# Voxtype Configuration # # Native Linux tray is enabled by default. Set [tray] enabled = false to opt out. # # Location: ~/.config/voxtype/config.toml # All settings can be overridden via CLI flags # Transcription engine: "whisper" (default) or "parakeet" # Whisper: whisper.cpp via whisper-rs (most compatible) # Parakeet: NVIDIA FastConformer via ONNX Runtime (requires --features parakeet) # engine = "whisper" # State file for external integrations (Waybar, polybar, etc.) # Use "auto" for default location ($XDG_RUNTIME_DIR/voxtype/state), # a custom path, or "disabled" to turn off. The daemon writes state # ("idle", "recording", "transcribing") to this file whenever it changes. # Required for `voxtype record toggle` and `voxtype status` commands. state_file = "auto" # Streaming recovery retains the full capture while a streaming backend runs. # If the backend errors or misses its stop deadline, Voxtype uses the configured # batch backend to deliver undelivered dictation. # [streaming_recovery] # enabled = true # stop_timeout_secs = 30 [hotkey] # Built-in hotkey using evdev (Linux input subsystem) # # Most users should leave this disabled and use compositor keybindings instead: # - Hyprland: bind/bindr in hyprland.conf # - Sway: bindsym --no-repeat/--release in config # - River: riverctl map / map -release # # Enable this if you're on X11, using a compositor without key-release support, # or prefer a dedicated key like ScrollLock. Requires 'input' group membership. enabled = false # Key to hold for push-to-talk (when enabled = true) # Common choices: SCROLLLOCK, PAUSE, RIGHTALT, F13-F24 # On macOS: F13-F20 supported (PrintScreen/ScrollLock/Pause arrive as F13/F14/F15) # Use `evtest` to find key names for your keyboard key = "SCROLLLOCK" # Optional modifier keys that must also be held # Example: modifiers = ["LEFTCTRL", "LEFTALT"] modifiers = [] # Activation mode: "push_to_talk" or "toggle" # - push_to_talk: Hold hotkey to record, release to transcribe (default) # - toggle: Press hotkey once to start recording, press again to stop # mode = "push_to_talk" # Modifier key for secondary model selection (evdev only) # Hold this key while pressing the hotkey to use secondary_model # model_modifier = "LEFTSHIFT" # Profile modifiers for context-aware post-processing (evdev only) # Maps modifier keys to named profiles defined in [profiles.*] sections # [hotkey.profile_modifiers] # LEFTSHIFT = "translate" # Shift+hotkey activates [profiles.translate] # RIGHTALT = "formal" # RightAlt+hotkey activates [profiles.formal] [audio] # Audio input device ("default" uses system default) # List devices with: voxtype info devices # (or `arecord -L`). This takes an ALSA device name, not a PulseAudio # or PipeWire source name -- names from `pactl list sources` will not work. device = "default" # Sample rate in Hz (whisper expects 16000) sample_rate = 16000 # Maximum recording duration in seconds (safety limit) max_duration_secs = 60 # Keep the audio of a recording whose transcription failed, so it can be # retried with `voxtype transcribe `. The 20 most recent are kept in # ~/.local/state/voxtype/failed-audio/, readable only by you. Set to false to # write nothing; a dictation whose transcription fails is then lost. # keep_failed_recordings = true # Pause MPRIS media players (Spotify, Firefox, etc.) when recording starts and # resume them when it stops. Talks D-Bus directly, so no playerctl binary is # required. Distros that ship playerctl-free (Omarchy Quattro, for example) can # enable this safely. # pause_media = false # Lower active media stream volume while recording, then restore it as soon as # recording stops. This is separate from pause_media; enable one behavior or # the other depending on whether you want media paused or only quieter. # duck_media = false # duck_media_volume_percent = 34 # [audio.dead_input] # Stop a recording whose input delivers only exact digital silence (every # sample zero) for the first 5 seconds. # enabled = true # window_ms = 5000 # threshold_dbfs = -100.0 # [audio.feedback] # Enable audio feedback sounds (beeps when recording starts/stops) # enabled = true # # Sound theme: "default", "subtle", "mechanical", or