GpGUGyFþ Ñ GFû öFõFß & d6ë8® ØFþ0Û...

4
- 1 - ᬯពᐇ⌧䝰䝕䝹䛾ᣑᙇ䛻ᇶ䛵䛟ᴦ᭤㛫㢮ఝᗘ䛾ホ౯ Evaluating Melodic Similarity based on an Extended Implication-Realization Model ▮⃝ḈᏊ *1 ୰㞞ᩄ *2 Ᏹὠ࿅Ṋோ *3 Sakurako YAZAWA *1 Masatoshi HAMANAKA 2 Takehito UTSURO *1 *1 ⟃ἼᏛᏛ㝔䝅䝇䝔䝮ሗᕤᏛ◊✲⛉▱⬟ᶵ⬟䝅䝇䝔䝮ᑓᨷ Graduate School of Systems and Information Engineering University of Tsukuba #1 *2 ி㒔ᏛᏛ㝔་Ꮫ◊✲⛉ *3 ⟃ἼᏛ䝅䝇䝔䝮ሗ⣔▱⬟ᶵ⬟ᕤᏛᇦ Department of Clinical System Onco-Informatics Kyoto University #2 Faculty of Engineering, Information and Systems University of Tsukuba #3 This paper proposes to measure similarity of melodies based on Implication-Realization Model (IRM), a music theory that abstracts music and then expresses music through symbol sequences based on information constituting the music such as pitch, rhythm, and rests and so on. This paper especially extends IRM so that the theory becomes much more appropriate to measuring similarity of melodies. More specifically, compared with the symbols of the original IRM, we introduce finer grained symbols by simply distinguishing up and down of interval directions and by dividing each most symbols of the original IRM into two extended symbols. Furthermore, we implement a parser which transforms tone sequence of an input melody into a sequence of the extended IRM symbols. The results of experimental evaluation through subjective human judgments show that the proposed extended IRM symbols outperform the original IRM symbols with respect to measuring similarity of melodies. 1. 