IMPROVEMENTS OF SYMBOLIC KEY FINDING METHODSSkip other details (including permanent urls, DOI, citation information)
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 3.0 License. Please contact firstname.lastname@example.org to use this work in a way not covered by the license. :
For more information, read Michigan Publishing's access and usage policy.
Page 1 ï~~IMPROVEMENTS OF SYMBOLIC Matthias Robine, Thomas Rod SIMBALS Pr( LaBRI niversity oi F-33405 Talence cec firstname.name( ABSTRACT Automatically estimating the key of a musical piece is an important part of a lot of musical applications such as music classification or music transcription. Existing methods rely on the comparison of pitch class profiles. Correlation computed between the input pitch profile and a key reference profile indicates the key of the musical piece tested. Other recent methods propose to consider the note inter1 1_ 1 1 1 T 1" per pro tai wi nal -A 4
Page 2 ï~~Based on Li and Huron observations , Madsen and Widmer recently adopted a different approach . Their method consists of using the information induced by the temporal order of notes. Indeed, two equal pitch class distributions may have different note transitions which may pri be] ca ref the imply different keys. Instead of simply taking the pitch occurrences into account, Madsen and Widmer thus con sider interval occurrences and map them in a 12 12 ma trix, giving a likelihood to each possible pitch class tran 3.2 sition. As in the pitch profile method, an input interval profile is compared to 24 key reference interval profiles, and the highest correlation induces the preferred key. The use of interval turned out to be very comparable (even slightly better in some cases) to pitch profiles. But since the two methods are successful on different pieces of music, it seems that they bring out different kinds of tonal information. It could be thus interesting to combine these two methods, in order to gather the more possible information. prc the Th prc fro SOl ter PROCEDURE FOR EXPERIMENTS 11
Page 3 ï~~Method Data Correct Key Errors MS use rel. nel. oth. pitch interval hybrid pitch interval hybrid 67.4 83.2 83.2 14.7 4.2 4.2 17.9 11.6 12.6 0.0 1.1 0.0 80.8 90.2 90.8 FF FF FF 71.0 72.4 75.7 7.0 7.0 4.7 16.5 15.6 15.7 5.5 5.0 3.9 81.4 82.3 85.0 w anl Table Results MIREX (M) FinFolk (FF) database. The results are obtained by pitch or interval correlation method, or by the sum of these correlations (hybrid method). They are given in % of the number of files Th its of me imj or with the MIREX Score (MS, see Section 3). We can see that the hybrid method gives the best results in terms of correct key detection, of errors induced, and also for the 5.2 MS value. On by de( COMPLEMENTARITY OF METHODS
Page 4 ï~~Data Corr. Correct Key Errors MS I rel. nel. oth. M no 83.2 4.2 12.6 0.0 90.8 M yes 89.5 4.2 6.3 0.0 93.9 FF no 75.7 4.7 15.7 3.9 85.0 FF yes 76.8 4.7 14.5 3.9 85.6 [q Table 2. Corrections (Corr.) of the retrieved key from the MIREX (M) or the FinFolk (F) database. The results are given in % of the number of files. They show that applying some flat profiles and other improvements (see Section 5) on the first retrieved key can correct some errors. E[A [(' CONCLUSION AND PERSPECTIVES In this paper we show that combining some different approaches for key finding can be more accurate than using each method separately. It is due to the complementary of the different methods used. For now, the combination adds the value of the scalar product for the pitch corre ['I I