Hotkeys, Wake Words, History: Build a Voice Workflow You Keep
Most people try voice input twice: once as a novelty, once after a wrist scare. Both times it fails for the same reason, not the transcription but the workflow. A setup survives daily use when capture is instant, hands-free modes exist for messy moments, and nothing you said is ever lost.
One trigger per mode
DictatorFlow runs as a daemon (dictatorflow --listen) that owns global hotkeys system-wide. Each register of speech gets its own key, so the computer never guesses whether your words were meant as prose or as an action:
- ctrl+shift+d dictate: transcribe and paste into whatever has focus.
- ctrl+alt+j command: process the utterance as an instruction instead of text.
- ctrl+shift+alt+space conversation: open a live back-and-forth session.
- ctrl+alt+r research: speak a question, dispatch it to a local coding agent, get markdown back.
Wake words for hands-busy moments
Hotkeys fail exactly when voice is most valuable: soldering, cooking, holding a child, whiteboarding. Enable voice activation and a local VAD plus wake-word detector runs entirely on-device; saying "flow" starts dictation without touching anything. Reserved wake words route to the right mode automatically, and because detection is local, nothing leaves the machine until you actually speak a sentence worth transcribing.
Every capture is written to local history the moment you finish speaking, audio first as WAV then compressed to Opus, transcript alongside as a text sidecar. If paste fails or the network drops, the recording still exists. Treat the history folder as a spoken clipboard: last week's standup notes are greppable.
Local or cloud, decide once
With an NVIDIA GPU, local mode runs Parakeet CTC 0.6B through ONNX Runtime on CUDA: private, offline, fast enough to feel instant. Cloud mode routes through a provider fallback chain for machines without the hardware. Pick per machine, not per sentence; a workflow you renegotiate daily is a workflow you abandon. Configure both in ~/.dictatorflow/config.json, verify with --test-local file.wav, and start with the installer.