Jev Voice
Turn local speech transcripts into typed macOS actions without asking Jev to generate command text.
What Jev decides
Which action to run, which supplied app, site, key, direction, target, or text span it uses, and whether special gates apply.
Local whisper.cpp transcribes an utterance, code supplies installed apps and candidate values, and one Jev request chooses an action and its arguments before deterministic macOS handlers execute it.
- Jev role
- Map a completed speech transcript or accessibility-tree observation onto a bounded macOS action and arguments.
- Input
- A local speech transcript, installed-app and site candidates, supplied text spans, or indexed accessibility controls for multi-step tasks.
- Output
- Typed Choice selections and Noul-style gates consumed by deterministic action code.
- Decision frequency
- One request for a single command, or once per step for a task that needs screen observation.
- Fallback
- Ignore utterances outside the active wake or push-to-talk window, reject unavailable values, and stop the bounded accessibility loop instead of inventing text or targets.
How it works
Overview
Jev Voice is a macOS voice assistant for opening applications and websites, searching, typing supplied text, pressing shortcuts, scrolling, changing media or volume, and other bounded operations. Audio is transcribed locally with whisper.cpp. Jev receives the transcript and code-built candidate values, then selects a typed action and typed arguments; it does not write free text.
Commands that require inspecting the current interface enter a separate accessibility-tree loop ported from jev-ultrafast. Wake-word and push-to-talk modes remain local until an accepted command is sent. The project requires macOS permissions and documents that destructive or misheard actions still need careful use; repository latency and cost figures are author-reported rather than independently reproduced here.
- Language
- Python
- License
- MIT
- Author
- Kevin Badi
Sources
Related projects
typesafe-computer-use
A native macOS computer-use loop built from local OCR, accessibility data, and bounded Jev actions.
What Jev decides
Which computer action, visible item, site, or off-screen accessibility control to use next.
Jev Chat Assistant
An Android conversation copilot that uses Jev to interpret chats and rank reply candidates without sending for the user.
What Jev decides
The likely intent, conversation risk, reply need, best next action, unresolved tension, and which of three candidate replies fits best.
Distill
Route coding-agent work, reasoning effort, and context retention through bounded Jev decisions.
What Jev decides
Which configured model and effort should handle a call, what evidence deserves attention, whether another attempt is useful, and what context should be retained.