Method, system, apparatus, medium and program product for vocal cord disease classification based on ResNet model of MFCC

Abstract

The invention provides a method, system and device for vocal cord disease classification based on a ResNet model of MFCC, a medium and a program product, and relates to the field of vocal cord disease classification. The method comprises the following steps: S1) collecting pronunciation of a subject, constructing a corresponding label, and preprocessing the pronunciation of the subject; s2) converting the standardized pronunciation data set into a three-dimensional Mel spectrum feature map; s3) a vocal cord disease classification model is constructed, the vocal cord disease classification model adopts a ResNet architecture, the three-dimensional Mel spectrum feature map is used as input, and an SE module is embedded in each level of residual unit of the ResNet; s4) using the loss function as supervision of the vocal cord disease classification model to obtain a trained vocal cord disease classification model; and S5) inputting the pronunciation of the to-be-classified testee into the vocal cord disease classification model.According to the method, through organic combination of convolution kernel design, attention mechanism embedding and a lightweight channel expansion strategy, efficient and accurate recognition of pathological voice is realized.

Description

Technical Field

[0001] This invention relates to the field of vocal cord disease classification, and in particular to methods, systems, devices, media, and program products for vocal cord disease classification based on the ResNet model of MFCC. Background Technology

[0002] Statistics show that the global incidence of voice disorders is around 10%, with hoarseness accounting for over 70% of these cases. However, traditional diagnosis relies on auscultation combined with laryngoscopy, which is subjective, time-consuming, and has a low rate of identifying early, subtle lesions. ResNet (Residual Network) is a deep convolutional neural network that solves the problems of vanishing gradients and degradation during deep network training by introducing "residual connections," enabling the network to learn complex features more effectively. Its core idea is to allow the network to learn residuals rather than direct mappings, thereby achieving deeper network structures. Compared to traditional convolutional neural networks, the biggest advantage of ResNet (Residual Network) is its ability to build deeper, more stable, and more accurate models.

[0003] This characteristic gives it a significant advantage in medical applications such as voice disorder diagnosis, where high sensitivity to features and stringent detail requirements are essential. In the field of intelligent voice disorder diagnosis, ResNet can extract key features from Mel spectrograms or audio images of speech, thereby achieving efficient and accurate multi-class classification. Its unique deep network structure can effectively capture subtle pathological differences in sound, which is crucial for complex medical audio analysis tasks. In recent years, with the rapid development of artificial intelligence technology, automatic classification methods based on sound signals have gradually become a research hotspot, but existing technologies still have the following shortcomings: (1) Data level: The sample size of vocal cord audio is small and the noise interference is strong (environmental sound). Traditional data augmentation methods cannot effectively cover the diversity of pathological sounds. (2) Feature processing: It is difficult to uniformly input the spectral features of audio of different lengths into the model, resulting in insufficient utilization of time dimension information. (3) Clinical diagnosis: Most of the existing models are for diagnosing whether a voice disease is present. There is still a gap in the refined diagnosis of various voice diseases.

[0004] In conclusion, there is an urgent need for a machine learning-based method to accurately classify vocal cord diseases. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a method, system, device, medium and program product for vocal cord disease classification based on the MFCC-based ResNet model, in order to solve many technical problems in the classification of vocal cord diseases in the prior art.

[0006] To achieve the above and other related objectives, a first aspect of the present invention provides a method for vocal cord disease classification based on a ResNet model using MFCC, comprising the following steps: Step S1) Collect the pronunciation of the subjects, construct corresponding labels, preprocess the pronunciation of the subjects, generate a standardized pronunciation dataset, and construct training set and validation set based on it; Step S2) Convert the standardized pronunciation dataset into a three-dimensional Mel-spectral feature map; Step S3) Construct a vocal cord disease classification model. The vocal cord disease classification model adopts the ResNet architecture, takes the three-dimensional Mel spectrum feature map as input, and embeds SE modules in each level of the ResNet residual unit. The SE module compresses, excites and recalibrates the features in sequence to make the features adaptively weighted. Then, the features are entered into the classification head for classification. Step S4) Use the loss function as supervision when training the vocal cord disease classification model to optimize the training network parameters and obtain the trained vocal cord disease classification model. Step S5) Input the pronunciation of the subject to be classified into the vocal cord disease classification model, and the vocal cord disease classification model outputs the corresponding vocal cord disease.

