MAI-Transcribe-2 Uyasebenza

I-MAI-Transcribe-2 iyimodeli enamandla kakhulu ye-Microsoft yokuguqula inkulumo ibe umbhalo isebenzisa i-AI, ehamba phambili ngokunemba, ngesivinini nangezindleko. Imodeli isekela izilimi ezingu-60, kanti isilinganiso se-word error rate esivivinyweni se-FLEURS singu-5.2% kuphela. Imodeli manje isinokuhlukaniswa kwezikhulumi nezitembu zesikhathi zegama negama, futhi inikeza izitayela ezimbili zokubhala: i-verbatim, egcina amaphutha okukhuluma, kanye ne-clean, esusa amagama okugcwalisa. Isekela nokutholwa kolimi okuzenzakalelayo nokushintshwa kwezilimi.

Ukubhala inkulumo ngezilimi eziningi: Isekela izilimi ezingu-60 futhi ingathola ngokuzenzakalelayo ulimi lokurekhodiwe ngaphandle kokucacisa ulimi kusengaphambili.
Ukuhlukaniswa kwezikhulumi: Ihlukanisa izikhulumi ezahlukene ekurekhodiwe okukodwa futhi inikeze umbhalo obhaliwe kusikhulumi esifanele.
Izitembu zesikhathi zegama negama: Ibeka indawo yesikhathi enembile yegama ngalinye, okwenza kube lula ukuvumelanisa imibhalo engezansi, ukusesha umsindo nokuhlela.
Izitayela zokubhala ezilungisekayo: Imodi ye-verbatim igcina amagama okugcwalisa namaphutha okukhuluma, kanti imodi ye-clean iyawasusa ukuze ikhiqize umbhalo ofundeka kalula.
Ukutholwa nokushintshwa kolimi okuzenzakalelayo: Isekela izingxoxo ezixuba izilimi ngokwemvelo, ithole futhi ivumelane nokushintshwa kolimi ngokuzenzakalelayo.
Ukubekwa phambili kwamagama angukhiye: Ungasetha uhlu lwamagama omkhakha ukuze uthuthukise ukunemba kokubona amagama athile.
Ukubhala ezindaweni ezinomsindo: Iyakwazi ukugcina ukunemba okuphezulu lapho kunomsindo ongemuva.

Isakhiwo esihlanganisiwe esisuka ekuqaleni kuye ekugcineni: Isebenzisa inethiwekhi eyodwa ye-neural ukwenza ngokuqondile ukuguqulwa kwezici zomsindo zibe uchungechunge lombhalo. Ihlanganisa ukubona inkulumo, ukuhlukaniswa kwezikhulumi nokutholwa kolimi kumodeli eyodwa, igweme ukunqwabelana kwamaphutha kwezinhlelo ezixhunyiwe futhi ifinyelele ukusebenza okuhle kakhulu ngesivinini nangokunemba.
Ukuqeqeshwa okuhlangene kwezilimi eziningi: Iqeqeshwa ngedatha enkulu yenkulumo nombhalo evela ezilimini ezingu-60. Imodeli ifunda izethulo zomsindo nemithetho yolimi eyabiwa yizilimi, ngakho ayigcini nje ngokumboza izilimi ezisetshenziswa kakhulu, kodwa igcina nokunemba okuphezulu ezilimini ezinezinsiza ezilinganiselwe futhi ikwazi ukuthola ukushintshwa kolimi ngokwemvelo ngesikhathi sangempela.
Ukuqopha okuqaphela umongo: Ngesikhathi sokuqopha, imodeli ayibikezeli isebenzisa uhlaka lomsindo lwamanje kuphela; ihlanganisa incazelo yomongo womhlaba wonke nokubekwa phambili kwamagama angukhiye angaphandle ukuze yenze ukucabanga okuhlangene. Ilungisa ngokuguquguqukayo ukusabalala kwamathuba kwamagama aphumayo, ithuthukise ukubonwa kwamagama ochwepheshe futhi isekele ukukhiqizwa okulawulekayo kwezitayela ze-verbatim ne-clean.

