Application of artificial intelligence CNN model in emotional recognition of instrumental music Authors: Yang Liu Published: 2025 ABSTRACT AbstractTo enhance the accuracy and expressiveness of emotional recognition in instrumental music, this study proposes a multi-layer music emotion recognition model. The model integrates Convolutional Neural Network (CNN), Bidirectional Gated Recurrent Unit (BiGRU), and Attention Mechanism, aiming to accurately capture complex emotional information in audio. First, the CNN is used to extract local emotional features of the audio. Then, the BiGRU is employed to model the contextual information of time series, strengthening the temporal continuity of emotional expression. Finally, the attention mechanism is introduced to dynamically focus on key emotional segments. To achieve multi-scale feature fusion, the model combines low-level audio features and high-level semantic features through weighted summation during the feature extraction stage. The experimental section is validated using three music emotion datasets, including two publicly available datasets Instrument Recognition in Musical Audio Signals (IRMAS) and Multitrack Dataset for Musical Audio (MedleyDB), as well as a large-scale dataset Database for Emotional Analysis of Music (DEAM), to comprehensively evaluate the performance, generalization ability, and robustness of the model. These datasets cover a large number of multi-category instrumental audio samples. The model is evaluated on three continuous emotional dimensions: Valence, Arousal, and Dominance. The experimental results show that the proposed model achieves Pearson correlation coefficients of 0.871, 0.832, and 0.784, respectively, which are better than those of the comparative models. In terms of Mean Squared Error (MSE), the values are 0.0187, 0.0208, and 0.0243, respectively, indicating higher prediction accuracy. In conclusion, the proposed fusion deep neural network model significantly improves the accuracy and generalization ability of emotional recognition in instrumental music. This study provides an effective method and practical inspiration for emotional modeling in complex musical environments.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-33623-1.Keywords: Music emotion recognition, Convolutional neural network, Bidirectional gated recurrent unit, Attention mechanism, Continuous emotional space, Feature fusionSubject terms: Engineering, Mathematics and computing, Neuroscience, Psychology, Psychology Introduction With the rapid development of digital music and multimedia technologies, music has not only served as a form of artistic expression but also become an important carrier for emotional communication and spiritual experience among people1–3. As a core topic in the field of music information retrieval, Music Emotion Recognition aims to automatically analyze and identify the emotional information embedded in musical works, thereby promoting the development of applications such as music recommendation systems, affective computing, and human-computer interaction4,5. Especially in instrumental music, the lack of semantic auxiliary information such as lyrics means that emotional expression more heavily relies on complex acoustic features like melody, rhythm, and timbre, which poses significant challenges for emotion recognition6–8. Traditional emotion recognition methods, which are mostly based on manually designed shallow audio features and simple machine learning algorithms, struggle to effectively capture the diversity and subtlety of musical emotions, leaving considerable room for improvement in recognition accuracy9–11. In recent years, with the rapid development of deep learning technologies, models based on Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) have demonstrated excellent performance in audio signal processing and emotion recognition12–14. CNN can effectively extract local time-frequency features of audio to capture rich emotional information. RNN, especially Bidirectional Gated Recurrent Unit (BiGRU), possesses the ability to model contextual relationships in time series, which helps capture the dynamic evolution of musical emotions15–17. Additionally, the introduction of attention mechanisms endows models with the capability to automatically focus on key audio segments, significantly improving the accuracy of emotion recognition18–20. However, most existing studies are still limited to single-level feature modeling, lacking effective fusion of low-level audio features and high-level semantic information. This limitation makes it difficult to comprehensively reflect the complex emotional expressions of instrumental music21–23. To address the above challenges, this study proposes a multi-layer fused music emotion recognition model that combines CNN, BiGRU, and Attention Mechanism to achieve high-precision emotion recognition for instrumental music. The model first extracts local time-frequency features through CNN to capture the basic emotional elements of music. It then uses BiGRU to capture emotional dynamics in sequential contexts, enhancing the model’s understanding of temporal dependencies. Finally, the attention mechanism guides the model to focus on key segments that contribute most to emotional expression, improving emotion discrimination ability. To further enhance the multi-scale characteristics of emotional expression, the model adopts a weighted fusion strategy to integrate low-level audio features and high-level semantic features, achieving a richer emotional representation. The experiments use the public Instrument Recognition in Musical Audio Signals (IRMAS) dataset and Multitrack Dataset for Musical Audio (MedleyDB) dataset, covering a rich and diverse range of instrumental music samples. To represent musical emotions in a more flexible and continuous manner, this study adopts Russell’s Circumplex Model of Affect. The model organizes emotions within a multi-dimensional continuous space composed of Valence, Arousal, and Dominance. This model is particularly suitable for instrumental music, as emotions in instrumental music typically flow and change over time rather than existing as discrete categories. By using this continuous emotional space, the model can capture subtle and complex emotional variations in instrumental works. The results show that the proposed model outperforms current mainstream comparative methods in terms of Pearson correlation coefficients and mean squared error metrics, verifying the model’s effectiveness and superiority. The main innovations and contributions of this study are as follows: (1) Construct and optimize a hybrid architecture that integrates CNN, BiGRU, and attention mechanisms. Through a multi-scale feature fusion mechanism, adaptive weighting of temporal and spectral emotional features is achieved, improving the model’s feature expression ability. (2) Introduce an emotional feature reconstruction module to redistribute and enhance emotional features in the latent space, thereby improving the model’s accuracy in recognizing the separability of complex music emotions. (3) Conduct systematic experiments and comparative analysis on multiple benchmark datasets. The results show that the proposed model outperforms existing methods in terms of robustness and generalization, and the improvement is statistically significant. The study is structured as follows: Section “Literature review” presents a review of research progress in the field of music emotion recognition, focusing on the applications of deep learning technologies in audio emotion analysis and their limitations. Section “Research model” elaborates on the design principles and technical implementation of the proposed multi-layer emotion recognition model, including the specific construction and functions of each module. Section “Experimental design and performance evaluation” introduces the experimental design, dataset characteristics, evaluation metrics, and analysis of experimental results to verify the model’s performance and advantages. Section “Conclusion” concludes the full text, discusses the research limitations, and prospects the potential future research directions