I-WaveNet
I-WaveNet, eyethulwe ngabakwa-DeepMind ngo-2016, bekuyinethiwekhi ye-neural ephumelelayo ekhiqiza umsindo ongahluziwe isampula eyodwa ngesikhathi, ekhiqiza inkulumo yemvelo emangalisayo nomculo.
Uhlolojikelele
It set the modern standard for high-fidelity text-to-speech.
I-Deep Dive
I-WaveNet iyimodeli ekhiqizayo ezenzakalelayo: ibikezela isampula ngayinye yomsindo ebekwe kuwo wonke amasampuli angaphambi kwayo, ngokuvamile kumasampuli angu-16,000 noma angu-24,000 ngomzuzwana. Ukuqanjwa kwayo okuyisisekelo kuyinqwaba ye-convolutions enwetshiwe ye-causal. I-Causal isho ukuthi imodeli ibheka emuva kuphela ngesikhathi, igcina ukuhleleka kwesizukulwane; ukunwetshwa kusho ukuthi isendlalelo ngasinye seqa inani elikhulayo lamasampuli, ngakho isitaki esinesizotha simboza izinkulungwane zamasampuli (inkambu ebanzi eyamukelayo) ngaphandle kwezindleko ezinkulu. Isekelwe ezicini zolimi noma i-mel-spectrogram, i-WaveNet ikhiqiza inkulumo eyimvelo kakhulu kunamavokhoda e-concatenative kanye ne-parametric ayandulele, evala isikhala esikhulu ekurekhodweni kwabantu futhi inika amandla izinguqulo zangaphambili zomsizi Google.
I-Technical Insight
I-convolutions enwetshiwe iyiqhinga elibalulekile: ngamanani okunwebeka angu-1, 2, 4, 8, njalo njalo, inethiwekhi kuphela amashumi ezendlalelo ezijulile ezingabheka izinkulungwane zamasampuli adlule, ithwebula yomibili imininingwane emihle ye-waveform kanye nesakhiwo eside se-prosodic. Okuphumayo kumodeli yevelu yesampula ngayinye njengokusatshalaliswa kwezigaba (ekuqaleni amaleveli angu-256 ngokuhlanganisa i-mu-law), kanye namayunithi okuqalisa afakwe esangweni kanye nezinsalela nokuxhumeka kokweqa kusimamisa ukuqeqeshwa kwalesi sitaki esijule kakhulu.
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 le-WaveNet
I-WaveNet yoqobo ibihamba kancane ngoba ukusampula kuyalandelana. Abalandelayo balungise lokhu: I-Parallel WaveNet ne-WaveRNN inike amandla ukuhlanganiswa kwesikhathi sangempela, futhi kamuva ama-vocodes asuselwa ku-flow- kanye nasekelwe ku-GAN njenge-WaveGlow ne-HiFi-GAN, kanye namavokhoda okusabalalisa, aphushe ikhwalithi nesivinini ngokuqhubekayo. Imibono ye-WaveNet ye-autoregressive, ne-dilated-convolution iphila kulezi zinhlelo futhi ibe nomthelela wezakhiwo ezingaphezu komsindo, iqinisa ifa layo ekumodeleni okukhiqizayo.
Ukuqaliswa Komhlaba Wangempela
Ikhiqiza amazwi azwakalayo emvelo Google Umsizi kanye Google Umbhalo-kuya-Inkulumo Yamafu
Isebenza njenge-neural vocoder eshintsha i-mel-spectrogram ibe ama-waveform kumapayipi e-TTS afana ne-Tacotron 2
Ihlanganisa upiyano wangempela nomculo wezinsimbi ovela kumsindo ongahluziwe
Ukuhlanganiswa kwezwi kwamathuluzi okufinyelela nokulandisa kwe-audiobook
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 WaveNet 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
Ukutholwa kwe-Audio Deepfake
Imibuzo evame ukubuzwa
What is WaveNet?
I-WaveNet, eyethulwe ngabakwa-DeepMind ngo-2016, bekuyinethiwekhi ye-neural ephumelelayo ekhiqiza umsindo ongahluziwe isampula eyodwa ngesikhathi, ekhiqiza inkulumo yemvelo emangalisayo nomculo. Isethe izinga lesimanje lokuthembeka okuphezulu kombhalo-kuya-enkulumweni.
I-WaveNet iwenza kanjani umsindo?
I-WaveNet i-autoregressive: ibikezela isampula ngayinye yomsindo ngokusekelwe kumasampuli afike ngaphambi kwayo.
Yiluphi uguquko olubalulekile lwe-convolutional ku-WaveNet?
Izimbangela ezinwetshiwe zivumela inethiwekhi ukuthi imboze izinkulungwane zamasampuli adlule kahle kuyilapho ibheka emuva kuphela ngesikhathi.
Kungani ukunwetshwa kusiza i-WaveNet?
Amazinga okunwebeka akhula ngokuphawulekayo anikeza inkambu ebanzi eyamukelayo ngaphezu kwezinkulungwane zamasampuli anezendlalelo ezimbalwa uma kuqhathaniswa.
Kwakuyini ukuhlehla okukhulu okusebenzayo kwe-WaveNet yasekuqaleni?
Ngoba isampula ngayinye incike kwedlule, isizukulwane besilandelana futhi ngenxa yalokho sihamba kancane, sigqugquzela i-Parallel WaveNet nabanye abalandelayo.
Epayipini lombhalo-kuya-enkulumweni, i-WaveNet ivamise ukusebenza njengeyini?
I-WaveNet ingathatha i-mel-spectrogram (isb., kusuka ku-Tacotron 2) futhi ikhiqize i-waveform yokugcina yomsindo wemvelo.