Emotion Recognition in Songs via Bayesian Deep Learning
Jeevan Singh Nayal, Abhishek Joshi, Bijendra Kumar · 2019
In the era where enormous amount of data is generated every moment through multimedia and Internet, songs are no exception. New songs are released through the Internet and make their way into digital music libraries. However, music information retrieval on these platforms can be really challenging, and in particular, the task of recognition of musical emotion is a popular research area. In this paper, we propose a novel method to recognize the emotion implicit in songs. To the best of our knowledge, ours is the first attempt to solve for emotion recognition incorporating Bayesian Deep learning technique. We obtain spectrograms from the audios to leverage both the time and frequency information and classify with a Bayesian Convolutional Neural Network (CNN). We demonstrate this approach by evaluating it on a benchmark dataset and achieve improved performance over traditional machine learning methods that have been used in the past for this task. Further, we provide a thorough analysis of our proposed approach and perform statistical significance test for comparison of proposed model against the baseline.