path to custom theme directory # theme = "default" # # Volume level (0.0 to 1.0) # volume = 0.7 [whisper] # Execution mode: "local" (default) or "remote" # - local: Use whisper.cpp locally via FFI # - remote: Send audio to OpenAI-compatible API endpoint # mode = "local" # Model to use for transcription (local mode) # Options: tiny, tiny.en, base, base.en, small, small.en, medium, medium.en, large-v3, large-v3-turbo # .en models are English-only but faster and more accurate for English # Or provide absolute path to a custom .bin model file model = "base.en" # Language for transcription # Use "en" for English, "auto" for auto-detection # See: https://github.com/openai/whisper#available-models-and-languages language = "en" # Translate non-English speech to English translate = false # Number of CPU threads for inference (omit for auto-detection) # threads = 4 # Load model on-demand when recording starts (true) or keep loaded (false) # When true, model is loaded when recording starts and unloaded after transcription # When false, model is kept in memory for faster response times (default) on_demand_loading = false # GPU memory isolation mode # Enable on laptops with hybrid graphics to let dGPU sleep between transcriptions. # gpu_isolation = true # Context window optimization for short recordings (disabled by default) # Speeds up transcription for clips under 22.5 seconds. Disabled by default # because some models (especially large-v3/turbo) may experience repetition # loops. Enable if you want faster transcription and don't experience issues. # context_window_optimization = true # --- Remote mode settings (used when mode = "remote") --- # remote_endpoint = "http://192.168.1.100:8080" # Required # remote_model = "whisper-1" # Model to request from server # remote_api_key = "sk-..." # Or use VOXTYPE_WHISPER_API_KEY env var # remote_timeout_secs = 30 # [parakeet] # Parakeet configuration (only used when engine = "parakeet") # Requires: cargo build --features parakeet # # Model name (from ~/.local/share/voxtype/models/) or absolute path # model = "parakeet-tdt-0.6b-v3" # # Model type: "tdt" (recommended, proper punctuation) or "ctc" (faster, char-level) # Auto-detected from model directory if not specified # model_type = "tdt" # # on_demand_loading = false [output] # Primary output mode: "type", "clipboard", or "paste" # - type: Simulates keyboard input at cursor position (requires wtype or ydotool) # - clipboard: Copies text to clipboard (requires wl-copy) # - paste: Copies to clipboard then simulates paste keystroke (requires wl-copy and wtype or ydotool) mode = "type" # Fall back to clipboard if typing fails fallback_to_clipboard = true # Keep the text of a transcription no output method could deliver, so it can # be recovered. The 200 most recent are kept in # ~/.local/state/voxtype/undelivered/, readable only by you, and the failure # notification names the file. Set to false to save no file and keep the text # out of the notification; an undelivered dictation is then lost. # keep_undelivered_text = true # Per-language XKB layout variants for multilingual dictation. # Use this with language arrays such as `language = ["en", "ru"]` when a # language needs a variant that should not apply to other languages. # Built-in layout defaults already include `ru = "ru"` in language_to_layout. # dotoolc does not work with variants and cannot receive these hints; voxtype # uses direct dotool for them. Switch your desktop layout to the target # layout/variant before dictating; dotool does not change the active layout. # [output.language_to_variant] # ru = "phonetic" # Delay between typed characters in milliseconds # 0 = fastest possible, increase if characters are dropped # On KDE, eitype uses an effective minimum of 1ms when this is 0 to prevent # long dictations from being silently truncated by KWin. type_delay_ms = 0 # Delay before typing starts (ms) # Allows virtual keyboard to initialize. Some users report this helps prevent # the first character from being dropped on text insertion. Try 100-200ms. # Note: When using compositor integration (via `voxtype setup compositor`), # best results come from not binding Escape in