HiFi-GAN na GAN Vocoders
HiFi-GAN nijwi ryibyara-rirwanya amajwi ahindura mel-spectrogramu mumajwi mbisi yamajwi hafi ako kanya, itanga imvugo-nziza ya studio yihuta cyane kuruta igihe nyacyo.
Incamake
It became the standard final stage of modern text-to-speech because it is fast, lightweight, and hard to distinguish from real recordings.
Kwibira cyane
Ijwi ni intambwe yanyuma mu miyoboro myinshi ya TTS: icyitegererezo nka Tacotron cyangwa FastSpeech gihanura mel-spectrogramu (ishusho yerekana inshuro nyinshi mugihe), kandi amajwi yuzuza ibyitegererezo nyirizina. Indangururamajwi zambere nka WaveNet yumvikanye neza ariko itanga amajwi sample-by-sample, bigatuma itinda kubabaza. HiFi-GAN, yasohowe na Kong, Kim, na Bae mu 2020, yasimbuye uwo muzingi wa autoregressive hamwe na generator imwe yo kugaburira imbere yatojwe nabi. Amayeri yingenzi ni ugukoresha ivangura ryinshi risuzuma amajwi ku munzani itandukanye no mu bihe bitandukanye, bigatuma generator ibona imiterere myiza hamwe nigihe cyikibanza neza. Igisubizo ni 22 kHz imvugo yashizwemo inshuro magana byihuse kuruta igihe nyacyo kuri GPU, hamwe nubwiza bwo guhangana nubutaka-ukuri amajwi.
Ubushishozi
Imashini itanga amashanyarazi ya HiFi-GAN yerekana urugero rwa mel-spectrogramme binyuze mumyanya ihindagurika, hamwe na Multi-Receptive Field bloks zivanze zivanga ingano yintanga zitandukanye hamwe niyaguka kugirango ifate imiterere itandukanye. Imiryango ibiri ivangura ikora igipolisi: Ivangura rishingiye ku bihe byinshi rihindura ikimenyetso cya 1D muri gride ya 2D kuri prima nka 2, 3, 5, 7, 11 kugirango zifate igihe cyigihe, kandi ivangura rishingiye ku bice byinshi risuzuma imiterere yumuraba ku myanzuro myinshi yatanzwe. Mel-spectrogramu hamwe nibihuye-bihombo bikomeza imyitozo ihamye.
Ingaruka z'Ingamba
Kugera no kugera
Itezimbere kugerwaho binyuze mu kwandukura, kuvuga, no guhuza amajwi.
Igiciro na bije
Amatsinda yibitangazamakuru arashobora kohereza amajwi yihuse hamwe na bije nto.
Umuvuduko n'igipimo
Sisitemu ireba abakiriya irashobora gutunganya imikoranire ivugwa murwego runini.
Kazoza ka HiFi-GAN na Vocoders ya GAN
Abavuza amajwi ba GAN bakomeje kuba bato kandi byihuse: ababakomokaho nka BigVGAN bongeraho ibikorwa birwanya anti-aliased kugirango bahuze muri rusange abaririmbyi, ibikoresho, n'indimi zitagaragara, naho UnivNet na Vocos biganisha kuri synthesis rusange, yose. Streaming na on-device variants ubu ikoresha amajwi imbere muri terefone no gutwi kubufasha buke bwihuse. Kwiyongera, gukwirakwiza amajwi no guhuza amajwi yerekana amajwi arimo gutondekwa muburyo bwa GAN-imwe ya generator imwe, bivanga ubudahemuka bwa diffusion hamwe numuvuduko wa GAN. Tegereza abavuga amajwi bazimangana muri rusange-intego ya neural amajwi ya codecs ikoresha imvugo numuziki.
Gushyira mu bikorwa Isi
Kubyara umusaruro uvugwa wabafasha basanzwe hamwe na porogaramu zo kugendana zikeneye ibisubizo nta gutinda kumvikana.
Guha imbaraga amajwi nyayo yo gukoroniza hamwe nibikoresho bya dubbing aho mel-spekrogramu ya clon ikozwe mumajwi asanzwe yumvikana.
Gutwara amajwi ya audiobook na podcast yo kuvuga inkuru ihuza amasaha yo kuvuga vuba kandi bihendutse.
Gukora nkicyerekezo cyimbere imbere yo kuririmba-amajwi synthesizers hamwe na demo yumuziki ukoresheje BigVGAN-yuburyo bwa majwi rusange.
Ingaruka & Kurinda
Gukoresha nabi amajwi no kwigira ibyago byiyongera mugihe uruhushya rubuze.
Ukuri kurashobora kugabanuka hejuru yimvugo, imvugo, cyangwa urusaku rwibidukikije.
Amajwi yubukorikori arashobora kwibeshya kumvugo yukuri nta kirango gisobanutse.
Igishushanyo mbonera
Shaka uruhushya rusobanutse rwo gufata amajwi, gukoroniza, no gukoresha.
Ikizamini cyiza mubiganiro bitandukanye hamwe nuburyo bwimbere.
Sobanura igihe umuntu agomba gusuzuma cyangwa kwemeza ibisubizo.
Andika amajwi yubukorikori kandi ugumane inyandiko zerekana kubazwa.
Komeza Ubushakashatsi
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the HiFi-GAN and GAN Vocoders quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
Ubuyobozi bukurikira
Iringaniza WaveGAN Vocoder
Ibibazo bikunze kubazwa
What is HiFi-GAN and GAN Vocoders?
HiFi-GAN nijwi ryibyara-rirwanya amajwi ahindura mel-spectrogramu mumajwi mbisi yamajwi hafi ako kanya, itanga imvugo-nziza ya studio yihuta cyane kuruta igihe nyacyo. Byahindutse icyiciro cya nyuma cyimyandikire igezweho-ku-mvugo kuko yihuta, yoroheje, kandi biragoye gutandukanya ibyanditswe byafashwe.
Nuwuhe murimo wibanze wijwi nka HiFi-GAN mumurongo winyandiko-mvugo?
Ijwi ni intambwe yanyuma ihuza ingero zifatika zifatika kuva mel-spectrogramu yakozwe na moderi ya acoustic.
Kuki HiFi-GAN yihuta cyane kurenza amajwi ya WaveNet?
WaveNet yabyaye amajwi icyitegererezo-by-icyitegererezo (autoregressively), mugihe HiFi-GAN itanga ibyokurya bitanga imbaraga imbere yumurongo umwe, ikagera kumuvuduko mwinshi-kuruta-igihe-nyacyo.
Niki HiFi-GAN's Multi-Period Discriminator ifasha cyane gufata?
Ivangura rya Multi-Period ivugurura imiterere ya 1D ihinduka 2D ikoresheje ibihe-byibanze kugirango umenye imiterere yigihe ihujwe nikibuga.
GAN abavuga amajwi bahuguwe 'muburyo butandukanye.' Ibyo bivuze iki hano?
Muburyo bwa GAN, generator yiga kubyara imiyoboro ifatika ihagije kugirango ibeshye imiyoboro ivangura yatojwe kuvuga nyayo uhereye kumajwi yubukorikori.
Ninde wacuranga nyuma wagura HiFi-GAN kugirango rusange muri rusange amajwi atagaragara, kuririmba, nibikoresho?
BigVGAN yapimye HiFi-GAN kandi yongeraho ibikorwa byo kurwanya anti-aliase buri gihe, itezimbere rusange mumajwi adakwirakwizwa nko kuririmba nibikoresho.