Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Interspeech2019&サテライト論文読み会(音声生成系) #interspeech2019jp

F1a39fc11df357f49ee8f05d1ab6efd5?s=47 gaku kotani
November 24, 2019

Interspeech2019&サテライト論文読み会(音声生成系) #interspeech2019jp

論文紹介
Y. Jiang et al., “The USTC System for Blizzard Challenge 2019,”
 in Blizzard Challenge Workshop, 2019.

F1a39fc11df357f49ee8f05d1ab6efd5?s=128

gaku kotani

November 24, 2019
Tweet

Transcript

  1. l5IF645$4ZTUFNGPS#MJ[[BSE$IBMMFOHFz  খ୩ַ
 ౦ژେֶᜊ౻ݚ% *OUFSTQFFDIαςϥΠτ࿦จಡΈճɹ࿦จ঺հ

  2. ࣗݾ঺հ ⾣খ୩ַʢͨ͜ʹ͕͘ʣ ɹὊ౦ژେֶᜊ౻େีݚڀࣨ%ʢֶ͔࢜Βʣ ɹὊ޷͖ͳ੠༏͸Ֆᖒ߳ࡊ ⾣ݚڀ෼໺ ɹὊ੠࣭ม׵ ɹὊ554΋͍ͨ͠ ɹɹՖᖒ߳ࡊͷ੠ʹͳΓ͍ͨͱ͍͏ΑΓɺՖᖒ߳ࡊͷ੠Λฉ͖͍ͨ 2

  3. ঺հ͢Δ࿦จ ⾣:+JBOHFUBM l5IF645$4ZTUFNGPS#MJ[[BSE$IBMMFOHF z
 ɹJO#MJ[[BSE$IBMMFOHF8PSLTIPQ  ɹ#MJ[[BSE$IBMMFOHF554ͷίϯϖʢࠓ೥ͷλεΫ͸.BOEBSJO554ʣ ɹ#$Ͱ࠷΋݁Ռ͕ྑ͔ͬͨνʔϜͷγεςϜઆ໌࿦จ ⾣બΜͩಈػ
 ɹᜊ౻ݚ΋65PLZPνʔϜͱͯ͠ࢀՃͨ͠ͷͰҰ൪ͷγεςϜ͕ؾʹͳΔ

    ɹ554ͷษڧ͠Α͏ 3 ޡղ΍ݸਓͷҙݟΛؚΜͰ͍·͢ ਖ਼֬ͳ৘ใ͸࿦จΛࢀর͍ͯͩ͘͠͞
  4. #MJ[[BSEDIBMMFOHF ⾣5BTL ɹὊ#VJMEBWPJDFGSPNUIFQSPWJEFEEBUB TVJUBCMFGPSFYQSFTTJWF554
 ɹɹGSPNQMBJOUFYUJOQVUz ɹὊλʔήοτ͸தࠃޠΛ஻Δ͓͡͞ΜʢZPVUVCFSΈ͍ͨͳʜ ʣ ɹὊ࣌ؒ෼ͷσʔλʢ෼ºVUUFSTʣ ɹɹL)[TBNQMJOH NQ

    ɹɹσʔλ͕ͪΐͬͱ͋ΕɿಡΜͰͳ͍ʗॻ͍ͯͳ͍ʗςΩετԻ੠ϖΞ͕ؒҧͬͯΔ ⾣645$͸Ϛϯύϫʔͱ֎෦σʔλͰղܾʜ ɹὊ΋͸΍ຊਓͳԻ੠߹੒ʢαϯϓϧ୳ͯ͠΋ແ͔ͬͨͰ͢͢Έ·ͤΜʜʣ 4 ࠓఱੋ෕౔ࢫɼૹ۳େՈҰ࿽ྱ෺ɻզૌᬐ౨ԋჴဗཱ৽࿕؀ྃʬၕཕՈ඀ʭɼ
 ࡞ູ৽ҰقతʬՈఉഎܠ੠ʭӁ඼ɼ䓟ՄҎࡏʮಘ౸ʯ"QQཬ໔ٮਵऔɻʢଓ͘ʣ
  5. 'SBNFXPSLPG645$TZTUFNGPS#$ ⾣NPEVMFT ɹὊ#&35CBTFEGSPOUFOEQSFEJDUJPO ɹὊ%VSBUJPONPEFMJOH ɹὊ("/CBTFENVMUJUBTLBDPVTUJDNPEFMJOH ɹὊ8BWF/FUCBTFEOFVSBMWPDPEFS 5 'MPXDIBSUPGUIFTZTUFN

