Skip to content

User guide and HUD controls

This guide covers how to position the desktop overlay, select audio devices, choose voice detection modes, and configure local Ollama models.

Operating the desktop HUD

The Elenchus overlay is a minimal, translucent window designed to stay in your peripheral vision during live calls without blocking other windows.

Moving and positioning

  • Click and drag anywhere on the HUD header or background to move it across monitors.
  • The window sets Qt.WindowStaysOnTopHint, keeping it visible when other apps (Discord, Zoom, Google Meet) have focus.

Status indicators

  • Green LED (#98ff38): Ready and listening for spoken statements.
  • Amber LED (#d97706): Transcribing audio or running fallacy inference.
  • Red LED (#ef4444): Warning state (such as an unreachable Ollama instance).

Audio waveform meter

The Monochart energy meter shows incoming audio amplitude in real time. Bars illuminate green once voice energy rises above the baseline noise floor.

Voice activity detection modes

Elenchus supports two listening modes depending on your setting:

  • Click the HUD action bar or press Space to start recording when your opponent speaks. Click again when they finish.
  • This avoids background noise, interruptions, or coughing from triggering unneeded model runs.

Continuous automatic VAD

  • The engine calculates real-time RMS energy automatically.
  • When speech remains above the noise threshold for 300 ms, recording starts. When silence persists for 700 ms, the audio slice sends to the transcriber.
  • Useful for hands-free listening.

Configuring audio capture

Capturing opponent audio with loopback

To analyze an opponent over Discord or a web browser on Windows:

  1. Set your system output to your normal headphones or speakers.
  2. In config.py (or through --device in the CLI), pick the loopback endpoint matching your output device.
  3. Elenchus captures the WASAPI loopback stream directly without needing virtual audio cable drivers.

Capturing microphone audio

To practice statements yourself or analyze an in-person room debate, select your physical USB microphone or headset input in config.py.

Ollama configuration

Elenchus communicates with your local Ollama daemon at http://localhost:11434.

Switching models

Specify your preferred model in config.py:

# Default model for concise fallacy detection
OLLAMA_MODEL = "qwen2.5-coder:3b"

Tested options:

  • qwen2.5-coder:3b (default): Fast inference and reliable on informal fallacy detection with concise phrasing.
  • qwen2.5:3b: Good general reasoning alternative.
  • llama3.2:3b: Lightweight alternative with low memory usage.
  • mistral:7b: Provides longer, more nuanced refutations at higher VRAM usage.

For benchmark comparisons and latency breakdown, see Runtime specifications. For the design rationale behind selecting coder models, see Why a coder model? Brevity.

Starting Ollama

Make sure Ollama is running before starting Elenchus: 

ollama list
If it is not running, start it in a separate terminal: 
ollama serve