[0007] In some embodiments of the first aspect of the present invention, step S1) further includes: Step S1a) Acquire the subject's vocalizations in a controlled acoustic environment with background noise controlled at 40–45 dB; Step S1b) Collect the following content: the six vowel syllables of Chinese / a / , / o / , / e / , / i / , / u / , and / ü / , with each vowel lasting for at least 3 seconds. Step S1c) Record in mono, at a sampling rate of 44.1kHz and a resolution of 16 bits, and truncate the audio to 3 seconds.

[0008] In some embodiments of the first aspect of the present invention, step S2) further includes: Step S2a) Perform short-time framing on the standardized pronunciation and apply a Hamming window to each frame of signal to reduce the impact of edge effects on spectral characteristics; Step S2b) Perform a fast Fourier transform on the processed signal to extract time-frequency domain features and obtain a spectral representation in both frequency and time dimensions; Step S2c) Maps the obtained spectrogram to the Mel scale, and extracts a spectral energy distribution that better matches the characteristics of human hearing through the Mel filter bank to obtain a two-dimensional Mel spectrogram; wherein, the two-dimensional Mel spectrogram includes a frequency dimension and a time dimension; Step S2d) Perform a logarithmic transformation on the spectral energy of each time frame in the two-dimensional Mel spectrogram to enhance the distinguishability of the low-energy part and compress the dynamic range; Step S2e) Stack the static Mel spectrogram with its first-order and second-order difference features to construct a three-dimensional Mel spectrogram; wherein the Mel spectrogram contains three dimensions: energy, frequency and time.

[0009] In some embodiments of the first aspect of the present invention, step S3) further includes: In step S3a), in each residual unit, the features of the three-dimensional Mel spectrum are first extracted through convolution and pooling operations, and then the weighted output feature map is obtained through the SE module. Step S3b) Utilize the residual linking structure to fuse the output feature maps of each residual unit in the vocal cord disease classification model; Step S3c) Use global pooling to compress the spatial dimension of the feature map fused by each residual unit, retaining the most important global features; Step S3d) Input the globally pooled feature map into the classification head, which includes a fully connected layer and a Dropout layer to achieve the final feature mapping to the label space.

[0010] In some embodiments of the first aspect of the present invention, step S3a) further includes: a three-level residual unit, wherein the first-level residual unit adopts a 3×3 convolution kernel, 32-batch normalization, and ReLU activation function; the second-level residual unit adopts a 3×3 convolution kernel, 64-batch normalization, ReLU activation function, and max pooling layer; and the third-level residual unit adopts a 3×3 convolution kernel, 64-batch normalization, ReLU activation function, and max pooling layer.

[0011] In some embodiments of the first aspect of the present invention, step S3) further includes: the SE module first performs a global average pooling operation on the input feature map to compress the two-dimensional spatial features into channel-level statistics to characterize the overall response information of each channel; then, the statistics are sequentially passed through two fully connected networks, wherein the first layer is processed by the ReLU activation function to capture the nonlinear relationship between channels, and the second layer is processed by the Sigmoid function to generate the weight coefficients corresponding to each channel, thereby realizing adaptive modeling of the importance of features of different channels; finally, the weight coefficients are multiplied with the input feature map channel by channel to obtain the weighted output feature map, thereby enhancing the acoustic features related to pathology and suppressing redundant information and noise features.

[0012] To achieve the above and other related objectives, a second aspect of the present invention provides a ResNet model based on MFCC for a vocal cord disease classification system. The vocal cord disease classification system includes: a data construction module for collecting the pronunciation of subjects, constructing corresponding labels, and preprocessing the pronunciation to generate a standardized pronunciation dataset for constructing training and validation sets; a feature map conversion module for converting the standardized pronunciation dataset into a three-dimensional Mel-ray spectrogram feature map; a model training module for running the vocal cord disease classification model, which adopts a ResNet architecture, using the three-dimensional Mel-ray spectrogram feature map as input, and embedding SE modules in each level of the ResNet residual units; the SE modules sequentially compress, excite, and recalibrate the features to adaptively weight them; subsequently, the features are entered into the classification head for classification; and a loss function is used as supervision during the vocal cord disease classification model training to optimize the network parameters, resulting in a trained vocal cord disease classification model; and a model inference module for inputting the pronunciation of the subjects to be classified into the vocal cord disease classification model, which outputs the corresponding vocal cord disease.