  6. 'SBNFXPSLPG645$TZTUFNGPS#$ ⾣1SFQSPDFTT
 ɹὊ.BOVBMBOOPUBUJPOT
 ɹɹ1JOZJO QSPTPEJDXPSECPVOEBSZ 
 ɹɹQSPTPEJDQISBTFCPVOEBSZ GPDVTQPTJUJPO ɹὊ&YUSBDUJPOPGGSBNFMFWFMBDPVTUJDGFBUVSFT ɹɹ.FMDFQTUSVN

    ' 767 ɹɹ)..BMJHONFOUGPSQIPOFNFCPVOEBSJFT
 6 ਓؒίετ͕͍͢͝ 'MPXDIBSUPGUIFTZTUFN
  7. 'SPOUFOENPEFM 7 ⾣5IFNBJOQSPDFEVSF ɹὊ5FYUQSPDFTTJOH ɹɹ4QFDJBMNBSLTDPOWFSTJPO ɹὊ(SBQIFNFUPQIPOFNFDPOWFSTJPO ɹɹ$IJOFTFDIBSBDUFSUPQJOZJO ɹὊ1SPTPEZQSFEJDUJPO ɹɹ#SFBLJOHBTFOUFODFJOUPQISBTFTGPDVTPOXPSET .VMUJUBTLNPEFMCBTFEPO#&35

  8. 'SPOUFOENPEFM 8 ⾣5IFNBJOQSPDFEVSF ɹὊ5FYUQSPDFTTJOH ɹɹ4QFDJBMNBSLTDPOWFSTJPO ɹὊ(SBQIFNFUPQIPOFNFDPOWFSTJPO ɹɹ$IJOFTFDIBSBDUFSUPQJOZJO ɹὊ1SPTPEZQSFEJDUJPO ɹɹ#SFBLJOHBTFOUFODFJOUPQISBTFTGPDVTPOXPSET JF"SBCJDOVNFSBMTUPWBMVFTPSTUSJOHT

    lzPSlUXP ٢ ٥ lzPSlpWFz
  9. 'SPOUFOENPEFM 9 ⾣5IFNBJOQSPDFEVSF ɹὊ5FYUQSPDFTTJOH ɹɹ4QFDJBMNBSLTDPOWFSTJPO ɹὊ(SBQIFNFUPQIPOFNFDPOWFSTJPO ɹɹ$IJOFTFDIBSBDUFSUPQJOZJO ɹὊ1SPTPEZQSFEJDUJPO ɹɹ#SFBLJOHBTFOUFODFJOUPQISBTFTGPDVTPOXPSET "TJOHMFDIBSIBWFTFWFSBMEJ⒎FSFOUQSPOVOT


    BDDPSEJOHUPVTBHFTNFBOJOHTJOTFOUFODFT 单 lE⒑OzPSlTIÁOzPSlDIÂOz
  10. 'SPOUFOENPEFM 10 ⾣5IFNBJOQSPDFEVSF ɹὊ5FYUQSPDFTTJOH ɹɹ4QFDJBMNBSLTDPOWFSTJPO ɹὊ(SBQIFNFUPQIPOFNFDPOWFSTJPO ɹɹ$IJOFTFDIBSBDUFSUPQJOZJO ɹὊ1SPTPEZQSFEJDUJPO ɹɹ#SFBLJOHBTFOUFODFJOUPQISBTFTGPDVTPOXPSET 5IFTUSFOHUIPGUIFCSFBL


    1SPTPEJDQISBTF - NBKPSCSFBL 1SPTPEJDXPSE - NJOPSCSFBL 'JHVSF&YBNQMFPGNBOEBSJOQSPTPEZTUSVDUVSF JO<Pan+, INTERSPEECH, 2019> $IBSBDUFS -FYJDPOXPSE 1SPTPEJDXPSE 1SPTPEJDQISBTF *OUPOBUJPOBMQISBTF
  11. 'SPOUFOENPEFM 11 ⾣5IFNBJOQSPDFEVSF ɹὊ5FYUQSPDFTTJOH ɹɹ4QFDJBMNBSLTDPOWFSTJPO ɹὊ(SBQIFNFUPQIPOFNFDPOWFSTJPO ɹɹ$IJOFTFDIBSBDUFSUPQJOZJO ɹὊ1SPTPEZQSFEJDUJPO ɹɹ#SFBLJOHBTFOUFODFJOUPQISBTFTGPDVTPOXPSET 'PDVTXPSETFNQIBTJ[FEXPSET

    UPESBXBUUFOUJPOUPQBSUJDVMBSXPSET
  12. 'SPOUFOENPEFM ⾣#&35CBTFEGSPOUFOENPEFM ɹὊPVUQVUT ɹɹ1PMZ@PVUQJOZJOQSFEJDUJPO ɹɹ-@PVUNJOPSCSFBLPSOPU ɹɹ-@PVUNBKPSCSFBLPSOPU ɹɹ4QDM@PVUTQFDJBMNBSLTDPOWFSTJPO ɹɹ'PDVT@PVUGPDVTPSOPU ɹὊ$IJOFTFDPSQVTJOOPWFMBOEOFXTpFMETUPQSFUSBJO 12

