Kodi mungagwiritse ntchito AI kupanga mbiri yatsopano kuchokera kwa ojambula omwe mumawakonda?
Kupita patsogolo kwaposachedwa pakuphunzira makina kwawonetsa kuti zitsanzo tsopano zimatha kumvetsetsa zovuta monga zolemba ndi zithunzi. OpenAI's Jukebox imatsimikizira kuti ngakhale nyimbo zimatha kutsatiridwa ndendende ndi neural network.
Nyimbo ndi chinthu chovuta kuchitsanzira. Muyenera kuganizira zinthu ziwiri zosavuta monga tempo, phokoso, phokoso ndi zovuta zina monga mawu, zida, ndi nyimbo.
Kugwiritsa ntchito zapamwamba makina kuphunzira njira, OpenAI yapeza njira yosinthira zomvera zaiwisi kukhala choyimira chomwe mitundu ina ingagwiritse ntchito.
Nkhaniyi ifotokoza zomwe Jukebox ingachite, momwe imagwirira ntchito, komanso zolephera zomwe zilipo paukadaulo.
Kodi Jukebox AI ndi chiyani?
Jukebox ndi neural net model yopangidwa ndi OpenAI yomwe imatha kupanga nyimbo ndikuyimba. Chitsanzochi chikhoza kupanga nyimbo zamitundu yosiyanasiyana ndi zojambula za ojambula.
Mwachitsanzo, Jukebox ikhoza kupanga nyimbo ya rock mumayendedwe a Elvis Presley kapena nyimbo ya hip hop monga Kanye West. Mutha kuyendera izi webusaiti kuti muwone momwe chitsanzocho chimagwirira ntchito pojambula nyimbo za ojambula omwe mumakonda ndi mitundu.
Chitsanzocho chimafuna mtundu, wojambula, ndi mawu kuti alowemo. Izi zikutsogolera chitsanzo chophunzitsidwa mamiliyoni a ojambula ndi mawu a nyimbo.
Kodi Jukebox imagwira ntchito bwanji?
Tiyeni tiwone momwe Jukebox amakwanitsira kupanga nyimbo zatsopano kuchokera pamitundu yophunzitsidwa mamiliyoni a nyimbo.
Encoding process
Ngakhale mitundu ina yamtundu wa nyimbo imagwiritsa ntchito deta yophunzitsira ya MIDI, Jukebox imaphunzitsidwa pa fayilo yeniyeni ya audio. Kuti akanikizire mawuwo kukhala pamalo owonekera, Jukebox amagwiritsa ntchito njira yojambulira yokha yotchedwa VQ-VAE.
Chithunzi cha VQ-VAE imayimira Vector Quantized Variational Autoencoder, yomwe ingamveke ngati yovuta, ndiye tiyeni tiyidule.
Choyamba, tiyeni tiyese kumvetsa zimene tikufuna kuchita pano. Poyerekeza ndi mawu kapena nyimbo zamapepala, fayilo ya audio yaiwisi imakhala yovuta kwambiri. Ngati tikufuna kuti chitsanzo chathu "chiphunzire" kuchokera ku nyimbo, tidzayenera kuchisintha kukhala choyimira choponderezedwa komanso chosavuta. Mu makina kuphunzira, timachitcha choyimira ichi a malo obisika.
An autoen kodi ndi njira yophunzirira yosayang'aniridwa yomwe imagwiritsa ntchito a neural network kuti mupeze zoyimira zobisika zosatsata mzere pakugawa deta. The autoencoder ili ndi magawo awiri: encoder ndi decoder.
The encoder amayesa kupeza malo obisika kuchokera pagulu la data yaiwisi pomwe a Chojambula amagwiritsa ntchito choyimira chobisika kuyesa kuchimanganso kuti chikhale chake choyambirira. The autoencoder kwenikweni imaphunzira momwe ingasinthire deta yaiwisi m'njira yochepetsera zolakwika zomanganso.
Tsopano popeza tadziwa zomwe autoencoder imachita, tiyeni tiyese kumvetsetsa zomwe tikutanthauza ndi "zosintha" autoencoder. Poyerekeza ndi ma autoencoder wamba, ma autoencoder osinthika amawonjezera malo obisika.
Popanda kudumphira mu masamu, kuwonjezera chotheka kumapangitsa kuti kugawa kobisika kukhale kophatikizika. Kusiyana kwakukulu pakati pa VAE ndi VQ-VAE ndikuti womalizayo amagwiritsa ntchito mawonekedwe obisika m'malo mopitilira.
Mulingo uliwonse wa VQ-VAE umayika pazolowera. Encoding yapansi panthaka imapanga kumangidwanso kwapamwamba kwambiri. Ma encoding apamwamba amakhalabe ndi nyimbo zofunika.
Kugwiritsa ntchito Transformers
Tsopano popeza tili ndi nyimbo zojambulidwa ndi VQ-VAE, titha kuyesa kupanga nyimbo m'malo ophatikizika awa.
