Children practice condition evaluation method, medium and system

Abstract

The invention provides a children practice situation evaluation method, medium and system, and belongs to the technical field of deep learning models.The children practice situation evaluation method includes the steps that children and standard playing audio signals are collected and preprocessed into time-frequency representation, a time window sequence is constructed to calculate window similarity, beat mark points are extracted to form a rhythm sequence, and the rhythm sequence is obtained; a rhythm perception analysis model is input to generate a rhythm feature representation vector, a multi-dimensional deviation vector is formed by applying multi-level time model analysis, a rhythm deviation tolerance range matrix is constructed, deviation similarity is calculated, and finally a rhythm performance score is output through a rhythm evaluation neural network model. According to the method, a convolutional neural network, a long-short term memory network and an attention mechanism are fused, and comprehensive quantitative evaluation of the playing rhythm performance of the children is realized through large-scale data set training and expert scoring verification.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of deep learning models, and in particular, relates to a method, medium and system for evaluating the practice of a child musician. BACKGROUND

[0002] In the field of music education, the evaluation of a child musician's performance is a key step in guiding learning progress. Traditional evaluation methods mainly rely on the subjective auditory judgment of music teachers, comparing the differences between the child musician's performance and the standard performance to assess the level of performance. With the development of technology, some automated evaluation systems have been applied to music education. These systems usually use simple audio signal comparison methods to calculate the similarity between the performance audio and the standard audio, or use basic feature extraction techniques to analyze the beat accuracy and stability of the performance audio. However, traditional automated evaluation methods have obvious limitations. The simple signal comparison-based method cannot distinguish between technical errors and artistic performance changes, resulting in overly mechanical evaluation results. The basic feature extraction technique is difficult to capture the multi-level structural features of rhythm performance, especially the micro-syllable time value processing and macro-phrase organization. Existing systems lack flexible definition of the allowable rhythm deviation range under different musical styles, and cannot provide reasonable evaluation standards for different difficulty levels and performance levels. These limitations make it difficult for existing technology to effectively evaluate the rhythm performance of child musicians and provide accurate quantitative feedback, and cannot provide comprehensive and objective rhythm performance evaluation results for child musicians. SUMMARY

[0003] Therefore, the present application provides a method, medium and system for evaluating the practice of a child musician, which can solve the technical problem of being unable to effectively evaluate the rhythm performance of a child musician and provide accurate quantitative feedback in the prior art.

[0004] The present application is implemented as follows: The first aspect of the present application provides a method for evaluating the practice of a child musician, comprising: collecting and preprocessing the child musician's performance audio signal and the standard performance audio signal; constructing a time window sequence and calculating the window similarity; extracting beat marker points to calculate the rhythm sequence; inputting the rhythm sequence and the window similarity into a rhythm perception analysis model to generate a rhythm feature representation vector; applying a multi-level time model analysis to generate a multi-dimensional deviation vector; constructing a rhythm deviation tolerance range matrix; calculating the deviation similarity; applying a rhythm evaluation neural network model to output a rhythm performance score; wherein the rhythm perception analysis model uses a convolutional long short-term memory network structure to analyze the rhythm sequence and generate a high-dimensional semantic-level rhythm feature representation, and the multi-level time model analyzes the time relationship at three levels of macro-rhythm structure, meso-beat group organization and micro-syllable time value.

[0005] Among them, the preprocessing step specifically converts the audio signal of the child's performance and the standard performance audio signal into time domain representation and frequency domain representation; the time window sequence refers to a continuous time segment sequence in which the child's performance audio signal and the standard performance audio signal are divided according to a fixed duration or an adaptive duration, the fixed duration is 20 milliseconds to 200 milliseconds, and the adaptive duration is dynamically determined according to the music beat structure.

[0006] Among them, the window similarity refers to calculating the similarity between the audio signal of the child's performance and the standard performance audio signal in time domain characteristics and frequency domain characteristics within each time window, and quantifying it through methods such as cross-correlation function, spectral distance or Mel-frequency cepstral coefficient distance. The window similarity is used to adjust the internal weight distribution of the rhythm perception analysis model.

[0007] Among them, the rhythm perception analysis model is trained based on a large-scale performance dataset and fine-tuned through style transfer to adapt to the performance characteristics of different styles; the rhythm feature representation vector contains implicit information such as rhythm pattern, dynamic changes, duration distribution and expressive processing. The dimension of the rhythm feature representation vector is determined by the rhythm perception attention mechanism and dynamically adjusted according to the length of the rhythm sequence.

[0008] Among them, the multidimensional deviation vector refers to the representation of the rhythm time deviation occurring in the performance of the children through a multidimensional vector, and each dimension of the multidimensional deviation vector corresponds to a different type of rhythm feature, including absolute time value deviation, relative rhythm proportion deviation, intra-beat structure deviation, rhythm stability deviation and expressive fluctuation deviation quantization parameters.

[0009] Among them, the rhythm deviation allowable range matrix refers to a multi-dimensional evaluation reference standard established according to different musical styles, repertoire difficulty and performance level, which defines the extent to which various rhythm deviations are technical errors or artistic expressions. The rhythm deviation allowable range matrix includes boundary thresholds and scoring weight coefficients.

[0010] The deviation similarity refers to a quantitative index obtained by calculating the cosine distance or Euclidean distance between the multidimensional deviation vector and the corresponding deviation vector of the standard performance template, which is used to measure the degree of similarity between the rhythm processing of the child's performance and the standard performance.

[0011] Among them, the rhythm evaluation neural network model refers to a deep neural network that integrates the attention mechanism and the residual connection structure, which generates a rhythm performance score for multi-dimensional deviation vectors and deviation similarity; the rhythm performance score includes sub-item scores and a comprehensive score in four dimensions: rhythm accuracy, rhythm stability, rhythm expressiveness, and rhythm technicality.

[0012] The specific structure of the rhythm perception analysis model is a hybrid architecture combining a multi-layer time convolution network and a bidirectional long short-term memory network, and includes a time domain feature extraction layer, a frequency domain feature extraction layer, a multi-scale convolution layer, a bidirectional long short-term memory layer, a self-attention mechanism layer and a fully connected output layer.

[0013] In the training process of the rhythm perception analysis model, a directional drift detection function is set for each convolution layer, the directional drift detection function judges whether there is a feature direction deviation by calculating a direction consistency index of the convolution output feature map, and when the direction consistency index exceeds a set threshold, a convolution adjustment mechanism is triggered.

[0014] The convolution adjustment mechanism uses a sparse matrix jitter enhancement method to perturb and reconstruct the convolution kernel parameters, and maintains the feature expression ability through dimension reduction reconstruction; the sparse matrix jitter enhancement method refers to a technique of introducing controlled random noise to sparsify the convolution weight matrix and increase the direction diversity.

[0015] The rhythm perception analysis model uses a window similarity adjustment method to adjust the internal feature weight distribution: for time segments with high window similarity, a lower weight is given to reduce their influence on the model output; for time segments with low window similarity, a higher weight is given to strengthen the model's learning of the difference area.

[0016] The weight distribution is realized through a dynamic attention mechanism, an attention mask matrix is generated according to the window similarity sequence, and the attention mask matrix is applied to the internal feature mapping process of the rhythm perception analysis model.

[0017] The second aspect of the present application provides a computer readable storage medium, the computer readable storage medium stores program instructions, the program instructions are run in the computer, and the program instructions are used to execute the above-mentioned music student practice condition evaluation method.

[0018] The third aspect of the present application provides a music student practice condition evaluation system, which includes the above-mentioned computer readable storage medium, the system is any one of a computer, a server and a single-chip microcomputer, the computer readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer readable storage medium.

[0019] The present application realizes comprehensive quantitative evaluation of the rhythm performance of a child by constructing a rhythm perception analysis model and a rhythm evaluation neural network model. The method can not only accurately capture the time domain and frequency domain differences between the child's performance and the standard performance, but also distinguish between technical errors and artistic performance changes, providing more accurate rhythm performance evaluation for the child. Through a multi-level time model, the present application can analyze the time relationship at three levels of macro rhythm structure, medium beat group organization and micro note duration, and comprehensively capture the rhythm processing characteristics of the performer at different time scales. Combined with the rhythm deviation tolerance range matrix, the system can flexibly define the evaluation standard of rhythm deviation according to different music styles, song difficulty and performance level, avoid mechanical evaluation, and respect the diversity of music performance. The present application solves the technical problem that the existing technology cannot effectively evaluate the rhythm performance of a child and provide accurate quantitative feedback, and provides comprehensive and objective rhythm performance evaluation results for the child. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A flowchart of the method of the present application. DETAILED DESCRIPTION

[0021] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0022] As shown in Figure 1 , it is a flowchart of a child practice condition evaluation method provided by the first aspect of the present application. The method includes the following steps:

[0023] S01, collecting a child performance audio signal and pre-processing to convert it into a time domain representation and a frequency domain representation;

[0024] S02, synchronously collecting a standard performance audio signal and pre-processing to obtain a standard time domain representation and a standard frequency domain representation;

[0025] S03, constructing a time window sequence and calculating the window similarity of the child performance audio signal and the standard performance audio signal in each time window;

[0026] S04, extracting the beat marker points in the time domain representation and the frequency domain representation of the child performance audio signal and calculating the adjacent beat interval to form a rhythm sequence;

[0027] S05, inputting the rhythm sequence and the window similarity to a rhythm perception analysis model to generate a rhythm feature representation vector;

[0028] S06, applying a multi-level time model to analyze the rhythm feature representation vector and quantizing it into a multi-dimensional deviation vector;

[0029] S07, constructing a rhythm deviation tolerance range matrix based on a performance style library;

[0030] S08, calculating a deviation similarity between the child performer's performance and a standard performance template based on the multi-dimensional deviation vector;

[0031] S09, outputting a rhythm performance score using a rhythm evaluation neural network model;

[0032] wherein the time window sequence refers to a sequence of continuous time segments obtained by dividing the child performer's performance audio signal and the standard performance audio signal according to a fixed time length or an adaptive time length, the fixed time length being 20 milliseconds to 200 milliseconds, and the adaptive time length being dynamically determined according to a musical beat structure.