    #&35CBTFEGSPOUFOENPEFM ۩ମతͳྔ͸࿦จʹࡌ͍ͬͯͳ͍ˍൃදͰฉ͍ͨ͸͕ͣͩ๨Ε·ͨ͠ʜ
  13. %VSBUJPONPEFMJOH ⾣"VUPSFHSFTTJWFNPEFMGPSRVBOUJpFEEVSBUJPOT
 ɹὊ2VBOUJGZJOHJOMPHEPNBJOTPGUNBY DSPTTFOUSPQZ ɹὊ$BVTBMFODPEFSPVUQVUTUIFQSFWJPVTFNCFEEJOHT ɹɹEVSBUJPO DPOUFYUDP[MJUUMFDPSSFMBUJPOXJUIEVSBUJPOT ɹὊ5IFDPOUFYUJOQVUJODMVEFT ɹɹQIPOFNFJEFOUJpDBUJPOT UPOFT

    
 ɹɹ5P#* UPLFOTBOECSFBLJOEJDFT TUSFTTqBHT BOETPPO ɹὊ1SFUSBJOFEPOBMBSHFNVMUJTQFBLFSDPSQVT 13
  14. %VSBUJPONPEFMJOH ⾣"VUPSFHSFTTJWFNPEFMGPSRVBOUJpFEEVSBUJPOT
 ɹὊ2VBOUJGZJOHJOMPHEPNBJOTPGUNBY DSPTTFOUSPQZ ɹὊ$BVTBMFODPEFSPVUQVUTUIFQSFWJPVTFNCFEEJOHT ɹɹEVSBUJPO DPOUFYUDP[MJUUMFDPSSFMBUJPOXJUIEVSBUJPOT ɹὊ5IFDPOUFYUJOQVUJODMVEFT ɹɹQIPOFNFJEFOUJpDBUJPOT UPOFT

    
 ɹɹ5P#* UPLFOTBOECSFBLJOEJDFT TUSFTTqBHT BOETPPO ɹὊ1SFUSBJOFEPOBMBSHFNVMUJTQFBLFSDPSQVT 14
  15. %VSBUJPONPEFMJOH ⾣"VUPSFHSFTTJWFNPEFMGPSRVBOUJpFEEVSBUJPOT
 ɹὊ2VBOUJGZJOHJOMPHEPNBJOTPGUNBY DSPTTFOUSPQZ ɹὊ$BVTBMFODPEFSPVUQVUTUIFQSFWJPVTFNCFEEJOHT ɹɹEVSBUJPO DPOUFYUDP[MJUUMFDPSSFMBUJPOJOEVSBUJPOWBMVFT ɹὊ5IFDPOUFYUJOQVUJODMVEFT ɹɹQIPOFNFJEFOUJpDBUJPOT UPOFT

    
 ɹɹ5P#* UPLFOTBOECSFBLJOEJDFT TUSFTTqBHT BOETPPO ɹὊ1SFUSBJOFEPOBMBSHFNVMUJTQFBLFSDPSQVT 15 
  16. %VSBUJPONPEFMJOH ⾣"VUPSFHSFTTJWFNPEFMGPSRVBOUJpFEEVSBUJPOT
 ɹὊ2VBOUJGZJOHJOMPHEPNBJOTPGUNBY DSPTTFOUSPQZ ɹὊ$BVTBMFODPEFSPVUQVUTUIFQSFWJPVTFNCFEEJOHT ɹɹEVSBUJPO DPOUFYUDP[MJUUMFDPSSFMBUJPOXJUIEVSBUJPOT ɹὊ5IFDPOUFYUJOQVUJODMVEFT ɹɹQIPOFNFJEFOUJpDBUJPOT UPOFT

    
 ɹɹ5P#* UPLFOTBOECSFBLJOEJDFT TUSFTTqBHT BOETPPO ɹὊ1SFUSBJOFEPOBMBSHFNVMUJTQFBLFSDPSQVT 16
  17. %VSBUJPONPEFMJOH ⾣"VUPSFHSFTTJWFNPEFMGPSRVBOUJpFEEVSBUJPOT
 ɹὊ2VBOUJGZJOHJOMPHEPNBJOTPGUNBY DSPTTFOUSPQZ ɹὊ$BVTBMFODPEFSPVUQVUTUIFQSFWJPVTFNCFEEJOHT ɹɹEVSBUJPO DPOUFYUDP[MJUUMFDPSSFMBUJPOXJUIEVSBUJPOT ɹὊ5IFDPOUFYUJOQVUJODMVEFT ɹɹQIPOFNFJEFOUJpDBUJPOT UPOFT