[0013] To achieve the above and other related objectives, a third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for vocal cord disease classification based on the MFCC-based ResNet model provided in the first aspect of the present invention.

[0014] To achieve the above and other related objectives, a fourth aspect of the present invention provides a computer program product comprising computer program code that, when executed on a computer, causes the computer to implement the method for vocal cord disease classification based on the MFCC-based ResNet model provided in the first aspect of the present invention.

[0015] To achieve the above and other related objectives, a fifth aspect of the present invention provides a computer device / apparatus / system / electronic terminal, including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the method provided in the first aspect of the present invention for vocal cord disease classification based on the ResNet model of MFCC.

[0016] As described above, the method, system, device, medium, and program products for vocal cord disease classification based on the MFCC-based ResNet model provided by this invention have the following beneficial effects: 1) This invention achieves efficient and accurate identification of pathological voices by organically combining convolution kernel design, attention mechanism embedding, and lightweight channel expansion strategy. It not only improves the technical accuracy and application breadth of artificial intelligence-assisted diagnosis, but also provides reliable technical support for early screening and remote diagnosis of clinical vocal cord diseases. It has good social value, economic prospects and strong practical application and promotion significance. Moreover, it can maintain high robustness and discrimination ability in medical audio scenarios with small samples and high noise.

[0017] 2) Furthermore, in terms of feature extraction, this invention innovatively employs a three-dimensional Mel spectrogram to map the speech signal into feature representations from three complementary perspectives: time-frequency, time-energy, and frequency-energy. This allows for the simultaneous preservation of the acoustic signal's temporal dynamics, frequency distribution, and energy variation characteristics. This multi-dimensional unfolding method enables a comprehensive characterization of fine-grained differences in pathological voices. Compared to traditional two-dimensional Mel spectrograms, this invention significantly improves feature discrimination capabilities and maintains good robustness even under small sample conditions. Validation examples show that on a pathological speech dataset collected under standard controlled conditions, the method of this invention achieves an overall classification accuracy of 86.25% for six types of vocal cord diseases, enabling fine-grained differentiation of six vocal cord states (see Table 1). This effectively overcomes the limitations of existing AI-assisted diagnostic methods in disease type identification, significantly expanding the applicability of diagnosis.

[0018] 3) Furthermore, the present invention enhances the local modeling capability of acoustic features: a 3×3 convolution kernel is used in the convolution feature extraction stage, which has a greater feature extraction capability than the commonly used 2×2 convolution kernel, and can cover a wider range of neighborhood information, thereby effectively capturing the feature differences of pathological voices and significantly improving the discrimination performance and diagnostic accuracy of the model.

[0019] 4) Furthermore, this invention improves the discriminative power of feature representation by introducing a channel attention mechanism (SE module) into the residual structure. This mechanism extracts the overall response of each channel through global average pooling and generates channel weights via nonlinear mapping, adaptively weighting the feature channels. This mechanism highlights key acoustic features relevant to pathology while suppressing irrelevant or redundant features, enabling dynamic optimization of the model during multi-stage feature extraction.

[0020] 5) Furthermore, this invention improves the lightweight nature of the model: by adopting a moderate channel expansion ratio (32→64→128) in the overall network design, the computational complexity and memory usage are significantly reduced while maintaining the feature representation capability. This makes it suitable for small sample training conditions and especially applicable to clinical scenarios in medical audio, ensuring the performance of the model in complex tasks. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the process of using the MFCC-based ResNet model for vocal cord disease classification, as provided in the example.

[0022] Figure 2 This is a flowchart illustrating the sub-steps of step S1 in the embodiment.

[0023] Figure 3 This is a flowchart illustrating the sub-step of step S2 in the embodiment.

[0024] Figure 4 This is a flowchart illustrating the sub-step of step S3 in the embodiment.

[0025] Figure 5 This is a schematic diagram illustrating the process of using the MFCC-based ResNet model for vocal cord disease classification, as provided in the example.

[0026] Figure 6 This is a core architecture diagram of the vocal cord disease classification model provided in the embodiment.

