound analysis program is stored on the computer-readable storage medium, and when the sound analysis program is executed by a processor, the steps of the sound analysis method described above are implemented.
[0028] The beneficial effects of the present invention are: firstly obtaining the parameters of the sound signal, and at the same time converting the sound signal into a Mel spectrogram, and using a neural network of a CNN architecture to process the Mel spectrogram, firstly using a first convolution kernel to extract the first feature of the Mel spectrogram, and then using a pooling layer to downsample the first feature to reduce the spatial dimension of the first feature map, and extract the features in the first feature, thereby reducing the computational complexity; then using a second convolution kernel to extract the second feature of the Mel spectrogram, and then using a pooling layer to downsample the second feature to reduce the spatial dimension of the second feature map, and extract the features in the second feature, introducing nonlinearity into the convolution kernel during the extraction process to avoid gradient explosion, and then using a fully connected layer to filter and screen the previously extracted features and then combine them, thereby obtaining a processed spectrogram, and combining the spectrogram with the parameters to generate the final sound analysis result, which overcomes the defects of single parameter or feature extraction, can achieve faster sound analysis and quickly obtain analysis results, and the analysis results are more accurate and intuitive. Combining the neural network with other sound signal parameters, a sound analysis method for intelligently evaluating sound is obtained, thereby quickly obtaining results. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. The drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative work:
[0030] Figure 1 A flow chart of a sound analysis method according to a preferred embodiment of the present invention;
[0031] Figure 2 A schematic diagram of a flow chart of an acoustic diagnosis method according to a preferred embodiment of the present invention;
[0032] Figure 3 A schematic diagram of the structure of a CNN neural network architecture of a preferred embodiment of the present invention;
[0033] Figure 4 A schematic diagram of a sound analysis method according to a preferred embodiment of the present invention;
[0034] Figure 5 A schematic diagram of the structure of a sound analysis system according to a preferred embodiment of the present invention;
[0035] Figure 6 Schematic diagram of the structure of a smart terminal according to a preferred embodiment of the present invention. DETAILED DESCRIPTION
[0036] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the following will be described clearly and completely in combination with the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are partial embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the protection scope of the present invention.
[0037] A sound analysis method based on a neural network in a preferred embodiment of the present invention; Figure 1 As shown, it is a flowchart of a sound analysis method based on a neural network provided by an embodiment of the present invention. The method can be executed by a device, and the device can be implemented by software and / or hardware.
[0038] Specifically, in this embodiment, the sound analysis method based on neural network includes:
[0039] S1: Get sound signal;
[0040] In a preferred embodiment of the present invention, the sound signal to be analyzed can be obtained through a sound receiving device, such as a sound collector or a preset recorded and stored sound signal. When obtaining the sound signal, try to collect the sound of a specific tone that is continuously emitted in a relatively quiet environment and save it. For example, let the sound of "a" in Chinese pinyin be continuously emitted in a relatively quiet environment, collect this sound and save it as the sound signal to be analyzed. If it is in a relatively noisy environment, it is necessary to filter the sound to remove the noise in the sound to obtain a relatively pure sound signal to be analyzed and perform subsequent analysis.
[0041] S2: Calculate the acoustic parameters of the acquired sound signal;
[0042] In a preferred embodiment of the present invention, the sound signal to be analyzed needs to be processed to calculate different parameters in the sound signal to be analyzed, including fundamental frequency, fundamental frequency deviation, fundamental frequency perturbation, signal-to-noise ratio, harmonic noise ratio and normalized noise energy.
[0043] The fundamental frequency is the frequency of the fundamental tone in the sound signal, which usually corresponds to the frequency of the vocal cord vibration when speaking. The fundamental frequency deviation refers to the difference between the actual measured value and the theoretical calculated value of the fundamental frequency in the sound signal.
[0044] The fundamental frequency is often calculated using the autocorrelation method. The specific calculation steps are as follows:
[0045] First, use formula ① to calculate the autocorrelation function R(t) of the sound signal:
[0046]
[0047] Where x[n] is the audio signal, N is the length of the signal, and t is the time delay.
[0048] Then find the first peak position t of the autocorrelation function max , and the fundamental frequency F0 is calculated by formula ②,
[0049]
[0050] The fundamental frequency perturbation refers to the rate of change of the sound wave frequency between adjacent cycles. Its calculation steps are as follows:
[0051] First, the base frequency is detected using the autocorrelation method or FFT, wherein the autocorrelation method is the same as the aforementioned autocorrelation function R(t), and will not be repeated here.
