Skip to content

Golos v0.1.0 — offline Russian dictation for Linux

Latest
Latest

Choose a tag to compare

View all tags
@geft1310-hue geft1310-hue released this 04 Aug 15:50
· 7 commits to main since this release
v0.1.0
4b8e26f

Hold the right Ctrl key, speak Russian, release it. Punctuated text is placed in your clipboard.

Website: https://geft1310-hue.github.io/golos/
English page: https://geft1310-hue.github.io/golos/en/

Everything runs on your own machine. No account, cloud processing, GPU or subscription is required. After the one-time model download, dictation works without internet access.

Install

git clone https://github.com/geft1310-hue/golos.git
cd golos
./install.sh

Requires Linux with an X11 session. The installer supports Debian, Ubuntu, Linux Mint and related distributions.

What you get

  • Automatic punctuation and numeric digits. GigaAM v3 emits punctuation, capitalization and digits as part of recognition; spoken punctuation commands are not required.
  • Three recognition variants, switchable in Settings. The accurate RNN-T model ships by default; the other variants download on demand.
  • Push-to-talk, not always-on. The microphone opens only while the configured key is held.
  • Clipboard output. Release the key, wait for recognition, then paste with Ctrl+V.

Measured on the development machine

Intel i7-2640M from 2011, two cores, 8 GB RAM:
Recognition speed RTF 0.31 — 10 s of speech in about 3.1 s
Cold model load 1.7–1.9 s
Memory, idle / model loaded ~65 MB / ~304 MB
Disk 217 MB for the default model
Network traffic while dictating zero

The reproducible scripts are in bench/. Accuracy comparisons and sources are documented on the website.

Privacy, and the one exception

Speech is not written to disk. Recognized text and recent-history entries stay in RAM and disappear when the process exits. Golos does not ask the clipboard manager to persist the text. The dictation process contains no networking code.

The stated exception: tools/fetch_models.py is a separate downloader used once during installation or when another model is selected.

Known limits

  • Linux/X11 only. There is currently no universal Wayland mechanism equivalent to the held global push-to-talk key interaction used by Golos.
  • Russian only.
  • Windows and Android versions are planned, not part of this release.

По-русски: README.ru.md

Assets 2
Loading