Music Emotion Recognition using Deep Neural Networks
No Thumbnail Available
Date
2021
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Uva Wellassa University of Sri Lanka
Abstract
Emotion is an integral part of music and a complex aspect of music that is not easily understood by
machines. The emotional aspect of music is further complicated by the fact that it is a subjective
experience that cannot be easily conveyed to machine. Although it is a complex problem, some
progress has been made in this area suggesting that it might be feasible to develop computation
models that can be used in real-world applications. Real-world applications of music emotion
recognition systems range from entertainment to healthcare. In this paper we introduce a deep
learning model that recognizes emotion in music from the audio signal. 1d and 2d convolution layers
with different kernel sizes have been tested. Adaptive pooling layers have also been used to extract a
fixed feature representation for the dense layers. We have also used trainable spectrogram extractors
to learn different representations of the audio. To address the lack of data for the task of music
emotion recognition we have also used the latest trends in audio data augmentation and converted it
for music data. Till now we have been able to achieve an accuracy of about 0.92 for the PMEmo
dataset and about 0.6 F-1 score from using the raw audio signal and 1D convolution layers to extract
features. Preliminary experiments show that using 1d convolutions with the combination of learnable
spectrograms performs satisfactorily. Further experiments are to be conducted using different
combinations of raw audio and calculated features. Different model architectures using recurrent
networks are also to be tested considering that audio has temporal relationship between each unit of
time. Finally, the work done in this study is mainly to explore the high dimensional feature space of
raw audio to extract features which can contribute to the recognition of emotion in music using
automated methods such as convolution and recurrent layers.
Keywords: Music Emotion Recognition; Deep Neural Networks; Music Data Augmentation; Arousal
and Valence Prediction
Description
Keywords
Computing and Information Science, Music, Neural Networks, Computer Science