[0052] The perturbation ratio is the ratio of the difference between the actual measured value and the theoretical calculated value, which is calculated using formula ③:
[0053]
[0054] The signal-to-noise ratio is used to evaluate the strength of the signal relative to the background noise, and is calculated using formula ④:
[0055] SNR=10×log10(Ps / Pn)④
[0056] Where: SNR stands for signal-to-noise ratio, Ps stands for signal power, and Pn stands for noise power.
[0057] Harmonic-to-noise ratio: also known as the signal-to-noise-distortion ratio, refers to the ratio of the effective value of the measured input signal to the total effective value of all noise and harmonic components (excluding DC components) below the Nyquist frequency.
[0058] The calculation requires preprocessing the sound signal, denoising and framing the sound signal, and then analyzing the signal. The signal is converted to the frequency domain using the Fast Fourier Transform (FFT), and then the main harmonic components and their frequency positions are identified. Harmonics and noise are then separated, and the frequency components are divided into harmonic parts and noise parts. Usually, regular frequency components are regarded as harmonics, while irregular parts are regarded as noise. The power of noise and harmonics is then calculated.
[0059] For harmonics, the calculation is:
[0060] P harmonic =∑ k |X(f k )| 2 ⑤
[0061] Among them, f k is the frequency of the harmonic;
[0062] For noise, it is calculated as:
[0063] P noise =∑ j |X(f j )| 2 ⑥
[0064] Among them, f j It's noise.
[0065] It can be concluded that the calculation method of harmonic noise ratio is:
[0066]
[0067] Normalized noise energy is the energy of the noise component calculated after taking into account the overall energy or power of the signal. By normalizing, you can better understand and analyze the relative impact of noise compared to the main components of the signal.
[0068] The calculation formula first needs to divide the sound signal into the signal part and the noise part. Usually, the main components of the sound signal can be extracted through a filter to separate the sound signal. Then the sound signal is short-time Fourier transform (STFT) and the spectrum is analyzed to identify the noise component. It includes the calculation of signal energy and noise energy, where the signal energy E s The calculation method is:
[0069]
[0070] Where x[n] is the signal sample and N is the signal length.
[0071] Noise Energy E n The calculation formula is:
[0072]
[0073] Where w[n] is the noise sample.
[0074] Finally, the normalized noise energy can be obtained by comparing the noise energy to the signal power:
[0075]
[0076] S3: convert the acquired sound signal into a Mel-spectrogram;
[0077] In a preferred embodiment of the present invention, the collected sound to be analyzed is converted into a corresponding mel spectrogram by a mel spectrogram method. The mel spectrogram can show the time domain and frequency domain information of the sound signal in one graph, thereby overcoming the defects of single parameter or feature extraction. The specific conversion of the sound signal into a mel spectrogram is a prior art, and this application does not elaborate on this part.
[0078] S4: Using a neural network to process the Mel-spectrogram image to obtain a processed spectrogram image;
[0079] In a preferred embodiment of the present invention, a neural network is used to process the converted Mel spectrogram to obtain different features of the sound, wherein a neural network with a CNN architecture is used to process the converted Mel spectrogram, and the neural network has an input layer, four convolutional layers, two pooling layers, two fully connected layers and an output layer. The converted Mel spectrogram is introduced through the input layer, and the features in the spectrogram are extracted by the convolutional layer, and then the extracted features are downsampled by the pooling layer to reduce the spatial dimension of the feature map, thereby reducing the computational complexity, and finally the extracted features are screened and combined by the fully connected layer to obtain a processed spectrogram image, and finally the processed spectrogram image is output by the output layer, and different features of the sound are combined to obtain a processed spectrogram image and output the result. In the learning process of the neural network, in order to prevent overfitting, a dropout is used, that is, some neurons are paralyzed by a defined probability, and this neuron does not participate in forward and backward propagation, just like being deleted in the network, while keeping the number of neurons in the input layer and the output layer unchanged, and then the parameters are updated according to the learning method of the neural network. In the next iteration, some neurons are randomly "paralyzed" so that they do not participate in training until the training is completed.