and application prospects. Literature review In recent years, the rapid development of AI and computer science has gradually transformed music emotion recognition into an important research area in music information retrieval24. The sounds produced by instrumental performances, with their unique immersive expression and beautiful timbre, have become an important outlet for performers to convey emotions25. Accurate emotion recognition in music has thus become a key focus of academic research. For example, Zhang et al.26 proposed a method called Modular Composite Attention Network for continuous music emotion recognition. Yang et al.27 introduced a music emotion classification method based on lyrics and comments. This method combined lyrics and comment texts to mine richer emotional information. There are also some application cases of AI in emotion recognition of sounds from instrumental performances. For example, Rajesh & Nalini28 collected a music dataset comprising four types of instruments: string instruments, percussion instruments, woodwind instruments, and brass instruments. They used Recurrent Neural Network (RNN) to identify emotions based on instrument categories and found that Mel Frequency Cepstral Coefficient (MFCC) features with deep RNN provided better performance for instrument-based emotion recognition. Bai29 developed a music education system that included four dimensions: web-based, mobile, web interface, and database. This system effectively utilized computer technology to build a vocal recognition and education framework. Bontempi et al.30 attempted to understand the phenomenon of human interpretation of music and the automatic generation of expressive performances. Greer et al.31 proposed a model for generating audio-music features that supported music understanding, leveraging self-supervised and cross-domain learning. Additionally, Siyal et al.32 developed and empirically tested a sound recognition model to examine the impact of different instruments on emotional expression. For example, Neware and Poulose33 deeply investigated the role of Recursive Feature Elimination (RFE) technology in enhancing the performance of music genre classification models. Based on the GTZAN dataset covering various audio features, they implemented multiple machine learning models including logistic regression, decision trees, random forests, gradient boosting, and Support Vector Machine (SVM). By applying RFE technology to effectively screen important features, they not only improved the classification accuracy of the models but also enhanced their interpretability. This approach provided insights for the optimal selection of emotional features in musical signals, contributing to the efficiency and accuracy of emotion recognition models. Poulose et al.34 designed and implemented a general music recommendation system that integrated various audio signal features. Using machine learning and deep learning models, the system performed multi-genre classification of music, covering 10 categories: blues, classical, country, disco, hip-hop, jazz, metal, pop, reggae, and rock. The system included a feature extractor, data preprocessing module, and classifier, meeting users’ diverse music recommendation needs. Their study provided references for multimodal feature fusion and classifier design in this research. In particular, it further validated the importance of considering multi-level features when handling complex musical emotion recognition tasks. Naveed et al.35 proposed an innovative Few-Shot Learning (FSL) framework for efficient classification of plant diseases. The framework used transfer learning for pre-training and combined Prototype Networks to implement class-prototype-based distance metric classification, suitable for scenarios with extremely few samples. This research field differed from musical emotion recognition. However, its FSL and transfer learning strategies were valuable. These strategies addressed the scarcity of annotated data in the music domain. They enhanced model generalization and enabled rapid adaptation to new emotional categories. Al-Maari et al.36 proposed a novel credit card fraud detection method based on ensemble machine learning, integrating Random Forest, Logistic Regression, and AdaBoost algorithms. By adopting a soft voting mechanism, the method effectively combined the advantages of each model to improve classification accuracy and reduce false positive rates. This ensemble learning approach involved multi-model fusion and classification optimization. These ideas provided insights for the fusion design in this study’s multi-layer emotion recognition model. Specifically, they guided the integration of different neural network structures and mechanisms. As a result, overall model performance and stability were enhanced. Hussain et al.37 proposed a machine learning framework based on Enhanced Random Forest (ERF) to analyze the relationship between the functions of mobile applications and user ratings. They used Natural Language Processing (NLP) technology to preprocess text data and combined machine learning models to achieve multi-dimensional performance evaluation and prediction. Although the research field was different, its methods of multimodal data processing and feature fusion provided valuable technical references for the integration of multi-source information in complex audio emotion recognition. In summary, existing literature fully demonstrates the extensive applications of artificial intelligence and machine learning technologies in the field of music emotion recognition and related domains. Through advanced models such as deep neural network, RNN, and attention mechanisms, researcher have made remarkable progress in extracting and understanding musical emotional features. Meanwhile, technologies like RFE, multimodal feature fusion, and ensemble learning have provided more accurate and stable solutions for emotion recognition. Additionally, the development of techniques such as few-shot learning and transfer learning has offered effective approaches to address the issues of data scarcity and insufficient annotation in the music domain. These achievements not only enrich the theoretical foundation of music emotion recognition but also provide technical support for the design and optimization of practical systems. However, current research still has several limitations. First, most methods analyze features at a single level, lacking deep fusion of low-level audio signals and high-level semantic emotional information. This makes it difficult to fully capture the complex expressions of emotions in instrumental music. Second, in the face of dynamic changes in different emotional dimensions, the time-series modeling capability of existing models still needs to be strengthened. Finally, issues such as data scarcity and label noise continue to restrict the generalization ability and application scope of models. Based on the above analysis, this study aims to propose a multi-layer music emotion recognition model that fuses CNN, BiGRU, and attention mechanisms. This model breaks through the limitations of traditional single-feature extraction to effectively capture multi-scale and multi-dimensional emotional information in instrumental music. By integrating advanced feature selection and fusion strategies, this study not only improves the accuracy and stability of recognition but also provides a new technical approach to address insufficient annotation and dynamic emotional changes. The design and experimental verification of this model strive to bridge the gaps in existing research and promote the advancement of music emotion recognition technology to a higher level. Research model Structure and feature modeling method of instrumental emotion recognition model based on CNN-BiGRU-attention Given the complexity and continuity of instrumental music sounds, to model the emotional features of instrumental sounds in both time and frequency domains, this study has improved the original CNN by integrating BiGRU and attention mechanisms to establish a comprehensive sound emotion recognition model. The reason for not using a single neural