I-HiFi-GAN ne-GAN Vocoder
I-HiFi-GAN iyivowuda yokukhiqiza eguqula i-mel-spectrogram ibe i-waveform yomsindo ongahluziwe cishe ngaso leso sikhathi, ikhiqize inkulumo yekhwalithi yesitudiyo ngokushesha kakhulu kunesikhathi sangempela.
Uhlolojikelele
It became the standard final stage of modern text-to-speech because it is fast, lightweight, and hard to distinguish from real recordings.
I-Deep Dive
Ivokhoda iyisinyathelo sokugcina kumapayipi amaningi e-TTS: imodeli efana ne-Tacotron noma i-FastSpeech ibikezela i-mel-spectrogram (isithombe esihlangene samafrikhwensi ngokuhamba kwesikhathi), futhi i-vokhoda igcwalisa amasampuli angempela e-waveform. Ama-vocoder asekuqaleni kwe-neural afana ne-WaveNet azwakala emuhle kodwa akhiqizwe amasampula omsindo ngesampula, okuwenza angasheshi kabuhlungu. I-HiFi-GAN, ekhishwe ngu-Kong, Kim, kanye no-Bae ngo-2020, ithathele indawo leyo luphu ye-autoregressive generator eyodwa yokuphakela phambili eqeqeshelwe ukulwa. Iqhinga layo eliyinhloko ukusebenzisa ababandlululi abaningi abahlulela umsindo ngezikali ezihlukene kanye namaphethini ezikhathi ezithile, okuphoqelela ijeneretha ukuthi ithole kokubili ukuthungwa okuhle kanye ne-pitch periodicity efanele. Umphumela uba inkulumo engu-22 kHz ehlanganiswe ngokushesha izikhathi ezingamakhulu kunesikhathi sangempela ku-GPU, ngomsindo weqiniso ophikisanayo wekhwalithi.
I-Technical Insight
Ijeneretha ye-HiFi-GAN iphakamisa isampula ye-mel-spectrogram ngokusebenzisa ama-convolutions aguquguqukayo, namabhulokhi astakiwe e-Multi-Receptive Field ahlanganisa osayizi abahlukahlukene be-kernel nokwenyuswa ukuze athwebule amaphethini amagagasi ahlukahlukene. Imindeni emibili ebandlululayo yenza umsebenzi wobuphoyisa: I-Multi-Period Discriminator iphinda ilolonge isignali ye-1D ibe amagridi e-2D ngesikhathi sokuqala njengo-2, 3, 5, 7, 11 ukuze ibambe i-pitch periodicity, futhi I-Multi-Scale Discriminator ihlola i-waveform ezinqumweni ezimbalwa ezehlisiwe. I-Mel-spectrogram nokulahlekelwa okuhambisana nesici kugcina ukuqeqeshwa kuzinzile.
I-Strategic Impact
Finyelela futhi ufinyelele
Ithuthukisa ukufinyeleleka ngokuloba, ukulandisa, nezixhumi ezibonakalayo zezwi.
Izindleko kanye nesabelomali
Amaqembu emidiya angathumela umsindo opholishiwe ngokushesha ngamabhajethi amancane.
Isivinini nesikali
Amasistimu abhekene nekhasimende angacubungula ukusebenzelana okukhulunyiwe ngesilinganiso esikhulu.
Ikusasa lamaVokhoda e-HiFi-GAN kanye ne-GAN
Amavokhoda e-GAN alokhu eba mancane futhi ngokushesha: inzalo efana ne-BigVGAN yengeza ama-activas aphikisana ne-aliased ukuze ajwayele ukucula kubaculi abangabonakali, izinsimbi, nezilimi, kuyilapho i-UnivNet ne-Vocos isunduza ekuhlanganiseni kwendawo yonke, i-All-band synthesis. Ukusakaza-bukhoma nokuhlukahluka okukudivayisi manje kusebenzisa izwi ngaphakathi kwamafoni nama-earbud kubasizi ababambezeleke kancane. Ngokukhulayo, amamodeli omsindo asabaleleyo nawokugeleza afakwa kumajeneretha ephasi elilodwa lesitayela se-GAN, okuhlanganisa ukwethembeka kwe-diffusion nesivinini se-GAN. Lindela ama-vocoder ukuthi afiphele abe amakhodekhi omsindo we-neural wenjongo evamile anika amandla kokubili inkulumo nomculo.