[0080] S5: combining the acoustic parameters and the processed spectrum image to obtain a sound frequency map and outputting it;
[0081] In a preferred embodiment of the present invention, the parameters in the sound signal are integrated with the processed sound spectrum image, and a specific sound frequency graph is displayed for output, so that the user can grasp the sound information more intuitively and accurately. The sound information can also be formed into a report for convenient observation and printing.
[0082] Through the above method, the neural network and other sound signal parameters are combined to obtain a sound analysis method for intelligent sound evaluation, so as to quickly obtain results.
[0083] Specifically, in the above embodiments, Figure 2 and Figure 3 As shown, the above step S4: using a neural network to process the Mel-spectrogram image to obtain a processed spectrogram image includes:
[0084] S40: extracting the first feature of the input mel-spectrogram image;
[0085] In a preferred embodiment of the present invention, since the size of the input Mel spectrogram is 512X512, two groups of 64 convolution kernels are first used to extract the first feature of the input Mel spectrogram. During the extraction process, a ReLU activation function is used in the convolution layer to introduce nonlinearity, converting linear features into nonlinear features to minimize gradient explosion.
[0086] Furthermore, each convolution is composed of multiple convolution kernels, and each convolution kernel has a parameter that needs to be adjusted continuously during the neural network training process, so the result will become y=ax+b. At this time, the convolution kernel is linear. Then when multiple convolution kernels appear in the convolution, multiple identical calculations are performed and accumulated. The result is y=a1x1+a2x2+…+a n x n +b1+b2+…+b n , and the neural network is still a linear model at this time, then adding nonlinearity to the neural network makes the neural network nonlinear to prevent the linearization effect. Other nonlinear functions can also be used to minimize gradient explosion.
[0087] S41: down-sampling the first feature to reduce the spatial dimension of the first feature;
[0088] In a preferred embodiment of the present invention, the maximum pooling is performed through the pooling layer, and the extracted first feature is downsampled to reduce the spatial dimension of the extracted first feature map, thereby reducing the computational complexity. At the same time, the pooling layer also has the function of further extracting features in the first feature and pooling the Mel spectrogram, that is, the feature range extracted by the convolution kernel after pooling becomes larger. For example, a 3X3 convolution kernel can only extract features within a 3X3 range in the Mel spectrogram before pooling. After the Mel spectrogram is pooled, it can extract features within a 6X6 range in the Mel spectrogram after pooling.
[0089] S42: extracting the second feature of the pooled Mel-spectrophone image;
[0090] In a preferred embodiment of the present invention, the converted mel spectrogram is further extracted using two groups of 32 convolution kernels to input the second feature of the mel spectrogram. The second feature is a larger feature structure than the first feature, that is, 64 convolution kernels extract small features in the mel spectrogram, but since some features are large, the 64 convolution kernels cannot extract the complete part of the feature. In this case, the pooling layer can be used to reduce the spatial dimension, and then the 32 convolution kernels are used to extract larger features so that the feature can be extracted by the 32 convolution kernels.
[0091] S43: down-sampling the second feature to reduce the spatial dimension of the second feature;
[0092] In a preferred embodiment of the present invention, the pooling layer performs maximum pooling to downsample the extracted second feature to reduce the spatial dimension of the extracted second feature map, thereby reducing the computational complexity. At the same time, the pooling layer also has the function of further extracting features from the second feature.
[0093] S44: combining the extracted first feature, the second feature and the downsampled feature to output a processed spectrum image;
[0094] In a preferred embodiment of the present invention, the extracted first feature and the second feature and the downsampled feature are screened and filtered through a fully connected layer, and then a higher level of abstraction and combination are performed to capture more complex patterns, and finally the processed spectrogram image is output after softmax separation.
[0095] Furthermore, in the process of feature extraction, each convolution kernel generates a feature layer. When multiple convolution kernels are used, multiple layers will appear. The fully connected layer is used to recombine multiple layers into one layer. Therefore, the features of each layer of the multi-dimensional image of the converted Mel spectrogram are extracted through the convolution layer and the pooling layer, and then the extracted features of each layer are expanded, filtered, and combined to form a two-dimensional image through the fully connected layer.