[0033] wherein the window similarity refers to a similarity between the child performer's performance audio signal and the standard performance audio signal in terms of time domain features and frequency domain features within each time window, quantified by methods such as cross-correlation function, spectral distance, or mel-frequency cepstral coefficient distance, and the window similarity is used to adjust the internal weight distribution of the rhythm perception analysis model.

[0034] wherein the rhythm perception analysis model refers to a rhythm representation extraction model trained using a convolutional long short-term memory network structure, which generates high-dimensional semantic-level rhythm feature representations by analyzing the time features in the rhythm sequence, and the rhythm perception analysis model is trained based on a large-scale performance dataset and fine-tuned by style transfer to adapt to different performance characteristics of different styles.

[0035] wherein the multi-level time model refers to a model that captures the performance characteristics of a performer in different time scales by analyzing the time relationships at three levels of macro-rhythm structure, meso-beat group organization, and micro-note duration, including rhythm stability, intra-beat structure processing, and note coherence.

[0036] wherein the rhythm feature representation vector refers to a high-dimensional feature vector obtained after processing by the rhythm perception analysis model, containing implicit information such as rhythm pattern, dynamic change, duration distribution, and expressive processing, and the dimension of the rhythm feature representation vector is determined by a rhythm perception attention mechanism and dynamically adjusted according to the length of the rhythm sequence.

[0037] wherein the multi-dimensional deviation vector refers to a multi-dimensional vector representation of the rhythm time deviation in the child performer's performance, each dimension of the multi-dimensional deviation vector corresponding to a different type of rhythm feature, including absolute duration deviation, relative rhythm proportion deviation, intra-beat structure deviation, rhythm stability deviation, and expressive fluctuation deviation quantization parameters.

[0038] The rhythm deviation tolerance range matrix is a multi-dimensional evaluation reference standard established according to different music styles, song difficulty and playing levels, and defines the degree to which various rhythm deviations belong to technical errors or artistic expression.

[0039] The deviation similarity is a quantitative index obtained by calculating the cosine distance or Euclidean distance between the multi-dimensional deviation vector and the corresponding deviation vector of the standard performance template, and is used to measure the closeness of the rhythm processing of the child performer and the standard performance.

[0040] The rhythm evaluation neural network model is a deep neural network that integrates attention mechanism and residual connection structure, and generates a rhythm performance score for the multi-dimensional deviation vector and the deviation similarity. The number of attention heads of the rhythm evaluation neural network model is determined by the dimension of the multi-dimensional deviation vector, the depth of the residual connection of the rhythm evaluation neural network model is determined by the evaluation accuracy requirement, and the sparse attention range of the rhythm evaluation neural network model is calculated from the rhythm sequence length.

[0041] The rhythm performance score is a quantitative evaluation result of the rhythm performance of the child performer, and includes four-dimensional sub-scores of rhythm accuracy, rhythm stability, rhythm expressiveness and rhythm technicality, and a comprehensive score.

[0042] The specific structure of the rhythm perception analysis model is a hybrid architecture combining a multi-layer time convolutional network and a bidirectional long short-term memory network, and includes a time domain feature extraction layer, a frequency domain feature extraction layer, a multi-scale convolutional layer, a bidirectional long short-term memory layer, a self-attention mechanism layer and a fully connected output layer. The time domain feature extraction layer uses a one-dimensional convolutional filter bank to extract rhythm features at different time scales, the frequency domain feature extraction layer uses a mel frequency cepstral coefficient transformation and a frequency domain convolutional network to extract timbre and pitch features, the multi-scale convolutional layer captures short-term and long-term dependencies through convolution operations with different kernel sizes, the bidirectional long short-term memory layer models the correlation between the context before and after the sequence and maintains long-term memory, the self-attention mechanism layer calculates the correlation strength between different time points and enhances the representation of key rhythm patterns, and the fully connected output layer maps high-dimensional features to a semantic-level rhythm representation space. The rhythm perception analysis model alleviates the gradient vanishing problem through residual connection and maintains information flow.

[0043] The rhythm perception analysis model sets a directional drift detection function for each convolution layer during the training process, the directional drift detection function judges whether there is a feature direction deviation by calculating the directional consistency index of the convolution output feature map, and when the directional consistency index exceeds a set threshold, a convolution adjustment mechanism is triggered, the convolution adjustment mechanism uses a sparse matrix jitter enhancement method to perturb and reconstruct the convolution kernel parameters, and the feature expression ability is maintained through dimensionality reduction reconstruction.

[0044] The directional consistency index is a centralized quantization value of the gradient distribution of the convolution layer output feature map in each direction, which is obtained by calculating the feature map gradient direction histogram entropy value or the principal component analysis characteristic value ratio, and the directional consistency index value range is 0 to 1, and the greater the value, the more serious the direction deviation.

[0045] The sparse matrix jitter enhancement method refers to a technology that introduces controlled random noise to sparse the convolution weight matrix and increase the direction diversity, including three steps of sparsity control, direction entropy maximization and weight normalization, which ensures to prevent feature space collapse without reducing model expression ability.

[0046] The rhythm perception analysis model uses the window similarity to adjust the internal feature weight distribution method: for the time segment with high window similarity, a lower weight is given to reduce its influence on the model output; for the time segment with low window similarity, a higher weight is given to strengthen the model learning of the difference area; the weight distribution is realized through a dynamic attention mechanism, an attention mask matrix is generated according to the window similarity sequence, and the attention mask matrix is applied to the internal feature mapping process of the rhythm perception analysis model.

[0047] The training data set of the rhythm perception analysis model is established, which specifically includes collecting standard performance audio of different difficulty levels and style genres from the International Standard Music Institute Performance Library, collecting performance samples of learners of different levels from global music education institutions, standardizing and preprocessing all audio including sampling rate unification and volume normalization, inviting professional music reviewers to score the rhythm performance of each audio in multiple dimensions and record comments, extracting time and frequency domain features in the audio as model input, labeling expert scores as supervision signals, establishing mapping relationship data pairs between audio features and expert scores, stratified sampling according to difficulty levels and style types to construct training set and validation set, and performing data enhancement processing on the training set including adding different degree of environmental noise, adjusting speed change and introducing random time offset operation to improve the robustness of the model.

[0048] The rhythm perception analysis model training step specifically includes first self-supervised training on large-scale unlabeled music data, adopting time mask prediction and rhythm consistency contrast learning two training tasks, the time mask prediction task randomly masks time segments in the input sequence and requires the model to restore the masked content according to the context, and the rhythm consistency contrast learning task requires the model to distinguish different performance versions of the same music segment and different music segments, and after training, supervised fine-tuning is performed on the labeled expert scoring dataset, a multi-task learning framework is used to simultaneously predict rhythm stability, expressiveness level, and technical proficiency multiple scoring dimensions, learning rate decay strategy is used during fine-tuning to avoid overfitting, early stopping mechanism is used to avoid overfitting, transfer learning method is used to construct a style rhythm evaluation model for different music styles, and finally the consistency level of the model evaluation result with the human expert score is verified through expert double-blind test, and the model architecture and parameter configuration are continuously optimized based on the verification result.

[0049] The step of directional drift detection and adjustment during the training process of the rhythm perception analysis model specifically includes calculating the direction consistency index of each convolution layer output feature map after forward propagation of each training batch, marking the layer as having a directional drift risk when the direction consistency index exceeds a preset threshold, applying the sparse matrix jitter enhancement method to generate adjusted convolution parameters for the marked layer, using the adjusted parameters to perform forward calculation again and verifying whether the direction consistency index is reduced to below the threshold, if it still exceeds the threshold, increasing the jitter intensity and repeating the adjustment until the requirement is met or the maximum number of attempts is reached, recording the impact of each adjustment on the model accuracy and retaining the adjustment result with the smallest accuracy loss, gradually recovering the accuracy loss in subsequent training, and reducing the direction consistency index threshold at the later stage of training to achieve more stringent drift control.

[0050] The step of training the rhythm evaluation neural network model specifically includes first constructing a training dataset, taking the multi-dimensional deviation vector, the deviation similarity, and the window similarity sequence as input features, taking the expert's rhythm performance score as a training label, dividing the training set and test set using the cross-validation method, using a weighted loss function combining the mean square error and cross-entropy loss as the loss function, using a stochastic gradient descent algorithm with a momentum term as the optimization method, dynamically adjusting the batch size and learning rate during training, and using weight regularization and random inactivation to prevent overfitting, evaluating the performance on the validation set after the model converges, analyzing the evaluation accuracy through the confusion matrix and mean square error index, and fine-tuning the model hyperparameters based on the evaluation results.

[0051] The specific implementation of the above steps is described in detail below.