network for modeling is that the audio signals of instrumental music have high temporal and complex frequency-domain structures. Moreover, emotional expression involves not only low-level acoustic features, such as pitch, timbre, and rhythm, but also high-level semantic patterns, such as the direction of the melody and the structure of musical sections. Therefore, a single neural network cannot fully capture all this information. CNN is inherently good at automatically extracting local spatial features from two-dimensional audio spectrograms and are particularly suitable for capturing short-term structures of audio signals along the time and frequency axes. BiGRU, as a lightweight variant of recurrent neural networks, retains the advantage of RNN in modeling sequential data. It can encode audio sequences in both forward and backward directions simultaneously, thereby capturing more comprehensive contextual relationships in the emotional changes of instrumental works. The attention mechanism can automatically focus on emotional inflection points, climaxes, or highly expressive sections in the performance of instrumental music. Therefore, the overall structure of our proposed model is designed as shown in Fig. 1: A comprehensive model for emotional recognition of instrumental sound. In the model shown in Fig. 1, raw instrumental music segments from the dataset first undergo noise suppression using spectral gating and zero-mean unit-variance normalization of the audio signals. The audio is then divided into fixed-length segments with overlapping to ensure temporal continuity. Next, Fourier transform and a Mel filter bank are used to process the audio segments, generating two-dimensional Mel spectrograms as the model input. These spectrograms are fed into a multi-layer CNN for local feature extraction. The convolutional layers in the CNN serve as feature extractors, accepting a musical emotional feature matrix of size IM×N. A 2D convolutional layer with K filters of size 3 × 33 is applied to capture fine-grained emotion-related textural features in both time and frequency dimensions. Subsequently, 2D batch normalization is performed on the convolutional output for standardization. The purpose of this step is to accelerate training convergence, suppress internal covariate shift, and thereby enhance the stability of the model. The equation is: and respectively represent the mean and standard deviation of the current small batch samples. and respectively represent the learnable scaling and translation parameters. ϵ is a tiny constant to prevent the divisor from being zero. This process activation function adopts Rectified Linear Unit (ReLU), and the expression is: Finally, the maximum pooling operation is used to down sample the feature map, to compress the feature dimension to retain the key elements in the emotional information conveyed by instrumental sound. The CNN model structure is shown in Fig. 2: Model structure of CNN. The CNN adopted in this study consists of three convolutional layers, each followed by batch normalization, ReLU activation function, and max pooling layer. The convolution kernel size is 3 × 3 for all layers, with the number of kernels set as 32 → 64 → 128. The “same” padding mode is used to maintain the feature map size unchanged. The features extracted by the CNN are then input into the BiGRU-Attention module for emotional modeling. Additionally, Section “Parameters setting” of this study supplements relevant hyperparameter settings, including learning rate (0.0001), batch size (32), number of training epochs (100), and optimizer (Adam), to enhance the reproducibility of the research. Subsequently, the output of the CNN is fed into the BiGRU structure. The BiGRU is utilized to analyze the semantic relationships between preceding and subsequent segments in the instrumental audio, including the emotional information conveyed in the musical features from the past and future contexts. Building on the RNN, the GRU model introduces gating mechanisms to address the issues of vanishing or exploding gradients, thereby enhancing the model’s performance. The BiGRU model further extends the GRU model by incorporating a bidirectional structure, which enables it to better capture bidirectional dependencies in sequential data. It consists of two independent GRU units: one processes the data in the forward direction of the time sequence, while the other processes it in the reverse direction. Through this bidirectional structure, the BiGRU model can capture both forward and backward information in the sequential audio data, thus better understanding and predicting patterns within the audio sequence. The structure of BiGRU model is shown in Fig. 3: The structure of BiGRU. To enhance the model’s ability to identify key emotional segments, this study introduces an attention mechanism after the BiGRU. This mechanism enables the model to focus on critical information within complex musical sequences, particularly elements such as melodic contours, rhythmic patterns, and harmonic structures that play decisive roles in emotional expression. By incorporating the attention mechanism, the model can more accurately capture and interpret emotional nuances in audio. In practice, the attention mechanism calculates the contribution of each time step’s hidden state to the overall emotional representation by learning its importance weights. First, a feedforward neural network is used to score the hidden state at each time step: represents the hidden state vector output by BiGRU at time step t. represents a weight matrix (a learnable parameter), which is used to linearly transform the hidden state. represents an offset vector (a learnable parameter), which is used for the offset of linear transformation. tanh(⋅): Hyperbolic tangent activation function, which is used to increase nonlinearity. V represents the projection vector (learnable parameter), which is used to map the result of nonlinear transformation into a scalar score. represents the “importance” score of hidden state . Then, the score is normalized by Softmax function, and the attention weight corresponding to each time step is obtained: represents the normalized attention weight and the contribution ratio of the t-th time step to the whole sequence representation. T represents the number of time steps of the input sequence. The Softmax operation ensures that all add up to 1. Finally, the hidden states of all time steps are weighted and summed according to the attention weight to form a context vector c as a new emotional feature representation: where c stands for context vector, that is, the final emotional representation that integrates all the hidden states of time steps and is weighted and averaged. comes from the attention weight of Eq. (12). represents the hidden state of each time step. With this attention mechanism, the model can significantly enhance its ability to identify emotional inflection points, climactic segments, and expressive fragments in instrumental music, while also reducing the interference of redundant information on recognition accuracy to a certain extent. The obtained context vector is then fed into a fully connected layer for further emotional mapping and prediction. The final output layer of the proposed model, based on Russell’s Circumplex Model of Affect, maps emotions to a continuous three-dimensional emotional space, predicting Valence, Arousal, and Dominance in this space. This model has been widely used in music emotion recognition research due to its ability to represent subtle emotional states and their temporal dynamics, making it particularly suitable for accurately modeling the emotional content conveyed by instrumental music. Multi-scale emotion expression strategy integrating low-level audio features and high-level semantic features The expression of emotions in instrumental music has distinct multi-level characteristics. It includes not only low-level perceptual features that can be directly extracted from the audio signal, such as rhythm, intensity, and pitch, but also high-level semantic features constructed from contextual relationships, such as the direction of the melody, harmonic