GMI44@Music and Language研究のサーベイ報告

2024年3月11日に行われた第44回GMIワークショップでの発表です．

Transcript

1. Music and Languageͷಈ޲ʹؔ͢ΔαʔϕΠ ࢁຊ ༤໵ʢஜ೾େֶʣ 2024/03/11 GMIϫʔΫγϣοϓ
2. ࣗݾ঺հ • ࢁຊ ༤໵ʢ΍·΋ͱ Ώ͏΍ʣ • ஜ೾େֶେֶӃ ࣉᖒݚڀࣨ ത࢜3೥ʢࠓ೥౓मྃ༧ఆʣ • ՎএςΫχοΫͷ෼ੳ͓Αͼݕग़ʹ͍ͭͯݚڀ • ৘ใΛ·ͱΊΔͷ͕޷͖ • https://github.com/yamathcy/awesome-music-informatics • https://github.com/yamathcy/ISMIR-2023-Papers • ౳… • ๺ݪઌੜʹ͓༠͍͍͖ͨͩɼܰ཰ʹൃදΛਃࠐ͠·ͨ͠ 2
3. ࠓ೔ͷ಺༰ • ࣗવݴޠॲཧ x Իָ৘ใॲཧͳݚڀʹ͍ͭͯͷαʔϕΠใࠂ • Իָ -> ςΩετ (Captioning, Question & Answering౳) • Իָ + ςΩετͷϚϧνϞʔμϧදݱֶश • Վࢺੜ੒ɼςΩετϕʔεͷԻָੜ੒ʹ͍ͭͯ͸ࠓճର৅֎ 3
4. ͭ·Βͳ͍ʁ • ࡢࠓͷେن໛ݴޠϞσϧʢLLMʣʹ͍ͭͯͲ͏ࢥ͍·͔͢ʁ • Ϣʔβʔͱͯ͠͸ศརʹͳͬͯॿ͔ͬͨ͜ͱ΋ଟ͍ • Ұํݚڀऀͱͯ͠͸Ϧιʔε࣋ͬͨ΋ΜউͪͷੈքʹͳΓ໘ന͘ͳ͘ͳͬͨɾ ΍ΕΔ͜ͱ͕গͳ͘ͳͬͨʁ • Իָʹ͍ͭͯ৘ใΛநग़ͯ͠؍࡯͍ͨ͠ਓʢat least ࢲʣʹͱͬͯ͸ɼɹɹɹɹɹɹ ͦͷͨΊͷࣗಈࣝผͷੑೳ্͕͕Δ͜ͱ͸Ή͠Ζ௥͍෩ʁ • ͳͲͳͲ… 4
5. Music and Language (:= M&L) • Իָ৘ใॲཧͱࣗવݴޠॲཧΛ݁ͼͭ ͚Δ෼໺ • Իָ͔ΒςΩετͰ৘ใΛநग़͢Δ ʢCaptioningɼ Q&Aʣ • ݴ༿ͰԻָΛ୳͢ʢMusic searchʣ etc… • Vision & Language౳ͱಉ͘͡ɼɹɹɹ ݴޠͱͷϚϧνϞʔμϧֶशͷҰ෼໺ 5 It’s a jazz music piece with slow tempo. Piano, muted-trumpet, tenor sax, bass, and drums are appeared. Captioning Question: Is this music a vocal song? Answer: No. It’s an instrumental song. Q&A ϛϡʔττϥϯϖοτͱ ςφʔαοΫε͕ϝϩσΟʔΛ ԋ૗͢Δδϟζָۂ Music search είΞ: 0.91
6. ͳͥࣗવݴޠ͔ʁ • 1. ΦʔϓϯϘΩϟϒϥϦʔͳ৘ใهड़͕Մೳ • ↔︎ λά෇͚౳ͷࣝผ໰୊Ͱ͸σʔληοτ ͷΫϥεʹറΓΛड͚Δ • ʢཧ࿦্͸ʣࡉ΍͔ͳχϡΞϯεΛهड़Ͱ ͖Δ • 2. ҉ʹ͞·͟·ͳ໰୊Λಉ࣌ʹղ͍͍ͯΔ • ↔︎ ࣝผ໰୊Ͱ͸ɼλεΫ͝ͱʹϞσϧΛ࡞ Βͳ͍ͱ͍͚ͳ͍ 6 ϩοΫ Ϊλʔ ܹ͍͠ 80’ … ͜ͷۂ͸80’ͷ ϩοΫιϯάͰ͋Γɼ ଎͍ςϯϙͰ ΪλʔͷԻͱͱ΋ʹ ܹ͘͠υϥϜ͕ ໐Γڹ͖·͢ɽ ɽɽɽ λά෇͚ Ωϟϓγϣχϯά ͜ͷۂ͸80’ͷ ϩοΫιϯάͰ͋Γɼ ΪλʔͷԻͱͱ΋ʹ ܹ͘͠υϥϜ͕ ໐Γڹ͖·͢ɽ ɽɽɽ Ϝʔυࣝผ δϟϯϧࣝผ ָثࣝผ
7. M&L͕ࠓΞπ͍ • ࠃࡍձٞISMIRʹ͓͍ͯ΋2021೥͝ΖΛڥʹM&Lܥݚڀ͕૿͑ͨ • ϕετϖʔύʔϊϛωʔτʹ΋ຖ೥ग़ݱ • 2023೥ॳΊʹMusicLM [Agostinelli 23] ͕ൃදɼݚڀࣄྫ΋ैͬͯ૿Ճ • ౰࿦จͰMusicCapsͱ͍͏Իָ-ΩϟϓγϣϯͷϖΞσʔληοτ΋ൃද • AudioυϝΠϯ͚ͩͰͳ͘ɼSymbolicυϝΠϯ·Ͱʹਐग़ • CLaMP [SWu 23] 7 <"HPTUJOFMMJ
   >
   .VTJD-.
   (FOFSBUJOH
   .VTJD
   'SPN
   5FYU
   "OESFB
   "HPTUJOFMMJ
   FU
   BM
   "S9JW
   
