A Deep Attentive Network for Music Emotion Recognition using the chorus part of the music

doi:10.21203/rs.3.rs-3130450/v1

Download PDF

Research Article

A Deep Attentive Network for Music Emotion Recognition using the chorus part of the music

https://doi.org/10.21203/rs.3.rs-3130450/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

In the domain of music information retrieval, music emotion recognition (MER) is an active area of research. The chorus part of the music is the most critical and emotional section. In this paper, we proposed an MER approach that combines Convolutional Neural Networks with Long Short-Term Memory to extract emotional features from the log-mel spectrogram of the music chorus part. Then, the features feed into an attention mechanism block to enhance helpful information. Also, we introduce a new dataset for the MER task that consists of 10,438 samples with valence/arousal labels. The results show that using chorus parts of music can improve the deep models' performance. Also, experiments show our proposed method achieves 15% and 40% relative improvement of valence and arousal, respectively, on the R2 score compared to the state-of-the-art and baseline models.

Music Emotion Recognition

Neural Network

Convolution Neural Network

Long Short-Term Memory

Attention Mechanism

In recent years, the online music market has achieved rapid development. To improve the efficiency of using music resources, they need to be organized based on label information, such as emotion and genre.

Manual assigning the label information to music can be time and labor-consuming. Therefore, automatically retrieving music information is necessary [1]. Music Information Retrieval (MIR) has been a rapidly growing area of research in recent years, driven by the need to process massive music resources automatically. The emergence of deep networks presents a new solution for MIR research, and researchers have tended to deep neural networks methods such as convolution neural network (CNN) and recurrent neural network (RNN) due to their strong learning ability and good portability [2, 3, 4].

Since emotion is one of the primary reasons people listen to music [5], it is essential to recognize the emotional labels of the music. Music emotion recognition (MER) is an important and challenging field of MIR research [1]. MER analyzes music features to map relationships between music features and emotion space and recognize the emotion expressed by music. Music chorus parts are the most representative and prominent part of musical structure, and human listeners can quickly understand where these parts are because they are the most repetitive and memorable parts of music [6]. These parts are the most important and emotional parts of music [7]. So using extracted features from these parts can improve music emotion recognition tasks.

Introducing an attention mechanism (AM) within deep learning is a notable achievement in essential data handling [8]. Deep neural networks based on attention have proven helpful in extracting information from multimedia content [9]. The use of the AM in deep neural network (DNN) architectures is a recent development aimed at increasing the importance of emotionally salient information [9].

This study proposes a method based on a deep attentive network and using extracted features from the music's chorus part to improve the MER process. We compare our approach with baselines and state-of-the-art methods on a database of 10,438 music labeled with arousal/valence representation. This database is built by using the Million Song Dataset Echo Nest mapping archive (MSDEN) [10] and the Deezer Mood Detection Dataset (DMDD) [11].

We can summarize the contributions of this paper as follows:

We propose a deep attentive learning solution for feature extraction in the MER task. The network extracts emotion-salient audio descriptors from the log-mel spectrograms.
We show the efficiency of processing the chorus part of the music in the MER task.
We introduce a new dataset for the MER task.

The rest of this paper is organized as follows. Section 2 briefly describes the MER task and the most recent studies implemented by DNN. In Section 3, the proposed method is presented in detail. The experiment setup and the benchmark dataset are described in Section 4. In Section 5, the performance analysis of the proposed method is discussed. Section 6 concludes the paper.

Many techniques have been developed over the last two decades with various datasets and features. A significant fraction of them was collected by Yang and Chen in [12] and Han et al. in [1].

In classical MER techniques [13, 1], first, audio is analyzed in small segments (frames) and represented in the time/frequency domain. The pre-processed signal is then utilized to extract audio characteristics that describe the input audio in many ways. Such a feature set necessitates previous knowledge to extract emotion-related information from the signal [14]. The features are then sent into a classifier, which maps the data to emotional classes. New models based on less feature engineering emerged due to deep learning advancements, particularly in speech recognition. This section presents a brief survey on emotion representation and deep MER.

2.1 Music emotion representation

An annotated emotional music database is needed to build a music emotion recognition system. A variety of mood representations have been used in the literature. There are two approaches [1], categorical and dimensional, for labeling emotions in music. In the categorical approach, emotions are characterized by discrete labels, such as sad, happy, angry, etc. In the second approach, emotions are represented in dimensional space.

