| Summary: | 碩士 === 輔仁大學 === 資訊工程學系碩士班 === 107 === Our main goal in this paper is to predict music emotion presented by two-dimensional coordinate system through deep learning algorithm. Within our algorithm, we quantize value of Valence and Arousal into 21 labels and do the prediction individually. We use 1000 Songs Dataset as our dataset, sample our music into segment with different length, and put these samples into different music translation methods. After the pre-processing, we use Convolution Neural Network as our machine learning model. Our training method is 10-cross fold validation and we get its mean value to evaluate the outcome of our model. We use Top-1, Top 3, and Fuzzy 3 as our evaluate method to evaluate overall results. In Top-1 evaluate, the accuracy of Valence is 54.7% and the accuracy of Arousal is 63.1%. In Top-3 evaluate, the accuracy of Valence is 82.2% and the Arousal is 81.8%. And in Fuzzy-3 evaluate, the accuracy of Valence is 82.2% and Arousal is 81.8%
|
|---|