   <48V
   >
   $-B.1
   $POUSBTUJWF
   -BOHVBHF.VTJD
   1SFUSBJOJOH
   GPS
   $SPTT.PEBM
   4ZNCPMJD
   .VTJD
   *OGPSNBUJPO
   3FUSJFWBM
   4IBOHEB
   8V
   FU
   BM
   *4.*3
   
   
8. ࠓ೔ͷτϐοΫ͸ओʹ2ͭ • Music to Text • Իָʹର͢Δ಺༰هड़จΛੜ੒ • Music-Language ϚϧνϞʔμϧදݱֶश • ԻָͱࣗવݴޠؒΛඥ෇͚ͨಛ௃ϕΫτϧͷ֫ಘ 8
9. Music to Text
10. ͦͷલʹ… • جૅతͳ༻ޠͨͪ • Encoder: ೖྗͱͳΔϝσΟΞ৘ใΛΑΓ௿࣍ݩͳϕΫτϧʹม׵ • Ex. BERT, RoBERTa౳ • Embedding: ϝσΟΞ৘ใͷີϕΫτϧදݱɽ্ۭؒͰͷϕΫτϧͷڑ ཭͕ҙຯతͳۙ͞Λࣔ͢ • Decoder: EmbeddingΛݩʹग़ྗͱͳΔϝσΟΞ৘ใͷܥྻΛਪ࿦ • Ex. GPT, LLaMA౳ 10
11. Music -> Text ᶃ Captioningܕ "VEJP
    &ODPEFS "VEJP
    &NCFEEJOH 5FYU
    %FDPEFS <4UBSU> 3PDL TPOH XJUI ʜ ೖྗ
    Իڹ৴߸ %FDPEFSೖྗ
    લͷεςοϓͷ୯ޠτʔΫϯ ग़ྗ
    ςΩετ 3PDL TPOH XJUI QJBOP ʜ ྆Ϟʔμϧ
    ༥߹ͷͨΊͷ
    ԿΒ͔ • ೖྗ: ԻָԻڹ৴߸ • Audio Encoderʹ௨͠ɼຒΊࠐΈϕΫτϧ ʢEmbeddingʣΛಘΔ • Text Decoderͷ಺෦ͰɼAudio embeddingͱ Text Decoderͷதؒग़ྗΛ༥߹ • Text Decoder͸ςΩετΛॱ࣍ੜ੒͢Δ • Muscaps [Manco 21] , LP-MusicCaps [Doh 23] etc. 11 <.BODP
    >
    .VT$BQT
    (FOFSBUJOH
    $BQUJPOT
    GPS
    .VTJD
    "VEJP
    *MBSJB
    .BODP
    FU
    BM
    
