Ingabe i-AI yombhalo ibe yinkulumo? [Ividiyo kanye nombuzo]

Impendulo emfushane: Umbhalo-ube-inkulumo ngumsebenzi wokuguqula umbhalo obhaliwe ube umsindo okhulunywayo; ukuthi ngabe "uyi-AI" kuncike endleleni owakhiwe ngayo. Amazwi esimanje, azwakala emvelo ngokuvamile aqhutshwa amamodeli okufunda komshini, kuyilapho izinhlelo ezindala zingase zithembele emithethweni noma ekuqoshweni okuthungiwe. Uma udinga ubufakazi, hlola ukuthi yini "engaphansi kwe-hood", hhayi nje ukuthi izwakala kanjani.

Izinto ezibalulekile okufanele uzicabangele:

Incazelo: I-TTS iwumgomo; i-AI iyindlela eyodwa engenzeka yokukufeza lokho.

Ukutholwa: Uma i-prosody kanye ne-pause zizwakala zingokwemvelo, kungenzeka ukuthi ziqhutshwa yimodeli.

Ukuhamba komsebenzi: Khetha ifu ngokwesilinganiso; khetha indawo ukuze uthole ubumfihlo kanye nezindleko ezingabikezelwa.

Ukufinyeleleka: I-TTS eqinile incike esakhiweni esihlanzekile: izihloko, izixhumanisi, ukuhleleka, umbhalo ohlukile.

Ukumelana nokusebenzisa kabi: Qinisekisa izicelo zezwi ezingavamile ngesiteshi sesibili, hhayi umsindo wodwa.

Izihloko ongase uthande ukuzifunda ngemva kwalesi:

🔗 Ingabe i-AI ingafunda umbhalo wesandla obhalwe ngokugqamile?
Indlela i-AI eqaphela ngayo ukubhala okunezinhlayiya kanye nemikhawulo evamile.

🔗 Inembile kangakanani i-AI namuhla?
Yini ethinta ukunemba kwe-AI kuyo yonke imisebenzi, idatha, kanye nokusetshenziswa kwangempela.

🔗 I-AI ithola kanjani ama-anomalies?
Incazelo elula yokubona amaphethini angavamile kudatha.

🔗 Indlela yokufunda i-AI isinyathelo ngesinyathelo
Indlela ewusizo yokuqala ukufunda i-AI kusukela ekuqaleni.

Kungani "I-AI Yombhalo Kuya Enkulumweni" izwakala idida kwasekuqaleni 🤔🧩

Abantu bavame ukubiza into ngokuthi “i-AI” uma izwakala kanje:

• ukuzivumelanisa nezimo

• umuntu

• "Kwenzeka kanjani lokho?"

Futhi ama-TTS anamuhla angazizwa kanjalo impela. Kodwa ngokomlando, amakhompyutha “ayekhuluma” esebenzisa izindlela eziseduze nobunjiniyela obuhlakaniphile kunokufunda.

Uma umuntu ebuza ukuthi i-Is Text to Speech AIisho ukuthini, lokho abakushoyo ngokuvamile ukuthi:

• "Ingabe ikhiqizwa yimodeli yokufunda komshini?"

• "Ingabe ifunde ukuzwakala njengomuntu ngedatha?"

• "Ingabe ingabhekana nokukhuluma amagama nokugcizelela ngaphandle kokuzwakala sengathi i-GPS inosuku olubi?"

Lezo zifiso zilungile. Aziphelele, kodwa zihloselwe kahle.