[0027] Figure 7 This is an architecture diagram of the SE module in the vocal cord disease classification model provided in the embodiment.

[0028] Figure 8 This is the confusion matrix in the verification example.

[0029] Figure 9 The ROC curves in the validation examples are shown below. Class 0 represents normal voice; Class 1 represents benign proliferative lesions; Class 2 represents malignant vocal cord tumors; Class 3 represents unilateral vocal cord paralysis with decompensation; Class 4 represents vocal cord sulcus; and Class 5 represents spastic phonation disorder.

[0030] Figure 10 A schematic block diagram of a ResNet model based on MFCC used in a vocal cord disease classification system, provided for an example.

[0031] Figure 11 A schematic block diagram of a computer device provided for an embodiment. Detailed Implementation

[0032] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0033] Before providing a further detailed description of the present invention, the nouns and terms used in the embodiments of the present invention are explained, and the nouns and terms used in the embodiments of the present invention are subject to the following interpretations: <1> MFCC: Mel-Frequency Cepstral Coefficients (MFCC) are a set of key coefficients used to construct the Mel cepstral spectrum. From a segment of a music signal, a cepstrum sufficient to represent that music signal can be obtained, and the Mel cepstral coefficients are the cepstrum (i.e., the spectrum of the spectrum) derived from this cepstrum. Unlike a typical cepstrum, the most distinctive feature of the Mel cepstral spectrum is that the frequency bands are uniformly distributed across the Mel scale. In other words, such frequency bands are closer to the non-linear human auditory system than the linear cepstrum representation commonly observed.

[0034] <2> ResNet: Residual Neural Network is a type of deep learning model. Its core principle is that each layer of the network does not directly learn the expected output, but instead learns the residual relationship between the input and the expected output. This type of network achieves identity mapping by skipping connections in certain network layers, and then adds and combines these connections with the outputs of the network layers.

[0035] In order to solve the technical problems described above, this invention introduces a new method: using a ResNet model based on MFCC to identify vocal cord diseases in sound.

[0036] Figure 1 and Figure 5 This document illustrates a flowchart of a method for vocal cord disease classification using a ResNet model based on MFCC, as described in an embodiment of the present invention. The method in this embodiment mainly includes the following steps:

[0037] Step S1: Collect the subjects' pronunciation, construct corresponding labels, and preprocess the subjects' pronunciation to generate a standardized pronunciation dataset, which is then used to construct the training and validation sets. The labels can be divided into normal voice and five vocal cord diseases: spastic phonation disorder, unilateral vocal cord paralysis decompensation, vocal cord sulcus, benign proliferative lesion, and malignant vocal cord tumor.

[0038] In the embodiments of the present invention, the specific implementation process of step S1 can be further divided into: Figure 2 The following steps are shown:

[0039] Step S1a) Acquire the subject's vocalizations in a controlled acoustic environment with background noise controlled at 40–45 dB. Specifically, vocalizations are performed using a headset condenser microphone, with the microphone positioned 3–4 cm away from the subject's mouth.

[0040] Step S1b) collects the six vowel syllables of Mandarin Chinese: / a / , / o / , / e / , / i / , / u / , and / ü / , with each vowel lasting at least 3 seconds. Specifically, the collected signals are transmitted to the computing device via an external audio interface.

[0041] Step S1c) Record in mono, at a sampling rate of 44.1kHz and a resolution of 16 bits, and truncate the audio to 3 seconds. Specifically, anonymization is performed during the recording, and the data is stored in a secure device with access control and saved as an uncompressed .wav file.

[0042] Step S2: Convert the standardized pronunciation dataset into a three-dimensional Mel-spectral feature map.

[0043] In the embodiments of the present invention, the specific implementation process of step S2 can be further divided into: Figure 3 The following steps are shown:

[0044] Step S2a: Perform short-time framing on the standardized pronunciation and apply a Hamming window to each frame to reduce the impact of edge effects on the spectral characteristics. Specifically, the window size for short-time framing is 2ms, and the frame shift is 10ms.

[0045] Step S2b: Perform a Fast Fourier Transform on the processed signal to extract time-frequency domain features and obtain a spectral representation in both frequency and time dimensions.