[0052] The specific implementation method of step S01 is to use a high-precision digital audio acquisition device to record the performance of the child. The sampling rate is set to 44.1kHz and the quantization accuracy is 24 bits. The collected audio signal is low-pass filtered to remove high-frequency noise, and then normalized to standardize the signal amplitude peak to 0dB. The pre-processed audio signal is converted into a time-frequency domain representation by short-time Fourier transform. The transformation adopts the Hamming window function with a window length of 1024 points and an overlap rate of 50%. The time domain representation is achieved by extracting features such as the envelope, zero-crossing rate, and energy change curve of the audio waveform, while the frequency domain representation is achieved by calculating spectral features such as power spectral density, Mel-frequency cepstral coefficients, and chromaticity features. The purpose of this step is to convert the original audio signal into time domain and frequency domain feature representations that are convenient for subsequent analysis, providing basic data for rhythm analysis.

[0053] The specific implementation method of step S02 is to use the same equipment and parameters as step S01 to collect the standard performance audio demonstrated by professional performers or teachers to ensure the consistency of the collection conditions. The standard performance audio undergoes the same preprocessing process as the Le Tong performance audio, including noise filtering, amplitude normalization, short-time Fourier transform and other processing to generate a standard time domain representation and a standard frequency domain representation. To ensure the effectiveness of the comparison, the speed difference between the standard performance and the Le Tong performance on the same piece of music should not exceed 15%. If this threshold is exceeded, the standard performance is adjusted through a time stretching algorithm. The time stretching is achieved using phase vocoder technology to adjust the performance speed while keeping the pitch unchanged. The purpose of this step is to establish an evaluation reference benchmark and provide a standard template for subsequent similarity calculations.

[0054] The specific implementation of step S03 is to construct a time window sequence based on the characteristics of the music. Two window construction strategies are used: a fixed duration strategy and an adaptive duration strategy. The fixed duration strategy evenly divides the audio into fixed-length time segments. For music segments with a simple rhythm, the fixed duration is set to 50 milliseconds; for music segments with a complex rhythm, the fixed duration is set to 30 milliseconds. The adaptive duration strategy dynamically determines the window length based on the beat structure of the music. For 4 / 4 music, the beat is subdivided into four subwindows; for 3 / 4 music, the beat is subdivided into three subwindows. Within each time window, the window similarity between the child's performance and the standard performance is calculated. The calculation method includes: time domain cross-correlation coefficient calculation, taking the maximum value of the cross-correlation function as the time domain similarity; frequency domain Euclidean distance calculation, taking the inverse of the Euclidean distance between spectral feature vectors as the frequency domain similarity; and Mel-frequency cepstral coefficient distance calculation, using a dynamic time warping algorithm to calculate the minimum deformation distance between the Mel-frequency cepstral coefficient sequences of the two signals. The final window similarity is obtained by weighted fusion of the above three similarities, with weight coefficients of 0.3, 0.3, and 0.4, respectively. The purpose of this step is to quantify the similarity between the children's performance and the standard performance in fine-grained time segments, providing local features for rhythm perception analysis.

[0055] The specific implementation of step S04 is to extract beat markers from the audio of the child performance by various methods. First, an envelope detection algorithm is used to extract the energy envelope of the audio signal, and the local peak points of the envelope are found to determine the possible beat positions; second, the autocorrelation function is used to analyze the periodicity of the audio signal, and the periodicity features of the audio signal are extracted; then the short-time energy analysis method is applied, the short-time energy change curve is calculated, and the energy mutation points are detected; finally, the frequency domain information is combined, and the time points with significant spectral changes are detected by spectrum flow calculation. The candidate beat points obtained by the above methods are fused by a confidence weighted voting mechanism, and the points with a confidence higher than the threshold value 0.7 are selected as the final beat markers. The time interval between adjacent beat markers is calculated to form a rhythm sequence, which contains the rhythm time value information in the child performance. The role of this step is to accurately extract the beat information from the audio signal and build a basic data structure for subsequent rhythm analysis.

[0056] The specific implementation of step S05 is to input the rhythm sequence obtained in step S04 and the window similarity sequence calculated in step S03 into the rhythm perception analysis model. The model uses a convolutional long short-term memory network structure, which includes multiple convolutional layers and long short-term memory network layers. First, the input data is processed by a one-dimensional convolutional layer to extract local time features, and the convolution kernel size is set to 3, 5 and 7, with 64 convolution kernels for each size; the convolution result is processed by batch normalization and then passed through a ReLU activation function; then it is input into a bidirectional long short-term memory network layer with 128 hidden units to capture long-term sequence dependencies; then the sequence data is processed by a self-attention mechanism with 8 attention heads to enhance the representation ability of important rhythm patterns; finally, it is mapped to a high-dimensional feature space through a fully connected layer to generate a rhythm feature representation vector. In this process, the window similarity is used to adjust the attention mechanism weight, and a higher attention weight is given to the area with low window similarity. The purpose of this step is to extract high-level semantic features of the rhythm sequence through a deep learning model and form a representation vector that comprehensively reflects the performance rhythm characteristics.

[0057] The specific implementation of step S06 is to apply a multi-level time model to analyze the rhythm feature representation vector obtained in step S05. The model analyzes the rhythm characteristics from three levels of macro, meso, and micro. The macro level analyzes the periodicity of the section structure and calculates the stability index of the rhythm pattern between adjacent sections; the meso level analyzes the beat group organization and evaluates the regularity and expressive change of the beat organization within the bar; the micro level analyzes the accuracy of note duration and calculates the deviation of the actual duration of each note from the theoretical duration. The analysis results are quantified as a multi-dimensional deviation vector, with 12 dimensions including absolute duration deviation, relative rhythm proportion deviation, intra-beat structure deviation, rhythm stability deviation, and expressive fluctuation deviation. The absolute duration deviation threshold is set to ±50 milliseconds, the relative rhythm proportion deviation threshold is ±5%, the intra-beat structure deviation threshold is ±30 milliseconds, the rhythm stability deviation uses the coefficient of variation, and the threshold is 0.1, and the expressive fluctuation deviation is determined based on expert scoring standards. The purpose of this step is to quantitatively analyze the performance rhythm characteristics from multiple time scales and generate a deviation vector that comprehensively reflects the rhythm performance.

[0058] The specific implementation of step S07 is to construct a rhythm deviation tolerance range matrix based on a large-scale performance database. First, collect professional performance recordings of different music styles (classical, jazz, pop, etc.) and classify them according to difficulty levels (beginner, intermediate, advanced); then hire professional reviewers to score the recordings and extract rhythm deviation features from professional performances; then statistically analyze the rhythm deviation distribution characteristics under different styles and difficulty levels; finally, establish a multi-dimensional tolerance range matrix, where each element of the matrix represents the tolerance range of a specific rhythm feature under specific style and difficulty conditions. For classical music, the absolute duration deviation tolerance range is ±30 milliseconds, and the relative rhythm proportion deviation tolerance range is ±3%; for jazz music, these ranges are widened to ±60 milliseconds and ±8%, respectively. The matrix also includes weight coefficients for each parameter, reflecting the importance of different rhythm features in overall evaluation. The purpose of this step is to establish an evaluation reference standard to distinguish between technical errors and artistic treatment, providing a basis for rhythm performance scoring.

[0059] The specific implementation of step S08 is to calculate the similarity between the multi-dimensional deviation vector obtained in step S06 and the corresponding deviation vector of the standard performance template. The calculation uses a weighted cosine distance method, taking into account the importance differences of different dimensions. First, normalize the multi-dimensional deviation vector and the standard deviation vector to unify the numerical range of each dimension; then determine the weight of each dimension according to the tolerance range matrix established in step S07; then calculate the weighted cosine similarity to obtain a deviation similarity index with a value range of 0 to 1, where 1 represents complete consistency and 0 represents complete difference. For key dimensions such as rhythm stability, the weight coefficient is set to 0.4; for performance detail dimensions, the weight coefficient is set to 0.2. The purpose of this step is to comprehensively evaluate the closeness of the child's performance and the standard performance in rhythm processing, and to provide quantitative basis for the final score.

[0060] The specific implementation of step S09 is to output the rhythm performance score by applying the rhythm evaluation neural network model. This model uses a deep neural network architecture that integrates attention mechanism and residual connection, containing 10 layers, each containing 128 neurons. The input layer receives the multi-dimensional deviation vector of step S06, the deviation similarity of step S08, and the window similarity sequence statistical features of step S03; the hidden layer uses a multi-head attention mechanism with the same number of heads as the dimension of the multi-dimensional deviation vector, i.e. 12; residual connection spans every two layers, effectively alleviating the gradient vanishing problem; the output layer generates four-dimensional sub-scores (rhythm accuracy, rhythm stability, rhythm expressiveness, rhythm technicality) and a comprehensive score. The score range is 0 to 100 points, where 0 to 60 points represents failing, 61 to 70 points represents passing, 71 to 80 points represents good, 81 to 90 points represents excellent, and 91 to 100 points represents outstanding. The purpose of this step is to generate the final rhythm performance evaluation result and provide objective and quantitative performance level feedback to the child.

[0061] The detailed structure of the rhythm perception analysis model is a hybrid architecture combining multi-layer time convolution network and bidirectional long short-term memory network. The time domain feature extraction layer is composed of 3 parallel convolution blocks, each containing one-dimensional convolution filters of different scales (3x1, 5x1, 7x1), with 64 filters to capture rhythm features at different time scales. The frequency domain feature extraction layer uses a 13-order mel frequency cepstral coefficient transformation combined with a two-dimensional convolution network to extract timbre and pitch features, with a convolution kernel size of 3x3 and 32 convolution kernels. The multi-scale convolution layer contains three one-dimensional convolution operations with different kernel sizes (3, 5, 9), which expand the receptive field range through dilated convolution technology to capture short and long-term dependencies. The bidirectional long short-term memory layer contains 2 layers with 128 hidden units in each layer to model the context correlation and maintain long-term memory. The self-attention mechanism layer uses an 8-head attention structure with 64 dimensions for each attention head to calculate the correlation strength between different time points and enhance the representation of key rhythm patterns through attention scores. The fully connected output layer contains 3 layers with 256, 128, and 64 neurons respectively to map high-dimensional features to semantic-level rhythm representation space. The model adds residual connections between each main layer to effectively alleviate the gradient vanishing problem and maintain information flow. In the feature fusion stage, a gating mechanism is used to dynamically adjust the fusion weight of time domain features and frequency domain features to enhance the model's adaptability to different types of rhythm patterns.