the submap. Some users have # had success with Escape bound by increasing this delay, but the most # consistent fix is to use F12 or another key instead. pre_type_delay_ms = 0 # Keystroke for paste mode (when mode = "paste") # Shipped default is "shift+insert": it works in terminal emulators as well # as GUI apps, unlike ctrl+v, which many terminals reserve for their own # bindings (readline's literal-insert, tmux/screen prefixes, etc.) instead # of treating as paste. If this key is removed from the file entirely, the # built-in default is "auto". # Examples: # paste_keys = "shift+insert" # Shipped default; universal paste shortcut # paste_keys = "auto" # ctrl+v in GUI apps, paste_keys_terminal in # # terminals (Hyprland, Sway and niri) # paste_keys = "ctrl+v" # Standard GUI shortcut # paste_keys = "ctrl+shift+v" # Some terminal emulators paste_keys = "shift+insert" # Keystroke paste mode sends when the focused window is a terminal, used by # paste_keys = "auto" (default: "ctrl+shift+v"). Setting it next to an # explicit paste_keys makes that keystroke terminal-aware too. # paste_keys_terminal = "ctrl+shift+v" # Compositor integration hooks # Use `voxtype setup compositor hyprland|sway|river` for automatic setup. # # pre_recording_command: Runs when recording starts. Switch to a submap/mode # where F12 can cancel recording/transcription. # pre_output_command: Runs before typing output. Switch to a submap/mode that # blocks modifier keys from triggering compositor shortcuts during typing. # post_output_command: Runs after typing output. Reset to normal mode. # # Example (Hyprland): # pre_recording_command = "hyprctl dispatch submap voxtype_recording" # pre_output_command = "hyprctl dispatch submap voxtype_suppress" # post_output_command = "hyprctl dispatch submap reset" [output.notification] # Show notification when recording starts (hotkey pressed) on_recording_start = false # Show notification when recording stops (transcription beginning) on_recording_stop = false # Show notification with transcribed text after transcription completes on_transcription = true # Show notification when no speech is detected and the recording is discarded on_no_speech = true # When text cleanup or post_process changed the transcription, show the edits # as a word diff in the notification instead of a plain preview show_edits = true # [output.post_process] # Pipe transcribed text through an external command for cleanup before output. # The command receives text on stdin and outputs processed text on stdout. # Useful for LLM-based text cleanup, grammar correction, filler word removal. # On any failure (timeout, error), falls back to original transcription. # # command = "ollama run llama3.2:1b 'Clean up this dictation. Fix grammar, remove filler words. Output only the cleaned text:'" # timeout_ms = 30000 # 30 second timeout (generous for LLM) [text] # Optional English-only trained cleanup. Download its model first with: # voxtype setup lard # The model is licensed CC BY-SA 4.0, separately from Voxtype's MIT code: # https://models.voxtype.io/text/lard-disfluency-v1.MODEL_CARD.md # Turning this on does not download anything. Without the model the stage is # skipped and voxtype doctor says so. # lard_disfluency = false # lard_model_path = "/path/to/lard-disfluency-v1.vxdf" # # The language you dictate in, for cleanup that depends on it (the stage above # is English-only). Leave it on "auto" unless your model handles several # languages and detects them automatically: an engine language that is set, # or an English-only model, already answers the question and wins over this. # It is not sent to the transcription engine. # language = "auto" # # Text processing options (word replacements, spoken punctuation) # # Enable spoken punctuation conversion (e.g., say "period" to get ".") spoken_punctuation = false # # Custom word replacements (case-insensitive) # replacements = { "vox type" = "voxtype" } # Remove common filler words while preserving word boundaries. filter_filler_words = true # Keep destructive restart collapsing opt-in: a false positive deletes words. collapse_restarts = false # Enable the cleanup stages whose guards leave ordinary