[0046] Step S2c: Map the obtained spectrogram to the Mel scale, and extract the spectral energy distribution that better matches the characteristics of human hearing through the Mel filter bank to obtain a two-dimensional Mel spectrogram; wherein, the two-dimensional Mel spectrogram includes a frequency dimension and a time dimension.

[0047] Step S2d: Perform a logarithmic transformation on the spectral energy of each time frame in the two-dimensional Mel spectrogram to enhance the distinguishability of the low-energy portion and compress the dynamic range, as shown in Formula 1:

[0048] in, This is represented as a transformed two-dimensional Mel-frequency spectrum. This is represented as a two-dimensional Mel-frequency spectrogram before transformation, where f represents the frequency index and t represents the time frame index. Represented as a constant to prevent zero values ​​( (Approaching infinity to 0).

[0049] Step S2e: Stack the two-dimensional Mel spectrogram with its first-order and second-order difference features to construct a three-dimensional Mel spectrogram. The three-dimensional Mel spectrogram can be used to characterize the dynamic changes of speech signals; the three-dimensional Mel spectrogram contains three dimensions: energy, frequency, and time. Specifically, the tensor structure of the three-dimensional Mel spectrogram is [energy × frequency × time].

[0050] In one example embodiment, Formula 1 and Formula 2 are used to calculate the first-order difference features of the two-dimensional Mel spectrogram. Second-order difference features of two-dimensional Mel spectrograms First-order difference features of two-dimensional Mel spectrograms As shown in Formula 2:

[0051] Where N is the size of the difference window, and n represents the time frame relative to the current time frame. t The offset frame number, the first-order difference feature of the two-dimensional Mel spectrogram It can characterize the rate of change of energy over time.

[0052] Then, the first-order difference features of the two-dimensional Mel spectrum Based on this, calculate the second-order difference characteristics of the two-dimensional Mel spectrum. Second-order difference features of two-dimensional Mel spectrograms As shown in Formula 3:

[0053] Second-order difference features of two-dimensional Mel spectrograms It can describe acceleration information that describes energy changes.

[0054] Finally, the two-dimensional Mel spectrum diagram First-order difference of two-dimensional Mel spectrogram Second-order difference of two-dimensional Mel spectrogram Stacking is performed along the channel dimension, as shown in Formula 4:

[0055] The first channel is a two-dimensional Mel spectrum, representing the original energy distribution; the second channel is the first difference of the two-dimensional Mel spectrum, representing the trend of energy change; and the third channel is the second difference of the two-dimensional Mel spectrum, representing the acceleration of change.

[0056] Step S3: Construct a vocal cord disease classification model. The vocal cord disease classification model adopts a ResNet architecture, using a 3D Mel-spectrum feature map as input. SE modules are embedded in each level of the ResNet residual unit. The SE modules sequentially compress, excite, and recalibrate the features, adaptively weighting them. Subsequently, the features are entered into the classification head for classification. The core architecture diagram of the vocal cord disease classification model is as follows: Figure 6 As shown.

[0057] In the embodiments of the present invention, the specific implementation process of step S3 can be further divided into: Figure 4 The following steps are shown: Step S3a: In each residual unit, features of the 3D Mel-ray spectrogram are first extracted through convolution and pooling operations, and then the weighted output feature map is obtained through the SE module. In a specific embodiment, it includes three levels of residual units. The first level residual unit uses a 3×3 convolution kernel, 32-batch normalization, and ReLU activation function; the second level residual unit uses a 3×3 convolution kernel, 64-batch normalization, and ReLU activation function. Compared to the first level residual unit, the second level residual unit also introduces max pooling to process the feature map; the third level residual unit uses a 3×3 convolution kernel, 64-batch normalization, and ReLU activation function. Similar to the second level residual unit, the third level residual unit also introduces max pooling to process the feature map. For illustration, max pooling is introduced in both the second and third level residual units to progressively downsample the feature map based on convolutional feature extraction, thereby achieving a higher level of feature abstraction. Max pooling can highlight the most significant local responses while reducing the spatial resolution of the feature map, thereby suppressing background noise and irrelevant details in the speech signal. By continuously using max pooling in the second and third level residual units, the model can not only form a multi-level feature representation from local details to global patterns layer by layer, but also maintain strong robustness in high-noise environments, thus significantly improving the recognition and classification performance of pathological voice differences. A 3×3 convolutional kernel is used in the convolutional feature extraction stage, which has a greater feature extraction capability than the commonly used 2×2 convolutional kernel, covering a wider range of neighborhood information, thus effectively capturing the feature differences of pathological voices and significantly improving the model's discriminative performance and diagnostic accuracy. Furthermore, a moderate channel expansion ratio is adopted in the overall network design, i.e., batch normalization from 32→64→128, which significantly reduces computational complexity and memory usage while maintaining feature representation capabilities. This allows the model to adapt to small sample training conditions, making it particularly suitable for clinical applications in medical audio, ensuring the model's performance in complex tasks.