structure, and emotional evolution of musical sections. Therefore, to fully capture the nuanced changes of musical emotions in the continuous emotional space, this study proposes a multi-scale fusion strategy. By integrating local perceptual features with global semantic information, this strategy enhances the model’s ability to recognize and discriminate musical emotional features. In the feature representation stage, Mel-spectrogram and Mel-Frequency Cepstral Coefficient (MFCC) is selected as the main low-level audio features. The reason for choosing these two features is mainly based on their physical meaning and effectiveness in music emotion recognition. Mel spectrograms can simulate the nonlinear perception of different frequencies by the human ear, mapping the frequency spectrum information of audio to the Mel scale, thus closer to the characteristics of human auditory perception. This enables the model to capture the emotional information carried by timbre, melody, and harmony changes. MFCC effectively extracts the main frequency envelope information of the audio by processing it with short-time Fourier transform and Mel filter bank, reflecting the timbre characteristics and subtle emotional changes of the instrument. Excluding other time-frequency features, this study mainly considers the following factors. Firstly, Mel spectrograms and MFCC have been widely validated in the fields of music information retrieval and emotion recognition, with strong representativeness and robustness. Secondly, this study aims to construct a deep model that can balance local audio features and global emotional expression. Excessive feature dimensions may lead to increased computational complexity and overfitting risks. Thirdly, through a multi-scale feature fusion strategy, the model has already integrated information of different frequency resolutions in the low-level feature stage, thus compensating for the potential shortcomings of a single feature. Therefore, in practice, this study combines Mel spectrograms and MFCC to ensure sufficient capture of key emotional features while also considering the trainability and generalization ability of the model. This study adopts an early fusion strategy to integrate the low-level perceptual feature with the high-level semantic feature . To eliminate interference caused by different feature scales, first normalize the two types of features separately: µ and σ are the mean and standard deviation of the corresponding features, respectively. The low-level feature represents the perceptual feature vector extracted from the audio signal, such as Mel spectrogram or MFCC. It mainly reflects timbre, melody, and local emotional details. The high-level feature represents the semantic feature vector extracted by the CNN-BiGRU-Attention model, mainly carrying global emotional information and temporal contextual relationships. ,,, are the mean and standard deviation of the corresponding features on the training set, respectively, used for normalization to eliminate dimensional differences and scale interference. After normalization, a unified feature vector is formed through weighted summation: The weight coefficients α,β∈[0,1] are the fusion weights of low-level features and high-level features, respectively, satisfying α + β = 1. During the training process, weights are automatically learned through gradient optimization to dynamically allocate the importance of different features to the final representation. This fusion strategy preserves the local emotional information of low-level features, and integrates the global contextual information of high-level semantic features. It thus enhances the diversity and structural consistency of feature representations and provides more comprehensive intelligence support for subsequent emotion recognition. During the fusion process, the CNN convolutional module is used to model local features of the input emotion feature maps, extracting short-term perceptual features of the audio in the two-dimensional time-frequency domain. These low-level features can effectively represent the emotional change trends of a music segment within a certain time window, such as pitch fluctuations, energy concentration, and rhythmic density. After the convolution operation of the CNN, ReLU activation and max-pooling are applied to further compress the feature dimensions and highlight key musical feature information, thus obtaining a sequence of local features. The local feature sequences are then fed into the BiGRU network for contextual modeling. BiGRU can capture the emotional evolution trajectories of music from both past and future contexts, demonstrating strong temporal modeling capabilities in music segments with nonlinear and multidimensional emotional changes. This fusion strategy enables the model to maintain stability in emotion discrimination while being responsive to subtle emotional differences, making it suitable for fine-grained emotion annotation tasks in complex music with multiple emotional dimensions. Integrating features from different levels also enhances the model’s generalization ability when dealing with diverse musical styles. Different styles of music often exhibit unique emotional tendencies in terms of rhythm, harmony, or melody, which are difficult to capture comprehensively with single-dimensional features. By integrating multi-source features, the model can construct a higher-dimensional cognitive space for music emotions, effectively enhancing the stability and robustness of recognition. Code availability The custom code used in this study, including programs related to emotion feature extraction, model training, and evaluation, has been submitted to the publisher as supplementary materials. Readers can obtain the complete code for reproducing the results or further research by downloading the supplementary files. The research code has no access restrictions. Experimental design and performance evaluation Datasets collection To comprehensively evaluate the performance of the proposed deep learning-based multi mechanism fusion model in instrumental music emotion recognition tasks, this study further introduces a large-scale Database for Emotional Analysis of Music (DEAM) dataset for experimental verification based on the Instrument Recognition in Musical Audio Signals (IRMAS) dataset and MedleyDB dataset. The IRMAS dataset provides stereo audio data, with a focus on four types of instruments: acoustic guitar, electric guitar, trumpet, and violin. The test set annotates specific high pitch instruments, providing an important experimental foundation for music information retrieval and emotion recognition. The MedleyDB dataset was released in 2014 and contains 122 multi-track music pieces, covering a variety of music styles and instruments. Each piece has been processed by professional audio engineers to ensure the purity of single instrument tracks, while providing complete mixed tracks to make it suitable for emotion recognition tasks. The DEAM dataset contains 1802 music clips, covering various styles such as pop, rock, classical, etc., and provides continuously annotated emotional dimensions Valence and Arousal, which can be used to verify the model’s generalization ability on large-scale and diverse music data. In order to ensure fairness and reproducibility in training, validation, and testing, all three datasets use 70% of the data for training, 15% for validation, and 15% for testing. The training, validation, and testing samples do not overlap with each other. During the training phase, data augmentation techniques such as time stretching, pitch shifting, and noise superposition are used to expand the diversity of training samples, thereby improving the robustness of the model to different audio transformations. To enhance the reliability of the experimental results, three different random seeds are used for repeated training in each experiment, and the final results are averaged. Hyperparameter tuning is performed on the validation set, including learning rate, batch size, and number of hidden layer units. The test set is only used for final evaluation to ensure the fairness and reproducibility