config.example.yml

# UltraWhisper Configuration
# Save this file to: ~/.config/ultrawhisper/config.yml

# Notification settings
notifications:
  visual_enabled: false
  audio_enabled: false

# Double tap activation (tap key twice quickly to toggle recording)
use_double_tap: false

# Push-to-talk settings for transcription mode (hold to record, release to transcribe)
push_to_talk:
  enabled: true
  key: Key.cmd  # Key.cmd, Key.ctrl, Key.alt, Key.shift, etc.

# Question mode push-to-talk settings (separate hotkey that auto-switches to question mode)
question_mode_push_to_talk:
  enabled: false  # Enable separate hotkey for question mode
  key: Key.alt    # Key.alt, Key.ctrl, Key.shift, etc. (different from transcription key)

# Hotkey settings (for non-push-to-talk mode)
hotkey:
  key: cmd
  modifier: Key.cmd

# Audio recording settings
audio:
  sample_rate: 16000
  channels: 1
  dtype: float32

# Whisper transcription settings
whisper:
  model_name: tiny  # tiny, base, small, medium, large-v3
  language: en

# LLM (Large Language Model) settings
llm:
  provider: openai  # openai, anthropic, or any OpenAI-compatible server
  model: gpt-4o
  api_key: your-api-key-here  # Get from https://platform.openai.com/api-keys
  base_url: https://api.openai.com/v1  # For OpenAI-compatible servers (Ollama, LMStudio, etc.)

  # Base prompt for text correction
  base_prompt: |
    You are correcting speech-to-text transcription errors. Fix grammar,
    spelling, and misheard words while preserving the original meaning.
    DO not mention anything about the transcript, only respond with corrected transcript.

  # Continue without LLM correction if service is unavailable
  skip_if_unavailable: true

  # MCP (Model Context Protocol) Servers - Optional
  # mcp_servers:
  #   - name: Example
  #     command: python
  #     args:
  #       - /path/to/mcp_server.py
  #     env:
  #       EXAMPLE_VAR: "value"

# Context detection (detects active application for smarter corrections)
context_detection: true

# Context-aware prompts for different applications
context_prompts:
  applications:
    code: Preserve code syntax, variable names, and technical terms. DO not use markdown formatting and do not use new lines or carriage returns.
    firefox: 'Format for web: use proper capitalization and punctuation for comments/posts.'
    chrome: 'Format for web: use proper capitalization and punctuation for comments/posts.'
    sublime_text: Preserve code syntax and technical terms.
    vim: Preserve code syntax, commands, and technical terms.
    gnome-terminal: Preserve commands, paths, and technical syntax. Don't add unnecessary punctuation.
    kitty: Preserve commands and technical syntax.
    alacritty: Preserve commands and technical syntax.
    slack: Keep casual tone, use appropriate punctuation for chat.
    discord: Keep casual tone for chat messages.
    obsidian: Format as markdown. Improve clarity and structure.
    libreoffice: Professional tone, proper grammar and punctuation.

  # Pattern-based prompts (matches against window title)
  patterns:
    - match: .*(Terminal|console).*
      prompt: Preserve command syntax and technical terms. Do not use newlines or carriage returns.
    # match a discord channel
    - match: .*kiwis-only.*
      prompt: Speak like a kiwi with kiwi specific lingo.  Do not change the meaning or add additional stuff, but really go over the top in kiwi-fying the language.
    # match game chat
    - match: .*steam_app_2357570.*
      prompt: If the transcript is a potentially mean message, convert it to a compliment
        or otherwise nice message.  Try to maintain as much of the original meaning
        and intention as possible, but if its just play mean, make it nice. DO not use ! or emdashes.
# Logging configuration
logging:
  level: info  # debug, info, warning, error
  log_context: true
  log_prompts: false  # Log prompts sent to LLM (useful for debugging)
  log_corrections: true
  redact_content: false  # Redact sensitive content from logs
  file: null  # Optional: path to log file

# Output settings
output:
  paste_mode: false  # Use clipboard paste instead of simulated typing
  typing_delay: 0.002  # Delay between typing each character (seconds)

# TTS (Text-to-Speech) settings
tts:
  provider: "system"  # "system", "openai", "elevenlabs", etc.

  # System TTS settings (Linux: espeak, macOS: say, Windows: SAPI)
  system:
    voice: null       # Use default system voice (or specify like "Alex" on macOS)
    rate: 200         # Words per minute (50-500)
    volume: 1.0       # Volume level (0.0-1.0)

  # Future cloud providers:
  # openai:
  #   voice: "alloy"    # alloy, echo, fable, onyx, nova, shimmer
  #   model: "tts-1"    # tts-1 or tts-1-hd
  #   api_key: "your-openai-api-key"
  # elevenlabs:
  #   voice_id: "your-voice-id"
  #   api_key: "your-elevenlabs-api-key"

# Mode configuration
modes:
  # Default mode on startup: "transcription" or "question"
  default: "transcription"

  # Question mode settings (conversational AI assistant)
  question:
    # Context-aware prompts for different applications
    context_prompts:
      default: |
        You are a helpful AI assistant. Provide concise, accurate answers to user questions.
        Keep responses brief unless more detail is specifically requested.

      # Add custom context prompts for different applications
      # code: |
      #   You are a coding assistant. Help with programming questions.

    # Question mode behavior
    output_response: false  # Whether to type/paste AI response (usually want this false for TTS)
    tts_enabled: true      # Whether to speak the response (requires TTS setup)
    conversation_history: 30  # Number of turns to remember (60 messages total)