Emotion Recognition in Classical Music - Presentation
3 likes1,064 views
This document proposes a two-stage approach to automatically identify emotions induced by Carnatic music using machine learning techniques. In the first classification stage, music samples are classified into emotional dimensions like devotion, pathos, and calmness. In the second clustering stage, music samples are grouped based on their emotional similarity as identified in the first stage. The results show some improvement over the baseline for classification but evaluations indicate that expanding the dataset with more labeled samples could further improve the accuracy of automatically identifying emotions in Carnatic music.
Emotion Recognition in Classical Music - Presentation
- 1. EMOTIONAL ANALYSIS OF CARNATIC MUSIC Ravi Kiran Arthi Ramachandran
- 2. AGENDA •INTRODUCTION. •OUR SOLUTION. •CONCLUSION.
- 3. INTRODUCTION •PROBLEM DEFINITION •CARNATIC MUSIC
- 4. PROBLEM DEFINITION. High • Framework for supporting Emotion level based Carnatic Music Retrieval. Goal. • Automatically identify emotions induced on listener. Problems • Cluster music samples based on emotional similarity. Explored. • Locate emotionally similar music samples.
- 5. CARNATIC MUSIC Origins • South Indian Classical Music. • Melodic component of a song. Raaga • Child Raagas are derived from Melakartha (Parent) Raagas. • Emotional component. Rasa • Navarasa – 9 Rasas. Raaga • Raagas induce emotions on listener. • Shankarabharanam Happiness Rasa.
- 6. OUR SOLUTION •REPRESENTATION OF EMOTIONS. •CORPUS. •TWO STAGE APPROACH. •CLASSIFICATION STAGE. •CLUSTERING STAGE.
- 7. REPRESENTATION OF EMOTIONS. •EXISTING APPROACH. •OUR APPROACH.
- 8. REPRESENTATION OF EMOTIONS – EXISTING. Thayer’s Model (Existing work) – 2d approach. Valence vs. Arousal.
- 9. REPRESENTATION OF EMOTIONS – OURS. Our model. • 10d approach. Dimensions. • Navarasas + Devotion. Points in • Emotional state of listener. • Emotional state = Mixtures of Navarasa. space. • Happy = Romantic + Peaceful. Dimension • {0-Absent, 1-Weak, 2-Moderate, 3-Strong}. • Example: <Devotion=3, Pathos=3, Values. Calmness=2>
- 10. CORPUS. •COMPOSITION. •EXPERT ANNOTATIONS.
- 11. CORPUS – COMPOSITION Existing Carnatic Music corpus – None. 109 music samples. Good mixture of Music samples. • Multiple emotional classes. • Multiple Raagas. • Multiple Artists. • Multiple Musical Instruments. Segments of 30s after initial 30s. • Beginning Slow exploration of notes.
- 12. CORPUS – ANNOTATIONS Labels. • Ratings on a scale of 0-3 for 10 dims. • Two annotators. Annotators. • Experts in Carnatic Music. Annotation • Easy and intuitive for Artists and Experts. Interface. • Google Docs + Esnips. Low Kappa • Average = 0.22 scores. Example. • <Devotion=3, Pathos=3, Calmness=2>
- 13. SOLUTION – TWO STAGE APPROACH.
- 14. TWO STAGE APPROACH. Classification Clustering Task. Task. Input = Ratings Input = Music for emotional Sample. dimensions. Output = Output = Ratings for Clusters of emotional emotionally dimensions. similar samples.
- 15. CLASSIFICATION STAGE. •TASK. T •RESULTS. •IMPROVEMENTS.
- 16. CLASSIFICATION STAGE – TASK • Predict emotions induced by given Why? Music sample. • Classification models trained for 4 Models. dimensions. • Dynamics, Rhythm, Pitch, Timbre, Tonal Features. • MATLAB (MIRToolbox) for feature extraction. Classification • SVM and Ripper. techniques. • WEKA for Classification.
- 17. CLASSIFICATION STAGE – RESULTS Process -10 fold Cross Validation. Metric - Accuracy (%). Baseline - Majority Class. Emotional Baseline (%) Best Best Best Dimension accuracy – accuracy – accuracy – Feature set SVM (%) Ripper (%) Bhakthi 71 Rhythm 71 71 Sringara 34.6 Timbre 45.8 43 Karuna 58 Dynamics 60 59 Shantha 42.9 Rhythm 49.6 54.3
- 18. CLASSIFICATION STAGE – IMPROVEMENTS. Improvements over baseline low. Sparse data. • Get more labeled samples. Segmentation technique. • Try initial, middle and end 30s segments rather than just one 30s segment.
- 19. CLUSTERING STAGE. •TASK. T •RESULTS. •IMPROVEMENTS.
- 20. CLUSTERING STAGE – TASK • Verify validity of emotional model. Why? • 2 clusters = Happy and Sad emotions. • Clusters are based on emotional similarity. Features • Ratings for 10 dimensions. Clustering • K-means and EM. techniques. • WEKA for Clustering.
- 21. CLUSTERING STAGE – RESULTS Qualitative evaluation. 2 Clusters. Purity of a cluster – Parent Raaga distribution. Raaga Distribution across Clusters. 25 20 Number of samples 15 10 5 0 Parent Raga Count of Cluster 0 Count of Cluster 1
- 22. CLUSTERING STAGE – IMPROVEMENTS. Use of a quantitative evaluation metric for purity. Continuous and finer grained ratings (0-10). Explore results for more than 2 clusters. Sparse data – Get more samples. More Expert annotations.
- 23. CONCLUSION.
- 24. CONCLUSION Novel technique Classification for representing results better than emotional states. baseline. Clustering results Music samples establish validity of clustered based representation of on emotional emotional states. similarity.