[0096] Through the above method, the parameters of the sound signal are first obtained, and the sound signal is converted into a Mel spectrogram, and the Mel spectrogram is processed using a neural network of a CNN architecture. The first convolution kernel is first used to extract the first feature of the Mel spectrogram, and then the first feature is downsampled using a pooling layer to reduce the spatial dimension of the first feature map, and the features in the first feature are extracted and the Mel spectrogram is pooled to reduce the computational complexity; then the second convolution kernel is used to extract the second feature of the pooled Mel spectrogram, and then the second feature is downsampled using a pooling layer to reduce the spatial dimension of the second feature map, and the features in the second feature are extracted. In the extraction process, nonlinearity is introduced into the convolution kernel to avoid gradient explosion, and then the fully connected layer is used to filter and screen the previously extracted features and then combine them to obtain a processed spectrogram. The spectrogram is combined with the parameters to generate the final sound analysis result, which overcomes the defects of single parameter or feature extraction, can achieve faster sound analysis and quickly obtain analysis results, and the analysis results are more accurate and intuitive. The neural network and other sound signal parameters are combined to obtain a sound analysis method for intelligently evaluating sound, so as to quickly obtain results.
[0097] Corresponding to a sound analysis method based on a neural network, the present invention also provides a sound analysis system based on a neural network, specifically, as Figure 3 and Figure 4 As shown, the neural network-based sound analysis system includes: a sound collection module 100 , an acoustic parameter analysis module 200 , an acoustic diagnosis module 300 and a report generation module 400 .
[0098] The sound collection module 100 is used to obtain and save sound signals;
[0099] Furthermore, the sound signal to be analyzed can be obtained through a sound receiving device, such as a sound collector or a preset recorded and stored sound signal, etc. When obtaining the sound signal, try to collect the sound of a specific tone that is continuously emitted in a relatively quiet environment and save it. For example, let the sound of "a" in Chinese pinyin be continuously emitted in a relatively quiet environment, collect this sound and save it as the sound signal to be analyzed.
[0100] The acoustic parameter analysis module 200 is used to analyze the acoustic parameters in the sound signal;
[0101] Furthermore, the sound signal to be analyzed needs to be processed to calculate different parameters in the sound signal to be analyzed, including fundamental frequency, fundamental frequency deviation, fundamental frequency perturbation, signal-to-noise ratio, harmonic noise ratio and normalized noise energy. The specific calculation method is the same as above and will not be repeated here.
[0102] The acoustic diagnosis module 300 converts the sound signal into a Mel-type spectrogram, and processes the converted Mel-type spectrogram using a CNN neural network to obtain a processed spectrogram;
[0103] Furthermore, the collected sound to be analyzed is converted into a corresponding Mel spectrogram by using a Mel spectrogram method. The Mel spectrogram can show the time domain and frequency domain information of the sound signal on a graph. Then, a neural network is used to process the converted Mel spectrogram to obtain different features of the sound. A neural network with a CNN architecture is used to process the converted Mel spectrogram. The neural network has an input layer, four convolutional layers, two pooling layers, two fully connected layers and an output layer. The converted Mel spectrogram is introduced through the input layer, and the features in the spectrogram are extracted by the convolutional layer. Then, the extracted features are downsampled by the pooling layer to reduce the spatial dimension of the feature map, thereby reducing the computational complexity. Finally, the extracted features are screened and combined by the fully connected layer to obtain a processed spectrogram image. Finally, the processed spectrogram image is output by the output layer, and the different features of the sound are combined to obtain a processed spectrogram image and the result is output.
[0104] The report generation module 400 combines the acoustic parameters with the processed spectrogram to obtain a sound frequency graph.
[0105] Furthermore, the parameters in the sound signal are integrated with the processed sound spectrum image, and a specific sound frequency graph is displayed for output, so that users can grasp the sound information more intuitively and accurately. The sound information can also be formed into a report for easy observation and printing.
[0106] Those skilled in the art will understand that Figure 6 The principle block diagram shown in the figure is only a block diagram of a partial structure related to the scheme of the present invention, and does not constitute a limitation on the smart terminal to which the scheme of the present invention is applied. The specific smart terminal may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0107] In an embodiment of the present invention, a smart terminal is provided. The smart terminal includes a memory, a processor, and a neural network-based sound analysis program stored in the memory and executable on the processor. When the neural network-based sound analysis program is executed by the processor, the following operation instructions are performed:
[0108] Acquire sound signals and save the sound signals;
[0109] Analyze acoustic parameters in sound signals;
[0110] Convert sound signals into Mel-spectrograms;
[0111] Introducing nonlinearity into the convolution kernel and extracting the first feature in the Mel-spectrogram image;
[0112] Down-sample the first feature to reduce the spatial dimension of the first feature, extract features from the first feature, and pool the Mel spectrogram;
[0113] Introducing nonlinearity into the convolution kernel and extracting the second feature in the pooled Mel-spectrogram image;
[0114] Downsampling the second feature to reduce the spatial dimension of the second feature, and extracting features from the second feature;
[0115] The extracted features are screened and filtered and combined into a processed spectrum image;
[0116] Combining the acoustic parameters with the processed spectrogram image to obtain a sound frequency map;
[0117] Output sound frequency graph.
