dictee.conf.example

# dictee.conf — Voice dictation configuration
# This file is generated by dictee-setup but can be edited manually.
# Location: ~/.config/dictee.conf
# After manual edits, restart the daemon:
#   dictee-switch-backend asr <backend>

# ============================================================================
# ASR (speech recognition)
# ============================================================================

# Active ASR backend: parakeet | canary | whisper | vosk
DICTEE_ASR_BACKEND=parakeet

# Audio source language (ISO 639-1 code)
DICTEE_LANG_SOURCE=fr

# Target language for translation (used by translation backends)
DICTEE_LANG_TARGET=en

# ── Whisper ──────────────────────────────────────────────────────────────────
# faster-whisper model: tiny, small, medium, large-v3-turbo, large-v3
# or a HuggingFace CTranslate2 repo: Systran/faster-distil-whisper-large-v3
DICTEE_WHISPER_MODEL=small

# Whisper transcription language (empty = use DICTEE_LANG_SOURCE)
#DICTEE_WHISPER_LANG=

# ── Vosk ─────────────────────────────────────────────────────────────────────
# Vosk model: language code (fr, en, de, es, it, pt, ru, zh, ja)
# The daemon automatically resolves to vosk-model-small-<lang>-x.xx
DICTEE_VOSK_MODEL=fr

# ── Audio source ─────────────────────────────────────────────────────────────
# Audio source: empty = default microphone, or app:<name> (e.g. app:firefox)
#DICTEE_AUDIO_SOURCE=

# ============================================================================
# Translation
# ============================================================================

# Translation backend: disabled | trans | libretranslate | ollama
DICTEE_TRANSLATE_BACKEND=disabled

# translate-shell engine (if backend=trans): google | bing
#DICTEE_TRANS_ENGINE=google

# LibreTranslate: local port and installed languages
#DICTEE_LIBRETRANSLATE_PORT=5000
#DICTEE_LIBRETRANSLATE_LANGS=de,en,fr,it,pt

# Ollama: translation model and forced CPU mode
#DICTEE_OLLAMA_MODEL=translategemma
#OLLAMA_NUM_GPU=0

# ============================================================================
# Push-to-talk (keyboard shortcuts)
# ============================================================================

# Mode: hold (press and hold) | toggle (press to start/stop)
DICTEE_PTT_MODE=hold

# PTT key (evdev code, 67=F9)
DICTEE_PTT_KEY=67

# Translation PTT key (0 = disabled)
#DICTEE_PTT_KEY_TRANSLATE=0

# Translation modifier: alt | ctrl | shift (with same PTT key)
#DICTEE_PTT_MOD_TRANSLATE=

# ============================================================================
# Display and feedback
# ============================================================================

# Copy text to clipboard
DICTEE_CLIPBOARD=true

# Speech animation overlay (Wayland layer-shell: KDE, Sway, Hyprland)
DICTEE_ANIM_SPEECH=true

# KDE Plasma widget (panel / taskbar)
DICTEE_ANIM_PLASMOID=false

# ============================================================================
# Text post-processing
# ============================================================================

# Enable post-processing (punctuation, rules, etc.)
DICTEE_POSTPROCESS=true

# Language-specific rules (false to disable)
#DICTEE_PP_ELISIONS=true
#DICTEE_PP_ELISIONS_IT=true
#DICTEE_PP_SPANISH=true
#DICTEE_PP_PORTUGUESE=true
#DICTEE_PP_GERMAN=true
#DICTEE_PP_DUTCH=true
#DICTEE_PP_ROMANIAN=true

# General rules
#DICTEE_PP_NUMBERS=true
#DICTEE_PP_TYPOGRAPHY=true
#DICTEE_PP_CAPITALIZATION=true
#DICTEE_PP_DICT=true
#DICTEE_PP_RULES=true
#DICTEE_PP_CONTINUATION=true
#DICTEE_PP_SHORT_TEXT=true