Impendulo esheshayo: iningi le-TTS yesimanje yi-AI - kodwa akuzona zonke ✅🔊

Nansi inguqulo ewusizo, engeyona eyefilosofi:

• I-TTS endala / yakudala: ngokuvamile ayiyona i-AI (imithetho + ukucutshungulwa kwesignali, noma ukuqoshwa okuthungiwe)

• I-TTS yemvelo yesimanje: ngokuvamile isekelwe ku-AI (amanethiwekhi e-neural / ukufunda komshini) [2]

"Ukuhlolwa kwezindlebe" okusheshayo (hhayi okungenakugwenywa, kodwa okuhle): uma izwi linayo

• ukuphumula kwemvelo

• ukuphimisela okubushelelezi

• isigqi esivumelanayo

• ukugcizelela okuhambisana nencazelo

...kungenzeka ukuthi kuqhutshwa yimodeli. Uma kuzwakala sengathi irobhothi lifunda imigomo nemibandela engaphansi kwendlu ekhanyayo, kungaba izindlela ezindala (noma isabelomali esinqunyiwe... akukho ukwahlulela).

Ngakho-ke... Ingabe i-AI yombhalo ibe yinkulumo? Kwemikhiqizo eminingi yesimanje, yebo. Kodwa i-TTS njengesigaba inkulu kune-AI.

Indlela umbhalo osebenza ngayo inkulumo (ngamazwi abantu), kusukela ku-robotic kuya ku-realistic 🧠🗣️

Izinhlelo eziningi ze-TTS - ezilula noma ezikhangayo - zenza uhlobo oluthile lwale phayiphi:

1. Ukucubungula umbhalo (okwaziwa nangokuthi “yenza umbhalo ukhulumeke”)
   Unweba igama elithi “Dkt.” libe elithi “dokotela,” uphatha izinombolo, izimpawu zokubhala, ama-acronym, futhi uzama ukungesabi.

2. Ukuhlaziywa kolimi
   Kuhlukanisa umbhalo ube yizici zokwakha inkulumo (njengemisindo ,amayunithi omsindo amancane ahlukanisa amagama). Yilapho igama elithi “irekhodi” (ibizo) vs “irekhodi” (isenzo) liba yi-opera ephelele.

3. Ukuhlela i-prosody
   Kukhetha isikhathi, ukugcizelela, ukuphumula, ukunyakaza kwephimbo. I-prosody ngokuyisisekelo umehluko phakathi kwe-"human" kanye ne-"monotone toaster."

4. Ukukhiqizwa komsindo
   Kukhiqiza i-waveform yangempela yomsindo.

Ukuhlukaniswa okukhulu kakhulu kokuthi “i-AI noma cha” kuvame ukuvela ekwakhiweni komsindo we-prosody +. Izinhlelo zesimanje zivame ukubikezela ukumelwa kwe-acoustic okuphakathi (ngokuvamile ama-mel-spectrogram) bese ziguqula lokho kube umsindo kusetshenziswa i- vocoder (futhi namuhla, leyo vocoder ivame ukuba ne-neural) [2].

Izinhlobo eziyinhloko ze-TTS (kanye nalapho i-AI ivame ukuvela khona) 🧪🎙️

1) Ukuhlanganiswa okusekelwe emithethweni / okuhlelekile (i-robotic yakudala)

Ukuhlanganiswa kwesikole sakudala kusebenzisa imithetho eyenziwe ngezandla kanye namamodeli e-acoustic. Kungaqondakala… kodwa kuvame ukuzwakala njenge-alien ehloniphekile. 👽
Akukubi kakhulu, kumane kulungiselelwe imikhawulo ehlukene (ubulula, ukubikezela, ukubalwa kwedivayisi encane).

2) Ukuhlanganiswa okuhlangene (umsindo "wokusika nokunamathisela")

Lokhu kusebenzisa izingcezu zenkulumo eziqoshiwe futhi kuzithunge ndawonye. Kungazwakala kukuhle, kodwa kuyaphuka:

• amagama angavamile angakuphazamisa

• isigqi esingavamile singazwakala singavamile

• Ukushintsha isitayela kunzima

3) I-Neural TTS (yesimanje, eqhutshwa yi-AI)

Izinhlelo ze-neural zifunda amaphethini kusuka kudatha futhi zikhiqiza inkulumo ebushelelezi futhi eguquguqukayo - ngokuvamile zisebenzisa i-mel-spectrogram → ukugeleza kwe-vocoder okukhulunywe ngakho ngenhla [2]. Lokhu ngokuvamile yilokho abantu abakushoyo ngokuthi “izwi le-AI.”