[0118] An embodiment of the present invention further provides a computer-readable storage medium, on which a neural network-based sound analysis program is stored. When the neural network-based sound analysis program is executed by a processor, the steps of any one of the neural network-based sound analysis methods provided in the embodiments of the present invention are implemented.
[0119] It should be understood that the present invention is described by some embodiments, and those skilled in the art will appreciate that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.
Claims
1. A sound analysis method based on a neural network, characterized in that: The steps include: Acquire sound signals and save the sound signals; Calculate acoustic parameters of sound signals; The sound signal is converted into a Mel spectrogram, and features in the Mel spectrogram are extracted through a CNN neural network, and the extracted features are combined to obtain a processed Mel spectrogram; The mel-spectrogram and acoustic parameters are combined to obtain the processed spectrogram.
2. The sound analysis method according to claim 1, characterized in that: In the process of extracting features from the Mel spectrogram by using the CNN neural network, the following is also included: Features in the mel spectrogram are extracted through at least one set of convolutional layers, and the extracted features are downsampled through a pooling layer. The extracted downsampled features are then combined by a fully connected layer to obtain a processed mel spectrogram.
3. The sound analysis method according to claim 2, characterized in that: The CNN neural network includes a first convolutional layer and a second convolutional layer, the first convolutional layer extracts a first feature of a mel spectrogram, the second convolutional layer extracts a second feature of the mel spectrogram, the pooling layer downsamples the first feature and the second feature and extracts features from the first feature and the second feature, and the fully connected layer combines the first feature, the second feature and the features extracted by the pooling layer to obtain a processed mel spectrogram.
4. The sound analysis method according to claim 3, characterized in that: The proportion of the second feature in the mel spectrogram is greater than the proportion of the first feature in the mel spectrogram; After the first convolution layer extracts the first feature from the mel spectrogram, the pooling layer downsamples the first feature and pools the mel spectrogram, and then the second convolution layer extracts the second feature from the pooled mel spectrogram, and the pooling layer downsamples the second feature again.
5. The sound analysis method according to claim 2, characterized in that: The convolution layer also includes the following steps in the process of extracting the features of the Mel spectrogram: Introduce a nonlinear function into the convolution layer so that the convolution layer can be used as a nonlinear function for feature extraction.
6. The sound analysis method according to claim 1, characterized in that: The CNN neural network also includes an input layer and an output layer. The CNN neural network includes: Some neurons will be paralyzed with a defined probability, and these neurons will not propagate, while the number of neurons in the input and output layers will remain unchanged, and then the parameters will be updated.
7. The sound analysis method according to claim 6, characterized in that: In the next iteration, the CNN neural network randomly paralyzes some neurons so that they do not participate in training. The trained CNN neural network can be obtained until the training is completed, and the trained neural network is used to process the Mel spectrogram.
8. A sound analysis system, characterized in that: include: A sound collection module, used to obtain and save sound signals; Acoustic parameter analysis module, used to analyze acoustic parameters in sound signals; The acoustic diagnosis module converts the sound signal into a Mel-type spectrogram, uses a CNN neural network to extract features from the converted Mel-type spectrogram, and combines the extracted features to obtain a processed spectrogram; The report generation module combines the acoustic parameters with the processed spectrogram to obtain the sound frequency graph.
9. An intelligent terminal, characterized in that: The intelligent terminal includes at least one memory, at least one processor, and a sound analysis program stored in the memory and executable on the processor. When the sound analysis program is executed by the processor, the steps of the sound analysis method according to any one of claims 1 to 7 are implemented.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a sound analysis program, and when the sound analysis program is executed by the processor, the steps of the sound analysis method according to any one of claims 1 to 7 are implemented.