Categorical models are problematic because there is no consensus on category numbers. There is no such problem in the case of dimensional models. The advantage of dimensional models is the reduced uncertainty compared to the categorical approach [15]. Also, the dimensional representation was used multiple times in the literature [16, 11, 17, 18].

Russell [19] proposed a two-dimensional model of valence and arousal. In the early research, Russell asked participants to categorize 28 emotion terms based on perceived similarities. Russell then used a statistical approach to arrange comparable emotion phrases in a circle (Figure 1). The arousal axis shows emotions ranging from calm to excitement, while valence shows the measure of displeasure to pleasure.

2.2 Deep learning in MER

Lidy and Schiner [20] used mel spectrogram and CNN for music genre and mood classification. The authors use two different architectures of CNN: one sequential and one parallel, which aim to capture both temporal and timbral information simultaneously. Dong et al. [2] introduce a bidirectional Convolutional Recurrent Sparse Network, which combines CNN and LSTM to strengthen the ability to learn features from the spectrogram. CNN can learn features adaptively, and RNN is more suitable for processing sequence data. Also, they optimize the emotional representation to make the training process faster. Jia [21] presented a DNN that uses a spectrogram and Low-level descriptors (LLD) as input. The model mainly comprises spectrogram + CNN-LSTM and LLDs + DNN. It extracts the spectrogram features from the two aspects of image features and timing features, and LLDs features are integrated into the network to improve the expression of emotional information. Chaudhary et al. [22] use three stacked CNN layers to learn emotional features from music spectrograms. Han et al. [23] use the idea of an Inception Module to extract features of different dimensions. Also, they use one-dimensional residual CNNs and Gate Recurrent Unit (GRU) to extract timing features.

Several multi-modal approaches were also made, trying to utilize both the audio and lyrics of the music tracks. A work worth mentioning is Delbouys et al. [24], who proposed a multi-modal music emotion prediction based on audio signals and lyrics. Mel spectrogram was used as input for audio features, and word embedding was used as input for lyrics features, which meant that each word was embedded in a continuous space, and the vectors corresponding to each word were stacked. Pyrovolakis et al. [18] used deep learning architectures to investigate and compare single-channel and multi-modal techniques for music mood identification. They suggested a multi-modal classification system based on CNN and BERT. They also demonstrated that correctly extracting and combining auditory information could improve the prediction target.

Although these earlier studies used various deep learning techniques to produce promising results in MER, there is still potential for development in this area [1]. In this research, we investigate how to detect music emotion using deep learning with an attention mechanism on the chorus part of the music.

In this section, we introduce the proposed method for MER. The method consists of two stages. At first, the music is pre-processed to extract the chorus part and represented concisely to feed our neural network. Afterward, the data is provided to our proposed network to extract features and make a classifier.

3.1 Data processing

The music chorus parts are extracted with Pychorus. Pychorus is a Python library to find choruses or interesting sections in pieces of music. The library is implemented based on Goto [6] with some modifications. The minimum length to consider the chorus is set to 15 seconds. Also, if the chorus is longer than 15 seconds, the first 15 seconds will be considered.

The extracted audio parts are not in a format that can be used immediately to train the systems developed. A log-mel spectrogram is computed for each music part and used as the input for the proposed network. The mel spectrogram uses a mel scale which is an attempt to imitate humans' perceptual hearing on musical pitches [25] and summarizes spectral information in a concise form. Log-mel spectrogram is computed by applying logarithmic transformation on the frequency. The motivation for applying the logarithmic transformation on the mel spectrogram is that humans do not hear loudness on a linear scale [26]. Figure 2 shows the visual difference between a sample music part's mel spectrogram and log-mel spectrogram.

The music log-mel spectrograms are computed via librosa^[1] with 64 mel-filters and 2048 window size, 2048 hop size, with a sampling frequency of 22050Hz. Since the mel scale's definition is influenced by a limited number of subjective psychoacoustic experiments, several methods exist in the literature [26, 27, 28]. Librosa employs the Slaney [26] method by default. The log-mel spectrogram was standardized to zero mean and unit variance before being fed into the network.

3.2 Model Description

The presented model combines CNN with LSTM, and AM is introduced into networks to enhance helpful information. CNN is applied to extract high-level invariant features from the log-mel spectrogram [2]. LSTM can effectively capture the context information of the input sequence and solve the problem of saving and transmitting sequence information [16]. AM gives different weights to the features, highlights the influence of the more critical factor, and helps the model make a more accurate classifier [21, 29]. Figure 3 shows the block diagram of the proposed network.