prose unchanged. repair_sentence_breaks = true format_numbers = true [osd] # On-screen display frontend. # - auto: Quickshell when qs and a Wayland session are available, else gtk4, # else native (default) # - gtk4: GTK4 + layer-shell frontend # - native: SCTK + wgpu + egui frontend # - quickshell: QML frontend with recipe/style customization frontend = "auto" # Quickshell style customization. These fields are ignored by gtk4/native. # style can be "default", an installed package name, or a package path. style = "default" # Palette source for Quickshell recipes. Leave unset to let a selected # package manifest choose its palette, falling back to Omarchy. # - omarchy: inherit colors from ~/.config/omarchy/current/theme/colors.toml # - fallback: use Voxtype's built-in palette # - package: use package colors when a style package provides them # - custom: use literal recipe colors # palette = "omarchy" # Quickshell host layout: compact, wide, minimal, tile, orb, or custom. layout = "compact" # Quickshell host frame. Background/border accept semantic color roles, # literal CSS colors, or "none". # [osd.frame] # background = "background" # border = "state" # glow = true # halo = true # Explicit trusted QML/style package path. Custom QML is trusted code. # plugin_path = "/path/to/voxtype-osd-package" # Declarative Quickshell visual recipe. Layers render in order. # [[osd.visual.layers]] # type = "bars" # shadow, waveform, bars, pulse, ring, meter, background, icon, label # source = "peak" # peak, rms, vad, state # color = "accent" # Omarchy token or literal color # order = 10 # gain = 1.0 # opacity = 1.0 # mirror = true [vad] # Filter silence-only recordings without requiring a separate VAD model. enabled = true backend = "energy" threshold = 0.5 min_speech_duration_ms = 100 [status] # Status display icons for Waybar/tray integrations # # Icon theme (or path to custom theme file): # Font-based (require specific fonts): # - "emoji" - Default emoji icons (🎙️ 🎤 ⏳) # - "nerd-font" - Nerd Font icons (requires Nerd Font) # - "material" - Material Design Icons (requires MDI font) # - "phosphor" - Phosphor Icons (requires Phosphor font) # - "codicons" - VS Code icons (requires Codicons font) # - "omarchy" - Omarchy distro icons # Universal (no special fonts needed): # - "minimal" - Simple Unicode (○ ● ◐ ×) # - "dots" - Geometric shapes (◯ ⬤ ◔ ◌) # - "arrows" - Media player style (▶ ● ↻ ■) # - "text" - Plain text ([MIC] [REC] [...] [OFF]) icon_theme = "emoji" # # Native Linux system tray: enabled by default, including omitted keys. # Disable in Status settings or via `voxtype config set tray.enabled false`. # [tray] # enabled = true # # On Omarchy, the daemon links the packaged Voxtype panel into # ~/.config/omarchy/plugins on first start. Set false to opt out; run # `voxtype setup omarchy` to link it by hand. # [omarchy] # link_settings_plugin = true # # Per-state icon overrides (optional, takes precedence over theme) # [status.icons] # idle = "🎙️" # recording = "🎤" # transcribing = "⏳" # stopped = "" # Dictation history: a record of every dictation, including failed and # cancelled ones, as one plain JSON line per dictation in # ~/.local/share/voxtype/history/YYYY-MM-DD.jsonl. Off here because this file # is also the system-wide fallback; turn it on with # `voxtype config set history.enabled true` or the tray's "Enable History". # [history] # enabled = false # retention_days = 30 # 0 keeps history forever # exclude_apps = ["KeePassXC", "Bitwarden", "1Password", "Proton Pass"] # record_app = true # which app the text went to # record_window_title = false # titles can hold subjects, names and URLs # Meeting mode: record and transcribe calls with `voxtype meeting start`. # Off here because this file is also the system-wide fallback; turn it on # with `voxtype config set meeting.enabled true`. docs/MEETING_MODE.md has # every option. # [meeting] # enabled = false # announce_recording = false # say "This meeting is being recorded..." into the call # announcement_voice = "kokoro" # or "jessica", an ElevenLabs voice downloaded once # announcement_volume = 0.5 # call volume, linear gain: 1.0 full, 0.5 about -6 dB, 0.3 about half as loud