21.500 hours CC-BY 4.0 Training Indimi zose

Insanganyamatsiko

voice audionlp

VAANI ni urutonde runini rw'amajwi y'indimi nyinshi rwakozwe n'Ikigo cy'Ubushakashatsi cya India (IISc), Bangalore, rufite amasaha angana na 21,500 y'amajwi yakusanyijwe ku baturage bagera ku 110,000 mu turere 120 mu ntara 22 za India. Rukubiyemo indimi 86 n'imivugire itandukanye ivugwa muri India, harimo indimi z'ingenzi zateganyijwe hamwe n'ubwoko bwinshi bw'indimi z'akarere n'iz'imiryango, hamwe n'amasaha 835 y'amajwi yanditswe.

Uru rutonde rwateguwe kugira ngo rushyigikire ibikorwa bitandukanye bijyanye n'ijwi n'ururimi harimo kumenya ijwi mu buryo bwikora, guhuza inyandiko n'ijwi, kumenya abavuga, kumenya ururimi, kunoza ijwi, no guteza imbere imiterere y'ururimi mu buryo bwinshi. Rwatangajwe munsi y'uburenganzira bwa CC BY 4.0, rugamije cyane cyane gutoza no kugerageza sisitemu za AI, rufite akamaro kanini ku bashakashatsi n'abateza imbere bakora ku ndimi z'Abahinde zifite ubushobozi buke kandi zitarahabwa agaciro.

Amateka n'Iterambere

VAANI (bisobanura "ijwi" cyangwa "ikiganiro" mu ndimi nyinshi z'Abahinde) yaremwe n'Ikigo cy'Ubushakashatsi mu Buhinde (IISc) i Bangalore mu rwego rwo gukemura icyuho kinini mu makuru y'ijwi ku ndimi z'Abahinde. Icyerekezo cy'indimi mu Buhinde ni kimwe mu byihariye ku isi, kigizwe n'indimi z'ibihumbi n'uduhanga twinshi, ariko igice kinini cy'amakuru y'ijwi asanzwe yibanda ku ndimi zifite ubushobozi bwinshi. VAANI yatejwe imbere kugira ngo yiyongereho ibikoresho byo gutoza mu rwego rw'izi ndimi zidakunze kubona ubushobozi, binyuze mu gukusanya amakuru mu buryo bwagutse mu turere tw'igihugu dutandukanye mu bijyanye n'ubutaka n'indimi.

Amakuru yakuwe ku bantu bagera ku 110,000 batandukanye mu turere 120 mu ntara 22 z'Ubuhinde, bigaragaza umuhate wo gufata impinduka z'ukuri z'akarere n'uduhanga aho kwishingikiriza ku nyandiko z'ubugenzuzi. Iyi dataset igizwe n'indimi n'uduhanga 86, kuva ku ndimi zikomeye zateganijwe nka Hindi, Tamil, Telugu, Bengali, Kannada, na Malayalam kugeza ku bwoko bwinshi bw'indimi z'akarere n'indimi z'imiryango nka Gondi, Santali, Kurukh, Wancho, na Tenyidie, n'izindi nyinshi.

Imiterere ya Dataset n'Ibikomeye Byayo

VAANI igizwe n'amasaha agera ku 21,500 y'amajwi muri rusange, ikayigira imwe mu makusanyirizo manini y'ijwi ry'indimi nyinshi zifite intego ku ndimi z'Abahinde. Muri aya, amasaha 835 yaranditswe, atanga inyandiko z'ukuri ku gice cy'iyi collection. Iyi dataset inakoresha ibice byinshi, ikayigira iyifashishwa mu bikorwa birenze ibijyanye n'ijwi bisanzwe.

Ibikomeye bya dataset birimo:

  • Guhagararira indimi n'uduhanga 86, harimo n'ubwoko bwinshi bw'indimi zifite ubushobozi buke n'indimi z'imiryango
  • Ibikorwa by'abantu bagera ku 110,000 baturutse mu bice bitandukanye by'ubutaka n'imibereho
  • Amasaha 21,500 y'amajwi yose hamwe n'amasaha 835 y'ijwi ryanditswe
  • Amashusho yakuwe mu turere 120 mu ntara 22 z'Ubuhinde
  • Yatanzwe munsi y'uburenganzira bwa CC BY 4.0, yemera gukoresha cyane hamwe no kwerekana umwimerere
  • Inkunga ku bushakashatsi bwose bw'uburyo bumwe n'uburyo bwinshi

Imirimo Ihamagarirwa n'Imikoreshereze

VAANI yateguwe kugira ngo ibe umutungo ku bikorwa bitandukanye bijyanye n'ijwi n'ibijyanye n'indimi. Abashakashatsi n'abateza imbere bashobora kuyikoresha mu gutoza no gupima sisitemu z'iyumviro ry'ijwi (ASR), gukora ijwi riturutse ku nyandiko (TTS), kumenya abavuga, no kumenya indimi. Iyi dataset inashyigikira imirimo ku kunoza ijwi no guteza imbere ibikoresho binini by'indimi (LLMs). Uburyo bw'ubwoko bw'indimi ikubiyemo butuma iboneka cyane mu gupima sisitemu zigenewe gukora mu buryo butandukanye bw'indimi mu Buhinde.

Uko harimo indimi nyinshi zifite ubushobozi buke n'uduhanga aho amakuru y'ijwi make cyangwa nta makuru ahari, VAANI ifite agaciro kanini ku bashakashatsi bita ku kubaka ikoranabuhanga ry'indimi rihamye. Itanga umusingi wo guteza imbere ibikoresho bya ASR n'ibijyanye n'ibisubizo by'indimi bishobora gufasha imiryango indimi zayo zagiye ziba mu iterambere ry'ikoranabuhanga rya AI.

Akamaro ku Ikoranabuhanga ry'Indimi z'Abahinde

Ingano n'ubwoko bw'indimi za VAANI bituma iba umusanzu ukomeye mu bushakashatsi ku ijwi ry'indimi nyinshi, cyane cyane mu rwego rw'indimi z'Amajyepfo y'Asia. Mu kwandika ubwoko bw'ijwi biva mu miryango y'imiryango, abahinzi, n'akarere hamwe n'indimi zivugwa cyane, iyi dataset ifata igice cy'umuco w'indimi z'Ubuhinde kitarakunze kugaragara mu bikoresho by'ikoranabuhanga. Uburenganzira bwayo bw'ubufunguro butuma ikorwa n'inzego z'uburezi, ibigo bya leta, n'abashakashatsi b'ubucuruzi bakora ku ikoranabuhanga ry'ijwi rihamye kandi rihagarariye.

Raporo