VoiceFlow settings guide

Recording hotkey

Config: hotkey.record | Default: Ctrl+Space

What it does: The key combination you hold (or tap, if toggle mode) to trigger recording.

Why change it: Your preferred hotkey conflicts with another app, or you want a combo that's easier to reach.

How to use: Open Settings, go to Hotkeys, click in the Record field, and press the keys you want (single key or combination). Try Alt+R, Ctrl+` (backtick, good for programmers), or CapsLock if you remap it.

Examples:

• Ctrl+Space: default, works in most apps
• Alt+V: mnemonic for "voice"
• Ctrl+Shift+D: "dictate", easy to remember

Trigger style

Config: trigger.style | Default: hold

What it does: How you activate recording: hold the key down while speaking, or tap once to start and tap again to stop.

Options:

• hold: record while key is pressed, stop on release. Instant feedback on when recording ends.
• toggle: tap once to start, tap again to stop. Hands-free, good for longer dictations.

When to use:

• Hold: quick bursts to Claude Code, chat messages, PR descriptions. You know exactly when recording stops.
• Toggle: hands-free note-taking while reading docs, long rambling thoughts, or meetings. No need to hold a key.

Why you'd switch: Your use case changed (now doing more hands-free work), or you want to reduce finger strain on long sessions.

Paste-again hotkey

Config: hotkey.paste_again | Default: Ctrl+Shift+V

What it does: Re-paste the last transcription at your cursor, without re-recording.

Why use it: You dictated something, pasted it, then clicked elsewhere. Now you want to paste the same text somewhere else without re-speaking.

Example: You dictate an error message to Claude Code, it gets pasted in the chat. Then you realize you also need it in a GitHub issue. Hit paste-again and it appears in the issue without re-recording.

Cancel hotkey

Config: hotkey.cancel | Default: Esc

What it does: Stop recording and discard the current capture (don't paste anything).

When you'd use it: You started recording, realized you want to rephrase, and want to bail without pasting. In hold mode, just release; in toggle mode, hit Esc to cancel before you tap to stop.

Default output mode

Config: output.mode | Default: clean

What it does: The cleanup mode that runs on every capture (unless you override it from the tray menu per-capture).

Options:

• raw: verbatim transcript, lightly punctuated. Nothing removed or reworded.
• clean: removes filler ("um", "uh", repeats), fixes capitalization, and improves punctuation. Works offline, instant.
• code: like clean, but also trims politeness and fixes programming terms ("dunder init" → __init__). Perfect for AI coding assistants.
• summary: condenses rambling speech into a tight, one-line brief. Needs an LLM.
• prompt: reformats your spoken words into a structured LLM prompt (Role, Task, Constraints, Context). Needs an LLM.
• systems: rewrites speech into formal Systems Engineering terminology per INCOSE/ISO 15288 (e.g. "make sure it works" → "verify compliance with the requirement"). Needs an LLM.

Which to pick:

• Writing most of the time? Use clean (default, works everywhere, instant).
• Dictating to Claude Code or Cursor? Use code (programming terms fixed, politeness trimmed).
• Taking rough notes you'll clean up later? Use raw.
• Want AI to condense your thoughts? Use summary or prompt (after configuring an LLM).

LLM provider

Config: llm.provider | Default: none

What it does: Points VoiceFlow at an AI model to power the summary and prompt modes. Completely optional.

Options:

• none: no LLM. summary and prompt fall back to a rule-based cleaner (offline, instant, lower quality).
• openai_compatible: any OpenAI-compatible endpoint. Examples: Ollama, LM Studio, Azure OpenAI, or even your own local server.
• anthropic: Claude API (requires API key; your text is sent to Anthropic).
• openai: OpenAI API (requires API key; your text is sent to OpenAI).

When to configure:

• You want better summaries: rambling spoken thoughts condensed into tight, readable text. Great for capturing half-formed ideas.
• You dictate prompts to AI: spoken instructions reformatted into structured role/task prompts, saving you keypresses.
• You want true local-only processing: use Ollama or LM Studio with a local model (careful: some model names ending in :cloud route through a remote server).

