1. Stochastic Text Models for Music Categorization Carlos Pérez-Sancho, José M. Iñesta, David Rizo Pattern Recognition and Artificial Intelligence group Department of Software and Computing Systems University of Alicante, Spain

2. SSPR 20082 Outline ► Introduction ► Music encoding  Melody  Harmony ► Experiments  Plain classification  Classifier ensembles  Hierarchical classification ► Conclusions

3. SSPR 20083 Introduction ► Premise: music content can be used to model musical style ► We use language modeling techniques to classify symbolic digital scores ► For that, digital scores need to be encoded into sequences of symbols

4. SSPR 20084 Music encoding ► Two different sources of information Melody Harmony

5. SSPR 20085 Melody encoding ► Polyphonic sequences are reduced to monophonic using skyline ► Pitch intervals and duration ratios are computed for each pair of consecutive notes ► Numeric values are encoded into ASCII symbols (2,×½) → Bf (1,×1) → AZ (-1,×2) → aF (-2,×1) → bZ (2,×1) → BZ (-4,×1) → dZ (-2,×1) → bZ (4,×1) → DZ Bf AZ aF bZ BZ dZ bZ DZ ( interval, duration ratio ) input string

6. SSPR 20086 Harmony encoding ► Chords are encoded as degrees relative to the tonality for transposition invariance ► Only chord changes are encoded Key: E flat VIm V I input string

7. SSPR 20087 Experiments ► Dataset: music from 3 genres and 9 sub-genres (around 60 hours of music) ► Classification techniques  Naïve Bayes  Language modeling (n-grams), classifying by lowest perplexity

8. SSPR 20088 Experimental setup ► First step: plain classification  80% of the dataset used  10-fold cross validation ► Second step: hierarchical classification using classifier ensembles  Weights of the ensembles adjusted using previous results  Remaining 20% of dataset used for validation

9. SSPR 20089 Classification results ► Best classification rates in the 3-classes problem were obtained using harmonic information ► When classifying sub-genres, melody usually performs better ► Naïve Bayes performs better most of the times ► No significant differences for different context sizes in n-grams. Chords (harmony)Melody 2-grams 3-grams 4-grams N.B.

10. SSPR 200810 Confussion matrix ► Misclassifications occur more frequently within broad domains ► Try to prevent intra-domain errors by using a hierarchical classifier ×100 %

11. SSPR 200811 Hierarchical classification

12. SSPR 200812 Hierarchical classification ► Harmony (chord progressions) is used at the first level ► Melody is used at the second level ► Instead of using single classifiers, an ensemble of classifiers is used at each level to increase robustness

13. SSPR 200813 Classifier ensembles ► Decisions are made by weighted majority vote ► Two weighting schemes  Linear best-worst weighting vote  Quadratic best-words weighting vote # errors weight

14. SSPR 200814 Hierarchical classification results ► With the remaining 20% of the dataset Single classifiers 2-grams3-grams4-gramsN.Bayes 3 classes Using harmony 87.888.4 87.8 9 classes Using melody 54.748.847.758.7 Best singleLinearQuadratic 3 classes 88.4 (melody 2-grams) 90.1 Hierarchical classification 9 classes 60.562.863.4 1st level (3 classes): harmony 2-grams 2nd level (3x3 classes): academic – melody NB jazz – melody NB popular – melody 4-grams Classifier ensembles

15. SSPR 200815 Conclusions ► Harmony and melody are suitable features for music genre classification ► Harmony is better for classifying broad musical domains, while melody is better for distinguishing sub-genres ► Misclassifications occur more frequently within broad domains ► Hierarchical classification and classifier ensembles outperformed the best single classifiers

16. Stochastic Text Models for Music Categorization Carlos Pérez-Sancho, José M. Iñesta, David Rizo Pattern Recognition and Artificial Intelligence group Department of Software and Computing Systems University of Alicante, Spain