of experimental results. In addition, this study designs cross dataset validation experiments, such as testing models trained on IRMAS on MedleyDB or DEAM, to further investigate the model’s generalization ability under different data sources and music styles. For evaluating the reliability of sentiment labels in the IRMAS dataset, this study introduces a manual annotation process. Three graduate students with a background in music psychology independently annotated 150 randomly selected instrumental pieces, including Valence, Arousal, and Dominance, using a 9-point continuous scoring system and referencing the Russell emotional circle model. By analyzing the consistency of annotations through Cohen’s k coefficient, the credibility of the original labels can be verified, providing a solid data foundation for model training and performance evaluation. Through the above measures, this study significantly improves the applicability, stability, and generalization ability of the model in complex instrumental music emotion recognition tasks while ensuring the scientific and reproducible nature of the experiments. Experimental environment The experimental environment is based on Ubuntu 20.04 operating system, and TensorFlow and PyTorch are used as deep learning frameworks, as shown in Table 1: Table 1Experimental environment setting.PackageConfigurationOperating systemUbuntu 20.04 LTSDevelopment frameworkTensorFlow 2.0, PyTorch 1.9ProcessorIntel Core i9-9900 KDisplay cardNVIDIA GeForce RTX 3080memory32GB DDR4 RAMHard disc1 TB SSD (Samsung 970 EVO)Python versionPython 3.8CUDA versionCUDA 11.1 (for PyTorch and TensorFlow acceleration)Dependency libraryNumpy, Scipy, Librosa, Matplotlib, Scikit-learn, Pandas Experimental environment setting. Parameters setting To optimize model performance, we systematically tuned key hyperparameters using Grid Search combined with five-fold cross-validation. The specific parameters and their ranges included: learning rate (0.0001, 0.001, 0.01), batch size (16, 32, 64), number of CNN filters (32, 64, 128), number of BiGRU hidden units (64, 128, 256), and dimensions of the attention layer along with dropout ratios. For each parameter combination, the model is trained on the training set. The Pearson correlation coefficients and mean squared errors are calculated on the validation set as performance metrics. By comparing the validation metrics of different configurations, this study selects the parameter combination with the best validation performance and stable training process as the final setup. Specifically, a learning rate of 0.001 balances convergence speed and accuracy; a batch size of 32 optimizes both training efficiency and model generalization. 64 CNN filters effectively extract local audio features. 128 BiGRU hidden units ensure model expressiveness while avoiding overfitting; and the attention mechanism parameters maximized focus on key emotional segments. The final parameter settings are detailed in Table 2. This configuration significantly enhanced model stability and prediction accuracy, providing robust support for subsequent experiments. Table 2Parameter setting.HyperparameterValueLearning rate0.001Batch size32Number of training rounds50optimizerAdamloss functionSparse Categorical CrossentropyConvolution layer filter size3 × 3Number of convolution layer filters64Pool size2 × 2Unit number of BiGRU layer128Attention mechanism layer2 Parameter setting. Performance evaluation In the performance evaluation, different indicators values for the three emotional dimensions of Valence, Arousal, and Dominance were calculated. Valence represents the positivity or negativity of an emotion. Positive emotions are pleasure, happiness, love, and satisfaction, while negative emotions are sadness, fear, anger, and disgust. Arousal represents the intensity or level of activity of an emotion. High arousal emotions indicate excitement, interest, and agitation, while low arousal emotions indicate calmness, tranquility, and dullness. Dominance represents the degree of control of an emotion. High dominance emotions typically indicate confidence, proactivity, and control, while low dominance emotions typically indicate helplessness, passivity, and being controlled. The evaluation indicators include Mean Squared Error (MSE), Mean Absolute Error (MAE), Pearson correlation coefficient (r), coefficient of determination (R²), accuracy, and F1 score. The calculations are as follows: In Eqs. (10)-(17), n represents the number of samples, represents the real value, represents the predicted value, and represents the average value of the real value. TP, TN, FP, and FN respectively represent true cases, true negative cases, false positive cases, and false negative cases in classification tasks. The experiments are conducted in an environment equipped with an NVIDIA RTX 3080 GPU. Multiple inference tests are performed on a single 4-second audio segment, and the average inference time is statistically approximately 28 milliseconds, indicating that the model possesses real-time processing capabilities. Furthermore, professional computing tools are used to calculate the number of floating-point operations (FLOPS) of the model structure. The results show that the computational load of this model is approximately 15 GFLOPS, reflecting low computational complexity. This result proves that the proposed model not only outperforms traditional methods in accuracy but also demonstrates good adaptability in terms of computational resource consumption, making it suitable for practical online music emotion recognition systems. To quantify the consistency of emotional judgments among the three annotators, Cohen’s κ coefficient is used for analysis. The experimental results (Table 3) show that the average Cohen’s κ values for the three dimensions of Valence, Arousal, and Dominance are 0.74, 0.76, and 0.72, respectively. All values fall within the “substantial agreement” range (0.61–0.80), indicating that the emotional labels in the dataset exhibit high subjective consistency and reliability. This provides a solid foundation for model training and evaluation. Table 3Parameter setting.Emotion dimensionCohen’s κ ValueValence0.74Arousal0.76Dominance0.72 Parameter setting. To verify the effects of spectral gating noise suppression and zero mean unit variance normalization on emotion recognition performance, this study compares the model performance before and after processing. The experiment is conducted on the IRMAS and MedleyDB datasets, with evaluation metrics including MSE, MAE, and Pearson correlation coefficient (r). The results are shown in Table 4. Table 4 suggests that after noise suppression and normalization, the MSE of the model significantly decreases in all three emotional dimensions, with Valence, Arousal, and Dominance decreasing by approximately 15.4%, 19.1%, and 15.1%, respectively; The Pearson correlation coefficients increases by approximately 3.6%, 2.9%, and 2.9%, respectively, indicating that the preprocessing step significantly enhances the signal-to-noise ratio and emotional expression ability of audio features. By combining audio segmentation and overlapping strategies, the model can better capture temporal continuous emotional changes, thereby improving the stability and robustness of emotion prediction. Overall, spectral gating noise suppression and normalization provide more robust input features for the CNN-BiGRU-Attention model, laying a solid foundation for subsequent feature extraction, multi-scale fusion, and final emotion recognition performance. Table 4Comparison of model performance before and after noise suppression and Normalization.Emotion dimensionMSE (Raw)MSE (Processed)r (Raw)r (Processed)Valence0.02210.01870.8410.871Arousal0.02570.02080.8090.832Dominance0.02860.02430.7620.784 Comparison of model performance before and after noise suppression and Normalization. To visually demonstrate the training dynamics and convergence of the proposed CNN-BiGRU-Attention model, this study plots learning curves of accuracy and loss during training and validation phases against the number of epochs, as shown in Fig. 4. The total number of training epochs is 