[論文紹介]Monoaural audio source spearation using deep convolutional neural networks

LVA/ICA 2017で発表された深層学習を用いた単一チャネル信号からの音源分離手法の論文紹介です。書誌情報は以下の通りです。
Chandna P., Miron M., Janer J., Gómez E. (2017) Monoaural Audio Source Separation Using Deep Convolutional Neural Networks. In: Tichavský P., Babaie-Zadeh M., Michel O., Thirion-Moreau N. (eds) Latent Variable Analysis and Signal Separation. LVA/ICA 2017. Lecture Notes in Computer Science, vol 10169. Springer, Cham
リンク：https://link.springer.com/chapter/10.1007/978-3-319-53547-0_25

Transcript

1. ࿦จ঺հ
   .POPBVSBM
   BVEJP
   TPVSDF
   TQFBSBUJPO
   VTJOH
   EFFQ
   DPOWPMVUJPOBM
   OFVSBM
   OFUXPSLT ຊ઒఩࠸
   1
2. ΞδΣϯμ w എܠ
   w Իݯ෼཭ͱ͸
   w Իݯ෼཭ख๏
   w ख๏
   w σʔλϑϩʔ
   w ϞσϧΞʔΩςΫνϟ
   w ࣮ݧ
   w ݁Ռ 2
3. എܠ 3
4. Իݯ෼཭ͱ͸ w ෳ਺ͷԻݯ͕ೖ͍ͬͯΔσʔλ͔Βཉ͍͠ԻݯͷΈΛ෼཭ɾநग़
   w Ի੠ೝࣝɺԕִձٞɺิௌثɺFUDʜ
   w ԻڹɾԻ੠৴߸ॲཧͷίΞٕज़ 4
5. ؍ଌ৚݅ w ຊ࿦จͷ૝ఆ͸୯Ұνϟωϧ৴߸ʢϞϊϥϧ৴߸ʣ
   w γϯάϧϚΠΫʹΑΔ࿥Ի
   w Իݯ෼཭ʹ͸࠷΋ࠔ೉͕ͩ࠷΋؆୯ͳ࿥Իํ๏ 5
6. Իݱ෼཭ͷैདྷख๏ w ࠓ·Ͱ͸/.'ͷΑ͏ͳߦྻ෼ղ͕ओྲྀ
   w 9
   
   "4
   
   /ͱ͍͏ઢܗϞσϧʢੵ࿨Ϟσϧʣ
   w ͜ͷΑ͏ͳ෼ղͷ࢓ํ͸ແݶʹ͋Δ
   
   ຊ࣭తͳ೉͠͞
   w "
   4
   /͕௚઀Θ͔Βͳ͍ͷͰ͍͔ʹϞσϦϯά͢Δ͔͕ॏཁ 6 IUUQTMPHNJKQUFDIBSUJDMFT
7. ਂ૚ֶशͷ୆಄ w ਂ૚ֶशͷద༻ʹΑͬͯඇઢܗؔ਺ͰϞσϧԽ͕ՄೳʢΑΓॊೈͳදݱೳྗʣ
   w 9
   
   G 4
   w ͔͠͠શ݁߹ωοτϫʔΫʢ.-1ʣ͸ܭࢉྔେ
   w ຊ࿦จͷ໨త͸ύϥϝʔλ਺ͷগͳ͍$//Λར༻͢Δ͜ͱͰϦΞϧλΠϜॲཧ ʹ΋௕͚ͨ//Ϟσϧͷߏங 7
8. w .-1ʹΑΔόΠφϦιʔε෼཭ʢ+34JNQTPO ʣ
   w (.. %//ͷ૊Έ߹Θͤख๏ʢ/VHSBIB ʣ
   w ωοτϫʔΫ͸.-1
   w &.ΞϧΰϦζϜͰ࠷దԽ ؔ࿈ݚڀ 8 IUUQTBSYJWPSHQEGQEG
9. ख๏ 9
10. σʔλϑϩʔ w //Λ༻͍ͨԻݱ෼཭ϑϨʔϜϫʔΫ
    w //ʹΑΔग़ྗͷ߹ܭ͕ݩͷࠞ߹Իݯͱ߹க͢ΔΑ͏ʹ5JNF'SFRVSODZ
    .BTLJOHΛར༻ʢ)VBOH ʣ 10
11. ϞσϧΞʔΩςΫνϟ w ΤϯίʔμσίʔμϞσϧ
    w 7FSUJDBM
    BOE
    )PSJ[POUBM
    $POWPMVUJPO
    -BZFS
    