# LLM post-processing (requires ollama)
#DICTEE_LLM_POSTPROCESS=false
#DICTEE_LLM_MODEL=gemma3:4b
#DICTEE_LLM_TIMEOUT=10
#DICTEE_LLM_CPU=false
#DICTEE_LLM_SYSTEM_PROMPT=default
#DICTEE_LLM_POSITION=hybrid

# ============================================================================
# Voice command suffixes
# ============================================================================

# Keyword appended after a voice command to distinguish it from dictation
# One per language (regex, empty = no suffix)
#DICTEE_COMMAND_SUFFIX_FR=finale?s?
#DICTEE_COMMAND_SUFFIX_EN=done
#DICTEE_COMMAND_SUFFIX_ES=listo
#DICTEE_COMMAND_SUFFIX_DE=weiter
#DICTEE_COMMAND_SUFFIX_IT=seguito
#DICTEE_COMMAND_SUFFIX_PT=pronto
#DICTEE_COMMAND_SUFFIX_UK=далі

# ============================================================================
# Notifications
# ============================================================================

# Enable desktop notifications (recording start/end, errors)
#DICTEE_NOTIFICATIONS=true

# Display transcribed text in notification
#DICTEE_NOTIFICATIONS_TEXT=true

# ============================================================================
# Audio context buffer
# ============================================================================

# Enable audio buffer (concatenates previous recording for context)
# Improves homophones and words split between two pushes
DICTEE_AUDIO_CONTEXT=true

# Max buffer duration in seconds (beyond this, the buffer is cleared)
DICTEE_AUDIO_CONTEXT_TIMEOUT=30

# ============================================================================
# GPU / CPU mode (advanced)
# ============================================================================

# Force CPU mode for all ASR backends, even if a GPU is detected.
# Truthy values: 1, true, yes (case-insensitive).
# Anything else (0, false, no, empty, unset) → GPU used if available.
# Use cases:
#   - laptop on battery (CPU mode is more power-efficient)
#   - GPU shared with other workloads (Ollama, games)
#   - debugging GPU-specific issues
# WARNING: Canary 1B is very slow on CPU (~5 s for 16 s audio, vs 0.65 s GPU).
#DICTEE_FORCE_CPU=0

# ============================================================================
# Parakeet quantization (advanced)
# ============================================================================

# Parakeet TDT model variant: int8 | fp32
# - int8: ~670 MB, faster on CPU with AVX-VNNI (-38 % latency measured),
#   negligible WER loss (~0-1 % depending on audio).
# - fp32: ~2.4 GB, best accuracy and fastest on GPU (CUDA Tensor Cores
#   don't benefit from int8 quantization in current ONNX Runtime).
# Recommended:
#   - CPU-only or low-VRAM GPU (< 4 GB): int8
#   - GPU with ≥ 4 GB VRAM: fp32
# The model files must be present on disk (~/.local/share/dictee/tdt/
# or /usr/share/dictee/tdt/). If both variants are installed, this var
# selects which one the daemon loads at startup.
#DICTEE_PARAKEET_QUANT=int8

# ============================================================================
# ORT threading (advanced)
# ============================================================================

# Number of ONNX Runtime threads for Parakeet/Canary transcription.
# Default: auto-detect (min(8, CPU thread count)).
# ⚠ DO NOT exceed your CPU's thread count (use `nproc` to check),
#   otherwise oversubscription → worse performance than default.
# Useful overrides:
#   - laptop on battery: low value (2-4) to save power
#   - container/VM with misconfigured cgroups
#   - shared workload (compilation, encoding) alongside dictation
# Valid range: 1 to 64
#DICTEE_INTRA_THREADS=8

# ============================================================================
# Debug
# ============================================================================

# Enable debug logs (shell + Rust + Python)
# Logs in /tmp/dictee-debug-<uid>.log
#DICTEE_DEBUG=false

# ============================================================================
# Diarization (internal)
# ============================================================================

# ASR backend used before switching to diarization mode (restored after)
# Managed automatically by dictee-switch-backend, do not modify
#DICTEE_PRE_DIARIZE_BACKEND=

# ============================================================================
# Internal marker — do not modify
# ============================================================================
DICTEE_SETUP_DONE=true