First, we fed the log-mel spectrograms input to one-dimensional CNNs. In one-dimensional CNN, the kernel slides along one dimension, which is suitable for dealing with time-series data. Pooling layers are typically used after convolution layers to perform nonlinear down sampling by taking the maximum or average value for each small subsection of the matrix. Our model comprises two consecutive one-dimensional convolution layers that use 32 and 16 filters with ReLU [30] as an activation function, a kernel size of 7, stride 1, and max pooling of size 4 and stride 4. We applied batch normalization [31] after each convolutional layer to stabilize the network. This technique normalizes the output of a previous layer by subtracting the average value of the batch and dividing it by the batch's standard deviation.

Using the feature sequence from the CNN blocks as the input to LSTM, the features of each time can be extracted through the LSTM unit. So, the extracted features are fed to an LSTM layer with 64 hidden states. Afterward, a self-attention block is applied to the output of the LSTM layer to obtain the importance of different features. The attention block is scaled dot-product attention [32] which the attention scores are calculated as the dot product of the queries and keys divided by the square root of the number of dimensions of the queries and keys to ensure they are stable.

Before feeding the features to the prediction module, we applied a flattened layer to transform LSTM output feature vectors into one-dimensional feature vectors. Finally, the feature vectors are fed into the prediction module to build a classifier. The prediction module consists of a fully connected layer with two nodes, dropout 0.5 and tanh, as its activation function. The dropout was employed to avoid the network over-fitting. Table 1 represents the network layers architecture for our proposed model.

Table 1: The network layer architecture of the proposed model

No.	Layer type	Detail of layer
1.	Input	The batch of music log-mel spectrograms 32×646×64 (batch×mel-filters×time)
2.	Convolution 1D	32 filters with kernel 7, stride 1, padding 3
3.	Batch Normalization 1D	Batch normalization with 32 channels
4.	ReLU	-
5.	Max Pooling	Kernel 4, stride 4
6.	Convolution 1D	16 filters with kernel 7, stride 1, padding 3
7.	Batch Normalization 1D	Batch normalization with 16 channels
8.	ReLU	-
9.	Max Pooling	Kernel 4, stride 4
10.	Long short-term memory	Number of hidden nodes 64, number of layers 1
11.	Self Attention	Scaled dot-product attention
12.	Flatten	Flattened to 1D tensor with 2560 neurons
13.	Tanh	-
14.	Dropout	keep probability 0.5
15.	Fully Connected (Output)	Fully Connected two neurons

4.1 Dataset

The proposed method needs music files to extract the chorus parts and compute the mel spectrograms. Since there was no suitable dataset with the chorus parts and Valence/Arousal labels, we built a dataset that fit our needs using DMDD and MSDEN. DMDD consists of a dataset of Valence/Arousal labels for 18,647 samples on the Million Song Dataset (MSD) [33]. MSDEN is mapping MSD song IDs to IDs of other services, such as Spotify.

To build our dataset, first, we mapped MSD IDs to Spotify IDs using the MSDEN dataset. Then we used spotDL[1] to download the music files of mapped songs. spotDL is an open-source tool that finds pieces by Spotify URL on YouTube if a match is found and downloads them along with album art, lyrics, and metadata. Finally, the music chorus parts are extracted by Pychorus, and the log-mel spectrogram is computed using librosa with the configuration stated in section 3.1.

Since some MSD IDs were unavailable on the Spotify IDs list, and some Spotify music was unavailable on YouTube, we cannot get all 18,647 samples in DMDD. Also, Pychorus cannot find a chorus for some music.

The above process finally led to a dataset including 10,438 samples. Due to rights restrictions, the music files did not include in the dataset. We released the extracted log-mel spectrograms (the chorus part and music's first 15 seconds), labels, MSD IDs, Spotify IDs, artists, track names, and the Python script we used to download music files^[2]. The test, train, and validation sets were separated in DMDD, and we also preserved each set for each sample. Table 1 lists the number of samples in each set. Figure 4 shows the distribution of data based on labels.

Table 2: Dataset details

Set	Number of Samples
Train	6213
Test	2026
Validation	2199

4.2 Experiment details

To experiment and analyze our method, the proposed method in this study is implemented by a Python script. The configuration of the experimental environment is listed in Table 2.