    *+$//
    
    <%PI
    >
    -1.VTJD$BQT
    --.#BTFE
    1TFVEP
    .VTJD
    $BQUJPOJOH
    4FVOHIFPO
    %PI
    FU
    BM
    *4.*3
    
    
12. Ϟʔμϧ༥߹ͷͨΊͷςΫχοΫ • Concat: ୯ʹ྆ϞʔμϧͷϕΫτϧΛ݁߹ • FiLM [Perez 18]: ҟͳΔϞʔμϧʹର͢Δ ΞϑΟϯม׵ͷ܎਺ΛܾΊΔ • Cross-attention: TransformerϕʔεͳΒ Ͱ͸ͷํ๏ɽ M&LͰ͸QueryͱKey&Value ΛͦΕͧΕҟͳΔϞʔμϧ͔ΒಘΔ 12 [Perez 18] FiLM Visual reasoning with a General Conditioning Layer. Perez et al. AAAI2018 https://vaclavkosar.com/ml/cross-attention-in- transformer-architecture
13. σʔληοτ: MusicCaps [Agostinelli 23] • ԻָͱΩϟϓγϣϯͷϖΞσʔληοτ • Իָ-ΩϟϓγϣϯͷσʔληοτͰ࠷ॳͷΦʔ ϓϯσʔλ • YouTubeϏσΦͷԻָΫϦοϓ5521݅ʹϓϩͷ ԻָՈ͕هड़จΛΞϊςʔγϣϯ, ܭ15࣌ؒ΄Ͳ • https://www.kaggle.com/datasets/googleai/ musiccaps • Music captioningͷֶशɾධՁʹ༻͍ΒΕΔ 13 "
    CBHQJQF
    FOTFNCMF
    QMBZJOH
    B
    RVJDL
    NFMPEZ
    JO
    VOJTPO
    FNQMPZJOH
    USJMMT
    UP
    PSOBNFOU
    JU YoutubeϏσΦͷ ԻΫϦοϓ Ωϟϓγϣϯ <"HPTUJOFMMJ
    >
    .VTJD-.
    (FOFSBUJOH
    .VTJD
    'SPN
    5FYU
    "OESFB
    "HPTUJOFMMJ
    FU
    BM
    "S9JW
     σʔλ͸
    গ͠଍Γͳ͍
14. طଘͷλά෇͚༻σʔλͷλά͔ΒٖΩϟϓγϣϯΛੜ੒͠σʔλΛ֦ு 14 • GPT-3.5 TurboʹΑͬͯɼλάͷΈͷ MagnaTagATune [Law 09], Million-song Dataset [Bertin-Mahieux 11]͔Βٖࣅతͳ ΩϟϓγϣϯΛੜ੒ • ্هͷૢ࡞ʹΑͬͯɼMusicCapsͷ200ഒҎ ্ͷ໿4000࣌ؒͷϖΞσʔλΛߏங • ٖΩϟϓγϣϯͰ͋ͬͯ΋ੑೳΛ্͛ΒΕͨ LP-MusicCaps [Doh 23] <%PI
    >
    -1.VTJD$BQT
    --.#BTFE
    1TFVEP
    .VTJD
    $BQUJPOJOH
    4FVOHIFPO
    %PI
    FU
    BM
    *4.*3
    
