config.example.toml

# voiceio configuration
# Copy to your config file and edit as needed.
# All values shown are defaults. Only override what you want to change.
#
# Config file location:
#   Linux/macOS: ~/.config/voiceio/config.toml
#   Windows:     %LOCALAPPDATA%\voiceio\config\config.toml

[hotkey]
key = "ctrl+alt+v"       # Combo to toggle recording. Examples: alt+v, ctrl+shift+r, ctrl+alt+v
backend = "auto"         # "auto", "evdev", "pynput", "socket"

[model]
name = "base"            # Whisper model: tiny, base, small, medium, large-v3
language = "en"          # ISO 639-1 code, or "auto" for auto-detection
device = "auto"          # "cpu", "cuda", or "auto"
compute_type = "int8"    # int8 (fastest), float16 (GPU), float32 (most accurate)
vocabulary_file = ""     # Path to vocabulary.txt (one term per line), empty = check default location

[audio]
sample_rate = 16000      # 16kHz is what Whisper expects
device = "default"       # Audio input device name or index
prebuffer_secs = 1.0     # Seconds of pre-buffer to capture (prevents clipping first syllable)
silence_threshold = 0.01 # RMS threshold for silence detection (lower = more sensitive)
silence_duration = 0.6   # Seconds of silence before triggering a streaming transcription pass
auto_stop_silence_secs = 5.0  # Seconds of sustained silence before auto-stopping recording
vad_backend = "silero"   # "silero" (neural net, recommended) or "rms" (simple amplitude)
vad_threshold = 0.5      # Silero speech probability threshold (0.0-1.0)

[output]
method = "auto"          # "auto", "xdotool", "ydotool", "wtype", "clipboard", "pynput"
streaming = true         # Stream text as you speak
min_recording_secs = 1.5 # Minimum recording duration before stop is accepted
cancel_window_secs = 0.5 # Double-press within this window cancels recording
punctuation_cleanup = true  # Auto-capitalize and fix spacing in Whisper output
number_conversion = true    # Convert spoken numbers to digits ("twenty five" → "25")

[commands]
enabled = true           # Recognize voice commands: "new line", "period", "scratch that", etc.
                         # Also includes "correct that" — flags the last word for review.

# Corrections dictionary: auto-replace misheard words in transcription.
# Managed via CLI: voiceio correct "wrong" "right"
# Stored in corrections.json alongside this config file.
# Say "correct that" during dictation to flag the last word for review.
# Run "voiceio correct --flagged" to see flagged words.

[feedback]
sound_enabled = true     # Play sounds on start/stop recording
notify_clipboard = false # Show desktop notification with transcribed text

[tray]
enabled = false          # Show system tray icon (requires: pip install voiceio[tray])

[daemon]
log_level = "INFO"       # DEBUG, INFO, WARNING, ERROR

[llm]
enabled = false                          # Use local LLM for grammar/spelling cleanup (requires Ollama)
model = ""                               # Model name (empty = auto). Examples: phi3:mini, llama3.2:1b, mistral
base_url = "http://localhost:11434"      # Ollama API URL
timeout_secs = 15.0                      # Max seconds to wait for LLM response

[autocorrect]
# api_key = ""                                   # API key (or set OPENROUTER_API_KEY / OPENAI_API_KEY env var)
# base_url = "https://openrouter.ai/api/v1"      # OpenAI-compatible API base URL
# model = "anthropic/claude-sonnet-4"             # Model for autocorrect analysis
# timeout_secs = 30.0                             # Max seconds to wait for API response

[tts]
enabled = false          # Enable text-to-speech (select text + ctrl+alt+s to hear it)
engine = "auto"          # "auto", "piper", "espeak", "edge-tts"
# hotkey = "ctrl+alt+s"  # Hotkey to trigger TTS (press again to cancel)
# voice = ""             # Engine-specific voice name (empty = default)
speed = 1.0              # Speech speed multiplier (0.5–2.0)
# model = ""             # Piper model name (empty = default en_US-lessac-medium)

[health]
auto_fallback = true     # Automatically switch to fallback backend on failure