Table 3: Experimental environment configuration

Parameter	Value
Operating system	Windows 11
CPU	Intel® Core™ i7-12700H
GPU	NVIDIA GeForce RTX 3060 (Laptop, 140W)
Python	3.10.6
PyTorch	2.0.1
PyTorch Compute Platform	CUDA 11.7
Pychorus	0.1
Librosa	0.10.0
spotDL	4.1.11

Since the target labels are dimensional, we have chosen the R2 score and root-mean-squared error (RMSE) as evaluation metrics. The R2 score ranges from -1 to -1, with a higher value being better. The smaller the RMSE, on the other hand, the better. A batch size of 32 samples and training up to 100 epochs was chosen for training the model. Moreover, the network's training uses the Stochastic Gradient Descent (SGD) optimization algorithm with a learning rate 0.01 and MSE as a loss function. To prevent over-fitting early stopping and L2 regularization are used with rates of 0.00001 and 0.0001, respectively.

4.3 Experiment results

To verify the advancement of our proposed method, we compared our approach with some state-of-the-art deep architectures and baselines. Since some research [18, 24] proposed multi-modal methods that use textual data (lyrics), these methods could not be directly compatible with ours. So, we have ignored textual data and compared the result with the audio-based section of their works.

The benchmark metrics were computed separately on the chorus part and the first 15 seconds of music to show the effect of using chorus parts as input. The result is summarized in Tables 4 and 5. Training time for deep methods is listed separately for training on CPU and GPU. As shown in the tables, our proposed model has obvious advantages over the others, which get lower RMSE and higher R2 scores.

Table 4: Performance comparison with baselines results on the chorus part (15 seconds) and music's first 15 seconds

Model	Input	MSE	R2 (V)	R2 (A)	R2 (Overall)	Training Time (sec)
SVR	First 15 sec	0.1087	-0.0151	0.0668	0.0258	552 (CPU)
SVR	Chorus part	0.1049	0.0118	0.1097	0.0608	573 (CPU)
Decision Tree Regressor	First 15 sec	0.1062	0.0097	0.0874	0.0485	1 (CPU)
Decision Tree Regressor	Chorus part	0.1058	0.0004	0.1069	0.0537	1 (CPU)
KNN	First 15 sec	0.1042	0.0178	0.1173	0.0675	1 (CPU)
KNN	Chorus part	0.1032	0.0124	0.1442	0.0783	1 (CPU)
Elastic-net LR	First 15 sec	0.1032	0.0276	0.1253	0.0765	40 (CPU)
Elastic-net LR	Chorus part	0.1029	0.0198	0.1411	0.0805	39 (CPU)
Gradient Boosting Regressor	First 15 sec	0.1060	0.0186	0.0799	0.0493	544 (CPU)
Gradient Boosting Regressor	Chorus part	0.1052	0.0157	0.1005	0.0581	458 (CPU)
AdaBoost Regressor	First 15 sec	0.1059	0.0155	0.0868	0.0511	515 (CPU)
AdaBoost Regressor	Chorus part	0.1047	0.0079	0.1182	0.0630	515 (CPU)
Our Model	First 15 sec	0.0992	0.0805	0.1402	0.1103	287 (CPU)
Our Model	Chorus part	0.0957	0.1207	0.1618	0.1412	251 (CPU)

Table 5: Performance comparison with state-of-the-art results on the chorus part (15 seconds) and music's first 15 seconds

Model	Input	MSE	R2 (V)	R2 (A)	R2 (Overall)	Training Time (sec)
Delbouys et al. (2018) (Audio Net Only) [24]	First 15 sec	0.1040	0.0300	0.1067	0.0684	49 (GPU) 88 (CPU)
Delbouys et al. (2018) (Audio Net Only) [24]	Chorus part	0.1026	0.0546	0.1046	0.0796	38 (GPU) 83 (CPU)
Chaudhary et al. (2020) [22]	First 15 sec	0.1016	0.0695	0.1054	0.0874	122 (GPU) 241 (CPU)
Chaudhary et al. (2020) [22]	Chorus part	0.1014	0.1014	0.0494	0.0841	121 (GPU) 238 (CPU)
Pyrovolakis et al. (2022) (Audio Net Only) [18]	First 15 sec	0.1048	0.0067	0.1186	0.06269	1650 (GPU) 3135 (CPU)
Pyrovolakis et al. (2022) (Audio Net Only) [18]	Chorus part	0.1043	0.0089	0.1260	0.0674	1797 (GPU) 5382 (CPU)
Han et al. (2023) [23]	First 15 sec	0.1087	0.0213	0.0222	0.0217	114 sec (GPU) 823 sec (CPU)
Han et al. (2023) [23]	Chorus part	0.1036	0.0442	0.0958	0.0700	112 sec (GPU) 820 sec (CPU)
Our Model	First 15 sec	0.0992	0.0805	0.1402	0.1103	70 (GPU) 287 (CPU)
Our Model	Chorus part	0.0957	0.1207	0.1618	0.1412	60 (GPU) 251 (CPU)