[0058] As an illustration, the architecture of the SE module (channel attention mechanism) can be as follows: Figure 7 As shown, the SE module consists of three main stages: Compression stage (Squeeze): The input feature map is compressed into channel-level statistics through global average pooling operation, which are used to characterize the overall response information of each channel; Excitation phase: The statistics are passed through two fully connected network layers in sequence. The first layer is processed by the ReLU activation function to capture the nonlinear relationship between channels, and the second layer is processed by the Sigmoid function to generate the weight coefficients corresponding to each channel, so as to achieve adaptive modeling of the importance of features of different channels. Reweighting stage: The weighting coefficients are multiplied channel by channel with the input feature map to obtain a weighted output feature map, thereby enhancing the acoustic features related to pathology and suppressing redundant information and noise features.

[0059] Step S3b: Utilize the residual linking structure to fuse the output feature maps of each residual unit in the vocal cord disease classification model. Specifically, the residual linking structure allows information to be directly passed to subsequent layers, thereby preserving the original features and preventing feature vanishing layer by layer.

[0060] Step S3c: Compress the spatial dimension of the feature map using global pooling, while retaining the most important global features.

[0061] Step S3d: Input the globally pooled feature map into the classification head, which includes fully connected layers and Dropout layers to achieve the final feature mapping to the label space. The labels of the classification head are: normal voice, spasmodic dysphonia, unilateral vocal cord paralysis decompensation, vocal cord sulcus, benign proliferative lesion, and malignant vocal cord tumor.

[0062] Step S4: Utilize the loss function as supervision in the vocal cord disease classification model to optimize the training network parameters, obtaining the trained vocal cord disease classification model. The loss function is the cross-entropy loss, as shown in Formula 5:

[0063] Where: Loss represents the loss function, N represents the number of samples, and C represents the number of classification categories. This represents the true label of sample i in category c; This represents the model's predicted probability that sample i belongs to category c.

[0064] Step S5: Input the pronunciation of the subject to be classified into the vocal cord disease classification model, and the vocal cord disease classification model outputs the corresponding vocal cord disease.

[0065] Figure 10 This is a schematic block diagram of a ResNet model based on MFCC used in a vocal cord disease classification system, as provided in an embodiment of the present invention. Figure 10As shown, the system includes: a data construction module 51, a feature map conversion module 52, a model training module 53, and a model inference module 54.

[0066] The data construction module 51 is used to collect the pronunciation of the subjects, construct corresponding labels, and preprocess the pronunciation to generate a standardized pronunciation dataset, which is then used to construct the training and validation sets. The feature map conversion module 52 is used to convert the standardized pronunciation dataset into a three-dimensional Mel-ray spectrogram feature map. The model training module 53 is used to run the vocal cord disease classification model. The vocal cord disease classification model adopts the ResNet architecture, takes the three-dimensional Mel-ray spectrogram feature map as input, and embeds SE modules in the residual units of each level of ResNet. The SE module compresses, excites, and recalibrates the features in sequence to adaptively weight the features. Then, the features are entered into the classification head for classification. The loss function is then used as supervision when training the vocal cord disease classification model to optimize the training network parameters and obtain the trained vocal cord disease classification model. The model inference module 54 inputs the pronunciation of the subjects to be classified into the vocal cord disease classification model, and the vocal cord disease classification model outputs the corresponding vocal cord disease.