[0062] The detailed steps for establishing the training dataset of the rhythm perception analysis model include: first, collecting more than 10,000 standard performance audios of different difficulty levels (from level 1 to level 10) and style genres (classical, jazz, popular, national, etc.) from international renowned music academies such as Juilliard School, Royal College of Music, etc. The audio length ranges from 30 seconds to 5 minutes. Collect 15,000 performance samples of learners of different levels (beginners, intermediate learners, advanced learners) from more than 300 music education institutions around the world, ensuring that the dataset covers learners of different ages and cultural backgrounds. Standardize the preprocessing of all audios, including sampling rate unification (converted to 44.1 kHz) and volume normalization processing (standardized to -23 LUFS). Hire 20 professional music reviewers (including music academy professors, professional performers, and experienced music education experts) to score the rhythm performance of each audio in multiple dimensions (accuracy, stability, expressiveness, and technicality, each dimension scored from 0 to 100) and record comments. Extract the time domain features (envelope line, zero crossing rate, energy curve, etc.) and frequency domain features (Mel-frequency cepstral coefficients, chroma diagram, spectral centroid, etc.) in the audio as model input, and ensure the consistency of the feature extraction process. Label the expert scores as supervision signals to establish the mapping relationship data pairs between audio features and expert scores. Stratified sampling according to difficulty level and genre type, dividing the dataset into training set, validation set and test set in the ratio of 8:1:1. Data augmentation processing is performed on the training set, including adding different levels of environmental noise (signal-to-noise ratio in the range of 10dB to 30dB), adjusting speed variation (±10% range) and introducing random time offset operation (±50 milliseconds range), improving model robustness and enhancing model generalization ability.

[0063] It should be noted that in the training process of deep neural networks, especially convolutional neural networks, the phenomenon of feature direction drift often occurs. The drift phenomenon is mainly caused by the uneven distribution of training data and the over-optimization of the model in certain feature directions. In particular, in music rhythm analysis, the differences between different styles, playing techniques and performance methods may cause the model to tend to overfit certain patterns. When the convolutional network level deepens, the feature map may gradually concentrate in a few directions, causing the so-called "feature space collapse", that is, the obtained features cannot reflect the actual situation, resulting in a decrease in the sensitivity of the model to subtle rhythm changes. This drift refers to the tendency of the convolutional layers in the network to over-emphasize certain specific direction features and ignore other direction information, resulting in the collapse of the feature representation space into a low-dimensional subspace. In the context of music rhythm evaluation, this drift may cause the model to be overly sensitive to certain types of rhythm features, while being less sensitive to other equally important rhythm performance features, thereby affecting the comprehensiveness and accuracy of the evaluation. Traditional neural network training methods usually cannot effectively detect and solve this feature direction drift problem, especially when dealing with music, which is highly complex and multi-dimensional.

[0064] The present application can monitor the direction distribution state of each convolutional layer output feature map in real time by introducing a directional drift detection function in the training process of the rhythm perception analysis model. The detection function accurately quantifies the degree of direction diversity of the feature space by calculating the direction consistency index of the feature map. When the direction consistency index exceeds the preset threshold, it indicates that the convolutional layer has a risk of feature direction deviation, at which time the system will automatically trigger the convolution adjustment mechanism. This active detection mechanism ensures that the feature extraction during model training maintains direction balance and avoids overfitting to certain rhythm patterns.

[0065] The convolution adjustment mechanism matched with the directional drift detection adopts a sparse matrix jitter enhancement method, which is a fine convolution parameter adjustment technique. This technique increases the direction diversity of feature representation by introducing controlled random noise to sparse the convolution weight matrix. Compared with traditional regularization methods, the sparse matrix jitter enhancement method not only prevents overfitting, but more importantly, maintains the sensitivity of the convolutional layer to multiple rhythm features, ensuring the rich expression ability of the feature space. Through dimensionality reduction reconstruction technology, the system maintains the semantic integrity of the features while increasing the direction diversity, avoiding the loss of effective information.

[0066] The combination of the directional drift detection and the convolution adjustment mechanism enables the rhythm perception analysis model to maintain the comprehensiveness and balance of feature extraction during the training process. Compared with the traditional neural network training method, the method can better capture the multi-dimensional features in the music rhythm and has appropriate sensitivity to different types of rhythm performance. This technical design greatly improves the model's understanding of complex music performance, making the evaluation results more consistent with the comprehensive judgment of human experts, providing more comprehensive and accurate rhythm performance evaluation for the young musicians, and further guiding the improvement of their playing skills and musical expressiveness.

[0067] The first core technical idea of the application is a rhythm perception analysis model based on a convolutional long short-term memory network structure. This model breaks through the limitations of traditional audio comparison methods and can not only capture the surface features of audio signals but also extract high-dimensional semantic rhythm features through deep learning. Traditional methods usually only focus on simple pitch matching or time alignment and cannot understand the semantic connotation of music performance. The rhythm perception analysis model of the application adopts a composite structure of time domain feature extraction layer, frequency domain feature extraction layer and multi-scale convolution layer, which can analyze rhythm features at different time scales and understand rhythm processing in music performance more comprehensively. Especially when dealing with artistic performance changes, the model can distinguish between intentional artistic processing and unintentional technical errors, providing a more detailed basis for evaluation.

[0068] The second core technical idea is a multi-level time model that analyzes the temporal relationship of the performance from three levels: macro-rhythm structure, meso-beat group organization and micro-note duration. Traditional evaluation methods often only focus on the rhythm accuracy of a single level and ignore the mutual relationship of rhythm processing at different time scales. The application can capture the rhythm processing features of the performer at different time scales through multi-level analysis, forming a more three-dimensional evaluation perspective. This hierarchical analysis method conforms to the natural process of human music cognition, making the evaluation results more consistent with the aesthetic judgment of human experts and avoiding the one-sidedness of mechanical scoring.

[0069] The third core technical idea is the rhythm deviation tolerance range matrix, which establishes a multi-dimensional evaluation reference standard according to different music styles, piece difficulty and performance level. Traditional methods usually use fixed standards to evaluate all types of music performance, which cannot adapt to the special requirements of different styles on rhythm processing. The rhythm deviation tolerance range matrix of the application defines the extent to which various rhythm deviations are technical errors or artistic expressions, making the evaluation standard adaptive. This flexible method of defining evaluation standards respects the diversity of music performance and avoids misjudging artistic processing as technical errors, improving the professional level of evaluation.

[0070] The synergy of the three core technical ideas produces significant technical advantages. The rhythm perception analysis model provides high-dimensional semantic feature representation, the multi-level time model provides a multi-scale analysis framework, and the rhythm deviation tolerance range matrix provides flexible evaluation criteria. The three work together to form a complete evaluation system. This synergistic mechanism enables the system to capture subtle technical differences while considering the diversity of artistic expression, achieving a balance between technical accuracy and artistic expressiveness. Compared to traditional methods that use a single dimension and mechanical evaluation, the present invention can provide more comprehensive and personalized evaluation results, better guiding the learning progress of the child and promoting the overall development of their musical performance ability.

[0071] The second aspect of the present application provides a computer-readable storage medium, wherein program instructions are stored in the computer-readable storage medium, and the program instructions are used to execute the above-mentioned child practice condition evaluation method when running in a computer.

[0072] The third aspect of the present application provides a child practice condition evaluation system, which includes the above-mentioned computer-readable storage medium. The system is any one of a computer, a server, or a single-chip microcomputer. The computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer-readable storage medium.

[0073] Specifically, the principle of the present application is that the technical principle of the present application is based on the cross-fusion of deep learning and music cognitive science, and the precise evaluation of the child's performance rhythm is realized by constructing a multi-level information processing architecture. The core principles include three aspects: signal processing and feature extraction, multi-level time modeling and rhythm perception analysis.

[0074] In the aspect of signal processing and feature extraction, the present application first preprocesses the child's performance audio and standard performance audio, converts them into time domain representation and frequency domain representation, and establishes an acoustic feature space. By constructing a time window sequence, the system calculates the similarity in each window to accurately locate the performance differences. At the same time, beat markers are extracted and rhythm sequences are calculated to provide a basic data structure for subsequent analysis. This multi-dimensional signal processing method can fully capture the time-frequency features of the performance audio, avoiding the limitations of single feature extraction methods.

[0075] In the aspect of multi-level time modeling, the present application analyzes the performance rhythm features at three levels: macro, meso, and micro. The macro level focuses on the overall rhythm structure, the meso level focuses on the beat group organization, and the micro level accurately analyzes the note duration. This hierarchical analysis method conforms to the natural process of human music cognition and can consider both technical accuracy and artistic expressiveness, breaking down complex performance into quantifiable multi-dimensional features. Combined with the rhythm deviation tolerance range matrix, the system can flexibly evaluate the rationality of rhythm deviation according to different musical styles, avoiding mechanical evaluation.