    <-BX
    >
    &WBMVBUJPO
    PG
    BMHPSJUINT
    VTJOH
    HBNFT
    5IF
    DBTF
    PG
    NVTJD
    UBHHJOH
    &
    -BX
    FU
    BM
    *4.*3
    
    <#FSUJO.BIJFVY>
    5IF
    NJMMJPO
    TPOH
    EBUBTFU
    5
    #FSUJO.BIJFVY
    FU
    BM
    *4.*3
    
15. Music -> Textܕ ᶄ Q&Aܕ • ೖྗ: ԻָԻڹ৴߸, ςΩετ • Իڹ৴߸ΛAudio Encoderʹ௨͠Audio embeddingΛಘΔ • ్தͰAudio EmbeddingͱTextଆͷதؒग़ྗͱ༥߹͢Δ • ೖྗςΩετ͸ • ᶃText Encoder Ͱॲཧ or • ᶄΠϯετϥΫγϣϯͱͯ͠Text Decoderʹ༩͑Δ • Text DecoderͰςΩετΛग़ྗ • ᶃBART-fusion [Zhang 22], ᶄMu-LLaMa [Liu 24] etc. 15 <;IBOH
    >
    *OUFSQSFUJOH
    4POH
    -ZSJDT
    XJUI
    BO
    "VEJP*OGPSNFE
    1SFUSBJOFE
    -BOHVBHF
    .PEFM
    *4.*3
    
    <-JV
    >
    -1.VTJD$BQT
    --.#BTFE
    1TFVEP
    .VTJD
    $BQUJPOJOH
    
    FU
    BM
    *$"441
    
    "DDFQUFE 
    
16. Music Understanding LLaMA [Liu 24] • Իָ৴߸ʹର͢ΔQ&AΛ࣮ݱ • ֶशࡁΈϞσϧΛ׆༻ • Audio Encoder: MERT [Li 24] • Text Decoder: LLaMA-2 [Touvron 23] • MERT͔Βͷग़ྗʹɼAdapterͱ ͍͏ػߏΛ௨ͯ͠ɼͦΕΛ Decoderͷ࠷ऴ૚ʹೖྗ 16 <5PVWSPO
    >
    -MBNB
    
    0QFO
    'PVOEBUJPO
    BOE
    'JOF5VOFE
    $IBU
    .PEFMT
    5PVWSPO
    FU
    BM
    "SYJW
    
    <-J
    >
    .&35
    "DPVTUJD
    .VTJD
    6OEFSTUBOEJOH
    .PEFM
    XJUI
    -BSHF4DBMF
    4FMGTVQFSWJTFE
    5SBJOJOH
    :
    -J
    FU
    BM
    *$-3
    
    
    
17. σʔληοτɿMusicQA [Liu 24] • LLMʹΑͬͯQ&AϖΞΛMusicCaps (Ωϟϓγϣϯ)΍MagnaTagATune (λ ά)͔Βੜ੒ • 2छྨͷํ๏Ͱ࣭໰จΛ࡞੒ • Closed-end: ΠϯετϥΫγϣϯͰ ࣭໰Λ༩͑ͯɼͦͷ౴͑Λੜ੒͢Δ • Open-end: Ϟσϧࣗ਎ʹ࣭໰Λੜ੒ ͤͯ͞ɼͦͷ࣭໰ͷ౴͑΋ੜ੒͢Δ 17 <-JV
    >
    -1.VTJD$BQT
    --.#BTFE
    1TFVEP
    .VTJD
    $BQUJPOJOH
    
