voxi types what you say, as you say it.

Name

voxi — standalone Linux voice input, continuous eager sentence streaming, and desktop typing engine

Synopsis

voxi command [subcommand] [flags]
voxi mode [batch|streaming|eager]
voxi record {toggle|start|stop|status}
voxi eager [--type] [--history] [--daemon] [--model name] [flags...]
voxi monitor [-w|--watch] [-i seconds] [-s sections]
voxi bench [--models list] [--backends cpu,gpu] [--record] [--file wav] [--json path]
voxi history {list|clear|copy ID|retype ID}
voxi config {get|set} type-delay-ms [MS]
voxi daemon modifier-service [--state-file path] [--interval dur]

Commands

mode [batch|streaming|eager]

Show the active voice-input mode, or switch to one of the three transcription strategies. With no argument, prints the current mode.

record {toggle|start|stop|status}

Control an active dictation recording session out-of-band, e.g. from a desktop hotkey or the GNOME extension's panel toggle.

eager [flags]

Run continuous eager sentence streaming dictation. Segments speech on pauses and transcribes phrases in real-time.

monitor [-w|--watch] [-i seconds] [-s sections]

Print or live-refresh a btop-style report covering transcription latency, audio RMS meters, CPU/GPU loads, and daemon health. Aliases: top, stats, resources.

bench [flags]

Benchmark transcription speed (RTF & speedup) for models declared in spec/models.yaml across cpu and gpu backends against standard reference samples.

history {list|clear|copy ID|retype ID}

Inspect and reuse local dictation transcripts. copy places text onto the clipboard; retype injects it directly into the active window.

config {get|set} type-delay-ms [MS]

Read or change typing speed in voxtype configuration to adjust keystroke pacing for legacy or latency-sensitive applications.

daemon modifier-service [flags]

Run the evdev modifier monitoring service directly (normally managed via voxi-modifierd system service).

Eager Options

--threshold N

Audio RMS energy threshold to trigger speech detection (default: 150).

--silence ms

Silence pause duration required to finalize a phrase utterance chunk (default: 800ms).

--pre-roll ms

Pre-speech circular buffer duration, preserving initial phonemes from clipping (default: 500ms).

--min-speech ms

Minimum speech duration to accept as a valid utterance rather than discarding as background noise (default: 200ms).

--max-window ms

Maximum window length before forcing transcription chunking even without a natural pause (default: 8000ms).

--type

Directly inject transcribed sentences into the active window via dotool(1).

--history

Log transcribed utterances into the local dictation history store.

--daemon

Run as a background daemon responding to record toggle controls.

--model name

Whisper model specification to load (default: small.en, with automatic base.en or CPU fallback).

Modifier Gating Architecture

voxi-modifierd reads physical modifier key states directly from the Linux kernel via evdev's EVIOCGKEY, completely independent of compositor or window manager events. When a modifier such as Super, Ctrl, Alt, or Shift is physically depressed, voxi buffers typed output and flushes it immediately upon release (sub-10ns gating). This prevents accidental window close commands or shortcut collisions during push-to-talk dictation while strictly ensuring zero non-modifier keylogging.