Ama-Neural Vocoders
I-neural vocoder iyimodeli eguqula ukumelela okuhlangene kwe-acoustic, ngokuvamile i-mel-spectrogram, ibe i-waveform ezwakalayo yangempela.
Uhlolojikelele
It is the final stage that gives modern text-to-speech and voice cloning their natural, human sound.
I-Deep Dive
Ukuhlanganiswa kwenkulumo yomdabu kusetshenziswa amavokhoda okucubungula isignali ayevame ukuzwakala ebhizi noma erobhothi. Ama-Neural vocoder afunda ukwakha kabusha amasampula omsindo angahluziwe avela ku-spectrogram ngokuqeqeshwa ngamahora okurekhoda kwangempela. I-WaveNet (DeepMind, 2016) yaba impumelelo, ibikezela isampula yomsindo eyodwa ngesikhathi kumasampuli angu-16,000+ ngomzuzwana, ikhiqiza inkulumo yemvelo emangalisayo kodwa ihamba kancane kakhulu. Amamodeli akamuva ahweba ngalelo bhodlela le-autoregressive ngesivinini: I-WaveGlow yasebenzisa isizukulwane esisekelwe ekugelezeni, i-Parallel WaveGAN ne-MelGAN yasebenzisa amanethiwekhi akhiqizayo ezitha, futhi i-HiFi-GAN yaba izinga elidumile ngokukhiqiza umsindo othembekile ongu-22kHz ngokushesha kakhulu kunesikhathi sangempela. Namuhla i-vokhoda cishe njalo iyingxenye yesibili yepayipi lezigaba ezimbili, elihambisana nemodeli ye-acoustic efana ne-Tacotron 2 noma i-FastSpeech ekhiqiza i-mel-spectrogram.
I-Technical Insight
I-mel-spectrogram ilahla ulwazi lwesigaba somsindo, igcina kuphela ukuthi amandla asakazwa kanjani kuwo wonke amabhendi efrikhwensi ngokuhamba kwesikhathi. Umsebenzi onzima wevokhoda ukusungula i-waveform ezwakalayo, ehambisanayo lapho ubukhulu bayo buhambisana nalokho okufakwayo. Amavokhoda asuselwa ku-GAN afana ne-HiFi-GAN asebenzisa ababandlululi abaningi abahlola isignali ezikalini ezihlukene kanye nezikhathi ezithile, iphusha ijeneretha ukuthi ikhiqize imininingwane emihle engokoqobo njengama-harmonics kanye nokudlula okubukhali kongwaqa.
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 Lama-Neural Vocoder
Amavokhoda aba mancane futhi ayashesha ukuze asebenze kumafoni nakumadivayisi ashumekiwe ngaphandle koxhumano lwamafu. Kuphinde kube nokuphushela kumavokhoda atholakala emhlabeni wonke ajwayelekile kunoma yisiphi isikhulumi, ulimi, ukucula, noma umsindo ongakhulumi ngaphandle kokuqeqeshwa kabusha. Ukuthambekela okuhambisanayo kugoqa i-vocoder ngokuqondile kumasistimu asuka ekupheleni kuya ekupheleni kanye namakhodikhi e-neural, kufiphaze umugqa phakathi kwezigaba ezihlukene ze-acoustic ne-waveform futhi kunciphise ama-artifact ethulwa ngokudlula ku-spectrogram emaphakathi.
Ukuqaliswa Komhlaba Wangempela
Ikhiqiza umsindo wokugcina okhulunywayo kuzisizi zokuguqula umbhalo ube inkulumo njengezifundi zesikrini nezinhlelo zokusebenza zokuzulazula
Ukukhiqiza amazwi aklanywe anomsindo wemvelo kumathuluzi okukopisha kanye ne-audiobook yokulandisa
Ukwakha kabusha amazwi okucula kumculo we-AI kanye nesoftware ye-virtual-vocalist
Inika amandla okukhiphayo kwezwi kudivayisi kuzipikha ezihlakaniphile namadivayisi afinyelelekayo ngaphandle kohambo oluya nokubuya lweseva
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 Neural 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-HiFi-GAN ne-GAN Vocoder
Imibuzo evame ukubuzwa
What is Neural Vocoders?
I-neural vocoder iyimodeli eguqula ukumelela okuhlangene kwe-acoustic, ngokuvamile i-mel-spectrogram, ibe i-waveform ezwakalayo yangempela. Kuyisigaba sokugcina esinikeza ukuhlanganisa umbhalo-kuya-enkulumweni wesimanjemanje kanye nezwi lokuhlanganisa umsindo wabo wemvelo, womuntu.
Uyini umsebenzi oyinhloko we-neural vocoder?
I-neural vocoder ithatha isici esihlangene se-acoustic, ngokuvamile i-mel-spectrogram, futhi ihlanganise i-wave wave yomsindo wangempela oyizwayo.
Iyiphi imodeli ka-2016 eyaba impumelelo eyenza ama-neural vocoder azwakale engokwemvelo?
I-WaveNet, evela ku-DeepMind ngo-2016, yakhiqiza isampula yomsindo ngesampula futhi yakhiqiza inkulumo yemvelo emangalisayo, iqala inkathi ye-neural vocoder.
Kungani i-WaveNet yoqobo yayihamba kancane ukukhiqiza umsindo?
I-WaveNet i-autoregressive: isampula ngayinye yomsindo incike kwedlule, ngakho-ke ukukhiqiza amashumi ezinkulungwane zamasampuli ngomzuzwana bekumba eqolo.
Yiluphi ulwazi olulahlwa yi-mel-spectrogram, okwenza umsebenzi wevokhoda ube nzima?
Ama-Mel-spectrogram agcina ubukhulu (amandla ngebhendi yefrikhwensi) kodwa isigaba sokwehla, ngakho-ke i-vokhoda kufanele yakhe kabusha i-waveform ehambisanayo esigaba sayo esizwakalayo.
Amavokhoda asuselwa ku-GAN afana ne-HiFi-GAN akuthuthukisa kanjani ukuba ngokoqobo?
I-HiFi-GAN isebenzisa ababandlululi abambalwa abahlola izikali zesikhathi ezihlukene kanye nezikhathi, iphusha ijeneretha ukuthi ikhiqize ama-harmonics angokoqobo kanye nesikhashana.