50. Learning curve analysis. The learning curves in Fig. 4 shows that the model’s accuracy rises rapidly in the early training stage (first 5 epochs), with both training and validation losses decreasing rapidly. This indicates that the model effectively learns emotional features from input data. As the number of training epochs increases, both training and validation accuracies gradually rise and stabilize, suggesting that the model progressively converges to a stable state. Throughout the training process, the training accuracy consistently exceeds the validation accuracy with a small gap, showing no obvious signs of overfitting. This demonstrates the model’s strong generalization ability to adapt well to unseen data. Both training and validation losses exhibit a continuous downward trend, indicating that the model continuously optimizes parameters to better fit the training data while improving performance on the validation set simultaneously. Additionally, the smooth variation of the learning curves reflects the stability of the training process, without significant fluctuations or divergence, indicating good convergence during training. This provides a strong guarantee for the reliability of the model’s subsequent performance and enhances confidence in its practical application effects. In summary, the learning curve analysis verifies that the proposed CNN-BiGRU-Attention model exhibits excellent convergence characteristics and good generalization ability during the training phase. It lays a solid foundation for the model’s outstanding performance in the emotional dimensions of Valence, Arousal, and Dominance. To deeply validate the contribution of each key module in the proposed CNN-BiGRU-Attention model to emotion recognition performance, ablation experiments were designed. By removing the CNN, BiGRU, and attention mechanism modules respectively, three simplified models are constructed and compared with the complete model. Model A (without CNN): The CNN part is removed, retaining only BiGRU and the attention mechanism to verify the role of CNN in local audio feature extraction. Model B (without BiGRU): The BiGRU is removed, retaining only CNN and the attention mechanism to examine the contribution of BiGRU in capturing temporal context relationships. Model C (without Attention): The attention mechanism is removed, retaining only CNN and BiGRU to test the auxiliary effect of the attention mechanism in focusing on key emotional segments. Pearson correlation coefficients (r) and mean squared errors (MSE) for three continuous emotional dimensions—Valence, Arousal, and Dominance—are used as performance metrics. The experimental results are shown in Fig. 5: Ablation experimental results. As Fig. 5 shows, the complete model outperforms all simplified models with removed modules in terms of correlation coefficients across all emotional dimensions, while significantly reducing MSE, reflecting the overall structural advantage. The CNN module is responsible for extracting local spatial features from two-dimensional Mel spectrograms. Removing CNN will weaken the model’s ability to perceive local audio information, resulting in a significant decrease in the prediction accuracy of Valence and Arousal, indicating the crucial role of CNN in capturing low-level audio features. Meanwhile, the absence of CNN makes the model insensitive to subtle changes in timbre and playing techniques, which are particularly important for expressing subtle emotions. The BiGRU module captures temporal dependencies and emotional trends through bidirectional encoding. Removing BiGRU will weaken the model’s ability to understand temporal context, making it difficult to track dynamic emotional changes, which is particularly evident in the Dominance dimension. In addition, the absence of BiGRU reduces the model’s ability to integrate before and after audio frames, making it easy to miss transitional changes in emotional intensity. The attention mechanism dynamically focuses on key emotional segments, filters out irrelevant information, and improves the recognition ability of melody contours, rhythm changes, and climax parts. Even though the model can maintain basic performance after removing attention mechanisms, the sensitivity and accuracy of overall emotion recognition decrease, especially in the recognition of rapid emotional fluctuations or music climax segments. The ablation experiment verifies the synergistic effect of CNN, BiGRU, and attention mechanism. The combination of the three enables the model to robustly capture complex musical emotional expressions, significantly improving the accuracy and stability of instrument music emotion recognition. To validate the effectiveness of selected audio features in emotion recognition tasks, this study compares five common low-level audio features under the same CNN-BiGRU-Attention network structure: Mel-spectrogram, MFCC, Chroma features, Spectral Contrast, and Tonnetz tonality features. Table 5 summarizes the average MSE performance of each feature across three emotional dimensions. Table 5MSE results of different low-level features under the same model.Feature typeValenceArousalDominanceChroma0.02980.03450.0371Spectral Contrast0.03120.03640.0392Tonnetz0.03060.03590.0385MFCC0.02130.02570.0294Mel-spectrogram0.01870.02080.0243 MSE results of different low-level features under the same model. From the experimental results, both Mel-spectrogram and MFCC exhibit significantly lower MSE values across all three dimensions compared to other features, demonstrating stronger emotion recognition capabilities. Specifically, the MSE values of Mel-spectrogram for Valence, Arousal, and Dominance are 0.0187, 0.0208, and 0.0243 respectively, outperforming MFCC’s 0.0213, 0.0257, and 0.0294. In contrast, Chroma, Spectral Contrast, and Tonnetz show higher errors across all emotional dimensions, indicating their relatively limited ability to capture musical emotional expressions. The primary reasons for this discrepancy are as follows: Chroma and Tonnetz emphasize tonality and pitch class distribution, which limits their effectiveness in capturing subtle dynamic emotional changes. Although Spectral Contrast reflects spectral intensity contrast, it has weaker modeling capabilities for temporal structures. By comparison, Mel-spectrogram and MFCC excel in preserving audio frequency structures and temporal dynamics, making them particularly suitable for integration with time-frequency joint modeling architectures such as CNN. Subsequently, the proposed combined model is compared with other baseline models, including traditional Support Vector Regression (SVR), single-structured CNN, Long Short-Term Memory network (LSTM), RNN, BiGRU, Transformer-based models, and the CNN-BiGRU model without the attention mechanism. Additionally, this study specifically compares it with the state-of-the-art music emotion recognition methods using advanced neural network architectures reported in recent literatures38,39. The experimental results of each model on the same dataset are shown in Figs. 6, 7, 8 and 9. Comparison of MSE performance of different models in three dimensions: Valence, Arousal and Dominance. Comparison of MAE performance of different models in three dimensions: Valence, Arousal and Dominance. Comparison of Pearson correlation coefficient (r) of different models. Comparison of determination coefficient values of different models. In Fig. 6, the traditional method SVR exhibited the poorest performance with MSE of 0.0321, 0.0365, and 0.0410 for Valence, Arousal, and Dominance, respectively, indicating its difficulty in handling complex instrumental emotion recognition tasks. Deep learning models generally outperformed SVR. The single-layer CNN achieves MSE values of 0.0269, 0.0312, and 0.0341, surpassing both LSTM and RNN. The LSTM yields MSE of 0.0276, 0.0309, and 0.0362, while the RNN shows 0.0284, 0.0321, and 0.0354—both lower than SVR but higher than CNN. Leveraging bidirectional temporal information, the BiGRU further reduces MSE to 0.0254, 0.0290, and 0.0325. The combined CNN + BiGRU model enhances performance further, with MSE dropping to 0.0232, 0.0272, and 0.0300. The