[0076] In the rhythm perception analysis aspect, the application constructs a rhythm perception analysis model based on a convolution long short-term memory network, which integrates time domain feature extraction, frequency domain feature extraction, multi-scale convolution and self-attention mechanism, and can extract high-dimensional semantic rhythm feature representation from the original audio. In order to solve the directional drift problem in neural network training, the application introduces a directional drift detection function and a sparse matrix jitter enhancement method to ensure that the model maintains directional diversity when capturing rhythm features and prevents feature space from collapsing. Combined with a multi-task learning framework, the system can simultaneously evaluate rhythm accuracy, stability, expressiveness and technicality in four dimensions, providing comprehensive evaluation.

[0077] In addition, the application adopts large-scale data set training and transfer learning strategy to ensure that the model adapts to different music styles and difficulty levels, and improves the universality and accuracy of evaluation. The model evaluation results are verified by expert double-blind test to ensure the high consistency of the evaluation results and the human expert scores.

[0078] A specific embodiment 1 of the application is provided below, and the specific implementation of each step in the embodiment 1 is described in detail as follows.

[0079] The specific implementation of step S01 is to use a high-precision digital audio acquisition device to record the performance process of a child musician, set the sampling rate to 44.1 kHz, and set the quantization precision to 24 bits. The collected audio signal is subjected to low-pass filtering to remove high-frequency noise, and then normalized to standardize the signal amplitude peak value to 0 dB. The low-pass filter adopts a Butterworth filter, and the cutoff frequency is set to 20 kHz, and the filter order is 4. The mathematical expression of signal normalization processing is as follows:

[0080]

[0081] In the formula, x(n) is the original audio signal; x norm (n) is the normalized audio signal; max(|x(n)|) is the maximum absolute value of the original signal; n is the sampling point index.

[0082] The preprocessed audio signal is converted into time-frequency domain representation by short-time Fourier transform, and the Hamming window function is used for the transform, the window length is 1024 points, and the overlap rate is 50%. The mathematical expression of short-time Fourier transform is as follows:

[0083]

[0084] In the formula, X(m, k) is the time-frequency representation; x(n) is the audio signal; w(n) is the Hamming window function; m is the frame index; k is the frequency index; N is the window length, which is 1024; j is the imaginary unit.

[0085] The Hamming window function is defined as:

[0086]

[0087] The time-domain feature extraction includes calculating the envelope E(m), the zero-crossing rate Z(m) and the energy variation curve P(m) of the audio waveform, and the calculation formula is as follows:

[0088] E(m) = max(|x(n)|), mL≤n<(m+1)L;

[0089]

[0090] In the formula, L is the frame length; shn(x) is a sign function, which takes 1 when x≥0 and -1 when x<0.

[0091] The frequency-domain feature extraction includes calculating the power spectral density S(m, k), the mel-frequency cepstral coefficient MFCC(m, p) and the chroma feature C(m, q), and the calculation formula is as follows:

[0092] S(m, k) = |X(m, k)| 2 ;

[0093]

[0094] In the formula, M(m, l) is the output of the mel filter bank; p is the cepstral coefficient index; L is the number of mel filters, which is 40; Ω q is the frequency index set corresponding to the qth chroma pair; q is the chroma index, which ranges from 0 to 11. The purpose of this step is to convert the original audio signal into time-domain and frequency-domain feature representations for subsequent analysis, providing basic data for rhythm analysis.

[0095] The specific implementation of step S02 is to collect standard performance audio demonstrated by professional performers or teachers using the same equipment and parameters as step S01, to ensure consistency of the collection conditions. The standard performance audio is preprocessed in the same way as the child performer's performance audio to generate standard time-domain and frequency-domain representations. To ensure the effectiveness of the comparison, the standard performance speed needs to be adjusted so that the speed difference between the child performer's performance on the same piece does not exceed 15%. The speed difference calculation formula is as follows:

[0096]

[0097] In the formula, Δv is the speed difference percentage; T s is the time length of each beat of the standard performance; T p is the time length of each beat of the child performer's performance.

[0098] When Δv > 15%, the standard performance is adjusted by time stretching algorithm, which is realized by phase vocoder technology, and the performance speed is adjusted while the pitch is kept unchanged. The mathematical expression of time stretching is as follows:

[0099]

[0100] In the formula, y(n) is the signal after time stretching; X(m, k) is the short-time Fourier transform of the original signal; φ y (m, k) is the corrected phase; h(n) is a synthesis window function; α is a time stretching factor, α = T p / T s ; R is a frame shift. The role of this step is to establish an evaluation reference, and to provide a standard template for subsequent similarity calculation.

[0101] The specific implementation of step S03 is to construct a time window sequence according to the music characteristics, and the window construction adopts a fixed length strategy and an adaptive length strategy. For the fixed length strategy, the window length W f is determined according to the rhythm complexity:

[0102]

[0103] For the adaptive length strategy, the window length W a is dynamically determined according to the music beat structure:

[0104]

[0105] In the formula, tempo is the tempo of the piece (unit: beats per minute); d is the subdivision number within a beat, d = 4 for four-four beat and d = 3 for three-four beat.

[0106] The window similarity of the student performance and the standard performance is calculated in each time window, including the time domain cross-correlation coefficient S time , the frequency domain Euclidean distance similarity S freq , and the mel-frequency cepstral coefficient distance similarity S MFCC :

[0107]

[0108] In the formula, x p (n) is the student performance audio signal; x s (n) is the standard performance audio signal; X p (i, k) and X s (i, k) are the short-time Fourier transforms of the student performance and the standard performance in the ith window, respectively; MFCC p (i) and MFCC s(i) the mel-frequency cepstral coefficient sequence of the child performance and the standard performance in the ith window, respectively; DTW is the distance calculated by dynamic time warping algorithm.

[0109] the final window similarity S w (i) is obtained by weighted fusion of the above three similarities:

[0110] S w (i) = α1S time (i) + α2S freq (i) + α3S MFCC (i).

[0111] In the formula, α1, α2 and α3 are weight coefficients, and take values of 0.3, 0.3 and 0.4 respectively, and satisfy α1+α2+α3=1. The purpose of this step is to quantify the similarity of the child performance and the standard performance in the fine-grained time segment, and to provide local features for rhythm perception analysis.

[0112] The specific implementation of step S04 is to extract beat markers from the child performance audio by various methods. First, an envelope detection algorithm is used to extract the energy envelope line of the audio signal, and the local peak points of the envelope line are used to determine the possible beat positions. The energy envelope E(n) is calculated as follows:

[0113] E(n) = |x(n)| 2 *h(n).

[0114] In the formula, x(n) is the audio signal; h(n) is the impulse response of a low-pass filter; * represents convolution operation.

[0115] Secondly, the periodicity of the audio signal is analyzed by using the autocorrelation function, and the calculation formula is as follows:

[0116]

[0117] In the formula, R(k) is the autocorrelation function; k is the time delay; N is the signal length.

[0118] Then, the short-time energy analysis method is applied, the short-time energy change curve is calculated, and the energy mutation point is detected:

[0119]

[0120] ΔE st (m) = E st (m) - E st (m-1).

[0121] In the formula, E st (m) is the short-time energy of the mth frame; L is the frame shift; ΔE st (m) is the short-time energy change rate.

[0122] Finally, combined with the frequency domain information, the time point when the spectrum changes significantly is detected through spectral flow calculation:

[0123]

[0124] Where SF(m) is the spectral flow of the m-th frame; X(m, k) is the short-time Fourier transform of the m-th frame.

[0125] The candidate beat points obtained by the above method are fused through the confidence weighted voting mechanism to calculate the confidence score C(t) of each time point:

[0126]

[0127] Where C E (t), C R (t), and C SF (t) is the confidence level obtained based on the energy envelope, autocorrelation function, short-time energy and spectral flow respectively; β1, β2, β3 and β4 are weight coefficients, which take values ​​of 0.3, 0.2, 0.3 and 0.2 respectively, and satisfy β1+β2+β3+β4=1.

[0128] Select the points with confidence higher than the threshold value 0.7 as the final beat markers to form a beat marker sequence B = {b1, b2, ..., b M}. Calculate the time intervals between adjacent beat markers to form a rhythm sequence T = {t1, t2, ..., t M-1}, where t i =b i+1 -b i The purpose of this step is to accurately extract beat information from the audio signal and build a basic data structure for subsequent rhythm analysis.

[0129] The specific implementation of step S05 is to compare the rhythm sequence T obtained in step S04 with the window similarity sequence S calculated in step S03. w Enter the rhythm perception analysis model. This model uses a convolutional long short-term memory network structure, which includes multiple convolutional layers and long short-term memory layers. Window similarity is used to adjust the attention mechanism weights. The attention weight matrix A is calculated as follows:

[0130]

[0131] Where A i，j is the element of attention weight matrix; S w (i, j) is the similarity between the i-th window and the j-th window; λ is the attention focus parameter, which is 5; N is the number of windows.

[0132] For the area with low window similarity, a higher attention weight is given, i.e. S w The smaller the (i, j) value is, the larger the A i，j value is. The rhythm perception analysis model outputs a rhythm feature representation vector F with a dimension of 128. The purpose of this step is to extract high-level semantic features of the rhythm sequence through a deep learning model, forming a representation vector that comprehensively reflects the performance rhythm characteristics.

[0133] The specific implementation of step S06 is to apply a multi-level time model to analyze the rhythm feature representation vector F obtained in step S05. This model analyzes rhythm characteristics from three levels: macro, meso, and micro. The macro level analyzes the periodicity of the section structure and calculates the stability index S macro :

[0134]

[0135] where K is the number of sections; P k is the rhythm pattern vector of the kth section; and sim is the cosine similarity function:

[0136]

[0137] The meso level analyzes the beat group organization and evaluates the regularity and performance changes of the beat organization within a measure. The beat group organization regularity R meso :

[0138]

[0139] where σ intra is the standard deviation of the beat interval within a measure; and σ inter is the standard deviation of the interval between measures.