Yini eyenza uhlelo lwe-TTS lube luhle (ngale kokuthi “wow, kuzwakala kungokoqobo”) 🎯🔈

Uma uke wahlola izwi le-TTS ngokuphonsa into efana nale:

"Angishongo ukuthi untshontshe imali."

...bese ulalela ukuthi ukugcizelela kushintsha kanjani incazelo ... usuvele uhlangane nokuhlolwa kwekhwalithi yangempela: ingabe kubamba inhloso, hhayi nje ukuphimisela?

Ukusethwa kwe-TTS okuhle ngempela kuvame ukuzwakala kahle:

• Ukucaca: ongwaqa abaqinile, abangenazo izinhlamvu ezithambile

• I-Prosody: ukugcizelela kanye nesivinini esihambisana nencazelo

• Ukuzinza: akushintshi ubuntu ngokuzumayo phakathi nesigaba

• Ukulawula ukuphimisela amagama: amagama, izifinyezo, amagama ezokwelapha, amagama omkhiqizo

• Ukubambezeleka: uma kusebenzisana, isizukulwane esihamba kancane sizizwa siphukile

• Ukusekelwa kwe-SSML (uma uchwepheshe): amacebiso okuphumula, ukugcizelela, kanye nokuphimisela [1]

• Amalungelo okuthola ilayisense nokusetshenziswa: kuyakhathaza, kodwa kunezinselelo ezinkulu

I-TTS enhle akuyona nje "umsindo omuhle." Ingumsindo osebenzisekayo. Njengezicathulo. Ezinye zibukeka kahle, ezinye zilungele ukuhamba, kanti ezinye zombili (ziyi-unicorn engavamile). 🦄

Ithebula lokuqhathanisa okusheshayo: “Imizila” ye-TTS (ngaphandle kwembobo yentengo yonogwaja) 📊😅

Intengo iyashintsha. Izibali ziyashintsha. Futhi imithetho "yezinga elikhululekile" ngezinye izikhathi ibhalwa njengemfumbe egoqwe kuspredishithi.

Ngakho-ke esikhundleni sokwenza sengathi izinombolo ngeke zisuke ngesonto elizayo, nansi indlela ehlala isikhathi eside:

(Uma ukhetha umzila: awukhethi “izwi elihle kakhulu,” ukhetha indlela yokusebenza. Yileyo ngxenye abantu abayithatha kancane.)

Lokho okushiwo yi-"AI" empeleni ku-TTS yanamuhla 🧠✨

Uma abantu bethi i-TTS “i-AI,” bavame ukusho ukuthi uhlelo lusebenzisa ukufunda komshini ukwenza okukodwa noma ngaphezulu kwalokhu:

• bikezela ubude besikhathi (ukuthi imisindo ihlala isikhathi esingakanani)

• bikezela amaphethini ephimbo/ithoni

• khiqiza izici ze-acoustic (ngokuvamile ama-mel-spectrogram)

• khiqiza umsindo nge-vocoder (ngokuvamile ye-neural)

• ngezinye izikhathi bakwenza ngezigaba ezimbalwa (ukuqala kusukela ekuqaleni kuya ekugcineni) [2]

Iphuzu elibalulekile: I-AI TTS ayifundi izinhlamvu ngokuzwakalayo. Iwukulingisa amaphethini okukhuluma kahle ngokwanele ukuze kuzwakale kuhlosiwe.

Kungani amanye ama-TTS engakabi yi-AI - futhi kungani lokho kungekubi 🛠️🙂

Ama-TTS angewona ama-AI angase abe yisinqumo esifanele uma udinga:

• ukuphimisela okuhambisanayo, okubikezelwayo

• izidingo zokubala eziphansi kakhulu

• ukusebenza okungaxhunyiwe ku-inthanethi kumadivayisi amancane

• ubuhle "bezwi lerobhothi" (yebo, kuyinto ethile)

Futhi: "okuzwakala sengathi kungumuntu kakhulu" akuhlali "kungcono kakhulu." Ngokuqondene nezici zokufinyeleleka, ukucaca + ukungaguquguquki kuvame ukunqoba ukulingisa okudrama.