䛿䛨䜑䛻 ᮏ✏䛷䛿㡢ᴦ⌮ㄽᬯពᐇ⌧䝰䝕䝹 [Narmour 1990] [Narmour 1992]䛻ᇶ䛵䛔䛯ᴦ᭤㛫㢮ఝᗘィ⟬䛻䛴䛔䛶㏙䜉 䜛䠊ᬯពᐇ⌧䝰䝕䝹䛷䛿ᴦ᭤䜢䝅䞁䝪䝹䛸䜀䜜䜛グ䛻ᢳ ㇟䛧⾲⌧䛩䜛䛣䛸䛜䛷䛝䜛䠊ᅇ䛿≉䛻䠈ᴦ᭤㛫㢮ఝᗘィ⟬ 䛻㐺䛧䛯ᬯពᐇ⌧䝰䝕䝹䛾ᣑᙇ䛻䛴䛔䛶㏙䜉䜛䠊 ᮏ✏䛷䛿䠈㡢ᴦ⌮ㄽ䛻ᇶ䛵䛔䛯ศᯒ⤖ᯝ䜢ᴦ᭤㛫㢮ఝᗘ ィ⟬䛻⏝䛔䜛᪉ἲ䜢ウ䛧䛯䠊ᴦ᭤㛫㢮ఝᗘィ⟬䛾◊✲䛿䛥 䛛䜣䛻⾜䜟䜜䛶䛚䜚, 䝴䞊䝄䞊䛻ዲ䜏䛾ᴦ᭤䜢㑅䜀䛫, 䛭䛾ᴦ ᭤ሗ䛛䜙䝴䞊䝄䞊䛾Ⴔዲ䜢⾲䛩䝧䜽䝖䝹䜢⏕ᡂ䛧ዲ䜏䛾ᴦ ᭤䜢㢮ఝᴦ᭤䛸䛧䛶㑅䜆ᡭἲ[Hoashi 2003]䜔ධຊ䛻㡢㡪≉ᚩ 㔞䜢⏝䛔䛯䝙䝳䞊䝷䝹䝛䝑䝖䜢ᵓ⠏䛩䜛䛣䛸䛷㢮ఝᴦ᭤䜢ᢳฟ 䛩䜛䝅䝇䝔䝮[Lampropoulos 2004], 㞳ᩓ䝣䞊䝸䜶ኚ䜢䛛䛡䛶 䝣䝺䞊䝈䛾䝟䝍䞊䞁䜢ぢ䛔䛰䛩ᡭἲ[Velardea 2013], ᴦ᭤䛻 䜎䜜䜛㡢➢䜢Ꮠ䛻ኚ䛧, 䛭䛾Ꮠ䛻ᑐ䛧䛶 N-gram 䜢㐺⏝䛩䜛䛣䛸䛷㢮ఝᴦ᭤䛾ᢳฟ䜢⾜䛖ヨ䜏[Doraisamy 2004] 䛺䛹ᵝ䚻䛺ᡭἲ䛜ᥦ䛥䜜䛶䛔䜛. 㡢ᴦ⌮ㄽ䜢⏝䛔䛯ᚑ᮶◊✲䛸䛧䛶䛿䠈㡢ᴦ⌮ㄽ GTTM[Lerdahl 1983]䛻ᇶ䛵䛔䛯ᴦ᭤㛫㢮ఝᗘィ⟬᪉ἲ䛜䛒 䛢䜙䜜䜛䠊GTTM 䛿䝯䝻䝕䜱䛻䜎䜜䜛㡢䛾㔜せᗘ䜢ィ⟬䛧䠈䝍 䜲䝮䝇䝟䞁䝒䝸䞊䛸䜀䜜䜛ᮌᵓ㐀䜢ᵓ⠏䛩䜛䛣䛸䛷䝯䝻䝕䜱䜢㝵 ᒙⓗ䛻ᤊ䛘ศᯒ,⌮ゎ䛩䜛㡢ᴦ⌮ㄽ䛷䛒䜛[Hamanaka 2007]GTTM 䛻ᇶ䛵䛔䛯ᴦ᭤㛫㢮ఝᗘィ⟬[Hirata 2013] GTTM 䛻䛚䛡䜛䝍䜲䝮䝇䝟䞁䝒䝸䞊䛜ᴦ᭤㛫䛷⮴䛧䛶䛔䜛㒊 ศ䛻ᑐ䛧䛶䠈䛭䛾䝍䜲䝮䝇䝟䞁䝒䝸䞊䛻䛔䛶䛔䜛㡢䛾㛗䛥䞉 䛾⨨䛻䛒䜛㡢䛾⥲䜢⏝䛔䛶㢮ఝᗘ䜢ィ⟬䛧䛶䛔䛯䠊䛣䛾ᡭ ἲ䛿䜋䜌䛨᭤䛾ሙ䛻䛾䜏㢮ఝィ⟬䛜䛷䛝䜛䜒䛾䛷䛒䜚䠈 䛨᭤䜢⣴䛩䜛ሙ䛻䛿䛸䛶䜒㐺䛧䛶䛔䜛䠊䛧䛛䛧ไ⣙䛜䛸䛶䜒 ཝ䛧䛔䛯䜑䠈ே㛫䛜⫈䛔䛯ሙ䛻䛂䛒䜛⛬ᗘఝ䛶䛔䜛䛃䛸ឤ䛨䜛 ᭤䛾⤌䛾㢮ఝᗘ䛜༑ศ㧗䛟䛺䜙䛺䛔䛸䛔䛖ၥ㢟䛜䛒䛳䛯䠊 ᮏ◊✲䛷䛿㡢ᴦ⌮ㄽᬯពᐇ⌧䝰䝕䝹䜢⏝䛔䛯ᴦ᭤㛫㢮ఝ ᗘィ⟬ἲ䜢ᥦ䛩䜛䠊ᬯពᐇ⌧䝰䝕䝹䜢⏝䛔䛯ඛ⾜◊✲䛸䛧 [Grachten 2005]䛿ᬯពᐇ⌧䝰䝕䝹䛷ᐃ⩏䛥䜜䛶䛔䜛ᵓ㐀䜢 ⏝䛔䜛ᡭἲ䜢ᥦ䛧䠈㢮ఝᗘ䜢ィ⟬䛷䛝䜛䛣䛸䜢♧䛧䛯䠊䛧䛛䛧䠈 ᬯពᐇ⌧䝰䝕䝹䛿ᮏ᮶䠈ᴦ᭤㛫㢮ఝᗘ䜢ィ⟬䛩䜛䛯䜑䛾㡢ᴦ ⌮ㄽ䛷䛿䛺䛔䛯䜑䠈䜸䝸䝆䝘䝹䛾ᵓ㐀䜢⏝䛔䛯䛰䛡䛷䛿䠈㢮ఝ ᗘィ⟬䛸䛔䛖ほⅬ䛷䛿⢭ᗘ䛜ୗ䛜䛳䛶䛧䜎䛖ၥ㢟䛜䛒䛳䛯䠊䛭䛣 䛷ᡃ䚻䛿䜸䝸䝆䝘䝹䛾ᬯពᐇ⌧䝰䝕䝹䛻ᑐ䛧䛶ᣑᙇ䜢⾜䛖䛣䛸 䛷ᴦ᭤㛫㢮ఝᗘィ⟬⢭ᗘ䛾䜢ᅗ䛳䛯䠊 ᡃ䚻䛿䜸䝸䝆䝘䝹ᬯពᐇ⌧䝰䝕䝹䠈䛚䜘䜃䠈ᣑᙇ䜢⾜䛳䛯 ᬯពᐇ⌧䝰䝕䝹䛾୧᪉䛷ᴦ᭤㛫㢮ఝᗘィ⟬䜢⾜䛳䛯䠊ィ⟬⤖ ᯝ䛻ᇶ䛵䛝㢮ఝᴦ᭤䜢䝕䞊䝍䝧䞊䝇䛛䜙㑅䜃ฟ䛧䠈⿕㦂⪅ ほᐇ㦂䜢⾜䛳䛯䠊䛭䛾⤖ᯝ䠈ᣑᙇ䜢⾜䛳䛯ᬯពᐇ⌧䝰䝕䝹䛷䛾 㢮ఝᗘィ⟬䛜䜸䝸䝆䝘䝹䛾ᬯពᐇ⌧䝰䝕䝹䜢ᅇ䛳䛯䠊 2. 䝯䝻䝕䜱䜈䛾ᬯពᐇ⌧䝰䝕䝹䛾䝅䞁䝪䝹 2.1 ᬯពᐇ⌧䝰䝕䝹 ᴦ᭤㛫㢮ఝᗘᐃ䝅䝇䝔䝮䜢ᵓ⠏䛩䜛䛻䛒䛯䜚䠈ᬯពᐇ⌧ 䝰䝕䝹䛸䛔䛖㡢ᴦ⌮ㄽ䛻╔┠䛧䛯䠊ᬯពᐇ⌧䝰䝕䝹䛸䛿㡢ᴦᏛ ⪅䛷䛒䜛 Eugene Narmour 䛻䜘䛳䛶ᥦၐ䛥䜜䛯⌮ㄽ䛷䛒䜛䠊䛣䛾 ⌮ㄽ㡢ኈ䛾㛵ಀ䜢䝛䝑䝖䝽䞊䜽䛾䜘䛖䛻ᤊ䛘䛶䝯䝻䝕䜱䜢グ㏙ 䛧䛶䛚䜚, 䛣䛾䝛䝑䝖䝽䞊䜽䜢ゎᯒ䞉ᢳฟ䛩䜛䛣䛸䛻䜘䜚䠈䝯䝻䝕䜱䛜 䛹䛾䜘䛖䛺ᵓ㐀䛻䛺䛳䛶䛔䜛䛛䜢ศᯒ䛩䜛䛣䛸䛜䛷䛝䜛䠊䜎䛯ᐇ 㝿䛾ゎᯒ䛿ᴦ᭤䜢ᵓᡂ䛩䜛㡢㧗䠈㡢⛬䠈䝸䝈䝮䜔ఇ➢➼䛾ሗ 䜢⏝䛔䛶ᴦ᭤䜢䝅䞁䝪䝹䛸䜀䜜䜛グ䜢⏝䛔䛶グ䜈䛸ᢳ ㇟䛧䛶⾲⌧䛩䜛䛣䛸䛜䛷䛝䜛 㐃⤡ඛ䠖▮⃝ḈᏊ䠈⟃ἼᏛᏛ㝔䝅䝇䝔䝮ሗᕤᏛ◊✲⛉ ༤ኈᚋᮇㄢ⛬䠈 㼟㼍㼗㼡㼞㼍㼗㼛㻬㼙㼡㼟㼕㼏㻚㼕㼕㼠㻚㼠㼟㼡㼗㼡㼎㼍㻚㼍㼏㻚㼖㼜 The 29th Annual Conference of the Japanese Society for Artificial Intelligence, 2015 2C5-OS-21b-6in

