Mozilla Common Voice 7.0 Voice Update

NVIDIA neMozilla vakaburitsa gadziriso kune yavo Common Voice datasets, iyo inosanganisira 182 yekutaura kwevanhu samples, kusvika 25% kubva kumwedzi mitanhatu yapfuura. Iyo data inoburitswa seruzhinji domain (CC6). Iwo akarongwa seti anogona kushandiswa mumashini ekudzidza masisitimu kuvaka mataurirwo ekutaura uye mamodheru.

Kuenzaniswa nekuvandudzwa kwekare, saizi yezvinyorwa zvekutaura muunganidzwa yakawedzera kubva pa9 kusvika 13.9 zviuru maawa ekutaura. Huwandu hwemitauro inotsigirwa hwakawedzera kubva pamakumi matanhatu kusvika makumi manomwe nematanhatu, kusanganisira kekutanga kutsigirwa kwemitauro yeBelarusian, Kazakh, Uzbek, Bulgarian, Armenian, Azerbaijani neBashkir. Iyo yakagadzirirwa yemutauro weRussia inovhara 60 vatori vechikamu uye maawa 76 ekutaura zvinhu (paiva nevatori vechikamu 2136 nemaawa 173), uye nokuda kwemutauro weUkraine - vatori vechikamu 1412 uye maawa 111 (paiva nevatori vechikamu 615 nemaawa makumi matatu).

Vanhu vanopfuura zviuru makumi manomwe neshanu vakatora chikamu mukugadzirira kwezvinhu muChirungu, vachiraira maawa e75 ekutaura kwakasimbiswa (paive nevatori vechikamu zviuru makumi matanhatu nemazana matanhatu nemaawa 2637). Sezvineiwo, mutauro uri munzvimbo yechipiri maererano nehuwandu hwe data yakaunganidzwa iRwanda, iyo 66 maawa akaunganidzwa. Izvi zvinoteverwa neGerman (1686), Catalan (2260) uye Esperanto (1040). Pakati peanonyanya kuwedzera saizi yedata rezwi mutauro wechiThai (920-kupetwa mubhesi, kubva pa840 kusvika kumaawa 20), Luganda (kubva pa12 kusvika maawa 250), Esperanto (kubva pa8 kusvika 80 maawa) uye chiTamil ( kubva pa100 kusvika ku840 hours).

Sechikamu chekutora chikamu muchirongwa cheCommon Voice, NVIDIA yakagadzirira akagadzirira-akadzidziswa mamodheru emuchina kudzidza masisitimu (anotsigirwa nePyTorch) zvichibva pane yakaunganidzwa data. Iwo mamodheru akagoverwa sechikamu chemahara uye akavhurika NVIDIA NeMo toolkit, iyo, semuenzaniso, yakatoshandiswa mu automated voice services yeMTS neSberbank. Mamodheru acho akagadzirirwa kushandiswa mukuzivikanwa kwekutaura, kusanganisa kwekutaura, uye masisitimu ekugadzirisa mutauro, uye anogona kubatsira kune vaongorori kuvaka masisitimu ekutaura-akaitwa nezwi, mapuratifomu ekunyora, uye otomatiki nzvimbo dzekufona. Kusiyana nemapurojekiti aimbove aripo, mamodheru akaburitswa haagumiri pakuzivikanwa mutauro wechiRungu uye anovhara mitauro yakasiyana-siyana, mataurirwo uye matauriro.

Ngatikuyeuchidzei kuti Common Voice project inotarisirwa kuronga basa rekubatana kuti riunganidze dhatabhesi yemaitiro ezwi iyo inofunga nezvekusiyana kwemazwi uye maitiro ekutaura. Vashandisi vanokokwa kutaura mazwi anoratidzwa pachiratidziro kana kuongorora mhando yedata rakawedzerwa nevamwe vashandisi. Iyo yakaunganidzwa dhatabhesi ine marekodhi emataurirwo akasiyana eakajairwa mitsara yekutaura kwevanhu inogona kushandiswa pasina kurambidzwa mumashini ekudzidza masisitimu uye mumapurojekiti ekutsvagisa.

Sekureva kwemunyori weVosk inoenderera mberi yekucherekedza raibhurari yekutaura, izvo zvisingabatsiri zveCommon Voice set idivi rimwe reizwi zvinhu (kunyanya kwevanhu vechirume 20-30 yemakore, uye kushomeka kwezvinhu nemanzwi evakadzi. , vana nevakura), kushaikwa kwekusiyana-siyana muduramazwi (kudzokororwa kwemazwi akafanana) uye kugoverwa kwezvakarekodhwa mumhando yeMP3 inoshatisa.

Source: opennet.ru

Voeg