Ukufinyeleleka kungenye yezizathu ezinhle kakhulu zokuba khona kwe-TTS ♿🔊

Le ngxenye ifanelwe ukugqanyiswa kwayo. Amandla e-TTS:

• izifundi zesikrini zabasebenzisi abayizimpumputhe nabangaboni kahle

• ukwesekwa kokufunda kwe-dyslexia kanye nokufinyeleleka kwengqondo

• izimo ezimatasa ngezandla (ukupheka, ukuhamba, ukukhulisa izingane, ukulungisa uchungechunge lwebhayisikili… uyazi) 🚲

Futhi nansi iqiniso eliyimfihlo: ngisho ne-TTS ephelele ayikwazi ukulondoloza okuqukethwe okungahlelekile.

Okuhlangenwe nakho okuhle kuncike esakhiweni:

• izihloko zangempela (hhayi “umbhalo omkhulu ogqamile ozenza isihloko”)

• umbhalo wesixhumanisi onenjongo (hhayi “chofoza lapha”)

• uhlelo lokufunda olunengqondo

• umbhalo ochazayo ohlukile

Isakhiwo esihlanganisiwe sokufunda izwi se-AI esisezingeni eliphezulu sisayinkimbinkimbi. Kusanda kuxoxwa.

Izimiso zokuziphatha, ukwenziwa kwezwi elifana nelabanye, kanye nenkinga yokuthi “linda - ingabe yibo ngempela?” 😬📵

Ubuchwepheshe bezinkulumo besimanje bunezindlela ezisemthethweni zokusebenzisa. Buphinde budale izingozi ezintsha, ikakhulukazi uma kusetshenziswa amazwi okwenziwa ukuzenza abantu.

Izinhlangano zokuvikela abathengi zixwayise ngokusobala ukuthi abakhwabanisi bangasebenzisa ukwenziwa kwe-AI voice cloning ezinhlelweni "eziphuthumayo zomndeni", futhi batusa ukuqinisekisa ngesiteshi esithembekile kunokuthembela ezwini [5].

Imikhuba ewusizo esisizayo (hhayi ukwesaba, nje… 2025):

• qinisekisa izicelo ezingavamile ngesiteshi sesibili

• setha igama lekhodi yomndeni lezimo eziphuthumayo

• phatha “izwi elijwayelekile” njengelingasabonakali ( okucasulayo, kodwa okungokoqobo)

Futhi uma ushicilela umsindo okhiqizwe yi-AI: ukudalulwa kuvame ukuba ngumqondo omuhle ngisho noma ungaphoqwanga ngokomthetho. Abantu abakuthandi ukukhohliswa. Abakuthandi.

Indlela yokukhetha indlela ye-TTS ngaphandle kokushintshashintsha 🧭😄

Indlela elula yokwenza izinqumo:

Khetha i-TTS yamafu uma ufuna:

• ukusetha okusheshayo nokukala

• izilimi eziningi namazwi

• ukuqapha + ukuthembeka

• amaphethini okuhlanganisa aqondile

Khetha okwasendaweni/okungaxhunyiwe ku-inthanethi uma ufuna:

• ukusetshenziswa ungaxhunyiwe ku-inthanethi

• imisebenzi yokusebenza yobumfihlo-kuqala

• izindleko ezingabikezelwa

• ukulawula okugcwele (futhi ukhululekile ngokushintshashintsha)

Futhi, iqiniso elilodwa elincane: ithuluzi elihle kakhulu ngokuvamile yilelo elifanelana nomsebenzi wakho. Akulona lelo elinesiqeshana sedemo esimnandi kakhulu.