Our model performed best when chorus parts were used. The outcome also demonstrates that using chorus parts can enhance other methods. In the case of arousal, we observed a 15% relative improvement in R2 of the best models and a 40% relative improvement in the case of the valence regressor. As a result, the valence regressor improved more than the arousal regressor. According to the results, our proposed model performs better than most stated methods regarding training time. Figures 5 and 6 represent R2 score and Loss (MSE) values for training and validation data.

In this paper, we have presented a new deep architecture for music emotion recognition from the chorus part of the music. Also, a new dataset was collected to extract the music chorus parts and compute the mel spectrograms. The proposed architecture is created by stacking two one-dimensional CNN layers for latent feature extraction and followed by a unidirectional LSTM to extract time-dependent features and a self-attention block to enhance the extracted features. A fully connected layer with two nodes is used for valence/arousal regression.

The results show that using chorus parts of music can improve the model performance. The experimental results also show that the proposed method outperforms state-of-the-art methods and baselines. Moreover, the training time of the model is significantly shorter than the other deep methods.

Future works could attempt to provide a multi-modal deep MER on chorus parts. Also, future works could use the proposed deep architecture for tasks other than music emotion recognition.

Correspondence

Elham Parvinnia – e-mail: [email protected]

Conflicts of Interest

The authors declare that they have no conflicts of interest.

Data Availability Statement

In case the article is accepted, the dataset will be released at the following address:

