choose remote or local

after you install the windows .exe, open settings → voice engine. you do not clone this repository or start a rumik server.

remote (https) — default after the .exe

the installed app already knows the public rumik-ai space address and calls it over https. that space is a hosted model, not something you run. audio overview sends text from your notebook and writes the wav files into your user-data folder.

a hugging face token is not required to start. anonymous use works until the public space hits its daily zerogpu quota. if that happens, paste your own token in settings → voice engine. the token stays on that computer. you still do not clone the repo.

you do not start a local https server“remote” means the desktop app calls hugging face. it does not mean you host rumik yourself.

local (this pc)

choose local only if you want on-device inference. you need nvidia drivers, a cuda-capable python 3 with torch, transformers, soundfile, accelerate, and bitsandbytes, then the official rumik-ai/rumik-oss-1 weights.

open settings → voice engine and select local (this pc).
copy the bind path and run the download command (or use copy command + open terminal).
refresh status or restart the app. mode should read local when cuda and weights are both present.

license for those weights is cc by-nc 4.0 (research / non-commercial). the model card is on hugging face.

when remote voice fails

the public space uses zerogpu and a daily anonymous quota. if the quota is exhausted, settings shows the error from that host. paste a hugging face token there, wait for quota to reset, or switch to local on an nvidia gpu. chat and the other studio tools still work.

what you hear

speakers are ira, aisha, siya, and zoya. pace is steady. accent follows the language of the line. the ui does not offer tone, accent, or pace pickers. debate uses a challenger and an advocate on alternating turns.

privacy terms docs