Ngamafuphi: Ingabe i-AI yokubhala umbhalo ibe inkulumo? 🧾✨

• Umsebenzi wokuguqula umbhalo ube yinkulumo: ukuguqula umbhalo obhaliwe ube umsindo okhulunywayo.

• I-AI iyindlela evamile esetshenziswa kuma-TTS anamuhla, ikakhulukazi kubantu abasebenzisa ama-realistic voices.

• Umbuzo uyinkimbinkimbi ngoba i-TTS ingakhiwa nge-AI noma ngaphandle kwayo.

• Khetha ngokusekelwe kulokho okudingayo: ukucaca, ukulawula, ukubambezeleka, ubumfihlo, ilayisense… hhayi nje ukuthi “wow, kuzwakala kungumuntu.”

• Futhi uma kubalulekile: qinisekisa izicelo ezisekelwe ezwini bese udalula umsindo wokwenziwa ngendlela efanele. Ukwethembana kunzima ukukuzuza futhi kulula ukukuqeda.

Isibonelo sangempela: Ukwakha umsebenzi we-TTS wesifundo esiku-inthanethi

Isimo

Cabanga ngomdali wezifundo omncane oku-inthanethi ofuna ukuguqula amanothi ezifundo abhaliwe abe yizinguqulo ezimfushane zomsindo zabafundi abakhetha ukulalela ngenkathi beya emsebenzini noma bebuyekeza. Lokhu kuyilungiselelo eliqanjiwe kodwa elingokoqobo: umdali oyedwa, izifundo ezingu-20, ngasinye sicishe sibe amagama angu-1,200, eshicilelwe kusayithi lokufunda lamalungu kuphela.

Umgomo awukona “ukukopisha” izwi lothisha noma ukwenza sengathi umsindo uqoshwa bukhoma. Umgomo ulula: ukulandisa kwesifundo okucacile nokuvumelanayo okulandela isakhiwo esibhaliwe, ukuphimisa amagama abalulekile ngendlela efanele, futhi kungahlolwa ngaphambi kokushicilela.

Ngenxa yokuthi lesi sihloko sesivele sichaza ifu uma kuqhathaniswa nokukhetha kwendawo, lesi sibonelo sisebenzisa indlela ehlanganisiwe: i-cloud TTS yomsindo wokugcina womphakathi, kanye ne-local/offline TTS ye-private drafts lapho umdali esahlela khona izinto zesifundo ezibucayi.

Okudingwa umsebenzi wokusebenza

• Hlanza umbhalo wesifundo ngezihloko ezifanele, amaphuzu ezinhlamvu, kanye nezigaba ezimfushane

• Uhlu lokuphimisela amagama, izifinyezo, kanye namagama obuchwepheshe

• Inothi lokudalula, njenge: “Inguqulo yomsindo ekhiqizwe ngombhalo ube yinkulumo futhi ibuyekezwe ngaphambi kokushicilelwa”

• Uhlu lokuhlola olulula lokubuyekeza ukuze kucace, kuphinyiselwe amagama, kuhambisane nesivinini, kanye nezigaba ezingekho

• Isitayela se-SSML esingakhethwa silawula uma ithuluzi elikhethiwe lisekela ukumiswa okwesikhashana, ukugcizelela, noma amacebiso okuphimisela

• Isinyathelo sokuvunyelwa ngumuntu ngaphambi kokuba umsindo udlale

Isibonelo semiyalelo

Sebenzisa lo myalelo lapho ulungiselela isifundo ngasinye se-TTS:

Guqula lesi sifundo sibe umbhalo-umbhalo ube yinkulumo ukuze kube nokulandisa okucacile kwezemfundo. Gcina incazelo ingashintshi, kodwa yenza kube lula ukulizwa ngokuzwakalayo. Hlukanisa imisho emide ibe mifushane. Maka lapho kufanele kube khona ukumiswa okufushane ngemva kwezihloko zesigaba. Maka noma yimaphi amagama angadinga ukubuyekezwa kokuphimisela, ikakhulukazi amagama, izifinyezo, amagama obuchwepheshe, noma amagama omkhiqizo. Ungangezi amaqiniso amasha. Ekugcineni, faka uhlu olufushane lwezinto umuntu okufanele azilalele ngaphambi kokushicilela.