    FU
    BM
    *$"441
    
    "DDFQUFE 
    Closed Open
18. Music & Languageͷ ϚϧνϞʔμϧຒΊࠐΈֶश
19. M&L ϚϧνϞʔμϧදݱֶश • ϚϧνϞʔμϧදݱֶश • EncoderʹΑͬͯಘΒΕΔEmbedding ͕ɼಉ֓͡೦༝དྷͰ͋Ε͹ڑ཭͕ۙ͘ ͳΔΑ͏ʹ഑ஔ͢ΔۭؒΛಘΔͨΊͷ ํ๏ • Audio EncoderͱText EncoderΛ࢖͍ɼ ྆ऀͷEmbeddingʹର͢Δڑ཭ʹؔ͠ ֶͯश͢Δ 19
20. ϚϧνϞʔμϧදݱֶश͕Մೳʹ͢Δ͜ͱ • จষʹΑΔԻָݕࡧ • จষͷText embeddingͱϥΠϒϥϦʹ͋ΔԻָͷ Audio embeddingؒͷ݁ͼ͚͕ͭͰ͖ΔΑ͏ʹ • จ຺৘ใΛ΋ͭԻָಛ௃ྔͷ֫ಘ • ͢Ͱʹdiscriminableͳಛ௃͕֫ಘͰ͖͓ͯΓɼɹ θϩγϣοτֶश΍ɼ௿ϦιʔελεΫ΁ͷసҠ ֶश͕ظ଴ • ୯ମͰ͸ੜ੒͸Ͱ͖ͳ͍͕ɼMusicLMͰ͸ɹɹɹ ςΩετϓϩϯϓτͱԻָͷඥ෇͚ͷϞδϡʔϧ ͱͯ͠Ұ෦ʹར༻͞Ε͍ͯΔ [Agostinelli 23] 20 ϛϡʔττϥϯϖοτͱ ςφʔαοΫε͕ϝϩσΟʔΛ ԋ૗͢Δδϟζָۂ
21. 21 • τϦϓϨοτֶश • ΞϯΧʔɼਖ਼ྫʢࣅ͍ͯΔ΋ͷʣɼෛྫʢҟͳΔ΋ͷʣΛ༻ҙ͠ɼ ΞϯΧʔ-ਖ਼ྫͷڑ཭ʹ ରͯ͠ΞϯΧʔ-ෛྫͷڑ཭Λେ͖͘͢Δ • θϩγϣοτԻָλά෇͚ [Choi 19], λάʹΑΔԻָݕࡧ [Won 21] ౳ • ରরֶश • ΞϯΧʔͱਖ਼ྫͷڑ཭Λ࠷খԽ͠ɼΞϯΧʔͱෛྫͷڑ཭Λ࠷େԽ • ը૾-ςΩετؒͷϞσϧCLIP [Radford 21]Ͱ੒ޭͨ͠࿮૊Έ • ྆ऀͷੑೳΛൺֱͨ͠ࣄྫ΋͋Γ; ରরֶशͷํ͕Ұ౓ʹ࠷దԽͰ͖Δαϯϓϧ͕ଟ͍ [Doh 23] ࣮ݱͷͨΊͷख๏͸͍͔ͭ͘ <$IPJ
    >;FSPTIPU
    -FBSOJOH
    GPS
    "VEJPCBTFE
    .VTJD
    $MBTTJ fi DBUJPO
    BOE
    5BHHJOH
    +
    $IPJ
    FU
    BM
    
    *4.*3
    
    <8PO
    >.VMUJNPEBM
    .FUSJD
    -FBSOJOH
    GPS
    5BHCBTFE
    .VTJD
    3FUSJFWBM
    .XPO
    FU
    BM
    *$"441
    
    <3BEGPSE
    >
    -FBSOJOH
    5SBOTGFSBCMF
    7JTVBM
    .PEFMT
    'SPN
    /BUVSBM
    -BOHVBHF
    4VQFSWJTJPO
    3BEGPSE
    FU
    BM
    "S9JW
    
