| |
| Acadêmico(a): Thomas Oelke Adriano |
| Título: MUSIC EMOTIONS INTEL: IDENTIFICADOR AUTOMÁTICO DE EMOÇÕES EM MÚSICAS |
| |
| Abstract: |
| This work presents the development of a new approach to the task of identifying emotions in music. While the traditional methods of emotion recognition in music are focused in identifying a singular emotion per music, this work proposes to recognize a set of emotions by music. To this aim, as training set, it was used the dataset of Mohammad (2013), 1000 Songs for Emotional Analyzes of Music. To extract the audio characteristics was used the Mel Frequency Cepstral Coefficients (MFCC), the Spectral Centroid (SC), the Zero Crossing Rate (ZCR), the chromagram e the tempogram. A Support Vector Regression (SVR) with RBF kernel was applied to learn the algorithm. To map emotionally the emotions the circumplex model of affect of Russel (1980) was used. The set of characteristics and the regression algorithm were chosen based on the work of Huq (2010). The result obtained with this work was measured both qualitatively and quantitatively. With the qualitative analysis, it could be observed that a representation of emotions using a set of emotions by music adheres better to the reality of the current music than a representation of only one emotion per music. For a quantitative analysis, a comparison was made between the minimum value of the expected accuracy with the result obtained by the proposed approach. The minimum accuracy value, which is, the baseline, was defined by the Mean Absolute Error (MAE) between the expected Valence and Alert values and their mean. The quantitative results were not as good as expected, however showed opportunity for improvements |
| |
|
