Mozilla yatulutsa zosintha zamaseti ake a Common Voice, zomwe zimaphatikizapo zitsanzo zamalankhulidwe kuchokera kwa anthu pafupifupi 200. Zambiri zimasindikizidwa pagulu la anthu (CC0). Maguluwa atha kugwiritsidwa ntchito pamakina ophunzirira kuti apange kuzindikira kwamawu ndi mitundu yophatikizika. Poyerekeza ndi zosintha zam'mbuyomu, kuchuluka kwa zolankhula zosonkhanitsidwa kwawonjezeka ndi 30% —kuchokera pa 13.9 mpaka 18.2 olankhulira. Chiwerengero cha zilankhulo zothandizidwa chawonjezeka kuchoka pa 67 mpaka 87.
Chilankhulo cha Chirasha chinaphatikizapo ophunzira a 2452 ndi maola a 193 olankhula (omwe kale anali 2136 ndi maola 173), chinenero cha Chibelarusi chinaphatikizapo 6160 otenga nawo mbali ndi maola 987 (omwe kale anali 3831 ndi maola 356), ndipo chinenero cha Chiyukireniya chimaphatikizapo maola 68 ndi maola 64 otenga nawo mbali ndi maola 66). Anthu opitilira 79,000 adatenga nawo gawo pokonzekera zida zachingerezi, kulamula maola 2886 amalankhulidwe otsimikizika (omwe kale anali otenga nawo gawo 75,000 ndi maola 2637).
Monga chikumbutso, pulojekiti ya Common Voice ikufuna kulinganiza zoyesayesa zogwirira ntchito kuti apeze nkhokwe ya ma tempulo a mawu omwe amajambula kusiyanasiyana kwa mawu ndi kalankhulidwe. Ogwiritsa ntchito amapemphedwa kuti azilankhula mawu omwe akuwonetsedwa pazenera kapena kuvotera mtundu wa data yomwe yawonjezeredwa ndi ogwiritsa ntchito ena. Zosungirako zomwe zasonkhanitsidwa, zomwe zimakhala ndi matchulidwe osiyanasiyana a mawu wamba a anthu, zitha kugwiritsidwa ntchito popanda malire pamakina ophunzirira makina ndi ntchito zofufuza. Malinga ndi Vosk, mlembi wa laibulale yozindikira zolankhula mosalekeza, zolephera za Common Voice dataset zikuphatikiza kukondera kwa mawu (kuchulukira kwa amuna azaka za m'ma 20 ndi 30, kusowa kwa zinthu zomwe zili ndi azimayi, ana, ndi okalamba), kusowa kwa mawu osiyanasiyana (kubwerezabwereza kwa mawu a MP), komanso kugawa kwa mawu omwewo.
Chochititsanso chidwi ndi kutulutsidwa kwa NVIDIA NeMo 1.6, chida chomwe chimapereka njira zophunzirira zamakina zopangira kuzindikira kwamawu, kaphatikizidwe ka mawu, ndi makina osinthira zilankhulo zachilengedwe. NeMo imaphatikizapo zitsanzo zokonzeka kugwiritsidwa ntchito, zophunzitsidwa kale zamakina ophunzirira makina motengera dongosolo la PyTorch, lopangidwa ndi NVIDIA pogwiritsa ntchito mawu a Common Voice komanso kuphimba zilankhulo zosiyanasiyana, katchulidwe ka mawu, ndi mitundu yolankhulira. Zitsanzozi zitha kukhala zothandiza kwa ofufuza omwe akupanga njira zoyankhulirana zoyatsidwa ndi mawu, nsanja zolembera, ndi malo oimbira mafoni. Mwachitsanzo, NVIDIA NeMo imagwiritsidwa ntchito pamawu omvera pa MTS ndi Sberbank. NeMo imalembedwa mu Python pogwiritsa ntchito PyTorch ndipo ili ndi chilolezo pansi pa chilolezo cha Apache 2.0.
Source: opennet.ru
