1. Tree structured and combined methods for comparing metered polyphonic music Kjell Lëmstrom David Rizo Valero José Manuel Iñesta CMMR’08 May 21, 2008

2. 2 Outline Objectives State of the art Tree representation of monodies and polyphonic songs Comparison of trees for obtaining similarities between songs Geometric methods Combination of methods Experiments Conclusions and work lines

3. 3 Melodic comparison (symbolic) Given the sequence of notes at the scores … Are those tunes the same?

4. 4 Target Polyphonic music comparison of whole songs

5. 5 Approaches to polyphonic comparison Convert into monophonic –Use sequence comparison Adapted text retrieval methods –PROMS: Clausen et al ‘00 –Doraisamy and Rüger ‘04: n-grams Geometric methods –Lubiw and Tanur ‘04 –Ukkonen, Lemström and Mäkinen ‘03 + CMMR’08 Session: MUSR: Music Retrieval papers

6. 6 Tree construction process (Rizo et al. ’03)  Based on the logarithmic nature of music notation  Each tree level is a subdivision of the upper level whole4 beats half2+2 quarter 4×1 8×½8×½eighth  Leaf labels can be any pitch magnitude  Rests are coded the same way as notes  Duration is implicitly coded in the tree structure......... F C EG 1 4/4 bar Initial time Duration Tree representation for monodies

7. 7  The complete melody (all bars) is a forest (all trees)  Bars can be grouped sequentially or hierarchically F C E G Representation of whole melodies A B C G Sequential grouping: CEGFABCG Tree representation

8. 8 Polyphonic tree representation Process repeated for each voice: replace single labels for sets {C,G} {C} {F}{F} C F CG G E {G} {C,G,E}{C,G,E} {C,F,G}{C,F,G} {C,E,F,G} Actually, the interval from the tonic is represented in the tree Using tree tonality guessing (rizo et al.’06) {0,7} {0} {5} {0,5,7} {0,4,5,7} Propagate from bottom using set union

9. 9 Polyphonic tree representation Better tree summarization: Use duration importance: rhythmic weights  Multiset Rhythmic weight = 2 h-l h = tree height l = node level {C=2,E=2,G=2} {C=1} {F=1} {C=1,F=1,G=2} {C=3,E=2,F=1,G=4} It has been tested to use the Krumhansl-Schmuckler profiles along with the rhythmic weights: worse results l = 1 l = 2 l = 3

10. 10 Comparing songs Compare songs = compare trees Approaches –Classical tree edit distances Shasha Selkow –Use only the information of the roots Sequence edit distance Longest Common Subsequence

11. 11 Tree comparison Use only information in the roots –Roots contain the summary of its children after propagation { C=0.3, E=0.1, G=1 }..... Bar 1Bar 2Bar 3Bar 4Bar N RootED and LCRS: -Let  be a tree level ot tree T, compose a sequence S  (T) with all nodes at that level in the forest -RootED and LCRS use  =1 -Distance between 2 songs A and B at a level  d(A,B,  a,  b )= stringDistance(S  a (A), S  b (B)) or d(A,B,  a,  b )= LCS(S  a (A), S  b (B)) Complexity with  = 1 O(|bars A | * |bars B |) SaSa { C=0.6, F=0.2 } {F=1, G=1,A=1, B=0.2} {C=0.3, E=0.2, G=0.5} { C=0.6, F=0.2 } Labels of the root of each tree

12. 12 Multiset substitution cost Define multiset as a vector: Index = interval from tonic Value = cardinality –E.g: {C=1, G=4, B=2} is defined as –[1,0,0,0,0,0,0,4,0,0,0,2] Multiset substitution cost between multisets X and Y represented by vectors v and w

13. 13 Graphical representations P1, P2, P3 algorithms from Ukkonen, Lemström, Makinen ‘03 P2v5, P2v6: indexed versions of P2 –Not published yet

14. 14 Method combination Dissimilarity measure for a method = distance between songs Combined dissimilarity measure = combination of distances between songs Combination = sum of normalized distances

15. 15 Experiments Corpora: –ICPS (68 files): 7 different polyphonic incipits: Schubert’s Ave Maria, Ravel’s Bolero, Alouette, Happy Birthday, Frère Jacques, Jingle Bells, When The Saints Go Marching In Covers made up of polyphonic piano files + “Band in a box” variations –VAR (78 files): Bach Goldberg variations Bach english suites variations Some Tchaikowsky variations

16. 16 Evaluation method Leave one out –All-against-all: each song S is compared with the rest of the songs, the result is an ordered list with the most similar songs first Accuracy –Top-recognition-rate (T RR n ): presence percentage of the a version of the song S among the top n slots Success rate = T RR 0 –Precision-at-|class| |class| = number of versions of the same song Times –Exclude preprocessing times: only performed once at startup of system Averages: all results are averages of all queries

17. 17 Results: ICPS Time and success rate Combined method: success rate

18. 18 Results: VAR Cuccess rate Combined method: again success rate

19. 19 Top-recognition-rate: ICPS Combined method gets a good result

20. 20 Top-recognition-rate: VAR Combined method is the best one: combined methods are more robust

21. 21 Conclusions and work lines Very hard task when MIDI files are real ones –Preprocess songs: Use automatic tonal analysis + tree propagation to remove non-important notes in songs Improve results by combining more different classifiers Tune the tree comparison measures: submitted Add LCS fast implementation from Hyyrö ‘04 Add confidence values to LCS Include meter extraction methods to build the trees Query MIDI

22. 22 END

23. 23 Melody = sequence of notes String representation + string distances –(Mongeau and Sankoff ‘90, Lemström 2000) GGAGCBGGGAGAGGCBB Symbols are combinations of pitch x rhythm Pitch can be: absolute pitch, pitch class, interval from tonic, interval, contour, high-def contour, nothing Rhythm can be: absolute, inter-onset interval, inter-onset ratio, contour, nothing e.g.: (G4,8)(G4,8)(A4,4)(G4,4)(C5,4)(B4,2)(G4,8)(G4,8) Best comparison results using intervals with no rhythm information

24. 24 String distances Drawbacks on the comparison without rhythm –Wrong results with: Same melodic distance and different rhythm:  edit distance Godfather theme Hungarian dance, Schubert  ≈ Too many ornament notes:  edit distance

25. 25  Propagation and prunning Tree construction process Max. prunning level defined s F F A G Rules (Rizo et al., 2003) Tree as initially coded from the score Tree representation

26. 26 The distance is computed as the cost of the operations to transform one tree into the other. TREE EDIT DISTANCE TREE EDIT DISTANCE (Zhang & Shasha, 1989) C G C A A C C C G C A C C A A t1t1 t2t2 d(t 1,t 2 ) Weighted operations of insertion deletion replacement Melodic similarity metrics Tree edit distance  O( |T 1 |  |T 2 |  h(T 1 )  h(T 2 ) ) Previous prunning process helps to overcome this complexity (Zhang & Shasha, “Simple fast algorithms for the editing distance between trees...”. SIAM J Comput., 8(6): 1245-1262. 1989) Tree representation