Transformer-based model demonstrates significant advantages in capturing long-range dependencies, with MSE values of 0.0228, 0.0265, and 0.0293—outperforming standalone CNN and BiGRU. The advanced models in references38,39 achieve MSE close to 0.022 across dimensions: Valence (0.0219 and 0.0223), Arousal (0.0257 and 0.0261), and Dominance (0.0281 and 0.0287), representing the current state of the art. In contrast, the proposed CNN-BiGRU-Attention model achieves MSE of 0.0187, 0.0208, and 0.0243 for Valence, Arousal, and Dominance—significantly lower than all other models. For example, compared to reference38, the Valence MSE decreases by ~ 14.6%, and Arousal by ~ 19.1%. This result indicates that the CNN-BiGRU architecture integrated with the attention mechanism more effectively captures emotional features in instrumental audio, enhancing the model’s sensitivity to critical emotional segments while reducing interference from redundant information. This significantly improves the accuracy and robustness of emotion recognition. Evidently, the model performs excellently in complex instrumental emotion recognition tasks, validating the scientific rationale and advancement of its design. In Fig. 7, the traditional SVR achieves mean absolute errors (MAE) of 0.141, 0.152, and 0.168 for Valence, Arousal, and Dominance, respectively. It shows substantially large overall errors that reflect its insufficient performance in emotion recognition. Among deep learning models, the RNN performs moderately with MAE values of 0.121, 0.132, and 0.145. The LSTM and BiGRU further reduces errors to 0.118/0.129/0.144 and 0.106/0.117/0.130, respectively, demonstrating improved modeling of sequential information. The standalone CNN excels in local feature extraction, yielding MAE of 0.111, 0.120, and 0.137. By integrating spatial and temporal features, the combined CNN + BiGRU model further decreases errors to 0.098, 0.109, and 0.122. Leveraging its capability to capture global dependencies, the Transformer model further reduces MAE to 0.094, 0.106, and 0.118. The state-of-the-art neural network models in References38,39 demonstrate excellent performance, with MAE values within the ranges of 0.092–0.093, 0.104–0.105, and 0.115–0.116 for Valence, Arousal, and Dominance, respectively, reflecting the current high standards in the field. The proposed CNN-BiGRU-Attention model further reduces MAE to 0.082, 0.095, and 0.107, representing approximately 11% and 9% reductions in Valence and Arousal dimensions compared to Reference38. This indicates that the introduction of the attention mechanism effectively enhances the model’s ability to capture key emotional segments, significantly improving emotion recognition accuracy and generalization. In summary, the model’s performance in MAE metrics further validates its advancement and superior capability in instrumental music emotion recognition tasks. In the Pearson correlation coefficient comparison of Fig. 8, SVR shows weak performance, especially in the Dominance dimension (0.652), indicating that traditional regression models have limited ability to capture emotional intensity. Although RNN demonstrates improvements, it still lags modern deep models. LSTM and BiGRU significantly enhance the capture of sequential information, with BiGRU’s bidirectional structure outperforming LSTM across all three dimensions. The CNN model also achieves good results through efficient extraction of spatial features. The CNN + BiGRU model, which integrates temporal and spatial features, further improves correlation coefficients, ranging from 0.838 to 0.804. Transformer, with its powerful global dependency modeling capability, slightly leads in all dimensions, and the models in References38,39, based on similar Transformer structures, show close performance. Notably, the proposed CNN-BiGRU-Attention model achieves the highest correlation coefficients in all dimensions, reaching 0.871 for Valence, 0.832 for Arousal, and 0.784 for Dominance. It fully demonstrates the sensitive capture and expressive ability of the multi-mechanism fusion model for complex musical emotions. In Fig. 9, the coefficient of determination (R²) reflects the model’s explanatory power for emotional variables. SVR performs worst, with R² values in all three dimensions falling below 0.52 (0.518 for Valence, 0.495 for Arousal, and 0.472 for Dominance), indicating limited ability to explain emotional variations. Although the RNN model shows some improvement, its R² (0.541/0.523/0.501) still lags most deep learning models. LSTM, BiGRU, and CNN exhibit relatively close R² performance, ranging from 0.57 to 0.66 across dimensions, suggesting that single models can explain emotional changes to some extent. The combined CNN + BiGRU model significantly improves R² (0.683/0.652/0.614), demonstrating that the fused architecture effectively enhances feature representation. Transformer-based models and those in References38,39 exhibit stronger explanatory power, with R² exceeding 0.64. Notably, the proposed model leads with the highest R² values of 0.742 (Valence), 0.665 (Arousal), and 0.583 (Dominance). It proves that the introduction of the attention mechanism provides an advantage in interpreting complex emotional data, enabling more accurate fitting of continuous emotional change curves. To further verify the statistical reliability of the proposed model’s superiority, this study conducts paired sample t-tests on the CNN-BiGRU-Attention model and various baseline models under the same dataset and evaluation conditions. The test results are shown in Table 6. Compared with the Transformer model with the best performance and the advanced neural network architecture in reference38, the proposed model achieves significant improvements in the three emotional dimensions of Valence, Arousal, and Dominance (p < 0.01). Meanwhile, to evaluate the stability and credibility of the model performance, this study further calculates the Pearson correlation coefficient and the 95% confidence interval (CI) of MSE. The results show that the CI range is narrow, indicating that the model performance improvement has consistency and statistical reliability. The CNN-BiGRU-Attention model shows significant advantages in three emotional dimensions compared to the latest reference model, and the p-values are all less than 0.01, indicating that the performance differences are statistically significant. The 95% CI fluctuation of its Pearson correlation coefficient is relatively small (± 0.014–0.018), and the MSE confidence interval is also stable (± 0.003–0.005), indicating that the model results have high stability and reproducibility. This further demonstrates the effectiveness and statistical reliability of the model fusion structure in instrument music emotion recognition tasks. Table 6Paired sample t-test results and 95% confidence intervals for different models.Emotional dimensionsModel comparisonMean difference (Δ)tpPearson r’s 95% CIMSE’s 95% CIValenceThe proposed model vs38+0.0504.72<0.0010.872 ± 0.0140.018 ± 0.003ArousalThe proposed model vs38+0.0474.35<0.0010.832 ± 0.0160.020 ± 0.004DominanceThe proposed model vs38+0.0413.890.0020.784 ± 0.0180.024 ± 0.005ValenceThe proposed model vs Transformer+0.0494.58<0.0010.872 ± 0.0140.018 ± 0.003ArousalThe proposed model vs Transformer+0.0454.26<0.0010.832 ± 0.0160.020 ± 0.004DominanceThe proposed model vs Transformer+0.0393.750.0030.784 ± 0.0180.024 ± 0.005 Paired sample t-test results and 95% confidence intervals for different models. To comprehensively measure the model’s performance under class imbalance, this study further introduces three indicators: accuracy, F1 score, and Area Under the Receiver Operating Characteristic Curve (AURoC), and compares them with the latest literatures38,39 across three emotional dimensions: Valence, Arousal, and Dominance. The results are shown in Table 7: Table 7Comparison of different models on ACCURACY, F1-SCORE, and AUROC indicators.ModelValence AccuracyArousal AccuracyDominance AccuracyValence F1-scoreArousal F1-scoreDominance F1-scoreValence AURoCArousal AURoCDominance AURoCReference380.810.780.760.740.710.690.780.750.73Reference390.830.800.780.760.730.710.800.770.75The