Jukebox amagwiritsa ntchito thiransifoma autoregressive kupanga zotulutsa mawu. Transformers ndi mtundu wa neural network womwe umagwira ntchito bwino ndi data yotsatizana. Chifukwa cha mndandanda wa zizindikiro, chitsanzo cha transformer chidzayesa kufotokozera chizindikiro chotsatira.
Jukebox imagwiritsa ntchito mtundu wosavuta wa Sparse Transformers. Mitundu yonse yam'mbuyomu ikaphunzitsidwa, chosinthiracho chimapanga ma code ophatikizika omwe amasinthidwa kukhala ma audio aiwisi pogwiritsa ntchito VQ-VAE decoder.
Artist and Genre Conditioning mu Jukebox
Mtundu wopanga wa Jukebox umapangidwa kuti ukhale wowongoka kwambiri popereka zidziwitso zowonjezera panthawi yophunzitsira.
Zitsanzo zoyamba zimaperekedwa ndi ojambula ndi zolemba zamtundu wa nyimbo iliyonse. Izi zimachepetsa entropy ya kulosera kwamawu ndikulola kuti mtunduwo ukhale wabwinoko. Zolembazo zimatithandizanso kuwongolera chitsanzocho mwanjira inayake.
Kupatula ojambula ndi mtundu, zizindikiro za nthawi zimawonjezedwa panthawi ya maphunziro. Zizindikirozi zimaphatikizapo kutalika kwa nyimboyo, nthawi yoyambira chitsanzo china, ndi kachigawo kakang'ono ka nyimbo yomwe yadutsa. Zowonjezera izi zimathandiza kuti chitsanzocho chimvetsetse machitidwe omvera omwe amadalira dongosolo lonse.
Mwachitsanzo, wojambulayo angaphunzire kuti kuwomba m'manja kwa nyimbo zamoyo kumachitika kumapeto kwa nyimbo. Chitsanzochi chingaphunzirenso, mwachitsanzo, kuti mitundu ina ili ndi zigawo zazitali kuposa zina.
Lyrics
Zitsanzo zokhazikika zomwe zatchulidwa m'gawo lapitalo zimatha kupanga mawu osiyanasiyana oimba. Komabe, mawu awa amakhala osagwirizana komanso osazindikirika.
Kuti athetseretu njira yoberekera ikafika pakupanga ma lyric, ochita kafukufuku amapereka zambiri pa nthawi yophunzitsira. Kuti athandize mapu a mawu amtundu wa nthawi yomwe amamvetsera, ofufuzawo adagwiritsa ntchito Mapazi kutulutsa mawu ndi NUS AutoLyricsAlign kuti mupeze mawu omveka bwino a mawu.
Zochepa za Jukebox Model
Chimodzi mwazolepheretsa zazikulu za Jukebox ndikumvetsetsa kwake nyimbo zazikuluzikulu. Mwachitsanzo, kachigawo kakang'ono ka masekondi 20 kakumveka kochititsa chidwi, koma omvera adzazindikira kuti nyimbo zobwerezabwereza zobwerezabwereza ndi mavesi palibe pomaliza.
Chitsanzocho chimachedwanso kupereka. Zimatenga pafupifupi maola 9 kuti mupereke mphindi imodzi yamawu. Izi zimachepetsa kuchuluka kwa nyimbo zomwe zingapangidwe ndikulepheretsa kuti chitsanzocho chisagwiritsidwe ntchito pazokambirana.
Potsirizira pake, ofufuzawo awona kuti deta yachitsanzo imakhala mu Chingerezi ndipo imasonyeza makamaka nyimbo zachizungu. Ofufuza a AI atha kuyang'ana kwambiri kafukufuku wamtsogolo pakupanga nyimbo m'zilankhulo zina ndi masitaelo anyimbo omwe si a Western.
Kutsiliza
Pulojekiti ya Jukebox ikuwonetsa kuthekera kwakukula kwamitundu yophunzirira makina kuti apange chithunzithunzi cholondola cha data yovuta monga ma audio aiwisi. Kupambana kofananirako kumachitika m'mawu, monga tawonera m'mapulojekiti ngati GPT-3, ndi zithunzi, monga zikuwonekera mu OpenAI's DALL-E2.
Ngakhale kuti kafukufuku m'derali wakhala wochititsa chidwi, palinso nkhawa zokhudzana ndi ufulu waluntha komanso momwe zitsanzozi zingakhudzire makampani opanga zinthu. Ofufuza ndi opanga ayenera kupitiriza kugwirizanitsa kuti atsimikizire kuti zitsanzozi zikupitirizabe kusintha.
Mitundu yanyimbo zamtsogolo zitha kukhala ngati chida cha oimba kapena ngati pulogalamu ya opanga omwe amafunikira nyimbo zamapulojekiti.
Siyani Mumakonda