    <%PI
    >
    508"3%
    6/*7&34"-
    5&9550.64*$
    3&53*&7"-
    %PI
    FU
    BM
    *$"441
    
    
22. τϦϓϨοτֶश 22 • 3ͭ૊ΈͷαϯϓϧΛ༻ҙ͠ɼɹɹɹ ͦΕͧΕAnchor, Positive sample, Negative sampleͱ͢Δ • Anchorʹର͠ྨࣅ౓͕ΑΓ͍ۙ΋ͷ ΛPositive sampleɼԕ͍΋ͷΛ Negative sampleͱ͠ɼͦΕͧΕͷɹ ڑ཭ͷࠩΛɼ͋Δ஋δʢϚʔδϯʣʹ ͳΔΑ͏ʹ࠷దԽ͢Δɽ
23. ࣮ݱํ๏ɿରরֶश • 2ͭͷҟͳΔϞʔμϧͷEmbeddingʹ ର͠ɼಉ͡σʔλϖΞͰ͋Ε͹಺ੵΛ େ͖͘ɼͦ͏Ͱͳ͚Ε͹খ͘͢͞Δ • ෳ਺ͷσʔλϖΞʹରͯ͠Ұڍʹֶश ͕Մೳ • ଛࣦؔ਺: InfoNCE Loss 23 I൪໨ͷαϯϓϧͷaudio embedding z^a_i, text embedding z^t_i ΞϯΧʔͷྨࣅ౓͕େ͖͘ͳΕ͹0ʹۙͮ͘
24. Ϣʔεέʔε 24 θϩγϣοτ෼ྨ ςΩετʹΑΔԻָݕࡧ • ࣄલʹֶश͍ͯ͠ͳ͍෼ྨλεΫͰ ͋ͬͯ΋ɼ֤Ϋϥε͝ͱʹจষΛ༻ҙ ͢Ε͹ɼ෼ྨ໰୊Λղ͘͜ͱ͕Ͱ͖Δ • ςΩετΫΤϦΛೖྗ͢ΔͱҰக౓ ͷߴ͍ԻָΛ୳͢͜ͱ͕Ͱ͖Δ
25. M&LͷରরֶशϞσϧ: MuLan [Huang 22] / MusCall [Manco 22] • Իָ (Audio) ͱςΩετͷϚϧν ϞʔμϧֶशϞσϧ • ͱ΋ʹରরֶश͕ϕʔε • (ISMIR 2022Ͱωλ͕ඃΔͱ͍͏ح੻) • θϩγϣοτ෼ྨʹΑΔԻָλά෇͚ɼసҠ ֶशʹΑΔԻָλά෇͚ɼԻָݕࡧʹ͓͍ ͯɼੑೳΛ֬ೝ 25 <.BODP
    >
    $0/53"45*7&
    "6%*0-"/(6"(&
    -&"3/*/(
    '03
    .64*$
    *
    .BODP
    FU
    BM
    *4.*3
     <)VBOH
    >
    .6-"/
    "
    +0*/5
    &.#&%%*/(
    0'
    .64*$
    "6%*0
    "/%
    /"563"-
    -"/(6"(&
    2
    )VBOH
    FU
    BM
    *4.*3
    
26. M&LͷରরֶशϞσϧ: CLaMP [SWu 23] • Իָ (Symbolic) ͱςΩετͷϚϧνϞʔμ ϧֶशϞσϧ • ରরֶश͕ϕʔε • M3ͱ͍͏ABC notationϕʔεͰָේͷࣗݾڭࢣ͋Γ ֶशϞσϧ΋ซͤͯఏҊ • WikiMusicDatasetͱ͍͏ָේʹର͢ΔΩϟϓγϣϯ σʔληοτΛఏҊ • ςΩετָේݕࡧΛՄೳʹ • θϩγϣοτ෼ྨ/సஔֶशͰྑੑೳ 26 <48V
    >
    $-B.1
    $POUSBTUJWF
    -BOHVBHF.VTJD
    1SFUSBJOJOH
    GPS
    $SPTT.PEBM
    4ZNCPMJD
    .VTJD
    *OGPSNBUJPO
    3FUSJFWBM
    4IBOHEB
    8V
    FU
    BM
    *4.*3
    