[0067] It should be understood that the specific processes by which each module performs the corresponding steps described above have been detailed in the above method embodiments, and will not be repeated here for the sake of brevity. It should also be understood that the module division in the embodiments of this invention is illustrative and merely a logical functional division; other division methods may exist in actual implementation. Furthermore, the functional modules in the various embodiments of this invention can be integrated into a single processor, exist as separate physical entities, or have two or more modules integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0068] Figure 11 A schematic block diagram of a computer device provided for an embodiment of the present invention. Figure 11 As shown, the computer device includes at least one processor 601, a memory 602, at least one network interface 603, and a user interface 605. The various components in the device are coupled together via a bus system 604. It is understood that the bus system 604 is used to implement communication between these components. In addition to a data bus, the bus system 604 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in… Figure 6 The general will label all buses as bus systems.

[0069] According to the method provided in the embodiments of the present invention, the present invention also provides a computer program product, the computer program product comprising: computer program code, which, when executed on a computer, causes the computer to perform... Figures 1 to 5The embodiments shown illustrate a method for vocal cord disease classification using a ResNet model based on MFCC in any of the examples.

[0070] According to the method provided in the embodiments of the present invention, the present invention also provides a computer-readable storage medium storing program code, which, when run on a computer, causes the computer to perform the above-described method. Verification Example

[0071] In the experimental verification, a self-collected pathological speech dataset was used for training and testing (training set: validation set: test set = 7:1.5:1.5), with a data ratio of male:female = 45.9%:54.1%. The vocal cord disease classification model achieved an overall accuracy of 86.25% in the disease classification task (as shown in Table 1). It still showed high robustness and classification performance in scenarios with small samples and high noise, verifying the effectiveness and practicality of the method of this invention. Among them, high noise refers to controlling the noise to within 45 dB, and the signal-to-noise ratio (SNR) is used as the quantitative indicator.

[0072] Table 1 shows the diagnostic performance of the vocal cord disease classification model provided by this invention on overall and various voice diseases. Overall, the model achieves an accuracy of 0.8625, a recall of 0.8336, an F1 score of 0.8339, and a specificity of 0.9714, indicating that the model has high reliability and robustness in the identification of multiple categories of voice diseases.

[0073] Figure 8 The confusion matrix of the vocal cord disease classification model provided by this invention is shown.

[0074] Figure 9 The ROC curve of the vocal cord disease classification model provided by this invention is shown; wherein, Class 0 is normal voice; Class 1 is benign proliferative lesion; Class 2 is malignant vocal cord tumor; Class 3 is unilateral vocal cord paralysis decompensation; Class 4 is vocal cord groove; and Class 5 is spastic voice disorder.

[0075] Table 1. Diagnostic performance of the vocal cord disease classification model on overall and various voice disorders.

[0076] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A method for vocal cord disease classification based on the ResNet model using MFCC, characterized in that, It includes the following steps: Step S1) Collect the pronunciation of the subjects, construct corresponding labels, preprocess the pronunciation of the subjects, generate a standardized pronunciation dataset, and construct training set and validation set based on it; Step S2) Convert the standardized pronunciation dataset into a three-dimensional Mel-spectral feature map; Step S3) Construct a vocal cord disease classification model. The vocal cord disease classification model adopts the ResNet architecture, takes the three-dimensional Mel spectrum feature map as input, and embeds SE modules in the residual units of each level of ResNet. The SE modules compress, excite and recalibrate the features in sequence to make the features adaptively weighted. Subsequently, the features are entered into the classification head for classification; Step S4) Use the loss function as supervision when training the vocal cord disease classification model to optimize the training network parameters and obtain the trained vocal cord disease classification model. Step S5) Input the pronunciation of the subject to be classified into the vocal cord disease classification model, and the vocal cord disease classification model outputs the corresponding vocal cord disease.

2. The method for vocal cord disease classification based on the ResNet model using MFCC as described in claim 1, characterized in that, Step S1 also includes: Step S1a) Acquire the subject's vocalizations in a controlled acoustic environment with background noise controlled at 40-45 dB; Step S1b) Collect the following content: the six vowel syllables of Chinese / a / , / o / , / e / , / i / , / u / , and / ü / , with each vowel lasting for at least 3 seconds. Step S1c) Record in mono, at a sampling rate of 44.1kHz and a resolution of 16 bits, and truncate the audio to 3 seconds.

