Abathuthukisi bomtapo wezincwadi weVosk bashicilele amamodeli amasha wokuqashelwa kwenkulumo yesiRashiya: iseva vosk-model-ru-0.22 kanye neselula yeVosk-model-small-ru-0.22. Amamodeli asebenzisa idatha yenkulumo entsha, kanye ne-neural network architecture entsha, eye yandisa ukunemba kokuqashelwa ngo-10-20%. Ikhodi nedatha kusatshalaliswa ngaphansi kwelayisensi ye-Apache 2.0.
Izinguquko ezibalulekile:
- Idatha entsha eqoqwe kuzipikha zezwi ithuthukisa kakhulu ukubonwa kwemiyalelo yenkulumo ekhulunywa kude.
- Uhlelo olusha lokukhipha umsindo luthuthukise kakhulu ukunemba kokuqashelwa kokurekhodwa kwe-wideband. Ngesikhathi esifanayo, ukunemba kokuqashelwa kocingo nakho kuye kwaba ngcono.
- Iphakheji yesandiso sesichazamazwi ikuvumela ukuthi wenze ngendlela oyifisayo ukuqashelwa kwamarekhodi obuchwepheshe ayinkimbinkimbi.
Ukuze uthole ukunemba okungcono kakhulu, kunconywa ukuthi ubuyekeze inguqulo ye-Wax ibe ngu-0.3.32. Ungase futhi ube nentshisekelo kuzici ezintsha zeVosk - ukuhlanganiswa ne-Unity, Nativescript, Jigasi. Amamodeli okuqaphela izilimi zesi-Kazakh nesi-Ukrainian. Imodeli yeseva idinga iphrosesa yesimanje kanye nememori engu-8GB ukuze isebenze. Imodeli yeselula ingasetshenziswa kumafoni naku-RaspberryPi 3+.
Source: opennet.ru