Performance Rendering for Piano Music with a Combination of Probabilistic Models for Melody and Chords

Transcription

1 IPSJ SIG Technical Repor Performance Rendering for Piano Music wih a Combinaion of Probabilisic Models for Melody and Chords Tae Hun Kim, 1 Saoru Fukayama, 1 Takuya Nishimoo 1 and Shigeki Sagayama 1 In his paper, we presen a mehod o generae human-like performance expression for polyphonic piano music wih a combinaion of probabilisic models for melody and chords o avoid daa sparseness problems. Probabilisic models and machine learning have been applied o solve he problem of generaing human-like expressive performance given a music score. In case of polyphonic music, however, i was difficul o make a racable model and a huge amoun of raining daa was necessary. The resuls of he experimens show ha he proposed mehod is able o generae flucuaions of performance parameers for polyphonic piano music such like human performers do. The resuls of subjecive evaluaions are also repored which indicae ha he generaed performance expression sounded human-like and have cerain degree of musicaliy J. Sundberg 1) S. Flossmann (performance conex) 1 Graduae School of Informaion Science and Technology, he Universiy of Tokyo 1 c 2010 Informaion Processing Sociey of Japan

2 IPSJ SIG Technical Repor 3) 4) G. Grindlay HMM(Hidden Markov Model) 5) Ingrid Haebler, CresMuse PEDB a) (1) b) (2) c) (3) Ingrid Haebler, CresMuse PEDB a) (4) b) (5) c) (6) Noe Index 1 2) 2 1 K. Teramura 1 N. Magaloff, R. Baik 2 6) 1 1 performance conex ( 1 ) ( 2 ) ( 3 ) ( 4 ) 2 c 2010 Informaion Processing Sociey of Japan

3 IPSJ SIG Technical Repor ( 1 ) ( 2 ) ( 3 ) ( 4 ) Tempo TempoFacor = log( Tempo scope ) (1) avg Tempo scope avg n n 3, n 2, n 1, n, n +1 Loudness melody Velociy = log( Velociy scope ) (2) avg Velociy scope avg n n 3, n 2, n 1, n, n +1 DuraionFacor melody Duraion score = log( Duraionreal Duraion score Duraion real ) (3) 1 - Pich - - Duraion score Duraion score NoeInerval I,II,III,IV NoeInerval I,II,III,IV NoeInerval I,II,III,IV DuraionRaio I,II,III,IV DuraionRaio I,II,III,IV DuraionRaio I,II,III,IV Meric I,II Meric I,II Meric I,II AricuaionMarks AriculaionMarks AriculaionMarks Pich Duraion score NoeInerval I,II,III,IV n (n 3, n 2), (n 2, n 1), (n 1, n ), (n, n +1) DuraionRaio I,II,III,IV Meric {very srong, srong, weak} Meric I,II n 1 n Meric AriculaionMarks saccao, accen, fermaa DiffOnseTime i = OnseTime o OnseTime i (4) OnseTime o OnseTime i Loudness chord i = log( Velociy i Velociy o ) (5) Velociy o Velociy i 3 c 2010 Informaion Processing Sociey of Japan

4 IPSJ SIG Technical Repor Pich Pich Pich Duraion score Duraion score Duraion score NoeDisance NoeDisance NoeDisance OuerNoe OuerNoe OuerNoe Piano Sonaa KV331, 1s Mov. Piano Sonaa KV331, 1s Mov. Piano Sonaa KV331, 1s Mov. Piano Sonaa KV331, 1s Mov. Hiroko Nakamura Norio Shimizu Ingrid Haebler Lily Kraus DuraionFacor chord i Duraion real o = log( Duraionreal i Duraion real o Duraion real i ) (6) 2 Pich Duraion score NoeDisance OuerNoe rue false 3.3 Condiional Random Fields(CRFs) 7) Sochasic Gradien Descen 8) Vierbi León Boou crfsgd CRFs k-means CresMusePEDB ver W. A. Mozar, Piano Sonaa, KV331, 1s Mov. 4 ( 3) 2 2 Vladimir Ashkenazy F. Chopin 14 ( 4) F. Chopin, Nocurne No. 10, Op. 32, 2nd Mov hp://leon.boou.org/projecs/sgd 2 hp:// 4 c 2010 Informaion Processing Sociey of Japan

5 IPSJ SIG Technical Repor Prelude Op. 28 No. 1, 4, 7, 15, 20 Eude Op.10-3, 10-4, Walz Op. 18, 34-2, 64-2, 69-1, 69-2 Nocurne No. 2 Op a) (4) b) (5) c) (6) Noe Index a) (1) b) (2) c) (3) W. A. Mozar Piano Sonaa KV 331-1(SNT331-1) F. Chopin Nocurne No. 10 Op. 32-2(NCT010) W. A. Mozar, Piano Sonaa KV545, 3rd Mov.SNT (deadpan) 25 1 SNT331-1 NCT SNT545-3 W. A. Mozar, Piano Sonaa SNT331-1 NCT010 SNT M.J. Pires W. A. Mozar Piano Sonaa, KV279-1, 279-2, 279-3, 331-1, 545-1, KV c 2010 Informaion Processing Sociey of Japan

6 IPSJ SIG Technical Repor 6 A B C D a) W. A. Mozar, Piano Sonaa, KV331, 1s Mov.b) W. A. Mozar, Piano Sonaa KV545, 3rd Mov.c) F. Chopin, Nocurne No. 10, Op. 32, 2nd Mov. d) 5 2 a) (1) b) (2) c) (5) d) (6) 5. 1) Sundberg, J. and e al.: Musical performance: A synhesis-by-rule approach, Compuer Music Journal, Vol.7, No.1, pp (1983). 2) Flossmann, S. and e al.: Expressive Performance Rendering: Inroducing Performance Conex, Proceedings of he 6h Sound Music and Compuing Conference(SMC), pp (2009). 3) Teramura, K. and e al.: Gaussian process regression for rendering music performance, Proceedings of he 10h Inernaional Conference on Music Percepion and Cogniion(ICMPC) (2008). 4) Vol.84, No.12 (2010). 5) Grindlay, G. and e al.: Modeling, analyzing, and synhesizing expressive piano performance wih graphical models, Machine Learning, Vol.65, pp (2006). 6) Parncu, R.: Accens and expression in piano performance, Perspekiven und Mehoden einer Sysemischen Musikwissenschaf (Niemöller, K. W., ed.), Peer Lang, Frankfur am Main, pp (2003). 7) Laffery, J. and e al.: Condiional random fields: probabilisic models for segmening and labeling sequence daa, Inernaional Conference on Machine Learning, pp (2001). 8) Boou, L.: Sochasic Gradien Learning in Neural Neworks, Proceedings of Neuro-Nîmes 91, Nimes, France, EC2 (1991). 6 c 2010 Informaion Processing Sociey of Japan