27. 27 • Captioning • MusicCaps https://www.kaggle.com/datasets/googleai/musiccaps • LP-MusicCaps-MC https://huggingface.co/datasets/seungheondoh/LP-MusicCaps-MC • LP-MusicCaps-MSD https://huggingface.co/datasets/seungheondoh/LP-MusicCaps-MSD • LP-MusicCaps-MTT https://huggingface.co/datasets/seungheondoh/LP-MusicCaps-MTT • MusicBench https://huggingface.co/datasets/amaai-lab/MusicBench • The song describer dataset https://github.com/mulab-mir/song-describer-dataset • WikiMuTe https://huggingface.co/datasets/davanstrien/WikiMuTe • Question-answering • Music-AVQA https://gewu-lab.github.io/MUSIC-AVQA/ • MusicQA https://crypto-code.github.io/MU-LLaMA-Demo/#dataget ࢖͑Δσʔληοτͷ·ͱΊ
28. ײ૝ a.k.a. ϞϠϞϠᶃ: ͳͥࣗવݴޠ͔ʁ • ԻָͱࣗવݴޠΛ݁ͼ͚ͭΔࢫΈ͸ɼԻ੠΍ը૾ͷͦΕͱ͸ҟͳΔʁ • ٕज़తʹɼ੝Γ্͕͍ͬͯΔࣗવݴޠॲཧͱ༥߹͢Δͷ͸໘ന͍ • ʮԻָͱʯ༥߹ͯ͠خ͍͜͠ͱ͸͋Δʁ • Իָͱࣗવݴޠ͸ͦ΋ͦ΋݁ͼ෇͚Ͱ͖Δͷ͔ • ࡢ೔ͷฏాઌੜͷ͓࿩ͷதʹ΋͋ͬͨ௨Γ྆ऀ͸ඇՄٯ • ໘ന͍, ͱ͍͏Ҏ্ʹԿΛ΋ͨΒͩ͢Ζ͏͔ 28
29. ײ૝ a.k.a. ϞϠϞϠᶄ: ࣗ෼ͷλεΫʹಛԽ͍ͤͨ͞৔߹ • ΩϟϓγϣϯʹೖΔ΂͖ཁ݅͸Կʁ • Իָʹ͸͞·͟·ͳଐੑΛ͍࣋ͬͯΔ • Ͳ͜·ͰΛهड़͢Δ΂͖͔ʁ • Իݯ͔ΒΘ͔Δ৘ใ (Acoustics)ɼͲͷΑ͏ʹফඅ͞Εͨʢ͞ΕΔʣ͔ͷ ৘ใ (Cultural)ɼϝλσʔλ৘ใ (Editorial)ʹେผͰ͖Δ [Pachet 05] • Ͳ͏ΩϟϓγϣϯΛ૊Ή΂͖͔ • ΩϟϓγϣϯΛهड़͢Δ୯Ґ͸ۂશମͰOK? • ࠁҰࠁͰঢ়ଶ͸มԽ͢Δ 29 [Pachet 05]Pachet, F., 2005. Knowledge management and musical metadata. Idea Group, 12.
30. ԻָͱࣗવݴޠॲཧͷϚϧνϞʔμϧֶशͷαʔϕΠ 30 • ಈ޲Λ঺հ • Իָ -> ςΩετ (Captioning, Question & Answering౳) • Իָ + ςΩετͷϚϧνϞʔμϧදݱֶश • ૉ๿ʹࢥ͍ͬͯΔٙ໰ΛͿͪ·͚ • ͦ΋ͦ΋࿦తͳ࿩΋ଟ͍ͱࢥ͏ͷͰɼٞ࿦ͷՐछͱͳΕ͹޾͍Ͱ͢ ·ͱΊ
31. ͋Γ͕ͱ͏͍͟͝·ͨ͠