Dala insiza ye-Azure: Kuphothali ye-Azure, dala insiza ye-AI Speech ukuze uthole ukhiye we-API nekheli le-endpoint lesifunda.
Xhuma ku-Foundry preview: Sebenzisa indawo yokungena ye-Microsoft Foundry public preview ethi https://ai.azure.com/catalog/models/MAI-Transcribe-2 ukuze ungene esibonakalayo sokusebenzisa imodeli.
Lungisa amapharamitha emodeli: Esicelweni se-API, cacisa model: "MAI-Transcribe-2" futhi uvule ukuhlukaniswa kwezikhulumi nezitembu zesikhathi zegama negama uma kudingeka.
Layisha umsindo ukuze ubhalwe: Thumela amafayela omsindo afana ne-WAV/MP3/MP4; imodeli izothola ulimi ngokuzenzakalelayo bese ibhala okuqukethwe.
Thola umphumela ohlelekile: Yamukela i-JSON ebuyiswayo equkethe umbhalo ophelele, izitembu zesikhathi zegama negama, amalebula ezikhulumi kanye namakhodi ezilimi atholwe ngokuzenzakalelayo.

Ukunemba okuhamba phambili emhlabeni: Isilinganiso se-word error rate ezilimini ezingu-60 ze-FLEURS singu-5.2% kuphela, kanti esiShayina sifinyelela ku-4.5%; idlula i-Gemini 3.1 Pro ne-Whisper v3-Large.
Isivinini esishesha kakhulu embonini: Isebenza ngesivinini esiphindwe ngo-403.6 kunesikhathi sangempela; umsindo wehora elingu-1 uqedwa cishe ngemizuzwana engu-9, okushesha ngokuphindwe ka-10 kune-GPT-Transcribe.
Izindleko eziphansi kakhulu embonini: Intengo yesikhathi esilinganiselwe ingu-0.10 USD/ihora, okuyehle cishe ngo-72% uma kuqhathaniswa nesizukulwane sangaphambilini.
Imisebenzi eyakhelwe ngaphakathi: Ihlanganisa ukuhlukaniswa kwezikhulumi, izitembu zesikhathi zegama negama, ukutholwa nokushintshwa kolimi okuzenzakalelayo kanye nokubhala ngezitayela ze-verbatim/clean, ngaphandle kokucubungula okwengeziwe.
Ukumbozwa kwezilimi eziningi ngaphandle komthungo: Isekela izilimi ezingu-60, ithole ngokuzenzakalelayo futhi ivumelane nezingxoxo ezixuba izilimi ngokwemvelo ngaphandle kokucacisa ulimi kusengaphambili.

Ukuhlolwa nokuhlaziywa kwekhwalithi ezikhungweni zezingcingo: Sebenzisa ukuhlukaniswa kwezikhulumi ukwehlukanisa abasebenzi basekhasimendeni namakhasimende, bese usebenzisa izitembu zesikhathi zegama negama ukukala isilinganiso senkulumo nokuthola ngokunembile izikhathi ezibalulekile.
Ukubhala ngobuhlakani imihlangano nezingxoxo: Guqula ukurekhodwa kwabantu abaningi kube umbhalo ohlelekile onamalebula ezikhulumi, bese unikeza ngokuzenzakalelayo izinto okufanele zenziwe nemithombo yezinqumo.
Ukubhala bukhoma kwama-agent ezwi: Yisebenzise njengengqimba yokufaka yokuzwa enokubambezeleka okuphansi, ihlanganiswe ne-MAI-Voice-2 Flash ukuze kwakhiwe umjikelezo wengxoxo yesikhathi sangempela osuka ezwini uye ezwini.
Imibhalo engezansi yemidiya nokwenziwa kwasendaweni: Izitembu zesikhathi zegama negama zehlisa kakhulu izindleko zokuvumelanisa imibhalo engezansi, kanti ukusekelwa kwezilimi ezingu-60 kusiza ukukhiqizwa kokuqukethwe komhlaba wonke.
Ubufakazi bokuhambisana nemithetho, umthetho nokulawulwa: Khiqiza amarekhodi anembile anokunikezwa kwesikhulumi nezitembu zesikhathi ukuze kuhlangatshezwane nezidingo zokucwaninga nokugcina ubufakazi ezimbonini zezezimali, ezempilo nakwezinye.