https://github.com/alitalaghat/music-emotion-dataset

D. Han, Y. Kong, J. Han and G. Wang, "A survey of music emotion recognition," Frontiers of Computer Science, vol. 16, no. 6, p. 166335, 2022.
Y. Dong, X. Yang, X. Zhao and J. Li, "Bidirectional Convolutional Recurrent Sparse Network (BCRSN): An Efficient Model for Music Emotion Recognition," IEEE Transactions on Multimedia, vol. 21, no. 12, pp. 3150–3163, 2019.
A. Aljanaki, Y.-H. Yang and M. Soleymani, "Developing a benchmark for emotional analysis of music," PloS one, vol. 12, no. 3, 2017.
J. Ramírez and M. J. Flores, "Machine learning for music genre: multifaceted review and experimentation with audioset," Journal of Intelligent Information Systems, vol. 55, no. 3, pp. 469–499, 2020.
T. Eerola and J. K. Vuoskoski, "A Review of Music and Emotion Studies: Approaches, Emotion Models, and Stimuli," Music Perception: An Interdisciplinary Journal, vol. 30, no. 3, pp. 307–340, 2012.
M. Goto, "A chorus section detection method for musical audio signals and its application to a music listening station," IEEE Transactions on Audio, Speech, and Language Processing, vol. 14, no. 5, pp. 1783–1794, 2006.
Y.-S. Huang, S.-Y. Chou and Y.-H. Yang, "Pop Music Highlighter: Marking the Emotion Keypoints," arXiv preprint arXiv:1802.10495, 2018.
D. Bahdanau, K. Cho and Y. Bengio, "Neural Machine Translation by Jointly Learning to Align and Translate," arXiv preprint arXiv:1409.0473, 2014.
E. Lieskovská, M. Jakubec, R. Jarina and M. Chmulík, "A Review on Speech Emotion Recognition Using Deep Learning and Attention Mechanism," Electronics, vol. 10, no. 10, p. 1163, 2021.
"Million Song Dataset Echo Nest mapping archive," 1 July 2016. [Online]. Available: https://labs.acousticbrainz.org/million-song-dataset-echonest-archive/. [Accessed 15 May 2023].
R. H. Rémi Delbouys, F. Piccoli, J. Royo-Letelier and M. Moussallam, "Music Mood Detection Based On Audio And Lyrics With Deep Neural Net," in International Society on Music Information Retrieval, 2018.
Y.-H. Yang and H. H. Chen, "Machine Recognition of Music Emotion: A Review," ACM Transactions on Intelligent Systems and Technology (TIST), vol. 3, no. 3, pp. 1–30, 2012.
J. S. Gómez-Cañón, E. Cano, T. Eerola, P. Herrera, X. Hu, Y.-H. Yang and E. Gómez, "Music Emotion Recognition: toward new, robust standards in personalized and context-sensitive applications," IEEE Signal Processing Magazine, vol. 38, no. 6, pp. 106–114, 2021.
R. Panda, R. M. Malheiro and R. P. Paiva, "Audio features for music emotion recognition: a survey," IEEE Transactions on Affective Computing, 2020.
S. Hizlisoy, S. Yildirim and Z. Tufekci, "Music emotion recognition using convolutional long short term memory deep neural networks," Engineering Science and Technology, an International Journal, vol. 24, no. 3, pp. 760–767, 2021.
J. Grekow, "Music emotion recognition using recurrent neural networks and pretrained models," Journal of Intelligent Information Systems, pp. 531–546, 2021.
R. Sarkar, S. Choudhury, S. Dutta, A. Roy and S. K. Saha, "Recognition of emotion in music based on deep convolutional neural network," Multimedia Tools and Applications, vol. 79, pp. 765–783, 2019.
K. Pyrovolakis, P. Tzouveli and G. Stamou, "Multi-Modal Song Mood Detection with Deep Learning," Sensors, vol. 22, no. 3, p. 1065, 2022.
J. A. Russell, "A circumplex model of affect," Journal of personality and social psychology, vol. 39, no. 6, p. 1161, 1980.
T. Lidy and A. Schindler, "Parallel Convolutional Neural Networks for Music Genre and Mood Classification," MIREX2016, 2016.
X. Jia, "Music Emotion Classification Method Based on Deep Learning and Explicit Sparse Attention Network," Computational Intelligence and Neuroscience, 2022.
D. Chaudhary, N. P. Singh and S. Singh, "Development of music emotion classification system using convolution neural network," International Journal of Speech Technology, pp. 571–580, 2020.
X. Han, F. Chen and J. Ban, "Music Emotion Recognition Based on a Neural Network with an Inception-GRU Residual Structure," Electronics, vol. 12, no. 4, p. 978, 2023.
R. Delbouys, R. Hennequin, F. Piccoli, J. Royo-Letelier and M. Moussallam, "Music Mood Detection Based On Audio And Lyrics With Deep Neural Net," in ISMIR, 2018.
K. W. Cheuk, K. Agres and D. Herremans, "The Impact of Audio Input Representations on Neural Network based Music Transcription," in International Joint Conference on Neural Networks (IJCNN), 2020.
J. Harrington and S. Cassidy, Techniques in Speech Acoustics, Springer, 199.
M. Slaney, "Auditory toolbox," Interval Research Corporation, Tech. Rep, 1998.
S. Young, G. Evermann, M. Gales, T. Hain, D. Kershaw, X. (. Liu, G. Moore, J. Odell, D. Ollason, D. Povey, V. Valtchev and P. Woodland, "The HTK book," Cambridge university engineering department, 2009.
D. O'Shaughnessy, Speech Communications: Human and Machine, 1999.
N. Niu, "Music Emotion Recognition Model Using Gated Recurrent Unit Networks and Multi-Feature Extraction," Mobile Information Systems, 2022.
V. a. H. G. E. Nair, "Rectified linear units improve restricted boltzmann machines," in Proceedings of the 27th international conference on machine learning (ICML-10), 2010.
S. Ioffe and C. Szegedy, "Batch normalization: accelerating deep network training by reducing internal covariate shift," in International Conference on International Conference on Machine Learning, 2015.
A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser and I. Polosukhin, "Attention Is All You Need," Advances in neural information processing systems, vol. 30, 2017.
T. Bertin-Mahieux, D. P. Ellis, B. Whitman and P. Lamere, "The million song dataset," 2011.

^[1] https://librosa.github.io (accessed on 20 December 2022)

^[2]https://github.com/spotDL/spotify-downloader

^[3] https://github.com/alitalaghat/music-emotion-dataset

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

A Deep Attentive Network for Music Emotion Recognition using the chorus part of the music

Status:

Version 1

Abstract

Figures

1 Introduction

2 Related work

2.1 Music emotion representation

2.2 Deep learning in MER

3 Proposed Method

3.1 Data processing

3.2 Model Description

4 Experiments & Results

4.1 Dataset

4.2 Experiment details

4.3 Experiment results

5 Conclusion

Declarations

References

Footnotes

Additional Declarations

Status:

Version 1