This thesis presents a new dataset of perceived and induced emotions for 200 audio clips. The gathered dataset provides users' perceived and induced emotions for each clip, the association of color, along with demographic and personal data, such as user's emotion state and emotion ratings, genre preference, music experience, among others. With an online survey we collected more than 7000 responses for a dataset of 200 audio excerpts, thus providing about 37 user responses per clip.
The focus of the thesis is the evaluation of classifying emotion states in audio with two existing algorithms. Regression algorithm is used to estimate valence and arousal ratings for audio. The Gaiatransform algorithm is used to classify audi clips in five mood clusters. Gaiatransform algorithm also provide probability of presence for six moods in song. Finally, the regression algorithm was used to analyze possible correlation between colors and mood in valence-arousal space.