3. The method for vocal cord disease classification based on the ResNet model using MFCC according to claim 1, characterized in that, Step S2) also includes: Step S2a) Perform short-time framing on the standardized pronunciation and apply a Hamming window to each frame of signal to reduce the impact of edge effects on spectral characteristics; Step S2b) Perform a fast Fourier transform on the processed signal to extract time-frequency domain features and obtain a spectral representation in both frequency and time dimensions; Step S2c) Maps the obtained spectrogram to the Mel scale, and extracts a spectral energy distribution that better matches the characteristics of human hearing through the Mel filter bank to obtain a two-dimensional Mel spectrogram; wherein, the two-dimensional Mel spectrogram includes a frequency dimension and a time dimension; Step S2d) Perform a logarithmic transformation on the spectral energy of each time frame in the two-dimensional Mel spectrogram to enhance the distinguishability of the low-energy part and compress the dynamic range; Step S2e) Stack the static Mel spectrogram with its first-order and second-order difference features to construct a three-dimensional Mel spectrogram; wherein the Mel spectrogram contains three dimensions: energy, frequency and time.

4. The method for vocal cord disease classification based on the ResNet model using MFCC as described in claim 1, characterized in that, Step S3 also includes: In step S3a), in each residual unit, the features of the three-dimensional Mel spectrum are first extracted through convolution and pooling operations, and then the weighted output feature map is obtained through the SE module. Step S3b) Utilize the residual linking structure to fuse the output feature maps of each residual unit in the vocal cord disease classification model; Step S3c) Use global pooling to compress the spatial dimension of the feature map fused by each residual unit, retaining the most important global features; Step S3d) Input the globally pooled feature map into the classification head, which includes a fully connected layer and a Dropout layer to achieve the final feature mapping to the label space.

5. The method for vocal cord disease classification based on the ResNet model using MFCC according to claim 4, characterized in that, Step S3a) further includes: a three-level residual unit, wherein the first-level residual unit uses a 3×3 convolution kernel, 32-batch normalization, and ReLU activation function; the second-level residual unit uses a 3×3 convolution kernel, 64-batch normalization, ReLU activation function, and max pooling layer; and the third-level residual unit uses a 3×3 convolution kernel, 64-batch normalization, ReLU activation function, and max pooling layer.

6. The method for vocal cord disease classification using a ResNet model based on MFCC according to claim 4 or 5, characterized in that, Step S3) further includes: The SE module first performs global average pooling on the input feature map to compress the two-dimensional spatial features into channel-level statistics, which are used to characterize the overall response information of each channel; then, the statistics are passed through two fully connected networks in sequence, where the first layer is processed by the ReLU activation function to capture the nonlinear relationship between channels, and the second layer is processed by the Sigmoid function to generate the weight coefficients corresponding to each channel, so as to achieve adaptive modeling of the importance of features of different channels; finally, the weight coefficients are multiplied with the input feature map channel by channel to obtain the weighted output feature map, thereby enhancing the acoustic features related to pathology and suppressing redundant information and noise features.

7. A ResNet model based on MFCC for a vocal cord disease classification system, characterized in that, include: The data construction module is used to collect the subjects' pronunciation, construct corresponding labels, preprocess the subjects' pronunciation, generate a standardized pronunciation dataset, and use it to construct training and validation sets; The feature map conversion module is used to convert standardized pronunciation datasets into three-dimensional Mel-spectrum feature maps; The model training module is used to run the vocal cord disease classification model. The vocal cord disease classification model adopts the ResNet architecture, takes the three-dimensional Mel spectrum feature map as input, and embeds the SE module in each level of the ResNet residual unit. The SE module compresses, excites and recalibrates the features in sequence to make the features adaptively weighted. Subsequently, the features are entered into the classification head for classification; then, the loss function is used as supervision when training the vocal cord disease classification model to optimize the training network parameters and obtain the trained vocal cord disease classification model. The model inference module inputs the pronunciation of the subject to be classified into the vocal cord disease classification model, and the vocal cord disease classification model outputs the corresponding vocal cord disease.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the method for vocal cord disease classification using the ResNet model based on MFCC as described in any one of claims 1 to 6.

9. A computer program product, characterized in that, The computer program product includes computer program code that, when run on a computer, causes the computer to implement the method of using the ResNet model based on MFCC for vocal cord disease classification as described in any one of claims 1 to 6.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method for vocal cord disease classification using the ResNet model based on MFCC as described in any one of claims 1 to 6.

Images