Abaphuhlisi bethala leencwadi leVosk bapapashe imodeli emitsha yokuqatshelwa kwentetho yesiRashiya: iseva vosk-model-ru-0.22 kunye neselula yeVosk-model-small-ru-0.22. Iimodeli zisebenzisa idatha entsha yentetho, kunye ne-neural network architecture entsha, eye yandisa ukuchaneka kokuqaphela nge-10-20%. Ikhowudi kunye nedatha ihanjiswa phantsi kwelayisensi ye-Apache 2.0.
Utshintsho olubalulekileyo:
- Idatha entsha eqokelelwe kwizithethi zelizwi iphucula kakhulu ukuqondwa kwemiyalelo yentetho ethethwa kude.
- Iskimu esitsha sokutsalwa komsindo siphucule kakhulu ukuchaneka kokuqondwa korekhodo lwe-wideband. Kwangaxeshanye, ukuchaneka kokuqaphela umnxeba kuye kwaphuculwa.
- Ipakethe yolwandiso lwesichazi-magama ikuvumela ukuba wenze ukuqondwa kweerekhodi ezinzima zobugcisa.
Ukuchaneka okulungileyo, kucetyiswa ukuba uhlaziye i-Wax version kwi-0.3.32. Unokuba nomdla kwizinto ezintsha zeVosk - ukudibanisa kunye noBumbano, iNativescript, iJigasi. Iimodeli zokuqaphela iilwimi zesiKazakh kunye ne-Ukraine. Imodeli yeseva idinga iprosesa yanamhlanje kunye ne-8GB yememori ukuze isebenze. Imodeli yeselula ingasetyenziswa kwiifowuni kunye ne-RaspberryPi 3+.
umthombo: opennet.ru