Rui Pedro Paiva
University of Coimbra
MOODetector: A System for Mood-based Classification
and Retrieval of Audio Music
2 MOODetector
Contents
 Research Goals
 Current Work
 Future Work
3 MOODetector
Research Goals
 Mood models
• Categorical vs Dimensional, Discrete vs Continuous
 Feature extraction
• Aud...
4 MOODetector
Current Work
 Automatic Creation of Mood Playlists
Mood model: Thayer model
5 MOODetector
Current Work
 Automatic Creation of Mood Playlists
Ground Truth
Yang’s annotations
 194 25-sec excerpts ...
6 MOODetector
Current Work
 Automatic Creation of Mood Playlists
Feature Extraction and Selection
Which features?
– Tim...
7 MOODetector
Current Work
 Automatic Creation of Mood Playlists
Feature Extraction and Selection
Audio features: 3 fra...
8 MOODetector
Current Work
 Automatic Creation of Mood Playlists
Playlist Generation
Closest songs to the seed in the T...
9 MOODetector
Current Work
 Automatic Creation of Mood Playlists
Preliminary Results
Regression
 10-fold cross validat...
10 MOODetector
Current Work
 Automatic Creation of Mood Playlists
Preliminary Results
Playlist quality
 SVR training a...
11 MOODetector
Current Work
 Mood Tracking
Ground Truth
Annotations by 2 subjects: quadrants only
Preliminary Results
...
12 MOODetector
Future Work
 Ground Truth
Clip-level and mood-tracking
 Feature Extraction
Propose good features for va...
Upcoming SlideShare
Loading in …5
×

MOODetector: A System for Mood-based Classification and Retrieval of Audio Music 2010

621 views

Published on

0 Comments
1 Like
Statistics
Notes
  • Be the first to comment

No Downloads
Views
Total views
621
On SlideShare
0
From Embeds
0
Number of Embeds
7
Actions
Shares
0
Downloads
17
Comments
0
Likes
1
Embeds 0
No embeds

No notes for slide

MOODetector: A System for Mood-based Classification and Retrieval of Audio Music 2010

  1. 1. Rui Pedro Paiva University of Coimbra MOODetector: A System for Mood-based Classification and Retrieval of Audio Music
  2. 2. 2 MOODetector Contents  Research Goals  Current Work  Future Work
  3. 3. 3 MOODetector Research Goals  Mood models • Categorical vs Dimensional, Discrete vs Continuous  Feature extraction • Audio, MIDI, lyrics(?)  Feature selection and evaluation • Feature relevance • Feature space reduction • Feature combinations  Knowledge extraction • Fuzzy rules  Mood tracking
  4. 4. 4 MOODetector Current Work  Automatic Creation of Mood Playlists Mood model: Thayer model
  5. 5. 5 MOODetector Current Work  Automatic Creation of Mood Playlists Ground Truth Yang’s annotations  194 25-sec excerpts manually annotated in the Thayer plane
  6. 6. 6 MOODetector Current Work  Automatic Creation of Mood Playlists Feature Extraction and Selection Which features? – Timing: Tempo, tempo variation, duration contrast – Dynamics: overall level, crescendo/decrescendo, accents – Articulation: overall (staccato/legato), variability – Timbre: Spectral richness, onset velocity, harmonic richness – Pitch (high/low) – Interval (small/large) – Melody: range (small/large), direction (up/down) – Harmony (consonant/complex-dissonant) – Tonality (chromatic-atonal/key-oriented) – Rhythm (regular-smooth/firm/flowing-fluent/irregular-rough) – Mode (major/minor) – Loudness (high/low) – Musical form (complexity, repetition, new ideas, disruption)
  7. 7. 7 MOODetector Current Work  Automatic Creation of Mood Playlists Feature Extraction and Selection Audio features: 3 frameworks Forward Feature Selection PCA Arousal-Valence Modeling Support Vector Regression  Grid parameter search
  8. 8. 8 MOODetector Current Work  Automatic Creation of Mood Playlists Playlist Generation Closest songs to the seed in the Thayer plane Mood trajectory
  9. 9. 9 MOODetector Current Work  Automatic Creation of Mood Playlists Preliminary Results Regression  10-fold cross validation, 20 repetitions  Forward Feature Selection: 53 features for Ar., 80 for Va.  R2 statistics  Yang: Ar = 58.3%, Va = 28.1%
  10. 10. 10 MOODetector Current Work  Automatic Creation of Mood Playlists Preliminary Results Playlist quality  SVR training and distance to the seed  Previously selected features  4-fold cross validation (75-25%)
  11. 11. 11 MOODetector Current Work  Mood Tracking Ground Truth Annotations by 2 subjects: quadrants only Preliminary Results Classification: SVM 4 classes (quadrants) Using only Marsyas: 48%
  12. 12. 12 MOODetector Future Work  Ground Truth Clip-level and mood-tracking  Feature Extraction Propose good features for valence Extract symbolic features from MIDI files MIDI versions of the audio songs Automatic music transcription  Feature Selection Feature combination: e.g., RRelieff Genetic Algorithms  Knowledge Extraction Neural-Fuzzy Systems

×