Skip to content

Releases

0.1.2 — hands-free dictation in the terminal

Choose a tag to compare

@rodri45l rodri45l released this 18 Sep 16:51

Hands-free dictation inside the terminal, with nothing to install.

Highlights

  • Bundled recorder (miniaudio, MIT-0) for macOS arm64/x64 and Linux x64 — no
    ffmpeg, no Homebrew, no calibration. Windows and Linux arm64 fall back to
    ffmpeg/parecord/arecord/sox.
  • Streaming capture: 16 kHz PCM is read from the recorder's stdout and the WAV
    is assembled in memory. No temp files exist at all, and a killed TUI can't strand
    a recorder holding the microphone.
  • Status indicator: a KITT-style scanner or an amplitude wave, showing grey /
    red / amber / violet for listening, speaking, transcribing and muted. f7
    switches style, f8 mutes.
  • Answer by voice: questions and permission prompts are answered out loud.
  • Voice kill switch: "stop" aborts the agent; "stop conversation mode" leaves
    hands-free mode — decided by the LLM, not keyword matching.

Fixes worth noting

  • ffmpeg on macOS buffers file output in ~256 KB blocks, which left the VAD blind
    for the first ~8 seconds of every capture. Capture now reads the stream.
  • A fresh opencode instance (no session yet) silently swallowed dictation. Voice now
    resolves a session on demand and follows the session you're looking at.
  • Dropped transcripts that contain no words at all ("."), and clips that are
    saturated (a knock) rather than speech.
  • Only one conversation loop can run per instance.

Requires opencode with TUI plugin support and an OpenAI-compatible STT endpoint
(local or cloud).