proposed model0.880.850.830.820.790.760.850.820.80 Comparison of different models on ACCURACY, F1-SCORE, and AUROC indicators. In Table 7, the accuracy, F1 score, and AURoC metrics of the proposed model are superior to two representative comparative models in three emotional dimensions. The improvement in the Dominance dimension is the most significant, with an accuracy increase of about 7% compared to reference39. This indicates that the proposed model has stronger performance in identifying complex and subtle emotional changes. The higher accuracy indicates that the model has good stability in overall prediction consistency, while the synchronous improvement of F1 score and AURoC further validates the model’s discriminative ability and robustness under class imbalanced data conditions. The improvement in comprehensive performance is mainly due to the synergistic effect of multi-scale feature fusion strategy and adaptive attention mechanism. It enables the model to focus more on emotional turning points and key segments, thus overcoming the problem of insufficient generalization of traditional single scale models in complex music emotion recognition. Discussion The results above demonstrate that the comprehensive music emotion recognition model proposed in this study performs exceptionally well across all evaluation metrics. This indicates that the combination of CNN’s ability to extract local features, BiGRU’s temporal modeling, and the dynamic focus on key emotional features provided by the attention mechanism significantly enhances the model’s generalization and stability when dealing with complex music emotion data. In comparison with similar studies, Blaszke et al.40 constructed CNN and RNN to detect up to 13 different instruments in polyphonic music. Chou et al.41 developed an AI-based social emotion classification model specifically designed to integrate three core positive social emotional elements-positive outlook, empathy, and problem-solving-into classical music. Jia42 proposed a music emotion recognition method based on CNN, which fused two parts of features and inputs them into a softmax classifier with a center loss function to recognize four types of music emotions. For example, Zhu et al.43 effectively improved the semantic understanding ability of medical question answering systems by combining knowledge embedding with Transformers, and demonstrated the potential of cross domain knowledge modeling in complex information parsing; Khan et al.44 emphasized that they achieved high-precision prediction of user emotions by integrating audio, visual, and physiological signals through a multi-scale and multimodal Transformer architecture. These studies indicate that Transformer and multimodal fusion methods have significant advantages in processing complex semantic or emotional information. In contrast, this study proposes a CNN-BiGRU-Attention fusion model for music emotion recognition tasks, which models the continuous emotion dimension in multi category instrument audio through multi-scale feature fusion and dynamic attention mechanism, effectively capturing fine-grained features of emotion changes in time series. Unlike the Transformer methods mentioned above, the proposed model emphasizes more on extracting local audio features and integrating temporal contextual information, thus achieving high accuracy and good generalization ability in music emotion recognition tasks. This indicates that even under similar deep learning frameworks, feature extraction and temporal modeling strategies designed for specific tasks can still significantly improve performance. However, this model structure was relatively simple and lacked dynamic modeling of emotional evolution. The proposed model not only effectively integrates low-level audio features with high-level semantic features but also leverages BiGRU to capture temporal dependencies in music sequences, making it more practically valuable. Conclusion This study addresses the complexity of emotional recognition in instrumental music by proposing a novel emotion recognition model integrating multiple artificial intelligence technologies. Based on a comprehensive review of existing methods, a multi-layered architecture is constructed, fusing CNN for local feature extraction, BiGRU for modeling temporal context dependencies, and an attention mechanism to focus on the most emotionally salient segments of music. Additionally, the model employs a feature fusion strategy to combine low-level audio signals with high-level semantic features, enabling multi-dimensional understanding of musical emotions. Experimental results demonstrate that using Russell’s emotional circumplex model as a continuous emotional space framework effectively aligns with the fluidity and subtlety of emotional expression in instrumental music. This not only enhances the precision and accuracy of emotion recognition but also provides a solid theoretical foundation for future research on modeling complex dynamic musical emotions. The model outperforms traditional approaches across three continuous emotional dimensions, particularly in capturing complex emotional expressions such as melodic development, rhythmic dynamics, and harmonic changes. Despite the promising experimental results, this study has several limitations. These datasets have relatively limited coverage in terms of musical styles and emotional types, failing to fully reflect the emotional differences of instrumental music under different cultures, styles, and expression methods. This somewhat restricts the model’s generalization ability and cross-scene application effects. Second, when facing musical segments with complex combinations of multiple instruments and intertwined timbres, the model’s recognition accuracy for subtle emotional changes still has room for improvement. Future research can further introduce more multi-dimensional and cross-cultural musical emotion datasets, and combine unsupervised or semi-supervised learning strategies to reduce reliance on high-quality emotional labels. Meanwhile, it can also explore multi-modal fusion paths, such as introducing lyric and musical score structure information into the model, to further enhance its ability to understand and express complex musical emotions. For instance, the generalization ability in small sample and cross style music emotion recognition still needs further improvement, and the model still faces certain challenges in separating emotional features in complex multi instrument mixed audio. Future research directions can introduce multimodal fusion strategies to combine audio signals with emotional cues such as lyrics and performer expressions, and simultaneously explore lightweight model structures to adapt to real-time music emotion recognition needs. Future research can also utilize self-supervised learning and transfer learning techniques to enhance the adaptability and robustness of the model in multi domain music data. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Author contributions Yang Liu: Conceptualization, methodology, software, validation, formal analysis, investigation, resources, data curation, writing—original draft preparation, writing—review and editing, visualization, supervision, project administration, funding acquisition. Data availability The datasets used and/or analyzed during the current study are available from the corresponding author Yang Liu on reasonable request via e-mail Liuyangcello@xmu.edu.cn. Declarations Competing interests The authors declare no competing interests. Ethics statement The studies involving human participants were reviewed and approved by Music Department, Art college, Xiamen University Ethics Committee (Approval Number: 2023.21000236). The participants provided their written informed consent to participate in this study. All methods were performed in accordance with relevant guidelines and regulations. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Data Availability Statement The datasets used and/or analyzed during the current study are available from the corresponding author Yang Liu on reasonable request via e-mail Liuyangcello@xmu.edu.cn.