Setup example (Ollama):

Config (config.toml):
[llm]
provider = "openai_compatible"
base_url = "http://localhost:11434/v1"
model = "mistral"

Then fire up Ollama, run ollama serve, and your summaries will use the Mistral model running on your machine.

Personal dictionary

Config: [dictionary] section in config.toml

What it does: A list of spoken phrases and what they should become in your output. Merges on top of the built-in programming dictionary (your entries win if there's a conflict).

Why use it: VoiceFlow keeps mishearing a term, or you want to add jargon specific to your domain, company, or stack.

How to add terms: Edit config.toml directly. Find or create the [dictionary] section and add pairs:

[dictionary]
"dunder main" = "__main__"
"postgres" = "PostgreSQL"
"my company name" = "MyCompanyName"
"async await" = "async/await"

Tips for good dictionary terms:

• Exact matches only. The term must match exactly as spoken (case-insensitive). "postgres" matches but "postgre" won't.
• Multi-word phrases work. You can map "machine learning model" → "ML model".
• Symbols are safe. Use any symbols you need: "double quote" → " works fine.
• Order doesn't matter. Longer phrases are processed first, so specific terms take precedence.

Real examples from v1:

• "dunder init" → __init__
• "fast api" → FastAPI
• "dot py file" → .py file
• "postgres connection" → PostgreSQL connection

When to add a term: After you notice VoiceFlow mishearing the same word three times. Add it to the dictionary and it's fixed from then on, in every mode.

Language

Config: speech.language | Default: auto-detect based on Windows locale

What it does: The language the speech engine listens for. Auto-detection is the default but slower and less accurate than explicitly setting a language.

Why change it: You work in a specific language and want faster, more accurate transcription.

Examples: en (English), fr (French), de (German), ja (Japanese). See Whisper's supported languages for the full list.

Performance: Setting a language is ~10-20% faster than auto-detect, with better accuracy for that language.

Speech engine

Config: speech.engine | Default: auto-configured on first run

What it does: The underlying transcription model and provider. VoiceFlow auto-configures on first run based on your hardware.

Options (v0.3.0):

• parakeet: NVIDIA's lightweight model. Faster, lower accuracy than Whisper. Bundled in the installer. Recommended for most users.
• whisper_tiny, whisper_base, whisper_small: OpenAI Whisper models, in order of speed vs. accuracy trade-off. Tiny is fast but less accurate; small is slower but more accurate. All bundled in the installer.

When to switch:

• Parakeet is mishearing you. Try whisper_base for better accuracy (trades speed).
• Transcription is too slow. Try parakeet or whisper_tiny for speed (trades accuracy).
• You need a specific language. Some models perform better in certain languages. Test both and see.

Start on login

Config: app.start_on_login | Default: false

What it does: Whether VoiceFlow launches automatically when you log into Windows.

Why enable it: You use VoiceFlow every day and want it running without remembering to start it. It sits quietly in the tray, takes no resources until you dictate.

How to set it: Right-click the tray icon → Settings → check Start on login. That's it; next restart, it'll launch automatically.

Tray icon click behavior

Config: app.tray_click | Default: menu

What it does: What happens when you left-click the VoiceFlow tray icon.

Options:

• menu: opens the context menu (mode switcher, Settings, Help).
• paste_again: re-pastes the last transcription (convenience shortcut).
• toggle_mode: cycles through modes without opening a menu.

Choose based on workflow:

• Default (menu): most flexible, covers everything in one click.
• paste_again: if you frequently re-paste the last dictation without changing it.
• toggle_mode: if you switch modes often and want a one-click cycle.

Capture history retention

Config: app.history_days | Default: 30

What it does: How many days of capture history VoiceFlow keeps (transcripts, timestamps, modes used). Older entries are automatically deleted.

Why care: History takes disk space (minimal) and can be a privacy concern if your machine is ever accessed by someone else.

Options:

• 0: no history, every capture is forgotten immediately.
• 7: keep a week of history.
• 30: default, keep a month.
• -1: keep forever.

Why you'd change it: Privacy-sensitive work (medical, legal transcripts) → set to 0. Research or reference work where you want to look back → set to -1.