[0140] The micro level analyzes the accuracy of note duration and calculates the deviation D micro :

[0141]

[0142] where t i is the actual duration of the ith note; is the theoretical duration of the ith note; and N is the number of notes.

[0143] The analysis result is quantized into a multi-dimensional deviation vector D = [d1, d2,..., d 12 ] T , which includes the following dimensions:

[0144] d1: absolute duration deviation, threshold ± 50 milliseconds;

[0145] d2: relative tempo ratio deviation, threshold value ±5%;

[0146] d3: intra-beat structure deviation, threshold value ±30 milliseconds;

[0147] d4: tempo stability deviation, expressed by coefficient of variation, threshold value 0.1;

[0148] d5: macro-periodicity deviation;

[0149] d6: meso-beat group deviation;

[0150] d7: micro-note coherence deviation;

[0151] d8: strong beat accuracy deviation;

[0152] d9: weak beat accuracy deviation;

[0153] d 10 : tempo acceleration trend deviation;

[0154] d 11 : tempo deceleration trend deviation;

[0155] d 12 : expressive fluctuation deviation.

[0156] wherein the coefficient of variation CV is calculated according to the formula:

[0157]

[0158] In the formula, σ is the standard deviation; μ is the mean value. The role of this step is to quantitatively analyze the performance tempo characteristics from multiple time scales, and generate a deviation vector that comprehensively reflects the tempo performance.

[0159] The specific implementation of step S07 is to construct a tempo deviation tolerance range matrix M based on a large-scale performance database. First, collect professional performance recordings of different music styles (classical, jazz, pop, etc.) and classify them according to difficulty levels (beginner, intermediate, advanced); then extract the tempo deviation characteristics in professional performances; then statistically analyze the tempo deviation distribution characteristics under different styles and difficulty levels; finally, establish the tolerance range matrix M, with dimensions of 12xSxL, where 12 is the dimension of the deviation vector, S is the number of styles, and L is the number of difficulty levels.

[0160] The matrix element M i，j，k represents the tolerance range of the i-th deviation dimension under the j-th style and the k-th difficulty level, and the calculation formula is as follows:

[0161] M i，j，k = [μ i，j，k - γσ i，j，k , μ i，j，k + γσ i，j，k];

[0162] where μ i，j，k is the average value of the i-th bias dimension at the j-th style and k-th difficulty level; σ i，j，k is the corresponding standard deviation; and γ is the tolerance range coefficient, taking a value of 2.

[0163] The tolerance range matrix also contains a weight coefficient matrix W, with the same dimension of 12xSxL, and the element W i，j，k represents the weight coefficient of the i-th bias dimension at the j-th style and k-th difficulty level, satisfying:

[0164]

[0165] For classical music, the absolute time value bias tolerance range is ±30 milliseconds, and the relative rhythm proportion bias tolerance range is ±3%; for jazz music, these ranges are relaxed to ±60 milliseconds and ±8%, respectively. The purpose of this step is to establish an evaluation reference standard to distinguish technical errors from artistic processing and provide a basis for rhythm performance scoring.

[0166] The specific implementation of step S08 is to calculate the similarity between the multi-dimensional bias vector D obtained in step S06 and the corresponding bias vector D s of the standard performance template. The calculation uses the weighted cosine distance method, taking into account the importance differences of different dimensions. First, normalize the multi-dimensional bias vector and the standard bias vector:

[0167]

[0168] Then determine the dimension weight W i，j，k according to the tolerance range matrix established in step S07; then calculate the weighted cosine similarity S D :

[0169]

[0170] where D norm，i and D s，norm，i are the i-th elements of the normalized bias vector and the standard bias vector, respectively; W i，j，k is the corresponding weight coefficient. The value range of S D is 0 to 1, where 1 represents complete consistency and 0 represents complete difference.

[0171] For key dimensions such as rhythm stability, the weight coefficient is set to 0.4; for expressive detail dimensions, the weight coefficient is set to 0.2. The purpose of this step is to comprehensively evaluate the closeness of the child's performance to the standard performance in rhythm processing and provide a quantitative basis for the final score.

[0172] The specific implementation of step S09 is to use the rhythm evaluation neural network model to output the rhythm performance score. The model adopts a deep neural network architecture that integrates the attention mechanism and residual connection, and contains a 10-layer structure, with each layer containing 128 neurons. The input layer receives the multi-dimensional deviation vector D of step S06 and the deviation similarity S of step S08. D And the window similarity sequence S in step S03 w The statistical characteristics of . The score generation process can be expressed as:

[0173] Score = f NN (D,S D , S w , stat );

[0174] Where, f NN is the neural network mapping function; S w，stat is the statistical characteristics of the window similarity sequence, including mean, variance, maximum value, minimum value, etc.

[0175] Specifically, the score is divided into four dimensions: rhythm accuracy Score acc , Rhythm Stability Score stab , Rhythmic Expression Score expr and rhythmic technical score tech :

[0176] Socre acc =f acc (D1, S D );

[0177] Score stab =f stab (D2, S D );

[0178] Score expr =f expr (D3, S D );

[0179] Score tech =f tech (D4, S D );

[0180] Where D1, D2, D3 and D4 are the sub-vectors related to accuracy, stability, expressiveness and technicality in the multi-dimensional deviation vector D respectively; acc 、f stab 、f expr and f tech is the corresponding scoring function.

[0181] Comprehensive score Scoretotal Calculated by weighted average:

[0182] Score total =γ1Score acc +γ2Score stab +γ3Score expr +γ4Score tech ;

[0183] Where γ1, γ2, γ3 and γ4 are weight coefficients, which are 0.3, 0.3, 0.2 and 0.2 respectively, and satisfy γ1+γ2+γ3+γ4=1.

[0184] Scores range from 0 to 100, with 0 to 60 indicating failing, 61 to 70 indicating passing, 71 to 80 indicating good, 81 to 90 indicating excellent, and 91 to 100 indicating outstanding. The purpose of this step is to generate the final rhythm performance evaluation results, providing objective and quantitative feedback on the children's performance level.

[0185] The detailed structure of the rhythm perception analysis model is a hybrid architecture that combines a multi-layer temporal convolutional network with a bidirectional long short-term memory network. The temporal feature extraction layer consists of three parallel convolutional blocks, each of which contains one-dimensional convolutional filters of different scales (3×1, 5×1, and 7×1). The total number of filters is 64, expressed as:

[0186] H t，i =σ(W conv，i *X t +b conv，i );

[0187] Where H t，i is the output feature map of the i-th convolutional block; W conv，i is the convolution kernel parameter; X t is the time domain input feature; b conv，i is the bias term; σ is the activation function, which uses the ReLU function; * represents the convolution operation.

[0188] The frequency domain feature extraction layer uses a 13th-order Mel-frequency cepstral coefficient transform combined with a two-dimensional convolutional network to extract timbre and pitch features. The convolution kernel size is 3×3 and the number of convolution kernels is 32, which can be expressed as:

[0189] H f =σ(W conv，f *X f +b conv，f );

[0190] Where H f is the output of the frequency domain feature extraction layer; W conv，f is the convolution kernel parameter; Xf is the frequency domain input feature; b conv，f is the bias term.

[0191] The multi-scale convolution layer captures short-term and long-term dependencies through one-dimensional convolution operations with different kernel sizes, denoted as:

[0192] H ms = [H ms，1 , H ms，2 , H ms，3 ];

[0193] H ms，i = σ(W ms，i *H cat +b ms，i );

[0194] where H ms is the output of the multi-scale convolution layer; H ms，i is the convolution output of the i-th scale; W ms，i is the convolution kernel parameter of the i-th scale; H cat is the concatenation of the time domain feature and the frequency domain feature; b ms，i is the bias term.

[0195] The bidirectional long short-term memory layer models the correlation of the sequence before and after and maintains long-term memory, denoted as:

[0196]

[0197] where LSTM f and LSTM b are the forward and backward long short-term memory networks, respectively; and are the forward and backward hidden states, respectively; h t is the concatenation of the bidirectional output.

[0198] The self-attention mechanism layer calculates the correlation strength between different time points, denoted as:

[0199] Q = W Q h;

[0200] K = W K h;

[0201] V = W V h;

[0202]

[0203] H att = AV;

[0204] where Q, K, V are the query, key, and value matrices, respectively; W Q , WK , W V are the corresponding parameter matrices; h is the bidirectional long short-term memory layer output; d k is the dimension of the key vector; A is the attention weight matrix; H att is the attention layer output.

[0205] The fully connected output layer maps the high-dimensional features to the semantic-level rhythm representation space, represented as:

[0206] F1 = σ(W1H att +b1);

[0207] F2 = σ(W2F1+b2);

[0208] F = W3F2+b3;

[0209] where W1, W2, W3 are weight matrices; b1, b2, b3 are bias vectors; F1, F2 are intermediate layer outputs; F is the final rhythm feature representation vector.

[0210] The model effectively alleviates the gradient vanishing problem and maintains information flow by adding residual connections between each main layer, and the residual connection is represented as:

[0211] H out = H in +f(H in );

[0212] where H in is the layer input; f is the layer mapping function; H out is the output after adding the residual connection.

