Isibuyekezo se-Mozilla Common Voice 7.0

I-NVIDIA ne-Mozilla bakhiphe isibuyekezo kudathasethi yabo Yezwi Elivamile, ehlanganisa amasampula enkulumo yabantu abangu-182, okukhuphuke ngo-25% kusukela ezinyangeni ezingu-6 ezedlule. Idatha ishicilelwa njengesizinda somphakathi (CC0). Amasethi ahlongozwayo angasetshenziswa kumasistimu okufunda omshini ukuze akhe ukunakwa kwenkulumo namamodeli okuhlanganiswa.

Uma kuqhathaniswa nesibuyekezo sangaphambilini, usayizi wento yenkulumo eqoqweni ukhuphukile ukusuka ku-9 kuya ku-13.9 wamahora wokukhuluma ayizinkulungwane. Inani lezilimi ezisekelwayo lenyuke lisuka ku-60 laya ku-76, okuhlanganisa okokuqala nokusekelwa kwezilimi zaseBelarusian, Kazakh, Uzbek, Bulgarian, Armenian, Azerbaijani kanye neBashkir. Isethi yolimi lwesiRashiya ihlanganisa abahlanganyeli abangu-2136 namahora angu-173 wezinto zokukhuluma (kwakunabahlanganyeli abangu-1412 namahora angu-111), futhi ngolimi lwase-Ukraine - abahlanganyeli abangu-615 namahora angu-66 (kwakunabahlanganyeli abangu-459 namahora angu-30).

Abantu abangaphezu kwezinkulungwane ezingama-75 babambe iqhaza ekulungiseleleni izinto ngesiNgisi, besho amahora angama-2637 enkulumo eqinisekisiwe (kwakukhona ababambiqhaza abayizinkulungwane ezingama-66 namahora angama-1686). Kuyathakazelisa ukuthi ulimi olusendaweni yesibili ngokwenani ledatha eqoqwe yiRwanda, okuqoqwe kuyo amahora angama-2260. Lokhu kulandelwa isiJalimane (1040), isiCatalan (920) nesi-Esperanto (840). Phakathi kwezinto ezikhula ngamandla usayizi wedatha yezwi ulimi lwesiThai (ukwanda okuphindwe ka-20 kwesisekelo, kusuka emahoreni ayi-12 kuye kwangama-250), isiLuganda (kusuka kwamahora angama-8 kuye kwangama-80), isi-Esperanto (kusuka kwamahora ayi-100 kuye kwangama-840) nesiTamil ( kusuka emahoreni angama-24 kuye kwangama-220).

Njengengxenye yokubamba iqhaza kwayo kuphrojekthi Yezwi Elivamile, i-NVIDIA ilungiselele amamodeli aqeqeshiwe asevele enziwe amasistimu okufunda ngomshini (asekelwa yi-PyTorch) ngokusekelwe kudatha eqoqiwe. Amamodeli asatshalaliswa njengengxenye yekhithi yamathuluzi yamahhala nevulekile ye-NVIDIA NeMo, okuthi, isibonelo, isivele isetshenziswa ezinsizeni zezwi ezizenzakalelayo ze-MTS ne-Sberbank. Amamodeli enzelwe ukusetshenziswa ekunakeni kwenkulumo, ekuhlanganiseni kwenkulumo, nasezinhlelweni zokucubungula ulimi lwemvelo, futhi angase abe usizo kubacwaningi abakha amasistimu engxoxo asebenza ngezwi, izinkundla zokuloba, nezikhungo zezingcingo ezizenzakalelayo. Ngokungafani namaphrojekthi atholakala ngaphambilini, amamodeli ashicilelwe awagcini nje ekuqashelweni kolimi lwesiNgisi futhi ahlanganisa izilimi ezihlukahlukene, iziphimiso kanye nezinhlobo zokukhuluma.

Ake sikukhumbuze ukuthi iphrojekthi Yezwi Elivamile ihloselwe ukuhlela umsebenzi ohlanganyelwe ukuze kuqoqwe isizindalwazi samaphethini ezwi acabangela ukuhlukahluka kwamazwi nezitayela zokukhuluma. Abasebenzisi bayamenywa ukuthi bakhulume imishwana yezwi eboniswe esikrinini noma bahlole ikhwalithi yedatha engezwe abanye abasebenzisi. Isizindalwazi esiqoqiwe esinamarekhodi okuphimisela okuhlukahlukene kwemishwana evamile yenkulumo yomuntu ingasetshenziswa ngaphandle kwemikhawulo ezinhlelweni zokufunda zomshini kanye namaphrojekthi ocwaningo.

Ngokusho kombhali womtapo wolwazi weVosk oqhubekayo wokuqashelwa kwenkulumo, ububi besethi ye-Common Voice wuhlangothi olulodwa lwezwi lezwi (ubukhulu babantu besilisa abaneminyaka engama-20-30 ubudala, kanye nokuntuleka kwezinto ezinamazwi abesifazane. , izingane kanye nabantu abadala), ukuntuleka kokuhlukahluka kusichazamazwi (ukuphindaphindwa kwemisho efanayo) kanye nokusatshalaliswa kokurekhodiwe ngefomethi ye-MP3 ehlanekezelwe.

Source: opennet.ru

Engeza amazwana