jev-voice
Talk to your Mac. Local whisper.cpp + one Jev (TypeSafe) call per command + macOS automation.
Talk to your Mac. Local whisper.cpp + one Jev (TypeSafe) call per command + macOS automation.
For a single voice command, local speech recognition produces an utterance; Jev receives that text, the frontmost app, installed-app candidates, and code-cut text candidates with a fan-out of typed questions. It selects action and argument candidates; code executes locally. Multi-step screen tasks use a separate one-request-per-step accessibility/browser loop.
Is this utterance a direct command to the computer?
The utterance transcript and candidate apps, sites, and actions are included in state. This chart shows fixed Noul questions from the single speculative request.
Does it request two or more actions in sequence?
The utterance transcript and candidate apps, sites, and actions are included in state. This chart shows fixed Noul questions from the single speculative request.
Energy-based voice detection gates local whisper.cpp transcription; wake-word and follow-up modes decide whether to route the utterance.
The command brain discovers installed apps and cuts possible text payloads from the utterance, then assembles action and argument questions.
Jev chooses from the candidates; code maps the answer to a macOS action. Multi-step tasks observe an accessibility/browser tree and ask Jev for a bounded next step.
The app records decisions and actions, supports dry-run for typed routing, and can speak a response.
The single-command path resolves action, arguments, and intent gates in one request while keeping candidate creation and execution in code. Screen tasks instead use observed indexed controls; Jev does not receive screenshots.