[0213] During the model training process, a directional drift detection function is set for each convolutional layer, which judges whether there is a feature direction shift by calculating the direction consistency index of the convolutional output feature map. The direction consistency index DCI is calculated as follows:

[0214] G x (i, j) = H(i+1, j)-H(i-1, j);

[0215] G y (i, j) = H(i, j+1)-H(i, j-1);

[0216]

[0217] Hist(θ) = {n1, n2,..., n B};

[0218]

[0219] where Gx and G y are the horizontal and vertical gradients, respectively; H is the output feature map of the convolution layer; θ is the gradient direction angle; Hist(θ) is the histogram of the direction angle; n k is the number of gradients in the kth direction interval; N is the total number of gradients; B is the number of histogram intervals, which is 8. The value range of DCI is 0 to 1, and the larger the value is, the more serious the direction deviation is.

[0220] When the DCI exceeds the set threshold value 0.7, the convolution adjustment mechanism is triggered, and the sparse matrix jitter enhancement method is used to disturb and reconstruct the convolution kernel parameters. The sparse matrix jitter enhancement method includes three steps of sparsity control, direction entropy maximization, and weight renormalization, which are mathematically expressed as follows:

[0221] Sparsity control:

[0222] M sparse = W conv ⊙ I sparse ;

[0223]

[0224] In the formula, W conv is the original convolution weight matrix; I sparse is the sparse mask matrix; s is the sparsity parameter, which takes a value in the range of 0.3 to 0.5; ⊙ represents the element product operation; rand(0, 1) is a random number between 0 and 1.

[0225] Direction entropy maximization:

[0226] W disturb = M sparse + α· N(0, σ 2 );

[0227] In the formula, W disturb is the disturbed weight matrix; N(0, σ 2 ) is Gaussian noise with a mean of 0 and a variance of σ 2 ; α is the disturbance intensity parameter, which is initially 0.1 and is dynamically adjusted according to the change of DCI; σ 2 is the noise variance, which takes a value of 0.01.

[0228] Weight renormalization:

[0229]

[0230] In the formula, W new is the reconstructed convolution weight matrix; ||W|| F represents the Frobenius norm of the matrix W.

[0231] The method for adjusting the internal feature weight distribution of the window similarity is a dynamic attention mechanism, and the attention mask matrix M is generated according to the window similarity sequence att :

[0232] M att (i,j)=1-S w (i,j);

[0233] The attention mask matrix is applied to the internal feature mapping process of the rhythm perception analysis model:

[0234] H′ att =H att ⊙M att ;

[0235] In the formula, H att is the original feature mapping; and H′ att is the adjusted feature mapping.

[0236] The steps for establishing the training data set of the rhythm perception analysis model specifically include collecting standard performance audios of different difficulty levels and style genres from the International Standard Music Academy Performance Library, and collecting performance samples of learners of different levels from global music education institutions. Standardized preprocessing is performed on all audios, including sampling rate unification and volume normalization processing. The sampling rate unification converts all audios to 44.1kHz, and the volume normalization processing standardizes all audios to -23LUFS (Loudness Unit Relative Full Scale). A professional music reviewer is invited to score the rhythm performance of each audio in multiple dimensions, including rhythm accuracy S acc , rhythm stability S stab , rhythm expressiveness S expr and rhythm technicality S tech , and the score of each dimension ranges from 0 to 100.

[0237] Time domain features X t and frequency domain features X f in the audio are extracted as model inputs:

[0238] X t =[E(n),Z(m),P(m)];

[0239] X f =[S(m,k),MFCC(m,p),C(m,q)];

[0240] The expert score label is used as a supervision signal to establish a mapping relationship data pair between the audio features and the expert score wherein X i =[X t，i , X f，i ] is the feature of the i-th sample, and Y i= [S acc，i , S stab，i , S expr，i , S tech，i ] are the corresponding expert scores.

[0241] According to the difficulty level and style type, the data set is divided into training set D train , validation set D val and test set D test in the ratio of 8:1:1. Data augmentation processing is performed on the training set, including adding different levels of environmental noise, adjusting speed changes, and introducing random time offset operations. The addition of environmental noise is represented as:

[0242] x noise (n) = x(n) + β·noise(n);

[0243] In the formula, x(n) is the original audio signal; noise(n) is the environmental noise signal; β is the noise intensity parameter, so that the signal-to-noise ratio is in the range of 10dB to 30dB; x noise (n) is the signal after adding noise.

[0244] The speed change is represented as:

[0245] x speed (n) = x(n / r);

[0246] In the formula, r is the speed change ratio, which is in the range of 0.9 to 1.1; x speed (n) is the signal after adjusting the speed.

[0247] The random time offset operation is represented as:

[0248] x shift (n) = x(n+δ);

[0249] In the formula, δ is the time offset, which is in the range of ±50 milliseconds; x shift (n) is the signal after time offset.

[0250] The steps of the rhythm perception analysis model training specifically include first performing self-supervised training on large-scale unlabeled music data, and using two training tasks of time mask prediction and rhythm consistency contrast learning. The time mask prediction task randomly masks the time segment in the input sequence and requires the model to restore the masked content according to the context. The mask generation process is represented as:

[0251]

[0252] In the formula, X(n) is the original input sequence; X masked (n) is the masked sequence; Ω maskMask is a special mask label.

[0253] The rhythm consistency contrastive learning task requires the model to distinguish different performance versions of the same music piece and different music pieces. The contrastive learning loss function L contrast is defined as:

[0254]

[0255] where f(x) is the feature representation of the model output; x i and are different performance versions of the same music piece; sim is the cosine similarity function; τ is the temperature coefficient, taking the value of 0.1; B is the batch size.

[0256] After self-supervised training, supervised fine-tuning is performed on the labeled expert rating dataset. A multi-task learning framework is used to simultaneously predict rhythm stability, expressiveness level, and technical proficiency dimensions. The multi-task learning loss function L MTL is defined as:

[0257]

[0258] where L k is the loss function of the kth task; λ k is the weight coefficient of the corresponding task, satisfying The loss function of each task uses mean square error:

[0259]

[0260] where y k，i is the true label of the ith sample on the kth task; is the corresponding predicted value; N is the number of samples.

[0261] During fine-tuning, a learning rate decay strategy is used:

[0262] η t = η0·(1+ γ·t) -p ;

[0263] where η t is the learning rate at step t; η0 is the initial learning rate, taking the value of 0.001; γ is the decay coefficient, taking the value of 0.0001; p is the decay exponent, taking the value of 0.75.

[0264] ​A rhythm evaluation model is constructed for different music styles by using a transfer learning method. A basic model is trained on mixed data of all styles, and then the model parameters are fine-tuned for the music style to be analyzed. Finally, the consistency level of the model evaluation results with the human expert scores is verified by double-blind expert test, and the consistency measure adopts the Pearson correlation coefficient r and the mean absolute error MAE:

[0265]

[0266] where y i is the expert score; is the model prediction score; and are the mean values of the expert score and the model prediction score, respectively; and N is the number of test samples.

[0267] The steps of directional drift detection and adjustment in the rhythm perception analysis model training process include calculating the direction consistency index DCI of each convolution layer output feature map after forward propagation before each training batch, marking the layer as having a risk of directional drift when the DCI exceeds the preset threshold 0.7, and applying the sparse matrix jitter enhancement method to generate adjusted convolution parameters for the marked layer. The method effect evaluation adopts the performance retention rate R perf :

[0268]

[0269] where Perf before and Perf after are the performance measures (such as accuracy or the reciprocal of the mean absolute error) of the model before and after adjustment, respectively. If R perf <95%, the jitter strength is increased and the adjustment is repeated until the requirement is met or the maximum number of attempts 5 is reached.

[0270] The steps of rhythm evaluation neural network model training include constructing a training data set, taking the multi-dimensional deviation vector D, the deviation similarity S D and the window similarity sequence S w as input features, and taking the expert's rhythm performance score as the training label. A weighted loss function L comb is used, which is a combination of mean square error and cross-entropy loss:

[0271] L comb = αL MSE +(1-α)L CE ;

[0272]

[0273] where α is the weight coefficient, taking the value 0.7; L MSE is the mean square error loss; and L CEis the cross-entropy loss; y i is the true rating of the ith sample; is the corresponding predicted rating; y i，c and are the true and predicted probability that the ith sample belongs to the cth rating category, respectively; N is the number of samples; C is the number of rating categories.

[0274] The optimization method uses a stochastic gradient descent algorithm with a momentum term, and the parameter update rule is:

[0275]

[0276] θ t = θ t-1 + v t t

[0277] where v t is the momentum at the tth step; μ is the momentum coefficient, taking the value of 0.9; η t is the learning rate; is the gradient of the loss function with respect to the parameter θ t-1 ; θ t is the updated parameter.

[0278] The batch size and learning rate are dynamically adjusted during the training process, and weight regularization and random deactivation are used to prevent overfitting. The weight regularization uses L2 regularization:

[0279]

[0280] where L reg is the loss function with regularization; λ is the regularization coefficient, taking the value of 0.0001; θ i is the model parameter.

[0281] The random deactivation probability is set to 0.3 and applied to the fully connected layer. After the model converges, the performance is evaluated on the validation set, the accuracy is analyzed and evaluated through the confusion matrix and mean square error indicators, and the model hyperparameters are fine-tuned based on the evaluation results.

