CS2 Voice Chat Translator

An always-on-top overlay that transcribes speech and shows the translation in the corner of the screen. Written for Counter-Strike 2 voice chat but it does not know or care what game is running.

Speech recognition runs locally with faster-whisper. Nothing about the audio leaves the machine unless you use the Google translation backend, which only ever sees the transcribed text.

Two modes

Push-to-translate (the default). Press F, say something in Turkish, press F again. The overlay shows it in English and Russian so you can paste it into text chat or just read it out. It is a toggle, not a hold, which sounds wrong until you try to hold a key down and talk while also playing.

Listen mode. Set PUSH_TO_TRANSLATE_ENABLED=false and it captures system audio through WASAPI loopback instead, segments it, and translates whatever your teammates are saying into Turkish. This one is more experimental. It works, but game audio and voice chat come down the same loopback stream, so gunfire ends up in the recogniser too.

Pipeline

Audio arrives as 16 kHz mono float blocks of 100 ms. An energy VAD accumulates them: RMS above the threshold means speech, and a segment is emitted once there has been 300 ms of trailing silence and the segment is at least 350 ms long, or once it hits the 3 second cap. That segment goes to whisper, the text goes to deep-translator, and the result is pushed to the Qt overlay through a signal.

The VAD is deliberately dumb. A neural VAD is more accurate but adds latency to every block, and at 100 ms blocks the RMS check is essentially free. If it misses quiet speech, drop VAD_ENERGY_THRESHOLD. If a noisy room keeps triggering it, raise it. 0.015 is a reasonable starting point.

condition_on_previous_text=False on the whisper call matters more than it looks. With it on, whisper carries context between segments and starts hallucinating continuations during silence.

What it does not do

No kernel driver, no DLL injection, no reading or writing game memory, no touching game files. It opens an audio device and draws a window. That is the whole footprint.

The overlay is a separate top-most click-through window, so the game has to run in borderless or windowed fullscreen for it to be visible. Exclusive fullscreen will cover it.

Game and platform terms of service still apply to you, not to me.

Setup

setup.bat

Creates a venv, installs the package with pip install -e . and copies .env.example to .env. There is also python launcher.py if you would rather click through a GUI installer.

Then:

run.bat

First run downloads the whisper weights, about 75 MB for base. On CPU, base with int8 keeps up fine. small is noticeably better on accented speech and roughly twice as slow. medium and up want a GPU.

Configuration

Everything lives in .env, parsed by config.py. The knobs you will actually touch:

Variable	Default	Notes
STT_MODEL_SIZE	base	tiny, base, small, medium, large-v3
STT_DEVICE	cpu	cuda if you have the CUDA runtime
STT_COMPUTE_TYPE	int8	float16 on GPU
PUSH_TO_TRANSLATE_HOTKEY	f	single char or a pynput key name
PUSH_TO_TRANSLATE_OUTPUT_LANGUAGE	ru	second output language
VAD_ENERGY_THRESHOLD	0.015	see above
OVERLAY_X / OVERLAY_Y	16	top-left corner in pixels
TRANSLATOR_PROVIDER	google	or libretranslate with a URL

Point LIBRETRANSLATE_URL at your own instance if you do not want text going to Google.

Layout

src/cs2_voice_translator/
  audio_capture.py    WASAPI loopback via soundcard
  mic_capture.py      microphone, used by push-to-translate
  vad.py              energy segmenter
  stt.py              faster-whisper wrapper
  translator.py       deep-translator / LibreTranslate
  overlay.py          PySide6 top-most window
  app.py              listen mode worker
  push_to_translate.py  hotkey mode worker
  config.py           .env parsing
  main.py             wiring

Requirements

Windows 10/11, Python 3.10+. A microphone for push-to-translate, any output device for listen mode.

License

MIT, see LICENSE.