    '$/ͷ૚ 11
12. w 7FSUJDBM
    $POWPMVUJPOBM
    -BZFS
    w ॎํ޲$POWPMVUJPO
    w ہॴతͳԻ৭ಛ௃ͷֶश͕Մೳ
    w /.'ͱ͸ҟͳΓ෼཭ιʔεڞ༗
    w )PSJ[POUBM
    $POWPMVUJPOBM
    -BZFS
    w ԣํ޲$POWPMVUJPO
    w 7FSUJDBM
    $POWͱ͸ҟͳΓ࣌ؒతมԽͷϞσϧԽ
    w ෼཭ιʔεຖʹଘࡏ͢Δ༷ʑͳ࣌ؒप೾਺ಛੑΛϞσϦϯάՄೳ ͭͷ$POWPMVUJPOBM
    -BZFS 12
13. 5JNFGSFRVFODZ
    TPGU
    NBTLJOH w //ग़ྗͷ༧ଌ݁ՌΛݩͷࠞ߹ιʔεʹ౳͘͢͠ΔΑ͏ͳ੍໿৚݅
    w
    w NBTLJOH
    MBZFS͸ܾఆ࿦తͳؔ਺ͳͷͰֶश͞ΕΔύϥϝʔλ͸ଘࡏ͠ͳ͍͕ɼ ωοτϫʔΫͷॏΈࣗମ͸ Λ༻͍ͯߋ৽͢Δ ∑ ˜ yn (f ) = ∑ mn (f )x(f ) = 1 × x(f ) = x(f ) ˜ yn (f ) 13
14. ύϥϝʔλͷֶश w "EBEFMUBʢ.%
    ;FJMFS
    ʣΛ࢖༻
    w 4(%ʢ//ֶशͷͨΊͷඪ४తͳΞϧΰϦζϜʣͷൃల൛
    w ֶश཰ͷࣗಈௐઅ͕Մೳ
    w ऩଋ͕ૣ͍ 14
15. ࣮ݧ 15
16. σʔληοτ w %FNJYJOH
    4FDSFUT
    %BUBTFU
    
    %4%
    w ۂͷϑϧτϥοΫσʔληοτ
    w υϥϜɾϕʔεɾϘʔΧϧɾͦͷଞͷτϥοΫͱ͍͏෼͚ํ
    w ֶशɾςετͦΕͧΕۂͣͭͰߏ੒ 16
17. ֶशςετσʔλͷՃ޻ w ݸʑͷιʔεԻݯͱͦΕΒ͕ࠞͬͨ͟ೖྗԻݯΛඵຖͷηάϝϯτʹ෼཭
    w ͜ΕΒʹରͯ͠୹࣌ؒϑʔϦΤม׵Λࢪ͢ʢ/ͷϋχϯά૭Λ࢖༻ʣ
    w
    ɿϋχϯά૭ؔ਺
    w ϑϨʔϜγϑτྔΛʢ
    Φʔόʔϥοϓʣʹ͢Δ͜ͱͰ͞Βʹ܇࿅σʔλ Λ૿΍͍ͯ͠Δʢ৑௕ੑ΋૿͢ʁʣ w(n) = 0.5 − 0.5 cos 2πn N − 1 (n = 0,...,N − 1) 17
18. ଛࣦؔ਺ͷΞδϟετ w ଛࣦؔ਺ʹ͸৐ଛࣦؔ਺Λ࢖༻ɿ
    w ϕʔεɾυϥϜɾϘʔΧϧ͸ۂશମʹҰ؏ͯ͠ଘࡏ͢Δ͕ɼͦͷଞָث͸େ͖͘มԽ͢Δ
    w ৐ଛࣦ//͸ฏۉతͳԻ৭ΫϥεΛֶश͠Α͏ͱ͢ΔͨΊɼָث͝ͱͷಛੑΛֶश͠ʹ͘͘ͳͬͯ͠·͏
    w ʮϘʔΧϧͱϘʔΧϧҎ֎ʯʮϕʔεͱϕʔεҎ֎ʯʮυϥϜͱυϥϜҎ֎ʯͷࠩΛද߲͢Λ௥Ճ
    w ʮͦͷଞʯ͸Ϊλʔ΍γϯηͷΑ͏ͳϋʔϞχοΫָثͰϘʔΧϧͱιʔε͕ࣅ͍ͯΔ͜ͱΛར༻
    w ҎԼ ͕ϘʔΧϧɼ ͕ͦͷଞʹରԠ Loss = ∑ ∥ ˜ yn − yn ∥2 ˜ y1 ˜ yN 18
19. ଛࣦؔ਺ͷΞδϟετ w ଛࣦؔ਺ʹ͸৐ଛࣦؔ਺Λ࢖༻ɿ
    w ϕʔεɾυϥϜɾϘʔΧϧ͸ۂશମʹҰ؏ͯ͠ଘࡏ͢Δ͕ɼͦͷଞָث͸େ͖͘มԽ͢Δ
    w ৐ଛࣦ//͸ฏۉతͳԻ৭ΫϥεΛֶश͠Α͏ͱ͢ΔͨΊɼָث͝ͱͷಛੑΛֶश͠ʹ͘͘ͳͬͯ͠·͏
    w ʮϘʔΧϧͱϘʔΧϧҎ֎ʯʮϕʔεͱϕʔεҎ֎ʯʮυϥϜͱυϥϜҎ֎ʯͷࠩΛද߲͢Λ௥Ճ
    w ʮͦͷଞʯ͸Ϊλʔ΍γϯηͷΑ͏ͳϋʔϞχοΫָثͰϘʔΧϧͱιʔε͕ࣅ͍ͯΔ͜ͱΛར༻
    w ҎԼ ͕ϘʔΧϧɼ ͕ͦͷଞʹରԠ Loss = ∑ ∥ ˜ yn − yn ∥2 ˜ y1 ˜ yN 19
20. ݁Ռ 20
21. ࣮ݧ݁Ռ w ධՁࢦඪɿ4%3ɾ4*3ɾ4"3ɾ*43ʢ7JODFOU ʣ
    w ஋͕େ͖͍΄Ͳੑೳྑ͍
    w .-1ͱ΄΅Ḯ৭ͳ͍݁ՌΛগͳ͍ॲཧ࣌ؒͰୡ੒ʢύϥϝʔλ࡟ݮͷޮՌʣ 21
22. σϞ w IUUQTXXXZPVUVCFDPNXBUDI W8X)Z/B%G& 22