Transcript of GpGUGyFþ Ñ GFû öFõFß & d6ë8® ØFþ0Û...

Page 1: GpGUGyFþ Ñ GFû öFõFß & d6ë8® ØFþ0Û o - kaigi.org · PDF file- 1 - / #'GpGUGyFþ Ñ GFû öFõFß & d6ë8® ØFþ0Û o Evaluating Melodic Similarity based on an Extended

- 1 -

Evaluating Melodic Similarity

based on an Extended Implication-Realization Model

*1 *2 *3 Sakurako YAZAWA*1 Masatoshi HAMANAKA2 Takehito UTSURO*1

*1

Graduate School of Systems and Information Engineering University of Tsukuba #1

*2 *3 Department of Clinical System Onco-Informatics Kyoto University #2 Faculty of Engineering, Information and Systems University of Tsukuba #3

This paper proposes to measure similarity of melodies based on Implication-Realization Model (IRM), a music theory that abstracts music and then expresses music through symbol sequences based on information constituting the music such as pitch, rhythm, and rests and so on. This paper especially extends IRM so that the theory becomes much more appropriate to measuring similarity of melodies. More specifically, compared with the symbols of the original IRM, we introduce finer grained symbols by simply distinguishing up and down of interval directions and by dividing each most symbols of the original IRM into two extended symbols. Furthermore, we implement a parser which transforms tone sequence of an input melody into a sequence of the extended IRM symbols. The results of experimental evaluation through subjective human judgments show that the proposed extended IRM symbols outperform the original IRM symbols with respect to measuring similarity of melodies.