Indlela yokuyihlola

Ngaphambi kokukhiqiza zonke izifundo ezingu-20, hlola izikripthi ezintathu zesampula:

1. Isifundo esisodwa esilula ngolimi olucacile

2. Isifundo esisodwa sobuchwepheshe esinama-afinyezo namagama angavamile

3. Isifundo esisodwa esinohlu, izihloko, nezixhumanisi ezingase zizwakale zingakhululekile uma zifundwa ngokuzwakalayo

Kuvivinyo ngalunye, lalela kanye ngaphandle kokufunda umbhalo, bese ulalela futhi ngenkathi ulandela isifundo esibhaliwe. Maka:

• Amagama aphinyiselwe kabi

• Imisho emide kakhulu ukuthi ingalandelwa endlebeni

• Izihloko ezingazwakali zihlukile ngokwanele

• Ukuphumula okungekho

• Noma iyiphi indawo lapho izwi lizwakala linamandla kakhulu, lithambile kakhulu, noma lidukisa

Umphumela omuhle uzwakala njengomlandisi ocacile oqondisa umfundi esifundweni. Umphumela omubi uzwakala njengomuntu ofunda ikhasi lewebhu ngaphandle kokuqaphela ukuthi izigaba, izibonelo, kanye nezixwayiso ziqala noma ziphela kuphi.

Umphumela

Umphumela obonisayo: Ngokusekelwe ekubekeni isikhathi izifundo ezintathu zesampula ngaphambi nangemva kokusebenzisa lo msebenzi.

Ngaphambi komsebenzi, ukulungiselela isifundo esisodwa samagama ayi-1,200 somsindo kwathatha cishe imizuzu engama-55: imizuzu engama-20 ukuhlanza umbhalo, imizuzu eyi-15 ukulungisa amagama angajwayelekile, imizuzu eyi-10 yokuvuselela umsindo, kanye nemizuzu eyi-10 yokubuyekeza indlela yokubiza amagama.

Ngemva kokudala uhlu lokuhlola lokubhala kanye nohlu lokubiza amagama olungasetshenziswa kabusha lwe-TTS, umsebenzi ofanayo uthathe cishe imizuzu engama-25 ngesifundo ngasinye: imizuzu engu-8 ukulungiselela umbhalo, imizuzu engu-7 ukukhiqiza umsindo, kanye nemizuzu eyi-10 yokubuyekezwa komuntu.

Ezifundweni ezingu-20, lokho kunganciphisa isikhathi sokukhiqiza kusukela emahoreni angaba ngu-18 kuya emahoreni angaba ngu-8 nemizuzu engu-20, okulinganiselwa ukuthi kuzokongwa amahora angu-9 nemizuzu engu-40. Umdali angakuqinisekisa lokhu ngokubeka isikhathi sesifundo ngasinye, ebala ukulungiswa kokuphimisela amagama, futhi alandelele ukuthi mangaki amafayela omsindo okudingeka avuselelwe ngaphambi kokuvunyelwa.

Yini engase ihambe kabi

Iphutha elivame kakhulu ukuphatha umsindo ongokoqobo njengolungile ngokwemvelo. Izwi lemvelo lisengalifunda kabi igama, leqe umongo, ligcizelele ngokweqile ibinzana elingalungile, noma lenze kube nzima ukulandela incazelo yobuchwepheshe.

Ubumfihlo bungenye ingozi. Izifundo ezisalungiswa, izibonelo zabafundi, noma izinto zezifundo ezikhokhelwayo akufanele zithunyelwe ethuluzini lamafu ngaphandle kokuthi umdali uhlole idatha yethuluzi kanye nemigomo yokugcina. Kuma-draft abucayi, ama-TTS endawo angase aphephe ngisho noma izwi lokugcina lingacacile kahle.