[0282] For a better understanding and implementation of the present application, the following provides an embodiment 2 of a specific application scenario of the present application: In this embodiment, 30 piano learners aged 8 to 12 are selected for the experiment. These learners are divided into a beginner group (1-2 years), an intermediate group (3-4 years), and an advanced group (more than 5 years) according to the learning period, with 10 people in each group. Three pieces of music of different difficulty are selected for the experiment: the "Prelude" for the beginner group, the "Turkish March" for the intermediate group, and the "Moonlight Sonata First Movement" for the advanced group. The researchers first collect the standard version played by professional piano teachers as a reference template, then record the audio of each learner's performance, and analyze it using the evaluation method of the present application. The evaluation process sets the sampling rate to 44.1 kHz, the quantization accuracy to 24 bits, the time window length to 40 milliseconds, and the overlap rate to 50%.

[0283] First, the collected audio signal is preprocessed, including low-pass filtering and amplitude normalization. Then, it is converted to a time-frequency domain representation by short-time Fourier transform, and relevant features are extracted. For window similarity calculation, the time domain cross-correlation coefficient weight α1 is set to 0.3, the frequency domain Euclidean distance similarity weight α2 is set to 0.3, and the Mel-frequency cepstral coefficient distance similarity weight α3 is set to 0.4.

[0284] In the beat marker extraction process, the envelope detection algorithm, autocorrelation function analysis, short-time energy analysis method, and spectral flux calculation are combined, and the confidence threshold is set to 0.7. The experimental results show that this combined method improves the beat detection accuracy by an average of 18.5% compared to a single method. Table 1 shows the accuracy comparison of various beat detection methods:

[0285] Table 1 Accuracy comparison of different beat detection methods

[0286] Detection method Primary group accuracy Intermediate group accuracy Advanced group accuracy Average accuracy Envelope detection algorithm 0.823 0.796 0.785 0.801 Autocorrelation function analysis 0.768 0.812 0.808 0.796 Short-time energy analysis 0.831 0.813 0.802 0.815 Spectral flux calculation 0.775 0.821 0.847 0.814 Combination method 0.942 0.958 0.973 0.958

[0287] The rhythm perception analysis model uses a convolutional long short-term memory network structure, containing 3 parallel convolutional blocks, 2 layers of bidirectional long short-term memory networks, and 8 heads of self-attention mechanisms. The model is pre-trained on a large-scale dataset and then fine-tuned for different musical styles. Table 2 shows the performance of the model on different style pieces:

[0288] Table 2 Evaluation performance of rhythm perception analysis model on different music styles

[0289] Music style Pearson correlation coefficient Mean absolute error Accuracy Classical 0.926 5.8 0.907 Jazz 0.892 7.3 0.884 Pop 0.913 6.2 0.896 Folk 0.885 7.9 0.873

[0290] For each learner's performance, the system generates a multi-dimensional deviation vector, containing 12 dimensions of rhythm features. Table 3 shows the deviation vector values and evaluation results of a certain intermediate group learner playing "Turkish March":

[0291] Table 3: Deviation vectors and evaluation results of the intermediate group learners playing the Turkish March

[0292] Deviation dimension Deviation value Allowable range Evaluation result Absolute time value deviation 35.7 ms ± 30 ms Slightly exceeded Relative tempo proportion deviation 2.8％ ±3％ Normal range In-bar structure deviation 22.4 ms ± 30 ms Normal range Tempo stability deviation 0.089 0.1 Normal range Macro periodicity deviation 0.082 0.1 Normal range Mesoscopic bar group deviation 0.115 0.1 Slightly exceeded Micro note coherence deviation 0.073 0.1 Normal range Strong beat accuracy deviation 18.3 ms ± 25 ms Normal range Weak beat accuracy deviation 31.2 ms ± 35 ms Normal range Tempo acceleration trend deviation 0.132 0.1 Significantly exceeded Tempo deceleration trend deviation 0.058 0.1 Normal range Performance fluctuation deviation 0.074 0.1 Normal range

[0293] Based on the analysis of the multi-dimensional deviation vectors, the system generated various scores. Table 4 shows the average scores of learners in three different groups:

[0294] Table 4: Average scores of learners in three different groups

[0295]

[0296]

[0297] The researchers also conducted a consistency test between the system scores and the professional teacher scores. They randomly selected 15 learners' performance recordings and asked five professional piano teachers to score independently, and then compared the scores with the system scores. The results showed that the average Pearson correlation coefficient between the system scores and the professional teacher scores reached 0.912, and the average absolute error was 5.3 points, as shown in Table 5:

[0298] Table 5: Consistency comparison between system scores and professional teacher scores

[0299] Scoring dimension Pearson correlation coefficient Mean absolute error Rhythm accuracy 0.923 4.8 Rhythm stability 0.935 4.5 Rhythm expressiveness 0.876 6.7 Rhythm technicality 0.912 5.2 Comprehensive score 0.915 5.3

[0300] Traditional methods of evaluating music practice mainly rely on subjective judgment by teachers or simple beat detection technology, which has problems such as inconsistent evaluation standards, complicated evaluation process, and inability to capture subtle rhythm characteristics. Traditional methods usually only focus on beat accuracy, ignoring important dimensions such as rhythm stability, expressiveness, and technicality, and the evaluation results often lack comprehensiveness and objectivity. The present invention uses multi-level time models and deep learning technology to comprehensively analyze rhythm characteristics from macro, meso, and micro levels, generating multi-dimensional deviation vectors, and achieving more objective and accurate evaluation. Experimental results show that the present invention improves the beat detection accuracy by 18.5% compared to traditional methods, and the consistency between the evaluation results and the professional teacher scores reaches 0.912, which is higher than the 0.75 of the traditional method (ordinary teacher and professional teacher score comparison method). In addition, the present invention can dynamically adjust the evaluation standards according to different music styles and difficulty levels, providing more personalized and targeted feedback, effectively promoting music learning for learners.

[0301] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 6, 7, 8, and 9.

[0302] Table 6: Variable explanation table (first part)

[0303]

[0304]

[0305] Table 7 Variable explanation table (second part)

[0306]

[0307]

[0308] Table 8 Variable explanation table (third part)

[0309]

[0310] Table 9 Variable explanation table (fourth part)

[0311]

[0312]

[0313] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A method for evaluating the performance of music children, characterized in that: include: Collect and pre-process the audio signals of the children's performance and the standard performance; construct a time window sequence and calculate the window similarity; Extract beat markers to calculate the rhythm sequence; input the rhythm sequence and window similarity into the rhythm perception analysis model to generate a rhythm feature representation vector; apply multi-level time model analysis to generate a multidimensional deviation vector; construct a rhythm deviation allowable range matrix; calculate deviation similarity; apply the rhythm evaluation neural network model to output a rhythm performance score; The rhythm perception analysis model uses a convolutional long short-term memory network structure to analyze rhythm sequences and generate high-dimensional semantic-level rhythm feature representations. The multi-level time model analyzes the time relationship at three levels: macro rhythm structure, meso beat group organization, and micro note duration.

2. The method according to claim 1, characterized in that The preprocessing step specifically converts the audio signal of the child's performance and the standard performance audio signal into time domain representation and frequency domain representation; the time window sequence refers to a continuous time segment sequence in which the audio signal of the child's performance and the standard performance audio signal are divided according to a fixed duration or an adaptive duration, wherein the fixed duration is 20 milliseconds to 200 milliseconds, and the adaptive duration is dynamically determined according to the music beat structure.

3. The method according to claim 2, characterized in that The window similarity refers to the calculation of the similarity between the audio signal of the child's performance and the standard performance audio signal in time domain characteristics and frequency domain characteristics within each time window, and is quantified by the cross-correlation function, spectral distance or Mel-frequency cepstral coefficient distance method. The window similarity is used to adjust the internal weight distribution of the rhythm perception analysis model.

4. The method according to claim 3, characterized in that The rhythm perception analysis model is trained based on a large-scale performance dataset and fine-tuned through style transfer to adapt to the performance characteristics of different styles; the rhythm feature representation vector includes rhythm pattern, dynamic changes, duration distribution and expressive processing, and the dimension of the rhythm feature representation vector is determined by the rhythm perception attention mechanism and dynamically adjusted according to the length of the rhythm sequence.

5. The method according to claim 4, characterized in that The multidimensional deviation vector refers to the representation of the rhythm time deviation occurring in the performance of the child musician through a multidimensional vector, wherein each dimension of the multidimensional deviation vector corresponds to a different type of rhythm feature, including absolute time value deviation, relative rhythm proportion deviation, intra-beat structure deviation, rhythm stability deviation and expressive fluctuation deviation quantification parameters.

6. The method according to claim 5, characterized in that The rhythm deviation tolerance matrix refers to a multi-dimensional evaluation reference standard established based on different musical styles, repertoire difficulty and performance level. It defines the extent to which various rhythm deviations are technical errors or artistic expressions. The rhythm deviation tolerance matrix includes boundary thresholds and scoring weight coefficients.

7. The method according to claim 6, characterized in that The deviation similarity refers to a quantitative indicator obtained by calculating the cosine distance or Euclidean distance between the multidimensional deviation vector and the corresponding deviation vector of the standard performance template, which is used to measure the degree of similarity in rhythm processing between the child's performance and the standard performance.

8. The method according to claim 7, characterized in that The rhythm evaluation neural network model refers to a deep neural network that integrates the attention mechanism and the residual connection structure, and generates a rhythm performance score for multi-dimensional deviation vectors and deviation similarity; the rhythm performance score includes sub-item scores and a comprehensive score for four dimensions: rhythm accuracy, rhythm stability, rhythm expressiveness, and rhythm technicality.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the method for evaluating the practice status of children according to any one of claims 1 to 8.

10. A system for evaluating the performance of music children's practice, characterized in that: The computer-readable storage medium according to claim 9 is included, the system is any one of a computer, a server, and a single-chip microcomputer, the computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

Images