Ngiyaqiniseka ukuthi sekwenzekile kuwe: unerekhodi elide lengxoxo noma ikilasi, futhi umcabango wokukubhala konke ubonakala njengomsebenzi omkhulu. Iqiniso liwukuthi, ukubhala ngesandla kuwukudonswa kwangempela futhi kuwukuchitha isikhathi okukhulu, ikakhulukazi lapho amathuluzi avamile amahhala edidaniswa namagama noma esungula idatha engekho.
Yilapho i-Whisper ingena khona, i-AI enamandla eyakhiwe yi-OpenAI elapha ukushintsha umdlalo. Akuyona nje enye uhlelo, kodwa imodeli yokuqaphela inkulumo ngokuzenzakalelayo (i-ASR) ehlaziya umsindo ngokunemba okumangalisayo, okusivumela ukuthi sithole umbhalo ohlelekayo ngemizuzu embalwa ngaphandle kokuhlupheka.
Iyini ngempela i-Whisper futhi isebenza kanjani ngaphakathi?
Kalula nje, i-Whisper uhlelo lwe-AI oluklanyelwe ngqo ukuguqula inkulumo ibe umbhalo. Ngokungafani nezinye izinhlelo zokusebenza ezivame ukwehluleka, lo modeli uhlaziya amaphethini omsindo, umongo, kanye nezici zesipikha ukuze ulethe umphumela wobungcweti. Enguqulweni yayo ye-v3, iqeqeshwe kumahora angaphezu kwesigidi omsindo, idlula kakhulu amahora angu-680.000 enguqulo yangaphambilini, okuholele ekunciphiseni kwamaphutha ngo-10% kuya ku-20%.
Enye yezinto ezimangalisa kakhulu ikhono layo lokusingatha iSpanishi, lapho inezinga lamaphutha elingaphansi kuka-5% . Ngaphezu kwalokho, ayinciphi nezinye izilimi; ingabhala izilimi ezingaphezu kwe-100, okuhlanganisa isiCatalan, isiBasque, isiGalician, isiJalimane, isi-Arabhu, nesiJapane, futhi ingakwazi ukubona ngisho nalapho isikhulumi sishintsha izilimi phakathi nomusho.
Izici zobuchwepheshe kanye nokuguquguquka
I-Whisper akuyona uhlelo lokusebenza oluvaliwe, kodwa imodeli yolimi esebenza njengesisekelo sezinye izinkampani ukudala amathuluzi azo zisebenzisa i-API. Ukuze ivumelane nanoma iyiphi ihadiwe, i-OpenAI iyikhiphe ngobukhulu obuhlukahlukene kuye ngamakhono omshini . Kunezinguqulo ezilula ezidinga ngaphansi kwe-1 GB ye-VRAM futhi zinamapharamitha ayizigidi ezingu-39, kuze kufike kumamodeli amakhulu anamapharamitha ayizigidigidi ezingu-1.550 adinga cishe i-10 GB ye-VRAM ukuze isebenze ngokugcwele.
Enye inzuzo enkulu yikhono layo lokuhumusha ukumiswa kwemvelo engxoxweni. Lokhu kusho ukuthi i-AI iyazi ukuthi kufanele ibeke kuphi i-comma noma i-period ngokusekelwe ekuthuleni kwesipikha, okusisindisela umsebenzi omningi wokuhlela kamuva. Ilungele ukuphatha imihlangano, ama-podcast, noma izingxoxo nabahlanganyeli abaningi, njengoba ihlonza futhi ihlukanise izipikha ngokuzenzakalelayo.
Indlela yokuyiqalisa: kusukela kwezobuchwepheshe kuya kokulula
Uma ungumuntu othanda ikhompyutha, ungaya ngqo ekhasini labo le-GitHub, ulande ikhodi yomthombo ovulekile, bese uyiqhuba endaweni yakho emshinini wakho ulandela imiyalelo yobuchwepheshe. Lokhu kufana kakhulu nokusetha umsizi wokuhlela we-AI , njengoba kudinga ulwazi oluthuthukisiwe futhi akuyona into elula kumuntu ongazi ukuthi uhlela kanjani.
Uma ungafuni ukwenza izinto zibe nzima, kunezinye izindlela ezilula kakhulu. Isibonelo, ungasebenzisa ipulatifomu ethi Replicate (replicate.com/openai/whisper) , ekuvumela ukuthi ulayishe amafayela akho bese ukhetha ukuthi iyiphi inguqulo yemodeli ofuna ukuyisebenzisa (njenge-v3) ngaphandle kokufaka noma yini kukhompyutha yakho. Ngale ndlela, noma ubani angasebenzisa amandla e-AI ukuthola uhlaka oluhlelekile ngesikhathi esinqunyiwe.
Ukuba nethuluzi elinjengaleli eduze kuyimpumuzo, ikakhulukazi kubanyathelisi noma abafundi ababevame ukuchitha amahora belalele isiqeshana somsindo esifanayo. Manje, njengoba benekhono lokucubungula izwi ngesikhathi sangempela futhi ngokunemba okucishe kube ukuhlinzwa , ukubhalwa kwezwi sekusuke emsebenzini oyisicefe kwaba inqubo ezenzakalelayo nephumelelayo esivumela ukuthi sigxile kulokho okubalulekile: ukuhlaziya okuqukethwe.