Kukhona futhi inkinga yokuthembana. Uma ikhosi isebenzisa ukulandisa okwenziwe ngezinto, abafundi akufanele bakholelwe ukuthi kuyirekhodi elibukhoma lomuntu. Ukudalulwa okufushane kugcina okulindelwe kucacile.

Ukudla okuwusizo

Ukuhamba komsebenzi okuhle kwe-TTS akukhona nje "ukunamathisela umbhalo, thola umsindo". Inguqulo enamandla ifaka isakhiwo esihlanzekile, ukulawula ukuphimisela amagama, ukubuyekezwa komuntu, kanye nokuhlolwa kwekhwalithi okulinganiselwe. Lowo umehluko phakathi komsindo owenziwe nge-AI ozwakala uwusizo kanye nomsindo owenziwe nge-AI ozwakala umangalisa imizuzwana eyi-10 yokuqala.

Imibuzo Evame Ukubuzwa

Ingabe i-AI yombhalo ibe yinkulumo, noma ingabe uhlelo olujwayelekile nje?

Umgomo wokusebenzisa umbhalo ube yinkulumo (TTS): ukuguqula umbhalo obhaliwe ube umsindo okhulunywayo. Ukuthi ngabe “yi-AI” kuncike endleleni esetshenziswa ngaphansi kwe-hood. Izinhlelo ezindala zingasekelwa emithethweni noma zithungwe ndawonye izingcezu eziqoshiwe, kuyilapho amazwi emvelo anamuhla evame ukuqhutshwa ukufunda komshini. Uma udinga ukuqiniseka, gxila kubuchwepheshe obusetshenziswayo kunokwahlulela ngomsindo kuphela.

Uma abantu bebuza ukuthi “Ingabe i-Text to Speech AI iyindlela yokubhala inkulumo,” yini abayibuzayo ngempela?

Esikhathini esiningi, babuza, “Ingabe kukhiqizwa imodeli yokufunda komshini?” noma “Ingabe kufundwe ukuzwakala njengomuntu kusuka kudatha?” Yingakho umbuzo ungazwakala ulula: I-TTS iyisigaba, hhayi indlela eyodwa. Emikhiqizweni eminingi yesimanje, amazwi emvelo kakhulu asekelwe ku-AI, kodwa kusenezindlela ezingezona eze-AI ezihlala zithembekile futhi zisebenza.

Ngingazi kanjani ukuthi izwi le-TTS likhiqizwa yi-AI ngokulalela nje?

"Ukuhlolwa kwezindlebe" kungasiza, kodwa akulona iphutha. Uma izwi liphethe ukuphumula kwemvelo, isigqi esibushelelezi, kanye nokugcizelela okulandela incazelo, cishe liqhutshwa yimodeli. Uma lizwakala liyisicaba, lihlukaniswe ngokuqinile, noma likhubeka ngenxa yokubekwa kwamagama, kungaba izindlela zokuhlanganisa ezindala noma isilungiselelo sekhwalithi ephansi. Ukuqinisekiswa okungcono kakhulu kusahlola indlela ebhalwe phansi yesistimu.

Isebenza kanjani ngempela i-AI text to speech yesimanje?

Izinhlelo eziningi zilandela umzila: zenza umbhalo ukhulumeke, zihlaziye amayunithi okuphimisela, zihlele i-prosody, bese zikhiqiza umsindo. Ukuhlukaniswa okukhulu kakhulu kwe-“AI vs not” kuvame ukuvela ekuhleleni i-prosody kanye nokukhiqiza umsindo. Izinhlelo eziningi zesimanje zibikezela izici ze-acoustic eziphakathi (ngokuvamile ama-mel-spectrogram) bese ziguqulwa zibe umsindo nge-vocoder. Ezinhlelweni eziningi namuhla, leyo vocoder i-neural.

Ingabe kufanele ngisebenzise i-cloud TTS noma ngisebenzise i-TTS endaweni yami kuphrojekthi yami?