Ukuqaliswa Komhlaba Wangempela
Ikhiqiza okukhiphayo okukhulunyiwe kwabasizi ababonakalayo nezinhlelo zokusebenza zokuzulazula ezidinga izimpendulo ngaphandle kokulibaziseka okuzwakalayo.
Inika amandla amathuluzi e-cloning wezwi lesikhathi sangempela kanye namathuluzi wokukopisha lapho i-mel-spectrogram ehlanganisiwe inikezwa umsindo ozwakalayo wemvelo.
Ukushayela i-audiobook kanye nezinkundla zokulandisa ze-podcast ezihlanganisa amahora okukhuluma ngokushesha naphansi.
Isebenza njengesiteji se-waveform ngaphakathi kwama-synthesizers okucula-izwi namademo omculo ngamavokhoda asemhlabeni wonke esitayela se-BigVGAN.
Izingozi & Guardrails
Ukusetshenziswa kabi kwezwi kanye nezingozi zokuzenza ongeyena ziyanda uma imvume ingekho.
Ukunemba kungase kwehle kuzo zonke izinhlobo zokuphimisela, izilimi zesigodi, noma izindawo ezinomsindo.
Umsindo wokwenziwa ungenziwa iphutha njengenkulumo eyiqiniso ngaphandle kokulebula okucacile.
Ukuqalisa Umhlahlandlela
Thola imvume esobala yokuthwebula izwi, ukuhlanganisa, nokusebenzisa kabusha.
Ikhwalithi yokuhlola kuzo zonke izipikha nezimo zangemuva.
Chaza ukuthi kunini lapho umuntu kufanele abuyekeze noma agunyaze okuphumayo.
Lebula umsindo wokwenziwa futhi ugcine amarekhodi atholakalayo ukuze aziphendulele.
Qhubeka Uhlole
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
Umhlahlandlela olandelayo
I-Parallel WaveGAN Vocoder
Imibuzo evame ukubuzwa
What is HiFi-GAN and GAN Vocoders?
I-HiFi-GAN iyivowuda yokukhiqiza eguqula i-mel-spectrogram ibe i-waveform yomsindo ongahluziwe cishe ngaso leso sikhathi, ikhiqize inkulumo yekhwalithi yesitudiyo ngokushesha kakhulu kunesikhathi sangempela. Kube yisigaba sokugcina esijwayelekile sokuguqula umbhalo ube inkulumo ngoba kuyashesha, kulula, futhi kunzima ukuhlukanisa nokurekhodwa kwangempela.
Imuphi umsebenzi oyinhloko wevokhoda efana ne-HiFi-GAN epayipini lokuguqula umbhalo ube inkulumo?
Ivokhoda yisigaba sokugcina esihlanganisa amasampula e-waveform yangempela asuka ku-mel-spectrogram ekhiqizwe imodeli ye-acoustic.
Kungani i-HiFi-GAN ishesha kakhulu kunevokhoda yangaphambili ye-WaveNet?
I-WaveNet ikhiqize isampula yomsindo ngesampula (ngokuzenzakalelayo), kuyilapho ijeneretha yokuphakela phambili ye-HiFi-GAN ikhipha i-waveform ngokudubula okukodwa, ifinyelela isivinini esishesha kakhulu kunesesikhathi sangempela.
Yini i-HiFi-GAN's Multi-Period Discriminator isiza ngokukhethekile ukuthwebula?
I-Multi-Period Discriminator ibumba kabusha i-1D waveform ibe yi-2D isebenzisa izikhathi ezinenombolo eyinhloko ukuze ithole ukwakheka kwezikhathi okuboshelwe ekuphakameni.
Amavokhoda e-GAN aqeqeshwe 'ngokuphambene.' Kusho ukuthini lokho lapha?
Ekusetheni kwe-GAN, ijeneretha ifunda ukukhiqiza amagagasi angokoqobo ngokwanele ukuze akhohlise amanethiwekhi ababandlululi abaqeqeshelwe ukusho okwangempela kumsindo wokwenziwa.
Iyiphi i-vocoder yakamuva enweba i-HiFi-GAN ukuze ihlanganise kangcono amazwi, ukucula, nezinsimbi ezingabonakali?
I-BigVGAN ikhulisa i-HiFi-GAN futhi yengeza ukusebenza kwezikhathi ezithile okunganqatshelwe, ithuthukisa ukwenziwa okuvamile kumsindo ongaphandle kokusabalalisa njengokucula nezinsimbi.