    
 ɹɹ5P#* UPLFOTBOECSFBLJOEJDFT TUSFTTqBHT BOETPPO ɹὊ1SFUSBJOFEPOBMBSHFNVMUJTQFBLFSDPSQVT 17 ۩ମతͳྔ͸࿦จʹࡌ͍ͬͯͳ͍͕࣌ؒҎ্ͱݴ͍ͬͯͨ͸ͣ
  18. "DPVTUJDNPEFMJOH 18 ⾣-45.3//BDPVTUJDNPEFMXJUI("/ ɹὊ5IFPWFSTNPPUIJOHQSPCMFN("/ ɹɹ%JTDSJNJOBUJOHJOTQFDUSBMEPNBJO POMZ41&  ɹɹ-FBTUTRVBSF("/ -4("/ 

    ɹὊ1SFEJDUFEGFBUVSFT' EJN.$&1 #"1
  19. 7PDPEJOH ⾣8BWF/FUCBTFEOFVSBMWPDPEJOH ɹὊ1SFUSBJOFEXJUIBNVMUJTQFBLFSEBUBTFU ɹɹ4QFBLFSFNCFEEJOHWFDUPSTBSFVTFEEVSJOHPOMZQSFUSBJOJOH ɹὊ$PNCJOBUJPOPG8BWF/FUT L)[BOEL)[WFST  ɹɹL)[8BWF/FUHFOFSBUFTCFUUFSMPXGSFRVFODZIBSNPOJD ɹɹ$PODBUFOBUJOHUIFJSBNQMJUVEFT ɹὊ"TJOHMFWBSJBODFCPVOEFE(BVTTJBOEJTUSJCVUJPONPEFMJOHSBXXBWFGPSN

    ɹɹJOUSPEVDFEJO$MBSJ/FU<1JOH > ɹɹSFMJFWFTUIFRVBOUJ[BUJPOOPJTF 19
  20. 7PDPEJOH ⾣8BWF/FUCBTFEOFVSBMWPDPEJOH ɹὊ$PNCJOBUJPOPG8BWF/FUT L)[BOEL)[WFST  ɹɹL)[8BWF/FUHFOFSBUFTCFUUFSMPXGSFRVFODZIBSNPOJD ɹɹ$PODBUFOBUJOHUIFJSBNQMJUVEFT 20 ࿦จʹઆ໌͕΄ͱΜͲͳ͍͕ɺߴଳҬͷҐ૬͸ϊΠζͰຒΊͯ
 (SJ⒏O-JNʢతͳ͜ͱʣΛͨ͠ͱฉ͍ͨؾ͕͢Δʢ͏Ζ֮͑ʜʣ

  21. 3FTVMUTPGFWBMVBUJPO 21 722 722 722 722 722 722 722 722

    722 722 722 722 722 722 722 722 722 722 722 722 722 722 722 722 722 722 n A M I W S Y Z E C G T H D V J N X K Q F O R B U L P 1 2 3 4 5 Mean Opinion Scores (All Listeners) System Score 361 361 361 361 361 361 361 361 361 361 361 361 361 361 361 361 361 361 361 361 361 361 361 361 361 361 n A M I W S Y Z E C G T H D V J N X K Q F O R B U L P 1 2 3 4 5 Similarity Scores Comparing to Original Speaker (All Listeners) System Score /BUVSBMOFTTTDPSFTPGTVCNJUUFETZTUFN 4JNJMBSJUZTDPSFTPGTVCNJUUFETZTUFN  
  22. 4VNNBSZ ⾣645$TZTUFNGPS#MJ[[BSE$IBMMFOHF ɹὊ#&35CBTFEGSPOUFOEQSFEJDUJPO ɹὊ"VUPSFHSFTTJWFEVSBUJPONPEFM ɹὊ("/CBTFENVMUJUBTLBDPVTUJDNPEFM ɹὊ8BWF/FUCBTFEOFVSBMWPDPEFS ⾣5IFWFSZIJHI.04TDPSFJTBDIJFWFEJOIJTUPSZPG#MJ[[BSE$IBMMFOHF ⾣ײ૝ ɹὊʢԻ੠αϯϓϧ͸༻ҙͰ͖ͳ͔͕ͬͨɺݱ஍Ͱฉ͍ͨײ͡ʣ·ΔͰຊਓ͔ͷΑ͏ͳ
 ɹɹԻ੠͕ੜ੒Ͱ͖͍ͯͯײಈ͕͋ͬͨ

    22