Khetha ifu uma ufuna ukusethwa okusheshayo, ukukala okulula, imenyu yezwi ebanzi nolimi, kanye namaphethini okuthembeka azinzile. Ama-API efu avame ukukalwa ngevolumu yombhalo kanye nezinga lezwi, ngakho-ke izindleko zingakhuphuka ngokusetshenziswa. Khetha i-TTS yendawo/engaxhunyiwe ku-inthanethi lapho ubumfihlo, ukusebenza okungaxhunyiwe ku-inthanethi, kanye nokusetshenziswa kwemali okubikezelwayo kubaluleke kakhulu kunokukhululeka kwe-plug-and-play. Indlela ehlanganisiwe ingakunikeza ikhwalithi yefu ngokubuyiselwa okungaxhunyiwe ku-inthanethi.

Iyiphi indlela engcono kakhulu yokwenza i-TTS isebenze kahle ukuze kufinyeleleke kumawebhusayithi noma kumadokhumenti?

I-TTS enamandla incike esakhiweni esihlanzekile, hhayi nje izwi "eliphezulu". Sebenzisa izihloko zangempela (hhayi nje umbhalo ogqamile omkhulu), umbhalo wesixhumanisi onenjongo, kanye nohlelo lokufunda olunengqondo. Engeza umbhalo ochazayo ukuze izithombe zingaphenduki zibe yizikhala ezithule, futhi ugweme amaqhinga okubeka aphazamisa indlela okuqukethwe okufundwa ngayo ngokuzwakalayo. Ngisho ne-TTS enhle kakhulu ayikwazi ukuxazulula isakhiwo esibi - izomane ilandise ama-tangles.

Ngingayinciphisa kanjani ingozi yokukhwabanisa ngokushintsha izwi noma izingcingo mbumbulu "eziphuthumayo zomndeni"?

Phatha izwi elijwayelekile njengelingaselona ubufakazi obuqinisekile ngokwalo. Umkhuba osebenzayo ukuqinisekisa izicelo ezingavamile ngesiteshi sesibili, njengokuthumela umyalezo ngenombolo eyaziwayo noma ukushaya ucingo ngendlela yokuxhumana ethembekile. Abantu abaningi baphinde babeke igama elilula lekhodi yomndeni ezimweni eziphuthumayo. Umgomo awukona ukwesaba - kuyisinyathelo sokuqinisekisa esisheshayo lapho izingozi ziphezulu.

Iyini i-SSML, futhi kufanele ngiyisebenzise nini ngombhalo ube inkulumo?

I-SSML iyindlela yokunikeza uhlelo lwe-TTS amacebiso engeziwe mayelana nendlela yokukhuluma umbhalo. Ingasiza ngokuma kancane, ukugcizelela, kanye nokuphimisela, ikakhulukazi amagama, izifinyezo, noma amagama obuchwepheshe. Uma wakha into exhumanayo noma ezwela uphawu, i-SSML ingathuthukisa ukuvumelana futhi inciphise ukufunda okungajwayelekile. Iwusizo kakhulu uma ukuphimisela okuzenzakalelayo kuseduze, kodwa kungasondele ngokwanele.

Izinkomba

1. I-W3C - Ulimi Lokumaka Lokuhlanganiswa Kwenkulumo (i-SSML) Inguqulo 1.1 - funda kabanzi

2. UTan et al. (2021) - Ucwaningo Lokwenziwa Kwenkulumo Yezinzwa (i-arXiv PDF) - funda kabanzi

3. I-Google Cloud - Intengo yombhalo ube inkulumo - funda kabanzi

4. I-OHF-Voice - I-Piper (injini ye-TTS ye-neural yendawo) - funda kabanzi

5. I-US FTC - Abakhwabanisi basebenzisa i-AI ukuthuthukisa izinhlelo "eziphuthumayo zomndeni" - funda kabanzi

Thola i-AI Yakamuva Esitolo Esisemthethweni Somsizi we-AI

Mayelana NATHI