bigvoice

Offline dictation · macOS

A small model.
A big voice.

Hold a shortcut in any app, speak, and your words land where your cursor is. Speech recognition runs entirely on your Mac: no account, no cloud, no subscription.

Download latest View source

macOS 14+ · Apple Silicon

The bigvoice window while listening: a live orange waveform, a 0:04 timer, and the live transcript with the newest word in orange.

01 — How it works

Two ways to get a word in.

Push to talk

Hold to speak. Release to insert.

⌃⌥Space

Hands-free

Press once to start. Press again to finish.

⌃⌥Return

esc cancels. Both shortcuts are yours to change in Settings.

The floating capsule while listening, with live orange bars and a 0:04 timer. The floating capsule after delivery, with a check mark.
The capsule floats above the Dock in every Space. It never takes focus and resizes to what it has to say.

02 — Style

Say it loosely. Send it ready.

Speech models write down every um, every doubled word, and the Tuesday you took back. Choose how much of that reaches the app, and how your words should sound in each one. Tone follows the app you're in: professional in Mail, relaxed in Messages, exact in your editor and terminal.

Verbatim

Exactly what the speech model heard. Nothing added, nothing taken away.

Clean

Ums, stutters, and self-corrections gone. Capitals and periods fixed. Instant, rules only.

Polished

Grammar and punctuation fixed by the language model already built into macOS. Your words and order stay yours.

Refined

Tightened into a ready-to-send message in the tone each app calls for.

The bigvoice Style page: a level control set to Polished, and cards for Messages, Email, and Documents showing what was heard, with fillers struck through, above the finished text in each app's tone.

A model can tidy your words, never change them. Every rewrite is checked against what you said; if it adds facts, answers your question, or drops part of it, bigvoice pastes the Clean version and tells you why.

03 — Why bigvoice

On your Mac. Off the cloud.

Two engines, one app

whisper.cpp on Metal for Whisper GGML models, and ONNX Runtime GenAI for streaming Nemotron Speech.

Reuses what you have

bigvoice finds speech models already on your Mac, including the one the GitHub Copilot app downloads, and uses them in place. Nothing is downloaded twice.

Live transcript

Words appear while you speak. The final pass after you stop is always authoritative.

Works in any app

Runs from the menu bar and pastes into the app you were using. It checks the focus and the inserted text before doing anything.

Never the wrong field

If focus changes while you dictate, bigvoice doesn't insert. Password fields are never filled. Auto-Return is off by default.

Clipboard respected

Your clipboard is restored after insertion, and a newer copy is never overwritten.

04 — Models

Use what you have before downloading more.

At launch bigvoice looks through Copilot's model cache, MacWhisper, superwhisper, VoiceInk, Handy, whisper.cpp and Hugging Face caches, and Spotlight. Need more? Install one of five MIT-licensed Whisper models (32 MB to 190 MB) in one click. Every download is SHA-256 verified.

The bigvoice Models page: a model already on this Mac and a catalog of Whisper models with disk sizes drawn to scale.

05 — Privacy

Your audio stays on this Mac.

06 — Install

Three steps. Then just talk.

  1. 1
    Unzip and move bigvoice.app to Applications.
  2. 2
    Open it and allow Microphone and Accessibility from the setup steps.
  3. 3
    Click into any text field, hold ⌃ ⌥ Space, and speak. Developer ID releases are notarized by Apple and open like any other Mac app; if macOS asks about a build, choose Open Anyway in Privacy & Security.
Download latest All releases