1. [Narmour 1990]

[Narmour 1992]

, ,

[Hoashi 2003]

[Lampropoulos 2004], [Velardea 2013],

, N-gram[Doraisamy 2004]

.

GTTM[Lerdahl 1983]GTTM

, [Hamanaka 2007] GTTM [Hirata 2013]

GTTM

[Grachten 2005]

2.

2.1

Eugene Narmour

,

The 29th Annual Conference of the Japanese Society for Artificial Intelligence, 2015

2C5-OS-21b-6in

Page 2: GpGUGyFþ Ñ GFû öFõFß & d6ë8® ØFþ0Û o - kaigi.org · PDF file- 1 - / #'GpGUGyFþ Ñ GFû öFõFß & d6ë8® ØFþ0Û o Evaluating Melodic Similarity based on an Extended

- 2 -

2.2

1

2.3

midi

1)k=1, 2)k=2, 3)k=2n+1 (1 n), 4)k=2n+2 (1 n), k

1) 2) 1 23)

4) 2

1, 18

The 29th Annual Conference of the Japanese Society for Artificial Intelligence, 2015

Page 3: GpGUGyFþ Ñ GFû öFõFß & d6ë8® ØFþ0Û o - kaigi.org · PDF file- 1 - / #'GpGUGyFþ Ñ GFû öFõFß & d6ë8® ØFþ0Û o Evaluating Melodic Similarity based on an Extended

- 3 -

2, IRM

3.

Seq1Seq2 2

python n-gram

4.

4.1 Essen

(http://www.esac-data.org/) 5000

50000 1.00 = 0,

0.05i< 0.05(i+1) (i=0,…,18), 0.95+0.01i< 0.95+0.01(i+1) (i=0,…,4) 25

2 125

15

5 4 32 1

4.2 3,4,5

1 Seq1 Seq2

1,

The 29th Annual Conference of the Japanese Society for Artificial Intelligence, 2015

Page 4: GpGUGyFþ Ñ GFû öFõFß & d6ë8® ØFþ0Û o - kaigi.org · PDF file- 1 - / #'GpGUGyFþ Ñ GFû öFõFß & d6ë8® ØFþ0Û o Evaluating Melodic Similarity based on an Extended

- 4 -

(a)

(b) 3, 0.95< 1

(a)

(b) 4. (0.65 < 0.95)

5.

[Narmour 1990] Eugene Narmour, “The Analysis and Cognition

of Basic Melodic Structures”, The university of Chicago press

[Narmour 1992] Eugene Narmour, “The Analysis and Cognition of Melodic Complexity”, The university of Chicago press

[Hoashi 2003] Hoashi K, Matsumoto K, Inoue N., “Personalization of User Content-based Music Retrieval on Relevance Feed Back” Proceedings of ACM Multimedia , pp110-119.

[Lampropoulos 2004] Lampropoulos A S, Sotiropoulos D N, Tsihrintzis G A, “Individualization of Music Similarity Perception via Feature Subset Selection”, IEEE, International Conference on System, Man and Cybernetics 2004

[Velardea 2013] Gissel Velardea, Tillman Weydeb, David Mereditha, "An approach to melodic segmentation and classification based on filtering with the Haar-wavelet " Journal of New Music Research Volume 42, Issue 4.

[Doraisamy 2004] Shyamala Doraisamy, Stefan Ruger “A Polyphonic Music Retrieval System Using N-Grams”, Proc. of ISMIR 2004

[Lerdahl 1983] Fred Lerdahl, "A Generative Theory of Tonal Music", The MIT Press

[Hamanaka 2007] Msatoshi Hamanaka, Keiji Hirata, Satoshi Tojo : ”FATTA: FULL AUTOMATIC TIME-SPAN TREE ANALYZER”, Proceedings of the 2007 International Computer Music conference

[Hirata 2013] Keiji Hirata, Satoshi Tojo, Masatoshi Hamanaka, ”Cognitive Similarity grounded by tree distance from the analysis of K. 265/300e “, Proceedings of CMMR 2013

[Grachten 2005] Maarten Grachten, Josep Lluis Arcos and Ramon Lopez de Mantaras, “Melody Retrieval using the Implication/Realization Model”, MIREX 2012 Symbolic Melodic Similarity Results

The 29th Annual Conference of the Japanese Society for Artificial Intelligence, 2015